Vocational education, occupational choice and unemployment over the professional career

Abstract

This study investigates the relationship between occupational choice and unemployment over the professional career with German administrative linked employer–employee data that track more than 800,000 graduates from vocational education over 25 years. Using short-run fluctuations in local and sectoral occupation-specific labor demand as instruments, it finds that choosing an occupation that later turns out to suffer from low or negative employment growth has a statistically and economically significant impact on unemployment over the professional career. On average, an unanticipated one-standard deviation decrease in occupation-specific employment growth raises unemployment by about 116 days over the life cycle.

1 Introduction

This study uses German administrative linked employer–employee data to document that occupational choice early in the professional career and long-term labor market outcomes (in particular unemployment) is strongly correlated. This correlation could in principle be explained by two competing hypotheses. First, it could be due to the sorting into separate occupations of individuals with different characteristics like ability, skills or motivation. Second, occupational choice early in the professional career itself might have causal effects on later outcomes. Under the second hypothesis, two otherwise identical individuals who simply by good or bad luck end up in two different training occupations might face very different career prospects.

Against this backdrop, this study examines whether an individual’s total amount of unemployment during the first 25 years on the labor market (lifetime unemployment according to Schmillen and Möller 2012) is causally affected by his or her occupational choice. The analysis focuses on graduates from Germany’s vocational education system. This is motivated by the fact that the system is highly structured, with a well-defined curriculum organized around about 300 officially recognized training occupations and therefore puts a particular emphasis on occupational choice. Moreover, data exist that allow a precise measurement of this occupational choice well before individuals’ actual labor market entry.

This study documents a correlation between occupational choice and lifetime unemployment and follows a three-pronged strategy to identify whether this correlation is due to sorting or instead represents a causal effect. First, it relies on high-quality linked employer–employee data that make it possible to control for a wide range of observable individual- and training firm-specific characteristics including location- and sector-specific fixed effects. Second, to minimize biases due to sorting into occupations by ability, skills, motivation or similarly unobserved individual characteristics, this study controls for individuals’ expectations of an occupation’s relative desirability with the help of long-term forecasts of occupation-specific labor demand. Third, to take account of both occupational sorting unrelated to individuals’ expectations and measurement error, it instruments the choice of a specific training occupation with Bartik-style short-run fluctuations in occupation-specific labor demand on the regional and sectoral level.

The ensuing analysis is based on an administrative linked employer–employee data set that contains the universe of social security records in Germany. From these records, information on the professional careers of 852,872 West German men and women who graduated from vocational education in 1981, 1982 or 1983 is extracted.¹ The data make it possible to identify the exact time and place of labor market entry—defined here as graduation from vocational education—for all these individuals and to track them for every day of the first 25 years of their professional careers.

The analysis demonstrates that the observed correlation between occupational choice and lifetime unemployment at least partly represents causal effects. More specifically, choosing an occupation that later turns out to see low or negative employment growth has statistically and economically significant impacts on lifetime unemployment, ceteris paribus. In the mean, an unanticipated one-standard deviation decrease in occupation-specific employment growth raises lifetime unemployment by about 116 days or 0.16 standard deviations. Again in the mean, the difference in the predicted lifetime unemployment between the first and ninth decile of unanticipated occupation-specific employment growth amounts to 536 days or 0.75 standard deviations.

Closely related studies include those by Fletcher and Sindelar (2009), Hanushek et al. (2016) and Schmillen and Möller (2012). Fletcher and Sindelar (2009) investigate how early occupational choice affects long-term health with an identification strategy similar to the one used here. Hanushek et al. (2016) demonstrate that vocational as compared to academic education is associated with gains in employment early in the professional career. However, a lack of adaptability to technological change means that these gains are at least partly offset by diminished employment later in life. Schmillen and Möller (2012) analyze the distribution and determinants of lifetime unemployment with German linked employer–employee data. In this context, they also investigate the long-term unemployment effects of the occupation pursued early in the professional career. However, as they rely on selection on observables, it might not be totally uncontroversial whether what they capture is in fact a causal relationship.

This study is also related to models that stress the influence of human capital on unemployment (cf. Ljungqvist and Sargent 1998, 2004, 2008). In this strand of literature, losing a job is interpreted as a sudden depreciation of human capital which (possibly together with other factors) might lead to long unemployment spells. Besides, it draws on the literature that shows the importance of occupation-specific human capital. For instance, Kambourov and Manovskii (2009, p. 63) “find that returns to occupational tenure are substantial. [...] Moreover, when occupational experience is taken into account, tenure with an industry or employer has relatively little importance in accounting for the wage one receives.” This study is also connected to the growing number of empirical studies that analyze whether labor market events or decisions early in the professional career have long-term effects. Examples include von Wachter and Bender (2006) who demonstrate substantial and long-lasting wage losses for some groups of young displaced workers, Raaum and Røed (2006), Kahn (2010) and Oreopoulos et al. (2012) who show that business cycle conditions at time of labor market entry have economically significant and persistent wage and employment effects and Gregg (2001), Mroz and Savage (2006) and Schmillen and Umkehrer (2017) who find that early-career unemployment can have long-term scarring effects.

The remainder of this paper is structured as follows: Sect. 2 sketches the theoretical background, discusses the identification strategy and provides background on Germany’s vocational education system. Section 3 introduces the linked employer–employee data set and presents some descriptive statistics. Section 4 contains the results of the multivariate analysis and elaborates on their robustness with regard to variations of the empirical setup and their validity for different subgroups. Section 5 concludes.

2 Conceptual considerations

2.1 Theoretical background

As discussed by Schmillen and Möller (2012), a standard neoclassical labor market model can easily explain how an investment in a disadvantageous kind of human capital early in the professional career can henceforth reduce an individual’s productivity and therefore depress his or her wages. In particular, assume that the human capital depreciation rate of newly laid-off workers depends on whether the human capital acquired in the previous job is still in demand at the time of separation. Then lay-offs caused by technical change or shifting trade patterns will lead to an especially strong depreciation of specific human capital acquired previously. Further assume that—as argued by Kambourov and Manovskii (2009), Schmillen and Möller (2012) and others—occupation-specific human capital is of crucial importance. Then individuals who early in their professional career acquire specific human capital in an occupation that later becomes obsolete will experience an especially pronounced loss of human capital if they are laid off.

Under the standard neoclassical labor market model’s assumption that labor markets are flexible, in spite of the lower productivity an individual who experiences a depreciation of occupation-specific human capital would be expected to find a different job without any great difficulties (at a lower wage). In the standard model, this individual would therefore not be expected to exhibit an elevated amount of lifetime unemployment.

In contrast, search and matching models by Ljungqvist and Sargent (1998, 2004, 2008) and others that link the depreciation of human capital and unemployment provide the theoretical basis for why a causal relationship between occupational choice and unemployment over the professional career might in fact exist. This causal relationship emerges mainly because of two mechanisms. First, assume that a welfare state exists that pays unemployment benefits in proportion to past earnings. As a consequence, individuals with highly depreciated human capital have relatively high reservation wages. This in turn results in difficulties in finding a new job with wages that are preferred to unemployment compensation. Second, assume that job search is associated with disutility. Then, individuals with depreciated human capital will exhibit reduced search intensities. As a consequence of the two mechanisms, there is a causal relationship between occupational choice and unemployment over the professional career.

Ultimately, the theoretical literature does not provide an unambiguous answer to the question whether occupational choice can have a causal impact on unemployment over the professional career. Instead, this question needs to be answered empirically.

2.2 Identification strategy

An empirical investigation of the link between occupational choice and unemployment over the professional career faces different potential sources of bias with sometimes ambiguous direction. The identification strategy pursued in this study starts with a very simple estimation equation and then sequentially removes one potential source of bias after the other.

In the first step, a simple linear regression is used to document the correlation between an individual’s occupational choice early in the professional career represented by the unanticipated occupation-specific employment growth $\Delta L^{\eta }_{i}$ and his or her lifetime unemployment $u_{i}$:

$$\begin{aligned} u_{i} = \alpha + \beta \Delta L^{\eta }_{i} + \epsilon _{i}. \end{aligned}$$

(2.1)

Index i denotes individuals and $\epsilon _{i}$ is an error term. $\alpha$ and $\beta$ are the regression coefficients with $\beta$ equal to the correlation between $u_{i}$ and $\Delta L^{\eta }_{i}$ corrected by the ratio of standard deviations of these variables.

The most straightforward way to represent occupational choice in Eq. 2.1 would have been with the help of dummy variables for individuals’ training occupations. However, this representation would have made an economic interpretation of results challenging. Therefore, Eq. 2.1 does not contain dummy variables for training occupations but instead the explanatory variable $\Delta L^{\eta }_{i}$. $\Delta L^{\eta }_{i}$ is constructed based on the occupation-specific employment growth from the year of labor market entry until 25 years later. An occupation with a higher employment growth over the observation period is deemed to be more advantageous or desirable ex post. Additionally, a correction is being made for individuals’ ex ante expectations about the employment prospects of different occupations. This minimizes the potential bias due to sorting into occupations by ability, skills, motivation or similarly unobserved individual characteristics.²

While no direct individual-level data on the formation of expectation by prospective trainees in the 1970s are available for the correction for individuals’ ex ante expectations, Germany’s Federal Employment Agency has long aimed to provide prospective trainees with relevant and reliable information on the employment prospects of different occupations. For this purpose, its regional offices have regularly invited students of different age groups to attend information sessions and the Agency has also published various guidelines and brochures (and nowadays makes much information available online, for instance through the Web site berufenet.arbeitsagentur.de). In order to assist its placement and advisory units with scientifically substantiated occupational forecasts, in 1967 the Agency also founded the Institute for Employment Research, an in-house research institute that has since broadened its areas of engagement.³

This study uses a long-term forecast of occupation-specific labor demand by Blüm and Frenzel (1975) that was directly commissioned by the Institute for Employment Research and published in its flagship book series (cf. Schmillen and Möller 2012). Blüm and Frenzel (1975) base their forecast of occupation-specific labor demand on a modified and adapted version of the manpower requirement approach developed by the OECD in the 1960s (cf. Parnes 1962). More specifically, the labor demand forecasts by Blüm and Frenzel (1975) are based on a detailed forecasting model of the German economy that partly assumes a continuation of past trends and partly incorporates expected structural change (in particular as a result of the dramatic changes in oil prices that occurred during the first energy crisis in the mid-1970s). Based on this forecasting model and additional assumptions on the substitutability of labor between occupations, labor demand forecasts from 1975 to 1990 are derived for a variety of occupations. While the forecast by Blüm and Frenzel (1975) does not represent direct individual-level data on the formation of expectation by prospective trainees in the 1970s, it arguably had a significant influence both on the Federal Employment Agency’s occupational guidance policy and on individuals’ expectation formations.⁴

Formally, $\Delta L^{\eta }_{i}$ is constructed as follows:

$$\begin{aligned} \Delta L^{\eta }_{i}&:= \Delta L^{\eta } + \nu _{i} \end{aligned}$$

(2.2)

$$\begin{aligned}&= \underbrace{\frac{L^{\eta }_{t+25}-L^{\eta }_{t}}{L^{\eta }_{t}}}_{\text {empl. growth}} - \underbrace{E_{1975}\left[ \frac{L^{\eta }_{t+25}-L^{\eta }_{t}}{L^{\eta }_{t}}\right] }_{\text {anticipated empl. growth}} + \nu _{i} \end{aligned}$$

(2.3)

$$\begin{aligned}&= \frac{L^{\eta }_{t+25}-E_{1975}[L^{\eta }_{t+25}]}{L^{\eta }_{t}} + \nu _{i} . \end{aligned}$$

(2.4)

In Eqs. 2.2, 2.3 and 2.4, subscript $t=\{1981,1982,1983\}$ denotes the labor market entry cohort and superscript $\eta = \{1,...,49\}$ one of 49 separate occupations as defined by Blüm and Frenzel (1975). Blüm and Frenzel (1975) base their classification of occupations on the 1975 edition of the Federal Employment Agency’s official Classification of Occupations (Klassifikation der Berufe 1975) in use from 1975 until 1988 and in slightly modified form until 2010. This classification differs from the International Standard Classification of Occupations (ISCO) but just as ISCO classifies workers according to the occupation they actually exercise at any given point in time and not necessarily the occupation they were trained in. However, for trainees the training occupation and the actual occupation are by definition the same. On its first three levels, the Classification of Occupations groups occupations into five Occupational Sectors (Berufsbereiche), 33 Occupational Sub-sectors (Berufsabschnitte) and 86 Occupational Groups (Berufsgruppen).⁵ The classification of occupations by Blüm and Frenzel (1975) is directly based on the Occupational Groups but subsumes some relatively similar Occupational Groups into larger groupings.

Equation 2.2 clarifies that individual-level expectations of occupation-specific employment growth $L^{\eta }_{i}$ are approximated by occupation-level data $L^{\eta }$. This results in a measurement error $\nu _{i}$ that is assumed to be independent, identically distributed and uncorrelated with $\Delta L^{\eta }$. $L^{\eta }_{t}$ in Eqs. 2.3 and 2.4 denotes the employment share of occupation $\eta$ in year t and $E_{1975}[L^{\eta }_{t+25}]$ the expectation in 1975 about the share of labor demand apportioned to occupation $\eta$ in year $t+25$. Equations 2.3 and 2.4 treat $L^{\eta }_{t+25}$ but not $L^{\eta }_{t}$ as a random variable and use the linearity of the expectation operator. Since Blüm and Frenzel (1975) forecast occupation-specific labor demand only until 1990, their predictions are linearly extrapolated to later years.

An example helps to illustrate the construction of $\Delta L^{\eta }$. Take the occupation health professionals (other than doctors or pharmacists) ($\eta =21$) and the labor market entry cohort 1981. According to the work by Blüm and Frenzel (1975), there were 228,000 health professionals in 1970 (or 0.86% of overall employment) and labor demand for this occupation was forecast to increase to 301,000 positions in 1990 (or 1.08% of total employment). Using the same forecast annual growth rates from Blüm and Frenzel (1975) for employment for health professionals and overall employment from 1970 to 1990 (1.47% annual growth for health professionals and 0.23% overall) to linearly extrapolate employment growth until 2006 leads to an anticipated number of 380,000 health professional positions in 2006 (or 1.24% of overall employment).

According to the administrative data used in this study, the relative share of employment for health professionals actually already reached 1.22% of total employment in 1981. By 2006, its share of total employment had grown to 1.75%. For the labor market entry cohort 1981, $\Delta L^{21}$ is therefore $(0.0175-0.0124)/0.0122=0.42$. Because both employment growth and anticipated employment growth take on values from $-\,100$% to $+\infty$ (employment can fall at most by 100% but can grow by more than a 100%), the range of $\Delta L^{\eta }$ exceeds $[-1;+1]$.

In a second step, the empirical investigation controls for a comprehensive set of socio-demographic and firm-specific variables. Covariates include an individual’s gender, graduation year, graduation age and education, and his or her training firm’s size and median wage. In some specifications, dummy variables for the training firm’s location (i.e., its Kreis or county) and sector are also included. As is the case for occupational choice, all control variables are being measured during vocational education before the actual labor market entry. They can thus be considered exogenous. The addition of a vector of control variables, $\mathbf x _{i}$, leads to the following estimation equation:

$$\begin{aligned} u_{i} = \alpha + \beta \Delta L^{\eta }_{i} + \mathbf x _{i}' \gamma + \epsilon _{i}. \end{aligned}$$

(2.5)

Can the coefficient $\beta$ in Eq. 2.5 be interpreted as the causal effect of occupational choice on lifetime unemployment? There are at least two reasons for doubt that this interpretation is correct. First, individuals might sort into training occupations according to unobserved characteristics unrelated to their expectations. Second, even though the study by Blüm and Frenzel (1975) was produced by the in-house research institute of Germany’s Federal Employment Agency and arguably had an influence both on the Agency’s occupational guidance policy and on individuals’ expectation formations, Eq. 2.2 makes it clear that in practice individuals’ expectations are only captured with some measurement error. As systematic expectations are already accounted for, the potential direction of a bias resulting from sorting is unclear. On the contrary, at least under the classical assumption of a measurement error that is independent, identically distributed and uncorrelated with $\Delta L^{\eta }$, the measurement error should lead OLS regressions to give estimators of $\beta$ that are inconsistent and downward biased (in absolute terms).⁶

In a third step, short-run fluctuations in occupation-specific labor demand on the regional and sectoral level are used as instruments for $\Delta L^{\eta }_{i}$. This strategy draws on Bartik (1991) and Fletcher and Sindelar (2009), aims to arrive at consistent estimates of $\beta$ in spite of the potential methodological concerns related to sorting and measurement error and assumes that the instruments are uncorrelated with the measurement error process, cf. Wooldridge (2002). Similar to Bartik’s (1991) shift-share approach, two instruments are defined as the unanticipated occupation-specific employment growth weighted by county- or sector-level employment, respectively:

$$\begin{aligned} \mathrm{Bartik}_\mathrm{c}&:= \sum _{\eta } \left( L^{\eta }_\mathrm{c} \Delta L^{\eta } \right) \quad {\text {and}} \end{aligned}$$

(2.6)

$$\begin{aligned} \mathrm{Bartik}_\mathrm{s}&:= \sum _{\eta } \left( L^{\eta }_\mathrm{s} \Delta L^{\eta } \right) . \end{aligned}$$

(2.7)

Here, c indexes counties and s sectors. $L^{\eta }_\mathrm{c}$ is the share of employment in occupation $\eta$ in county c. $L^{\eta }_\mathrm{s}$ is defined accordingly. The resulting estimation equations are:

$$\begin{aligned} u_{i}&= \alpha + \beta {\hat{\Delta L^{\eta }_{i}}} + \mathbf x _{i}' \gamma + \epsilon _{i} \quad {\text {and}} \end{aligned}$$

(2.8)

$$\begin{aligned} {\hat{\Delta L^{\eta }_{i}}}&= \delta + \lambda _{1} \mathrm{Bartik}_{\mathrm{c},i} + \lambda _{2} \mathrm{Bartik}_{\mathrm{s},i} + \mathbf x _{i}' \theta + \sigma _{i}, \end{aligned}$$

(2.9)

with $\alpha$, $\beta$ and the vector $\gamma$ as the regression coefficients in Eq. 2.8 and $\delta$, $\lambda _{1}$, $\lambda _{2}$ and the vector $\theta$ in Eq. 2.9. $\epsilon _{i}$ and $\sigma _{i}$ are the error terms in Eqs. 2.8 and 2.9, respectively. ${\hat{\Delta L^{\eta }_{i}}}$ is the predicted value of $\Delta L^{\eta }_{i}$.

As detailed in Angrist et al. (1996), a suitable instrument needs to be relevant, ignorably assigned and excluded. For the county average of the unanticipated occupation-specific employment growth, relevance is established almost mechanically. If a specific county presents prospective trainees with many training opportunities in occupations that later turn out to exhibit strong employment growth, this represents an occupation-specific local labor demand shock. In consequence, it is more likely that a trainee from this county will choose an occupation with strong employment growth. Arguably, the instrument is also ignorably assigned. Following for example Gregg’s (2001) reasoning and the fact that prospective trainees entering vocational training in Germany tend to be very young and can only expect a modest allowance (cf. Sect. 2.3), the choice of location at labor market entry can be seen as exogenous. Any remaining concerns about sorting of individuals into certain counties before the start of vocational education are addressed with the help of the repeated cross-sectional design of the data set. This allows the inclusion of fixed effects for the training firms’ counties so that only the variation from within-county short-run labor demand fluctuations over 1981, 1982 and 1983 is being used. Finally, the instrument can be considered excluded. This is because it is independent of local occupation-specific labor supply shocks and because time-varying patterns of economic conditions, the accumulation of skills and the dynamism of matching processes early in the professional career prevent it from influencing lifetime unemployment through channels other than occupational choice.

The second instrument can be deemed relevant, ignorably assigned and excluded for very similar reasons. The main difference in the suitability of the instruments might be that the inclusion of fixed effects for the training firms’ sectors is likely to be even more important than the addition of county fixed effects. Without these sectoral fixed effects, the interaction between occupational and sectoral choice and the sorting of individuals into sectors according to their unobserved characteristics might have been of considerable concern.

If the identifying assumptions hold, the approach described so far will rid estimates of any potential biases from unobserved individual characteristics, sorting and measurement error. However, a regression of lifetime unemployment on the unanticipated occupation-specific employment growth also poses another, somewhat more technical challenge: almost 30% of the individuals in the sample are not unemployed for a single day during the 25 years after their graduation from the vocational education system. This is the typical case of a corner solution outcome as defined by Wooldridge (2002). As a consequence, OLS or simple IV estimates are biased and inconsistent because of a correlation between the regressors and the error term. To take account of the corner solution outcome, the discussion emphasizes Tobit models (cf. Tobin 1958) and Smith and Blundell’s (1986) conditional maximum likelihood estimator for a Tobit model with continuous endogenous regressors. Outputs from standard OLS and IV regressions are also presented but mostly as a reference point and because standard diagnostic tests are more readily available for them.

2.3 Germany’s vocational education system

Vocational education systems play an important role in Austria, Switzerland and some other European economies. In many other parts of the world, including in the USA and the UK, there has long been a discussion about whether to strengthen the importance of education programs that directly prepare people for specific trades, crafts or careers (cf. for example Heckman 1993 or Neumark 2002). Germany’s vocational education system usually combines apprenticeships in a company and instruction at a school and is therefore a dual system. It is regularly described as the model vocational education system. As a majority of young people go through the system before transitioning from school to work, it is also of central importance for the German labor market.

For the cohorts considered here, official information on the percentage of individuals that went through the vocational education system is not readily available. Therefore, Table 1 uses the administrative data described in more detail below (more specifically the SIAB, a representative 2% sample of the universe of individuals who have received unemployment benefits and/or been employed subject to social security contributions in Germany) to list the proportion of all individuals entering the West German labor market from 1980 to 1989 that at the time of their labor market entry had previously spent time in the vocational education system. The table shows that throughout the 1980s between 48 and 68% of labor market entrants had spent time in vocational education. In 1981, 1982 and 1983 the share amounted to 55.9, 62.3 and 68.2%, respectively. These numbers are consistent with evidence by Bundesinstitut für Berufsbildung (2006) that shows that in 1993—the earliest year for which official data are readily available—65.8% of all 16- to 24-year-olds in Germany were in the vocational education system.

Table 1 Share of labor market entrants with prior vocational training experience

As for instance explained in Hippach-Schneider et al. (2007), across most of Germany there are three main types of (nonvocational) secondary schools: general high schools (Hauptschulen) that usually end after grade nine and prepare students for blue collar occupations, intermediate high schools (Realschulen) that run up to grade ten and prepare students for clerical occupations and academic high schools (Gymnasien) that prepare students for higher education.⁷ While all three main types of schools offer distinct school-leaving certificates, the access to the German vocational education system is not formally linked to any of these. For instance, according to Bundesinstitut für Berufsbildung (2006) in 1970 80% of students in vocational education and training had no school-leaving certificate or one from a general secondary school and 19% had a certificate from an intermediate secondary school. 1% had a certificate from an academic high school (Abitur) that would have also entitled them to pursue higher education.

The vocational education system is organized across the entire territory of Germany in a largely standardized form. However, trainees’ age when they enter or leave the system is not uniform. One cause of heterogeneity is the different schooling paths that students take before they enter the vocational education system. A second cause is that the duration of vocational training varies. While training usually lasts for 3 years, there are number of occupations where training can be completed after 2 years. These occupations tend to lead to blue collar jobs (like alteration tailor, warehouse operator or machine operator). A third cause is that, oftentimes, trainees with Abitur may be entitled to finishing their training early. Finally, for some trainees failure to pass final examinations or other reasons may lead to longer-than-usual training periods.

Importantly, the vocational training system is highly structured, with well-defined curricula organized around about 300 officially recognized training occupations. These include a wide array of both blue and white collar occupations and range from doctor’s assistant to optician to oven builder. While training content is highly structured, the matching of trainees and training firms takes place in a decentralized fashion. Usually, prospective trainees first decide on their intended training occupation. Next, they submit applications to training firms. Then, training firms make a selection among applicants. Finally, trainees and training firms enter into training contracts. Once the contracts are signed, the trainees are entitled (and obliged) to enroll part time in a vocational school.

During training, trainees receive an allowance from their training firm and have to pay social security contributions. The allowance depends on what is negotiated through collective bargaining but tends to be rather modest, about a third of trainees’ expected starting earnings after the completion of training. While prospective trainees are in principle free to search for a training firm across Germany, their young age at the beginning of training and the modest allowance paid during training means that in practice their geographic mobility both before and during training is generally low. Indeed, in a sample of trainees similar to the one used here Schmillen and Umkehrer (2017) observe that 97.6% of trainees do not significantly change their place of residence during their training.⁸

3 Data and descriptive evidence

This study relies on German linked employer–employee data. They are created by merging two data sets: first, the Integrated Employment Biographies (IEB, cf. Dorner et al. 2010) and, second, the Establishment History Panel (BHP, cf. Hethey-Maier and Seth 2010). Both are administrative data sets provided by the Institute for Employment Research.

The version of the IEB used here contains the universe of all individuals who received unemployment benefits and/or were employed subject to social security contributions in the Federal Republic of Germany for at least one day between 1975 and 2008. Only spells of dependent employment not covered by social security—like those of civil servants or family workers—and spells of self-employment are not included. All in all, the IEB cover about 80% of Germany’s total workforce. They encompass detailed longitudinal information on employment status, wages, socio-demographic and firm characteristics to the exact day. Because Germany’s social security agencies use the underlying administrative data to compute social security contributions and unemployment benefits, they are highly reliable. In the context of this study, another important advantage of not relying on survey but on administrative data is that one does not need to worry about panel mortality or nonresponse.

The BHP encompasses all German establishments that employ at least one worker who is subject to social security contributions on June 30th of any given year. Compared to the IEB, the BHP contains more detailed firm-specific variables. Among these are an establishment’s sector and its geographic location. Information on the number of employees and their median wage is also included. The different cross sections of the BHP can be combined to form a panel. The merger of the IEB and BHP through common firm identifiers makes it possible to control for observable characteristics not only on the individual level but also on the level of the training firm.

This study focuses on those individuals that start their professional career after graduating from Germany’s vocational education system or more precisely form the dominant part of the vocational education system that combines apprenticeships in a company and vocational education at a school (the dual system of vocational education). Therefore, the sampled trainees receive an allowance from their training firm, have to pay social security contributions and are included in the linked employer–employee data set. The information on trainees captured in the data set is in particular related to the type of training and the nature of the firm providing the training. Since it is available for the time before the actual labor market entry, any problems that might be caused by unobserved initial conditions are avoided.

Table 2 contains summary statistics for this study’s two key variables. These are lifetime unemployment and the unanticipated occupation-specific employment growth, $\Delta L^{\eta }_{i}$. Information on individuals’ unemployment spells is extracted from the IEB. Unemployment is defined as a period during which an individual is registered with the Federal Employment Agency and receives unemployment benefits—which in Germany can be drawn on a daily basis. This definition reflects data availability and the fact that Germany’s official unemployment rate is based on the number of individuals registered as unemployed, an overwhelming portion of whom are eligible to receive unemployment benefits.⁹ At the same time, this particular definition also has some drawbacks. These include that it is not strictly identical to the International Labour Organization’s (ILO) established definition of unemployment. The ILO’s definition relies on survey instead of administrative data and defines somebody as unemployed who is out of work, actively looking for work and available for work. Both under- and overcounting are possible. On the one hand, an individual might be drawing unemployment benefits but not actively looking for work. On the other hand, a different individual might be actively looking for work but not entitled to unemployment benefits. In the sensitivity section below, an alternative definition of unemployment is used to make sure results are robust to variations in the definition of unemployment.

Table 2 Summary statistics for lifetime unemployment and unanticipated occupation-specific employment growth

Using the receipt of unemployment benefits to define unemployment spells has the important consequence that because regulations concerning unemployment benefits have varied during the last decades, it is difficult to compare the length of unemployment periods from different points in time. For instance, until 1981 individuals could claim unemployment benefits after 6 months of work subject to social security contributions within a 3-year window. Since 1982 the minimum vesting period has been 12 months. Also, in the 1980s the maximum length of unemployment benefit receipt for older workers was increased in several steps from 12 months to up to 32 months. In turn, some of this increase was reversed in 2004. To assure that all individuals in the sample were subject to the same institutional environment throughout their professional career and also that results are not driven by cohort effects, the analysis is restricted to individuals that finished their first vocational training in 1981, 1982 or 1983.

Table 2 shows that the average sampled individual is unemployed for 439 days (almost 15 months) during the first 25 years after graduation from the vocational education system. This is equivalent to an average unemployment rate of slightly under 5%. This is qualitatively not too different but somewhat lower than the average of the official unemployment rate for West Germany for the period 1981–2008 as reported by the German Council of Economic Experts and depicted in Table 3. The slight discrepancy between the data sources can most likely be attributed to the fact that the measure of unemployment used in this study only encompasses periods where unemployment benefits are received and might thus fail to fully take account of all official unemployment periods.

Table 3 Official unemployment rate for West Germany

According to Table 2, the standard deviation of lifetime unemployment is 713 days (around 2 years). Confirming similar observations by Schmillen and Möller (2012) and Schmillen and Umkehrer (2017), the table makes it clear that the distribution of lifetime unemployment is highly skewed to the right: over the entire observation period almost 30% of individuals in the sample are never registered as unemployed. At the same time, 20% of individuals are registered as unemployed for at least 700 days (almost 2 years) and 10% for 1,172 days (3.2 years) or longer.

$\Delta L^{\eta }_{i}$ is constructed using the Sample of Integrated Labour Market Biographies (SIAB), a representative 2% sample of the IEB. The variable is then added to the linked employer–employee data set created from the merger of the IEB and BHP. As a consequence, information on unanticipated occupation-specific employment growth is valid not only for the specific sample used in this study but also for the German labor market as a whole. Table 2 demonstrates that with respect to the individuals in the linked employer–employee data set, $\Delta L^{\eta }_{i}$ is on average negative over the observation period; mean unanticipated occupation-specific employment growth amounts to $-\,13$% over 25 years. The variable’s standard deviation is 41%. For 10% of individuals in the sample, the overall unanticipated employment growth for their training occupation is lower than $-\,72$%, while for another 10% it is higher than $+\,41$%.

For illustrative purposes, Table 4 summarizes the five occupations with the highest and lowest employment growth, anticipated employment growth and unanticipated employment growth between 1983 and 2008. The three variables are constructed as per Eqs. 2.3 and 2.4.¹⁰

Table 4 Occupations with highest and lowest employment growth between 1983 and 2008

A number of noteworthy patterns emerge from Table 4. First, the list of occupations with the highest and lowest employment growth to a large extent reflects long-run structural change in Germany, including the seminal decline of some manufacturing sectors like the textile and leather goods industries and of mining as well as the rise of the service sector. Also, two relatively generic occupations, no or other occupations and unskilled laborers, are among the occupations with the strongest employment growth over the observation period. This reflects that over time the 1975 edition of the official Classification of Occupations faced increasing challenges to adequately capture an ever-changing labor market and occupational structure. Eventually, these challenges led to the introduction of a new Classification of Occupations in 2010.

Second, the list of occupations with the highest and lowest anticipated employment growth makes it clear that the forecasts by Blüm and Frenzel (1975) were not far off the mark in anticipating the main elements of structural change, including the tertiarization of Germany’s economy and labor market. For instance, spinners, miners and textile makers all appear among both the five occupations with the lowest employment growth and the five occupations with the lowest anticipated employment growth.

Third, the list of occupations with the highest and lowest unanticipated employment growth shows an interesting mixture of occupations where Blüm and Frenzel (1975) correctly forecast the direction of the trend in employment growth but under- or overestimated the extent of the trend and occupations where not even the direction of the trend was correctly anticipated. For instance, even though legal professions are both among the three occupations with the strongest employment growth and anticipated employment growth over the period from 1983 to 2008, they are also the occupation with the most significantly negative unanticipated employment growth. In contrast, social workers are not among the five occupations with the highest anticipated employment growth but registered among the top five occupations in terms of both employment growth and unanticipated employment growth.

As mentioned above, in addition to lifetime unemployment and $\Delta L^{\eta }_{i}$ a number of other variables are included in the multivariate analysis in Sect. 4 as controls. These are an individual’s gender, graduation year, graduation age and education as recorded in the IEB, and his or her training firm’s size and median wage as extracted from the BHP. Some regressions also include fixed effects for the training firms’ locations and/or sectors as again extracted from the BHP.

Gender is an important variable in every regression related to labor market outcomes. However, since men and women exhibit strikingly different rates of labor force attachment, it is controversial whether they should be included in the same regression of lifetime unemployment (cf. Schmillen and Möller 2012 or Schmillen and Umkehrer 2017). In this study, the baseline regressions include both genders but results from regressions run separately for men and women are also presented.

It is a stylized fact that education is closely related to the occurrence of unemployment. Thus, controlling for education is important and the regressions in this study include a dummy variable that measure whether an individual holds an Abitur. As mentioned above, for the cohorts considered here this is the case for only a very small subset of those pursuing vocational education. One would expect individuals with an Abitur to be faced with a lower amount of lifetime unemployment, ceteris paribus.

As noted in Sect. 1, Kahn (2010) and others stress the long-term effects of business cycle conditions at time of labor market entry. Therefore, the year of graduation from vocational education is used as an additional control variable. Because unemployment rates in West Germany were lower in 1981 than in 1982 and 1983 one would expect individuals who graduated from vocational education in 1981 to experience comparatively less lifetime unemployment.

Age at graduation from vocational education is another individual-level control variable that is included in all regressions. The overwhelming majority of sampled individuals were aged between 18 and 20 when they graduated from vocational education. Therefore, the following five age groups are considered: 17 years and younger, 18, 19, 20 and 21 years and older. Given that a control variable for whether an individual obtained a school-leaving certificate from an academic high school is also included in all regressions, a monotonically increasing relationship between age at graduation from vocational education and subsequent unemployment might be expected.

Two control variables related to individuals’ training firms are also used in all regressions. First, the size of the training firm, measured by the number of its employees. Second, the training firm’s medium wage. Generally speaking, in Germany the influence of labor unions is strongest in large companies. Thus, the size of the training firm proxies employees’ bargaining power which in turn is related to a higher proportion of individuals being hired by their training firm after graduation from vocational education, fewer lay-offs and a lower risk of unemployment. Accordingly, individuals trained by a larger firm are expected to face a smaller amount of lifetime unemployment, ceteris paribus. Similarly, individuals trained by a firm that pays higher wages and is thus more innovative and productive are expected to face a smaller amount of lifetime unemployment, ceteris paribus.¹¹

4 Results

4.1 Baseline estimates

Tables 5 and 6 summarize the outputs of eight different estimates of the conditional expectation function of lifetime unemployment. Even though the focus is on the impact of occupational choice on unemployment over the professional career, coefficients for control variables are also displayed. Besides, for all IV regressions the table contains first-stage coefficients and outputs for first-stage F statistics, Hansen J statistics for the endogeneity of the unanticipated occupation-specific employment growth and difference-in-Sargan tests for the instruments’ exogeneity. To correct for possible intra-regional correlations, standard errors are clustered at the county level.

Table 5 Different estimates of lifetime unemployment—OLS and Tobit regressions

As a starting point, in column (1) of Table 5 lifetime unemployment is regressed on the unanticipated occupation-specific employment growth and a constant. This is an estimation of Eq. 2.1. The regression suggests that a 100% increase in the unanticipated occupation-specific employment growth is associated with a reduction in the expected amount of lifetime unemployment by about 73 days. This relationship is statistically significant.

Table 6 Different estimates of lifetime unemployment—IV and Tobit IV regressions

Columns (2) and (3) show that the picture remains practically unchanged if one turns to Eq. 2.5 and controls for the full set of observable characteristics mentioned above, i.e., an individual’s gender, education and age at graduation from vocational education as well as the year of his or her graduation and observable characteristics of the training firm. Even if county and sector fixed effects are also included, the relationship between the unanticipated occupation-specific employment growth and lifetime unemployment remains negative. Apparently, selection based on observable individual or training firm characteristics hardly biases estimates that do not account for it.

As mentioned above, nearly 30% of the men and women who graduated from Germany’s vocational education system in 1981, 1982 or 1983 were not unemployed for a single day during the 25 years after labor market entry. To address the resulting corner solution outcome, the OLS regressions of columns (1) to (3) of Table 5 are supplemented by a Tobit model. In column (4), results are depicted for a regression that again contains the full set of individual and training firm characteristics and county and sector dummies. Instead of the Tobit model’s coefficients, the marginal effects on the observed amount of lifetime unemployment are directly displayed. As is common in the literature, the table contains the average marginal effects. The Tobit specification exhibits a slightly larger marginal effect of the unanticipated occupation-specific employment growth on lifetime unemployment than the comparable OLS regression. However, this difference is not statistically significant.

As discussed above, a conservative interpretation of the results from both the OLS and the Tobit models would probably classify them as providing descriptive evidence. In other words, they document a negative correlation between the unanticipated occupation-specific employment growth and lifetime unemployment. However, from Table 5 alone it remains somewhat dubious if there is a causal relationship between the two variables. To investigate whether such a relationship exists, the regressions summarized in Table 6 turn to estimation Eqs. 2.8 and 2.9 and instrument occupational choice with short-run fluctuations in occupation-specific labor demand on the regional and sectoral level.

Column (1) of Table 6 uses only the first instrument, column (2) only the second one and column (3) both. Effects of the unanticipated occupation-specific employment growth on lifetime unemployment are negative for all three regressions. However, if only one of the instruments is used statistical significance, precision of estimates and also the strength of the first stage are somewhat of a concern. When both instruments are jointly used, the first stage is strongly identified (as indicated by the F statistics) and the second-stage effect is relatively precisely measured as well. Additionally, the Hansen J statistics cannot reject the exogeneity of the instruments.

A comparison of the output summarized in column (3) of Table 6 with that of column (3) of Table 5 shows that the coefficient associated with the unanticipated occupation-specific employment growth continuous to be negative in the IV regression. In absolute terms, it is larger in the IV than in the OLS regression. At first glance, this might appear puzzling. However, systematic sorting into occupations according to unobserved individual characteristics is already accounted for in column (3) of Table 5 as long as it is correlated with individuals’ expectations. At the same time, as discussed above any bias caused by measurement error in $\Delta L^{\eta }_{i}$ is expected to lead to a coefficient that in absolute terms is smaller in the OLS than in the IV regression. Put differently, the coefficient associated with $\Delta L^{\eta }_{i}$ is likely larger in the IV than in the OLS regression because of measurement error. Also of note, the coefficient in the IV model is relatively imprecisely estimated and the Difference-in-Sargan exogeneity test indicates that $\Delta L^{\eta }_{i}$ is not actually endogenous.

To take account of the corner solution outcome, column (4) of Table 6 estimates Eqs. 2.8 and 2.9 with Smith and Blundell’s (1986) estimator for a Tobit model with endogenous regressors. Ultimately, this regression appears most appropriate for controlling for all potential sources of bias that might distort a regression of lifetime unemployment on occupational choice. Hence, it represents the preferred specification. The regression in column (4) confirms a statistically significant relationship between occupational choice and lifetime unemployment. Thus, one can conclude that choosing an occupation that later turns out to see low or negative employment growth has a long-term causal impact on an individual’s expected amount of unemployment.

In terms of the economic significance, a 100% increase in the unanticipated occupation-specific employment growth lowers the expected amount of lifetime unemployment by 284 days. Given that according to Table 2 the standard deviation of the unanticipated occupation-specific employment growth is 0.41 (or 41%), a one-standard deviation increase in this variable reduces unemployment by about 116 days over the professional career. This represents an average reduction in lifetime unemployment of 0.16 standard deviations.

Table 7 presents a different perspective on the economic significance of the impact of occupational choice on lifetime unemployment. It lists the predicted amount of lifetime unemployment for all deciles of $\Delta L^{\eta }_{i}$, holding all other control variables at their mean. Results reflect the highly nonlinear relationship between $\Delta L^{\eta }_{i}$ and lifetime unemployment. For instance, between the first and second decile of $\Delta L^{\eta }_{i}$, the predicted amount of lifetime unemployment decreases by 104 days, but between the fourth decile and the median of $\Delta L^{\eta }_{i}$ it varies by only 14 days. All in all, the table confirms an economically significant relationship between this study’s two key variables: the difference in the predicted amount of lifetime unemployment between the first and ninth decile of $\Delta L^{\eta }_{i}$ amounts to 536 days or 0.75 standard deviations.

Table 7 Predicted lifetime unemployment according to unanticipated occupation-specific employment growth

The economic significance of the impact of $\Delta L^{\eta }_{i}$ on lifetime unemployment is also sizable when compared to the marginal effects of some of the more important control variables (which of course should not be interpreted as capturing causal effects). For instance, predicted lifetime unemployment is 129 days lower for individuals with an Abitur as compared to those without and 162 days lower for individuals in a training firm with more than 1000 workers as compared to those in a training firm with 25 workers or less, ceteris paribus. A one-standard deviation increase in the training firm’s median wage lowers predicted lifetime unemployment by 29 days.

Apart from providing a benchmark for the economic significance of the impact of occupational choice on lifetime unemployment, the control variables’ coefficients in Table 6 demonstrate generally strong, significant and robust correlations between initial conditions and later labor market outcomes. Since men and women exhibit strikingly different rates of labor force attachment, one should probably abstain from reading too much into the relationship between gender and lifetime unemployment shown in Table 6. In contrast, some of the other control variables’ coefficients may be easier to interpret. The preferred specification of column (4) of Table 6 shows that in line with expectations having a higher graduation age is associated with more lifetime unemployment, ceteris paribus. Also in line with expectations and as already mentioned, holding a high-school diploma and graduating from a training firm with a larger workforce and higher median wage are negatively related to lifetime unemployment.

4.2 Sensitivity checks

Table 8 reports the outcomes of eight sensitivity checks that evaluate whether the finding of a statistically and economically significant impact of occupational choice on unemployment over the professional career is robust to variations of the empirical setup. The table’s reference point is the regression reported in column (4) of Table 6, i.e., the Tobit IV estimation that uses both instruments and controls for the full set of covariates.

Table 8 Different estimates of lifetime unemployment—Tobit and Tobit IV robustness regressions

In column (1) of Table 8 standard errors are clustered at the sector level. This exercise relates to the fact that the regression contains two sets of dummy variables and it is a priori unclear which of these should be the basis for clustering standard error.¹² Column (2) checks whether results are robust to the exclusion of all individuals who are not found in the data during the last four years of the observation period. This is to make sure that baseline estimates are not biased by individuals with especially advantageous or disadvantageous occupational choices being more or less likely to retire early from the labor market. Column (3) excludes all individuals who are seasonally employed during more than four calendar years out of the 25 years after their labor market entry. This is to assure that results are not purely driven by individuals who “only” have a very elevated amount of lifetime unemployment because they are seasonally employed during a large portion of their professional career.¹³

Column (4) of Table 8 evaluates the effects of altering the definition of unemployment. As an alternative to the definition of unemployment used so far it relies on the concept of nonemployment introduced by Fitzenberger and Wilke (2010). All periods not spent in employment that follow an employment spell and contain at least one spell of receiving unemployment benefits are counted as nonemployment. In column (5), instead of the county average of the unanticipated occupation-specific employment growth, the average by labor market district (Arbeitsmarktregion) is used as an instrument. Labor market districts are the administrative classification used by Germany’s Federal Labor Agency. Arguably, they are more adequate for labor market analyses than the more general-purpose counties.¹⁴

In column (6) of Table 8, trainees with an Abitur, i.e., a school-leaving certificate from an academic high school that would also enable them to pursue higher education, are excluded. This is to make sure results are not driven by the inclusion of this relatively small and peculiar group. In column (7), the sample is restricted to individuals that change occupations at least once within the first 25 years on the labor market. This sensitivity check aims to test whether changing one’s occupation is a viable way to avoid the effects of occupational choice on lifetime unemployment. Finally, in column (8) the assumption underlying Eq. 2.1 and other estimation equations that there is a linear relationship between the unanticipated occupation-specific employment growth and lifetime unemployment is relaxed. To account for possible nonlinearities, a more flexible specification is adopted where the unanticipated occupation-specific employment growth enters both as a linear term and in squared form. For computational reasons and in contrast to all other sensitivity checks, the specification in column (8) is estimated with a Tobit instead of a Tobit IV model.

As Table 8 demonstrates, results are qualitatively very robust to all eight alternative specifications presented here. No relevant coefficients ever change their signs. As is to be expected for instrumental variable regressions, the precision of the second-stage estimates is sometimes an issue. However, the second-stage estimates’ orders of magnitude stay within the same band. Maybe most notably among the various sensitivity checks of Table 8, column (7) shows that even though there is substantial mobility between occupations (656,367 out of 852,872 sampled individuals change their occupation at least once) this mobility does not reduce the importance of the unanticipated employment growth in the occupation in which the training took place. Put differently, even for those individuals that are mobile across occupations the unanticipated employment growth of the original training occupation has a causal impact on lifetime unemployment. In addition, column (8) indicates that the assumption of a linear relationship between $\Delta L^{\eta }_{i}$ and lifetime unemployment seems to be a good approximation for the relationship between the two variables, but that in fact the impact of $\Delta L^{\eta }_{i}$ becomes somewhat stronger with increases in $\Delta L^{\eta }_{i}$.

4.3 Subgroup analysis

In Table 9, the outcomes of eight regressions are displayed that evaluate whether the finding of a significant impact of occupational choice on unemployment over the professional career is homogenous across a number of different subgroups. The table’s reference point is again the regression reported in column (4) of Table 6, i.e., the Tobit IV estimation that uses both instruments and controls for the full set of explanatory variables.

Table 9 Different estimates of lifetime unemployment—Tobit IV regressions for subgroups

In columns (1) and (2) of Table 8 results are presented separately for men and women. In columns (3) to (6) the sample is divided into four groups according to individuals’ age at the start of vocational training. Group one consists of individuals aged 15 or younger, group two and three of individuals aged 16 and 17, respectively, and group four of individuals aged 18 or above. In columns (7) and (8), the sample is split into two groups according to individuals’ training occupations. The split relies on a simplified version of the highest level of the 1975 edition of the Classification of Occupations and groups individuals into those pursuing a career either in manufacturing or technical occupations or in services occupations.¹⁵

Table 9 demonstrates that the finding of a significant impact of occupational choice on unemployment over the professional career is present for both men and women but not for all other subgroups. In particular, the impact of occupational choice is negative for all age groups at the start of vocational training but only significantly so for individuals aged 18 and above. In addition to that, the impact is only negative and significant for individuals trained in services occupations but not for those striving for a career in manufacturing or technical occupations.

It can only be speculated why occupational choice at the onset of the professional career plays no or little role for younger trainees and those pursuing training in manufacturing or technical occupations. With regard to differences between trainees of different age groups, many older trainees might be more rigidly focused on a certain career path than younger ones. This might explain why older trainees are more likely to experience unemployment as a consequence of an unanticipated drop in occupation-specific employment growth.

In Germany, training in manufacturing and technical occupations has traditionally focused on a core of relatively rigid and occupation-specific skills. In contrast, training in services occupations has tended to impart more general and fungible skills such as bookkeeping and accounting, as well as customer-focused skills and other noncognitive skills that are relatively easily transferable. At first glance, it might therefore appear puzzling that unanticipated drops in occupation-specific employment growth have no apparent impacts on trainees in manufacturing or technical occupations. At the same time, Acemoglu and Pischke (1996, p. 29) note that as far as vocational training in Germany during the late 1970s and mid-1980s is concerned “the manufacturing industry [...] is the sector where training is of the highest quality and the net costs borne by the [training] firms are the largest.” In effect, the high quality of training in the manufacturing industry might have shielded trainees in manufacturing and technical occupations from negative consequences of unanticipated drops in occupation-specific employment growth.

5 Conclusions

This study analyzed the relationship between occupational choice and unemployment over the professional career. It relied on a German administrative linked employer–employee data set to document that the two are highly correlated. Using short-run fluctuations in local and sectoral occupation-specific labor demand as instruments, the main finding was that the occupational choice of graduates from vocational education has a statistically and economically significant impact on unemployment over the professional career. Graduates from vocational education who start their professional career in an occupation that in subsequent years exhibits an unanticipated drop in employment growth suffer from more unemployment in the long term. A one-standard deviation decrease in the unanticipated occupation-specific employment growth raises unemployment by about 116 days over the professional career, ceteris paribus.

This finding is hard to reconcile with the hypothesis that observed correlations between occupational choice during vocational education and unemployment over the professional career are entirely due to the sorting into occupations of individuals according to their observed or unobserved ability, skills, motivation or other characteristics. Instead, it appears that at least for individuals that go through vocational education the occupational choice at the beginning of the professional career can itself have long-run unemployment effects. Two otherwise identical individuals who simply by good or bad luck end up in two different training occupations may face very different career prospects.¹⁶

In terms of policy implications, the finding that having good or bad luck when choosing a training occupation at often very young age can lead to significantly different career prospects might serve to emphasize the importance of the public provision or financing of skills development initiatives. More specifically, the finding demonstrates that such initiatives arguably need to go beyond foundational education prior to the entry into the labor market. Instead, options for lifelong learning need to be supported over the entire professional career—either in the form of on-the-job training, specialized classroom-based retraining or a combination thereof.

This study’s main finding is in line with results by Kambourov and Manovskii (2009) and Schmillen and Möller (2012) that highlight the importance of occupation-specific human capital. It is also consistent with theoretical models by Ljungqvist and Sargent (1998, 2004, 2008) that establish a causal connection between human capital and unemployment. Finally, it lends tentative support to those like Heckman (1993) and Schneider and Zimmermann (2010) who criticize vocational education systems for forcing people to choose rather rigid occupations very early in life. However, only a direct comparison of the relative labor market advantages of vocational as compared to academic education could really corroborate this supposition. Such an exercise is not possible with the data set and identification strategy used here but would constitute interesting follow-up work.

Another interesting direction for follow-up work would be to differentiate between the impacts of occupational choice on unemployment during different phases of the professional career. Recent research by Jaimovich and Siu (2012), Foote and Ryan (2015) and others shows that changes in relative demand for certain occupational skills in general and for mid-level skills imparted by vocational education in particular are more likely to occur during recessions than to evolve gradually. Therefore, a distinction of the impacts between periods of economic expansion and recession would be particularly worthwhile.

Appendix A: sample selection, data cleansing and summary statistics

As mentioned in Sect. 2, this study focuses on 852,872 individuals that entered the West German labor market in 1981, 1982 or 1983. All of these men and women graduated from Germany’s vocational education system. East Germans are excluded, because their employment history has only been recorded in the IEB data since the early 1990s. Thus, all analyses are based on a group that is fairly homogeneous with regard to experience, training and background.

Table 10 Summary statistics for explanatory variables

While the information contained in the administrative linked employer–employee data set can generally be considered highly reliable, it is not completely free of questionable information. For example, the IEB contain a small number of occupational codes that have been documented as erroneous. In order to ensure valid and undistorted results, all main and control variables including lifetime unemployment and the unanticipated occupation-specific employment growth were carefully checked and implausible data points were replaced with missing values.

Table 10 presents summary statistics for all variables. It shows for instance that women comprise 42% of the sample and that only 3% of sampled individuals held an Abitur upon graduation from vocational education. Concerning the age at graduation from vocational education, 8% of individuals were 17 years old or younger, and 25% were 18 years old, 29% 19 years old, 20% 20 years old and 18% 21 years old or older. Almost half of trainees were in training firms with less than 25 workers (a reflection of the strong role of small- and medium-sized firms in Germany’s vocational education system), 10% in firms with 25–50 workers, 9% in firms with 51–100 workers, another nine percent in firms with 101–250 workers, 7% in firms with 251 to 500 workers, 6% in firms with 501–1000 workers and 10% in firms with 1000 workers or more. The median daily wage of the training firm was euro 41.80.