Local Effects of Education and Age Groups on Unemployment in Germany

Busch, Fabian; Ochsen, Carsten Article — Published Version Local Effects of Education and Age Groups on Unemployment in Germany Growth and Change Provided in Cooperation with: John Wiley & Sons Suggested Citation: Busch, Fabian; Ochsen, Carsten (2025) : Local Effects of Education and Age Groups on Unemployment in Germany, Growth and Change, ISSN 1468-2257, Wiley Periodicals, Inc., Hoboken, NJ, Vol. 56, Iss. 1, https://doi.org/10.1111/grow.70011 This Version is available at: https://hdl.handle.net/10419/313752 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. http://creativecommons.org/licenses/by/4.0/ Growth and Change - ORIGINAL ARTICLE OPEN ACCESS Local Effects of Education and Age Groups on Unemployment in Germany Fabian Busch 1 | Carsten Ochsen 2 1 Department of Economics, University of Rostock, Rostock, Germany | 2 Department of Labour Economics, University of Applied Labour Studies, Schwerin, Germany Correspondence: Carsten Ochsen ([email protected]) Received: 9 November 2023 | Revised: 10 September 2024 | Accepted: 11 November 2024 Keywords: ageing | human capital | labour mobility | regional unemployment | spatial interactions ABSTRACT This article provides a comprehensive analysis of how regional changes in the age and education distribution of the labour force affect local and neighbourhood unemployment rates. Based on theoretical considerations, we argue that differences in job search, separation, and commuting are key factors in group differences, and therefore, changes in relative group size affect the level of unemployment. The empirical analysis focuses on local labour markets in Germany, using a dynamic spatial panel data model. According to the estimates, an increasing proportion of young and/or low‐educated workers raises local unemployment, while larger proportions of older prime‐age and/or highly educated workers raise unemployment in neighbouring labour markets. As a result, the recent ageing and education developments in the German labour force have led to a 25% reduction in the unemployment rate. JEL Classification: R23, R12, J11, J24, J61 1 | Introduction The unemployment rate is often used as an indicator of the overall economic situation. However, it can also be seen as an average unemployment risk that takes into account different groups of people with varying levels of education and age. For instance, those with lower levels of education are more likely to be unemployed than those with higher levels of education. 1 A change in their group shares changes the overall unemployment rate, even if the unemployment rate of both groups does not change. Similarly, younger workers are more likely to be unemployed than older workers. 2 Over the past 2 decades, the OECD average has shown that unemployment rates decrease with increasing age and education levels, regardless of the size of the groups and economic cycles. This means that changes in the distribution of age and education groups are crucial factors that affect the unemployment rate in the long term. The aim of this article is to analyse how regional changes in the age and education distribution of the labour force affect local and neighbourhood unemployment rates. 3 Based on theoretical considerations, we argue that differences in job search, separation, and mobility in terms of commuting are key factors in group differences, and therefore, changes in relative group size affect the level of overall unemployment. Both age and education have a compositional impact on unemployment in the local and neighbouring regions. Our hypothesis is that the recent changes in the distribution of education and age in the German labour force account for substantially reducing the unemployment rate. Our research considerably contributes to the literature by empirically assessing compositional effects on unemployment at the regional level. We consider a spatial econometric approach and data on the local distribution of age and education. An ageing process and an increased education level of the labour This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited. © 2025 The Author(s). Growth and Change published by Wiley Periodicals LLC. Growth and Change, 2025; 56:e70011 1 of 20 https://doi.org/10.1111/grow.70011 force characterise the period under study. Our dynamic space‐ time panel data model (spatial Durbin model) reveals that a rising share of youth and/or low‐educated workers increases local unemployment in Germany. However, these groups do not affect unemployment in the neighbouring labour market. Conversely, a larger share of older prime‐age and/or highly educated workers in the surrounding area raises the unemployment rate in the local region. These groups have statistically weak negative effects on unemployment in their home area. In conclusion, the ageing of the labour force and the rising education level reduce overall unemployment. Furthermore, the current changes in age and education groups are almost equally important and account for an unemployment rate reduction of about 25%. These findings have practical implications for policymakers and labour market analysts in understanding and addressing regional unemployment. The article is organised as follows: Section 2offers a literature overview, and Section 3provides some theoretical considerations. Section 4describes the data and the econometric approach and reports and discusses the estimated results. Section 5concludes. 2 | Literature Review During the early 1970s, the young population increased in many developed countries. Today, we are witnessing a substantial rise in the proportion of older workers, which has implications for the labour market when younger and older workers are imperfect substitutes (Easterlin, Wachter, and Wachter 1978). 4 While Easterlins' cohort crowding hypothesis primarily focuses on marriage, fertility, wages, and labour market participation, Perry (1970) focussed on the relationship between population group size and employment. Based on macroeconomic data, the subsequent scholarly discussion concludes that a larger share of the working‐age population youth raises the overall unemployment rate. This is because the young often have the highest unemployment level among age groups. 5 In contrast, Shimer (2001) concludes that the labour supply of many young people can reduce the overall unemployment rate using US state‐level panel data. He argues that a higher share of young individuals in the working‐age population can lead to an increase in job creation due to their higher search intensity. A similar result is found in Nordström Skans (2005), who uses Swedish data and concludes that younger workers can benefit from belonging to a large cohort in terms of reduced unemployment. Garloff, Pohl, and Schanne (2013) conducted a study to analyse the impact of smaller labour market entry cohorts on unemployment and the direct effect of the age structure on unemployment in West Germany. They found that if the age distribution of the labour force remained unchanged, the unemployment rate would have been higher. As the size of the younger generation entering the labour market in Germany has been decreasing, the demographic change could enhance job opportunities and reduce the unemployment rate. Ochsen (2021) analyses the local effect of the age distribution of the working‐age population on unemployment. Using US county‐level data, he applies a dynamic space‐time panel data model and considers different age groups in the local and neighbouring regions. The results provide strong evidence that (spatial) age group changes are an important long‐term driver of overall unemployment change. Ageing of the working‐age population reduces overall unemployment, and the present changing age structure leads to a long‐term reduction of the US unemployment rate. A major benefit of education is that the unemployment risk decreases with an increasing education level (Mincer 1991). For example, Acemoglu (1998,2002a,2002b,2003) discusses the role of technological change and wages for shifts to more skilled and less unskilled workers. Consequently, the distribution in an economy shifts continuously towards more educated workers, similar to the ageing process. When employers prefer college graduates for jobs that require a high school degree, the relative demand for college graduates rises. Early literature that finds empirical evidence for this are Teulings and Koopmanschap (1989), Howe (1993), and van, Ours, and Ridder (1995). The evidence that larger groups of better‐educated individuals crowd out less‐educated individuals also has consequences for the low‐educated unemployed, as, for example, Wolbers (2000), Arberg (2003), Gesthuizen and Wolbers (2010), and Abrassart (2015) point out. Therefore, creating low‐skilled jobs may not necessarily improve the employability of low‐skilled workers. The literature discussed analyses either the effects of ageing or education on unemployment. An exception is a study by Biagi and Lucifora (2008), who examine the effects of demographic and educational changes on the evolution of unemployment rates for a panel of European countries. Their research findings suggest that demographic and education changes affect young and adult workers and more or less educated individuals differently. The study reveals that changes in the population age structure (baby bust) positively relate to youth unemployment rate, whereas changes in the educational structure (education boom) reduce unemployment among the more educated. Finally, national‐level data are not appropriate for covering within‐country mobility. In small local regions, spatial mobility (in terms of commuting) is related to local labour market tightness. In addition, spatial mobility is different for age groups (younger workers is more mobile than older workers) and education groups (the highly educated are more mobile than low‐ educated workers). 6 The effects of a changing age structure in the local labour force on unemployment in a spatial interaction model are considered only in the study by Ochsen (2021). Concerning educational distribution and spatial mobility, Kulu, Lundholm, and Malmberg (2018) report that changes in population composition, mainly increased enrolment in higher education, account for much of the rising spatial mobility using data for Sweden. Using French data, Lemistre and Moreau (2009) find that returns to spatial mobility for men increase with education. 3 | Theoretical Considerations Shifts in the labour force to larger shares of educated and older workers affect the unemployment rate when group‐specific 2 of 20 Growth and Change, 2025 unemployment rates differ. To analyse this, we divide the labour force into two groups, Group 1 and Group 2. Age Group 1 represents the younger workers (y), and age Group 2 represents the older workers (o). For education, we distinguish between low‐educated (μ) as Group 1 and high‐educated (h) as Group 2. For simplicity, the labour force consists of these two groups only (either age or education), with a labour force share of pfor the first group and 1 −pfor the second group. Workers are either employed or unemployed; if they are unemployed, we assume they are seeking a new job. The aggregated unemployment rate uconsists of the group‐specific rates weighted at the respective labour force share: u=pu 1 þ(1 −p)u 2 . In the standard search and matching framework, equilibrium unemployment is explained by two flow rates: the separation rate sand the job finding rate f(θ). While sis the risk of job loss and the corresponding flow to unemployment, f(θ) is the probability of an unemployed person finding new employment (with market tightness θ). Given that u,u 1 , and u 2 are in equilibrium, we have: u∗=pu∗ 1+(1−p)u∗ 2=ps1 s1+f1(θ)+(1−p)s2 s2+f2(θ)(1) It is easy to see that a rising pincreases (decreases) u* when u∗ 1>u∗ 2(u∗ 1<u∗ 2). In addition, for a given distribution of the groups, the unemployment rate rises if at least one group's unemployment rate increases. Finally, the flow rates into and out of unemployment can explain the group‐specific unemployment rate change. To point out that group‐specific flows and unemployment rates matter, we use aggregated stock and flow data provided by the German Federal Employment Agency. Table 1provides average values for youth (15–24 years) and older workers (50–64 years) as well as for low (no apprenticeship) and high‐educated workers (academic education) for the period 2008–2018 using monthly data. u∗ iis calculated according to Equation (1) and u i is the usual unemployment rate, calculated using the group‐ specific stock of the unemployed divided by the group‐specific stock of the labour force. The separation risk and job‐finding rate are above average for youth. Separation is above the average for the low‐educated workers, but job finding is below. This difference is the reason for the vast differences between these two groups in terms of the unemployment rate. Older workers and those who are highly educated have separation rates below the average. Concerning job finding, these two groups differ; the highly educated are above the average, and older workers are below. Again, this difference explains the difference in unemployment rates. Another important finding is that the age groups have a clear, dynamic pattern. Compared to older workers, youth is more often unemployed but also faster reemployed. In contrast, highly educated workers benefit from favourable flow rates, while the low‐educated suffer from adverse flow rates concerning the unemployment rate. Within all groups, both average unemployment rates in Table 1are very similar, notwithstanding that the equilibrium unemployment rate is a simplified concept. Hence, the flow rates are useful in explaining what happens if the share of age or education groups changes. For example, from an aggregated perspective, we can argue that less low‐educated and more high‐educated workers mean less separation and faster job finding, which decreases the unemployment rate. However, more older and less younger workers result in less separation but also slower job finding. Labour market dynamics obviously decline, but the effect on the unemployment rate is ambiguous. 7 The effects are more complex when considering regions (e.g., counties) and allowing workers to commute between neighbouring regions. 8 Younger workers are regional more mobile than older workers, and highly educated workers are more mobile than low‐educated workers. To accommodate this empirical observation, we focus on regional labour market interactions. The search rate σ=uþeis the sum of unemployed and employed job seekers divided by the labour force, with e≤1−u. From a regional perspective, it is obvious that people apply not only for jobs in their home region but also in surrounding regions. In this case, workers commute between their home and workplace region. We refer to commuting and inter‐regional searches as mobility; hence, this definition excludes moves from one region to another. To maintain the model's simplicity, we consider job seekers and vacancies only from the local region land regions adjacent to l, which we treat as one homogenous region, n. The tightness (θ) of the local labour market is given by TABLE 1 |Average flow rates and unemployment rates in Germany. Separation rate Job finding rate u∗ iu i Youth 0.038 0.416 8.4 8.1 Older workers 0.018 0.172 9.8 10.2 Low educated 0.073 0.216 25.7 25.7 High educated 0.011 0.25 4.3 4.3 All 0.023 0.228 9.4 9.5 Note: Monthly data are taken from statistics of the Federal Employment Agency. Job‐finding rates are calculated as the ratio of flows from unemployment to employment in the previous month, and separation rates are calculated as flows from employment to unemployment in the previous month. Equilibrium unemployment rates are calculated according to Equation (1), and the (normal) unemployment rate is calculated as the number of unemployed divided by the labour force. Period: January 2008 to December 2018. 3 of 20 θl=vl/(ul+el+u ∼n+e ∼n)=vl/(σl+σ ∼n), and the tightness of the adjacent districts' labour market is given by θn=vn/(un+en+u ∼l+e ∼l)=vn/(σn+σ ∼l), where v l (v n ) denotes the local (neighbourhood) vacancy rate and ~ represents spatial search activities. All job seekers apply for jobs in their home region. Because younger workers and more educated workers are more mobile, the number of regional mobile job applicants depends on the age and education structure of the job seekers. Only some of the older and low‐educated job seekers from neighbouring regions apply for jobs in the local region. We refer to σl=plσl 1+(1−pl)σl 2and σn=pnσn 1+(1−pn)σn 2as local search rates and σ ∼n=[pnσn 1+(1−pn)σn 2α]Ln/Lland σ ∼l=[plσ1 l+(1−pl) σ2 lα]Ll/Lnas spatial search rates. All workers resident in the local region, L l , are normalised to 1. The rate σ ∼n, related to the labour force in the local labour market, has the same denominator as σ l .σ ∼nand σ n differ because they are related to different labour force sizes, σ ∼nto L l and σ n to L n . The share of Group 2 (low‐educated or older workers) job seekers is larger in their resident region. The mobility weighting factor α, with 0 ≤α<1, accommodates the limited spatial mobility of older and low‐educated workers; hence, σn 2>σn 2α. The differences between σ ∼land σ l are analogous. This affects the distribution of the job seekers available to local firms: plσl 1 σl+σ ∼n+pnσn 1 σl+σ ∼n≡pl. Hence, the job seeker structure depends on the group distribution (education or age) of the labour force in both regions. Job seekers from the local region find, on average, new employment at the rate fl i(θ l ,pl)+fn i(θ n ,pn) because of the spatially mobile search activities. From this, it follows that the spatial correlation of unemployment rates is positive, and the (spatial) correlation of vacancy and unemployment rates is negative. Separations can differ across groups and regions. Finally, the local labour force, L l , can be subdivided into three groups: local unemployed u l , residents employed in the local region ω l,l , and residents employed in the neighbour region ω l,n . Since L l =1, we have u l þω l,l þω l,n =1. The local unemployment rates of a group evolve according to separation and job finding, with i=[y,o] or i=[μ,h]: ˙ ul i=sl i(1−ωl,n i−ul i)+sn iωl,n i−fl i(θl,pl)ul i−fn i(θn,pn)ul i. sl i(1−ωl,n i−ul i)is the group‐related flow into unemployment from local employment. sn iωl,n iis the group‐related flow into local unemployment from jobs in the neighbouring region. On the right‐hand side, the last two terms are the probabilities of transition into a new job in the local and neighbouring labour market. With ˙ ui=0 and the summation of the two unemployment rates weighted at the respective local population proportions, p l and (1 −p l ), we obtain the local equilibrium unemployment rate: ul=ul 2+pl(ul 1−ul 2) =sl 2+(sn 2−sl 2)ωl,n 2 sl 2+fl 2(θl,pl)+fn 2(θn,pn) +pl(ul 1−ul 2), includes spatial and (spatial) group effects. The second term on the fraction line indicates that local unemployment increases as the number of spatially mobile workers increases and sn 2>sl 2, with ωl,n 2as the share of residents employed in the neighbouring region (n). There are two channels for the group‐related effects: the first effect is ‘hidden’ in the (spatial) job‐finding rates, and the second is related to the differences in group‐related unemployment rates. This second term disappears if ul 1=ul 2. For ul 1>ul 2(ul 1<ul 2), an increasing proportion of Group 1 workers increases (decreases) overall separation and unemployment. The first effect contains group‐related matching efficiency and mobility effects on the neighbouring labour market. This effect means that the more Group 1 workers are in the neighbouring region, the lower the local market tightness and, hence, the lower the probabilities of transition into a new job for local workers (plis the share of job seekers available to local firms). Thus, the proportion of group‐specific workers in the local and surrounding labour markets is important to the local unemployment rate. In the empirical part, we estimate regional panel data with a spatial panel model to analyse this empirically. 4 | Empirical Analysis This section analyses the relationship between changes in age and education composition of the labour force and the unemployment rate using data for Germany. The previous section shows the stock and flow data provided by the German Federal Employment Agency (Table 1). In this section, we first discuss the development of the groups that are considering using OECD data for Germany. The education groups are now differentiated by the ISCED classification, which is slightly different from the German Federal Employment Agency classification. Figure 1 provides the evolution of two age group shares and two education group shares for 1999–2018. The shares of prime‐age workers (25–49 years) and medium‐educated (ISCED 3–4) are not displayed. The trend in the data shows that the shares of the younger and low‐educated decline over time, while the shares of older and highly‐educated workers increase. Given that the unemployment rates of older and highly educated workers are lower, these shifts must decrease the overall unemployment rate. Focussing only on younger or low‐educated workers, as typically done in the literature, is misleading because trends in other groups are not considered. The evolutions of German unemployment rates for age and education groups are provided in Appendix A(Figures A1 and A2). Based on this comparison, analysing the relationship between different age and education groups and unemployment appears 4 of 20 Growth and Change, 2025 meaningful. Since the analysis of macroeconomic data would provide no substantial new findings, regional data will be applied because they allow for considering a more differentiated pattern. Therefore, the econometric analysis will utilise county‐ level data (NUTS‐3 level). 4.1 | Data and Econometric Framework We use the German Sample of Integrated Labour Market Biographies (SIAB‐7514), a random sample drawn from the Integrated Employment Biographies (IEB). This data source entails individual data on labour market biographies. 9 Covering 16 years from 1999 to 2014, we converted the raw data into monthly segments. Hence, we analyse the period from January 1999 until December 2014. We computed our variables at the individual level and aggregated them at the administrative district level. This gives us a strongly balanced panel consisting of 402 cross‐section units (counties) and 77,184 observations for different shares of the local labour force. We refer to Data Description section in Appendix Afor a detailed description of the editing process. Table 2describes how the variables are generated and provides a summary statistic for these variables. We will use the age group shares in different combinations to deal with different reference groups. The two different types education and schooling will be used separately. Education is related to formal education after schooling and consists of no apprenticeship, apprenticeship, and academic degree. Schooling is related to the last school leaving certificate and is separated into no certification, certification without a university entrance qualification, and high school degree. In both cases, the reference in the regressions is the medium group (in terms of schooling, it is certification without a university entrance qualification, and in terms of education, it is apprenticeship). We consider only employed and unemployed individuals of the working‐age population between 15 and 64 years. This is related to the retirement age in Germany. In addition, we do not consider other individual characteristics like gender or work experience. Also, we abstain from interacting age with education groups, such as subdividing youth into three groups: youth without apprenticeship, youth with apprenticeship, and youth with academic degrees. 10 This is caused by the limited number of individuals at the regional level. In some cases, we would not have enough individuals within a specific subgroup for our econometric analysis. Therefore, we decided to use only age groups and education groups in the labour force. We are primarily interested in the effects of a change in group compositions in the local and neighbouring regions on local unemployment. For example, the local youth share captures group‐specific job finding and separation in the local region. In addition, we consider the effect of changes in youth share in the neighbouring region on local unemployment. Since the young in both regions hold, on average, the same job‐relevant characteristics, we do not argue that, for example, younger workers in neighbouring regions are more productive than those in the local region. This is not possible because, in the estimates, every share is considered as a local region and as a neighbour region (I am the neighbour of my neighbour). However, the neighbouring region's youth needs to be spatially mobile (in terms of commuting) to work in the local region. 11 This is why we want to consider youth share effects of both the local and the neighbouring regions on local unemployment in the estimates. In addition, we also want to control for other age groups to differentiate between these groups and the reference group of older workers. For education, we argue in the same way. Here, we expect that the low‐educated group is less mobile. Since this group is often small, we use the medium‐educated group as a reference. To consider all these aspects, we use a flexible spatial econometrics approach: The spatial Durbin model. In the following, we describe the structure of the model. To account for additional unobserved time and spatial varying effects at the local level, time lagged and spatial lagged effects of the dependent variable ln u it (unemployment rate in logarithm) FIGURE 1 |Labour force share for age and education in Germany. 5 of 20 are considered (Equation 2). To generate spatially lagged counterparts, we constructed a spatial weight matrix, W, that indicates the contiguity of regions and defined contiguity between two regions as those that share a common border. The matrix has the entry 1 if two regions share the same border and 0 otherwise. Then, we row normalise W, which ensured that all weights were between 0 and 1 and that weighting operations can be interpreted as an average of the neighbouring values. ln u i,t−1 is the time lagged dependent variable and γthe autoregressive time dependence parameter. Wln u it generates the average values of the regions adjacent to region i, and λis the spatial dependence parameter—the spatial lagged effect of the dependent variable. Wln u i,t−1 is the combined spatial and time lagged dependent variable, and πis the spatio‐temporal diffusion parameter. The inclusion of the spatial and time lagged dependent variable could serve as a control for omitted variables or at least reduce omitted variable bias (LeSage and Pace 2009). We will discuss the issue of endogeneity in Section 4.4. To sum up, we consider a spatial and time dynamic model that is also known as the dynamic spatial Durbin model (with time and fixed effects): ln uit =γln ui,t−1+λW ln uit +πW ln ui,t−1 +α1lnagegroup1it +β1Wlnagegroup1it +α2lnagegroup2it +β2Wlnagegroup2it +α3lnedugroup1it +β3Wlnedugroup1it +α4lnedugroup2it +β4Wlnedugroup2it +ci+θt+eit (2) where ln u it , ln agegroup it , ln edugroup it and e it are stacked Tn �1 column vectors, Wis a row normalised n�nspatial weights matrix that is nonstochastic and generates the spatial dependence between cross‐sectional units, c i are regional, and θ t are time effects. ln agegroup1 is the share of the first age group (e.g., youth) in the local region, and Wln agegroup1 is the average share of the same group in the neighbouring regions. The same applies to the second age group and the two education groups (edugroup). 12 The bias‐corrected quasi maximum likelihood approach provided by Yu, de Jong, and Lee (2008) is considered for the dynamic models. 13 The effects of the time and spatial lagged dependent variable will not be discussed below. 14 However, these lags help afterwards to calculate the dynamic long‐run effects. In all regressions, county‐cluster robust standard errors are considered. The parameters αand βin Equation (2) cannot be interpreted as elasticities or partial derivatives due to spillover effects. 15 Therefore, we first provide the estimated coefficients and, subsequently, the resulting elasticities. Because of their limited mobility, not all older workers or less educated workers in the neighbouring region apply for jobs in the local region, and therefore, the spatial group shares serve mostly as a proxy variable for mobility in terms of commuting. For example, let us assume that prime‐age workers are more attractive to firms than other age groups. An increase in the neighbouring prime‐age share induces more job applications at firms in the local region. This, in turn, decreases search costs and increases the vacancy rate. However, this also decreases the local market tightness and the probability of transitioning into a new job for local job seekers. This effect is likely larger than the effect on vacancies (more jobs). With respect to the parameter β, we expect a positive effect. In contrast, the parameter αis TABLE 2 |Variable description and basic statistics. Variable Description Obs Mean SE Min Max Unemployment rate Number of unemployed individuals in the district labour force to all individuals in the district labour force 77,184 0.0996 0.055 0.008 0.450 Age groups Youth Number of individuals in the district labour force aged 15–24 to all individuals in the district labour force aged 15–64 77,184 0.069 0.018 0.012 0.152 25–39 years Number of individuals in the district labour force aged 25–39 to all individuals in the district labour force aged 15–64 77,184 0.356 0.052 0.217 0.512 25–49 years Number of individuals in the district labour force aged 25–49 to all individuals in the district labour force aged 15–64 77,184 0.666 0.045 0.476 0.785 40–49 years Number of individuals in the district labour force aged 40–49 to all individuals in the district labour force aged 15–64 77,184 0.310 0.032 0.184 0.411 Education/schooling groups No Apprenticeship Number of individuals in the district labour force without vocational education to all individuals in the district labour force 77,184 0.134 0.057 0 0.337 Academics Number of individuals in the district labour force with an academic education to all individuals in the district labour force 77,184 0.115 0.063 0.016 0.518 No Graduation Number of individuals in the district labour force with no graduation to all individuals in the district labour force 77,184 0.013 0.010 0 0.090 High School Number of individuals in the district labour force with university entrance qualification to all individuals in the district labour force 77,184 0.183 0.092 0.030 0.691 Note: Data are taken from the SIAB‐7514 and aggregated to a balanced monthly county level from January 1999 to December 2014. 6 of 20 Growth and Change, 2025 negative if the local share of prime‐age workers increases, and this age group is more attractive to firms overall. If the spatial effects of the considered group structures are essential, we have to consider the bias on αif we neglect β. Let ω be the parameter for the local effect when the spatial effect is neglected. The standard result is then ω=αþβδ, where δ measures the covariance of the local and the spatial age or education structure. The latter is positive in the data, and we expect βto be positive, which yields a positive bias on ω. Concerning the statistical relevance of our estimates, we provide county‐cluster robust standard errors to control for heteroskedasticity, serial correlation and cross‐sectional dependence in the residuals. We consider the False Positive Risk (FPR) to provide information for statistical evidence. In contrast to p‐ value, the FPR measures the probability of the null hypothesis being true (Colquhoun 2017,2019). For a discussion of the misinterpretation of p‐value, see, for example, Wasserstein and Lazar (2016). We consider FPR =0.05 (equals p‐value of 0.0034) and FPR =0.01 (equals p‐value of 0.0005). For the computation of the FPR, we refer to Computation of the False Positive Risk section in Appendix A. 4.2 | Results Table 3provides different basic specifications of Equation (2). For reasons of comparison, regressions (1), (5), and (9) are simple fixed and time effects models. In regressions (2), (6), and (10), only the spatial lagged dependent is considered (γ=π=0), while in (3), (7), and (11), the time lagged effect is also included (π=0). In (4), (8), and (12), all lagged effects of the dependent are considered. Since all spatial and time lagged effects provide strong empirical evidence, we prefer (4), (8), and (12) as the best specification. 16 We consider only one group in these estimates because the focus here is on model specification in general. However, a short discussion of the results will help us relate our findings to the existing literature. The estimated elasticity of the standard fixed and time effects estimates can be compared with the long‐run elasticities provided below. While (1) and (5) are in line with the literature, regression (9) provides no reliable estimates. This is because the share of those who graduate without a school leaving certificate is very small (about 1%). Therefore, the results of regression (9)–(12) will not be interpreted further. Concerning regressions (2)–(4), we find empirical evidence for the local effect of the youth share on unemployment. Due to the strong empirical evidence for lagged dependent effects, we rely most on regression (4). In this case, the spatial effect of the youth share provides no empirical evidence. The regressions (6)–(8) consider formal education. Independent of the specification, we find that the local share of those in the labour force with no apprenticeship is positively related to the unemployment rate, while the spatial effect is negative. Hence, compared to the reference group, this group is less attractive to firms. TABLE 3 |Basic results for age, education, and schooling. Dependent variable: log unemployment rate Reference group Age 25–64 At least apprenticeship At least secondary education Considered groups (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) Log youth 0.193 b 0.123 b 0.037 b 0.039 b (0.016) (0.013) (0.003) (0.003) W(log youth) 0.078 a −0.016 −0.005 (0.024) (0.006) (0.005) Log no apprenticeship 0.316 b 0.253 b 0.079 b 0.078 b (0.021) (0.018) (0.005) (0.005) W(log no apprenticeship) 0.020 −0.042 b −0.030 b (0.031) (0.008) (0.007) Log no graduation 0.002 0.001 0.001 0.001 (0.003) (0.002) (0.001) (0.001) W(log no graduation) 0.009 a 0.001 0.001 (0.003) (0.001) (0.001) Spatial lag (λ) 0.477 b 0.162 b 0.241 b 0.478 b 0.166 b 0.244 b 0.499 b 0.172 b 0.243 b Time lag (γ) 0.798 b 0.809 b 0.789 b 0.800 b 0.804 b 0.813 b Spatial‐time lag (π)−0.116 b −0.113 b −0.112 b Within R 2 0.808 0.158 0.919 0.919 0.814 0.601 0.920 0.920 0.803 0.133 0.919 0.918 Observations 77,184 77,184 76,782 76,782 77,184 77,184 76,782 76,782 77,184 77,184 76,782 76,782 Note: Spatial lag, time lag, and spatial‐time lag refer to the dependent; all regressions include fixed and time effects; county‐cluster robust standard errors are in parentheses; period: monthly data for January 1999 to December 2014; balanced county‐level panel. a FPR ≤0.05. b FPR ≤0.01. 7 of 20 Table 4considers groups of different ages and formal education (without apprenticeship, with apprenticeship, academic degree). In addition to the youth share, a second and third age group is added with different age group ranges and reference groups. We do this to additionally control for possible interactions between the education and age groups. 17 The educational reference is always the group with apprenticeship. 18 While we find empirical evidence for the local youth share effect in all specifications, the spatial effect is not statistically relevant. For the age group 25–39 years, we find neither empirical evidence for the local nor spatial effect. 19 When the reference age group is 50–64, the age group 25–49 has a positive spatial effect (regression (3)). In regression (4), we subdivide this age group into 25–39 and 40– 49 years. Here, we find positive empirical evidence only for the spatial effect of the age group 40–49 years. Older prime‐age workers (40–49 years) seem to be more mobile and competitive than the reference (50–64 years) and affect the market tightness. 20 In addition, when we consider the direction of all age cohort effects between 25 and 49 years, the local effects are negative, while the spatial effects are positive. This finding indicates that this age group is more attractive than the reference cohort. We find the opposite direction of the effects for the young, which means that the reference is more attractive to firms than young workers. TABLE 4 |Results for age and education. Dependent variable: log unemployment rate Reference apprenticeship and age group 25–64 40–64 50–64 Considered groups (1) (2) (3) (4) Log youth 0.036 b 0.036 b 0.034 b 0.034 b (0.003) (0.003) (0.003) (0.003) W(log youth) −0.013 −0.011 −0.007 −0.010 (0.006) (0.007) (0.006) (0.007) Log 25–39 0.002 −0.003 (0.010) (0.012) W(log 25–39) 0.010 0.040 (0.020) (0.021) Log 25–49 −0.014 (0.020) W(log 25–49) 0.119 b (0.038) Log 40–49 −0.005 (0.011) W(log 40–49) 0.072 b (0.021) Log no apprenticeship 0.077 b 0.077 b 0.078 b 0.078 b (0.005) (0.005) (0.005) (0.005) W(log no apprenticeship) −0.030 b −0.030 b −0.030 b −0.029 b (0.008) (0.008) (0.008) (0.008) Log academics −0.009 −0.009 −0.009 −0.011 (0.006) (0.006) (0.006) (0.006) W(log academics) 0.049 b 0.050 b 0.044 b 0.041 b (0.009) (0.009) (0.009) (0.009) Spatial lag (λ) 0.240 b 0.240 b 0.240 b 0.239 b Time lag (γ) 0.796 b 0.796 b 0.795 b 0.795 b Spatial‐time lag (π)−0.120 b −0.120 b −0.122 b −0.122 b Within R 2 0.919 0.920 0.920 0.920 Observations 76,782 76,782 76,782 76,782 Note: Spatial lag, time lag, and spatial‐time lag refer to the dependent; all regressions include fixed and time effects; county‐cluster robust standard errors are in parentheses; period: monthly data for January 1999 to December 2014; balanced county‐level panel. a FPR ≤0.05. b FPR ≤0.01. 8 of 20 Growth and Change, 2025 participation in programs of active labour market policies (in the data since 2000). These data come from different sources and are merged in the IEB. The suffix 7514 stands for the panel version covering 1975 until 2014, with both years included in the panel. Our main goal was to reshape the raw data in a way that enabled us to distinguish between employment and unemployment at the administrative district level (cross‐section) monthly (time unit) throughout the entire period of the analysis. We use the territorial allocation from December 31, 2014 for our analysis, which means we have 402 cross‐ section units (Kreise). We analyse the period from January 1999 to December 2014, 192 months. Defining employment and unemployment was the first step towards achieving this particular panel structure. Since the SIAB‐7514 consists of individual data in its raw version, we define an individual as employed if and only if the individual reports the characteristic attribute 101 in the employment status variable. In turn, an individual is deemed unemployed if and only if it reports one of the characteristic attributes 1,2,31,32,41,51 in the employment status variable. 26 To aggregate the individual data to a monthly administrative district level, we had to edit the original panel substantially. In the first step, we convert the raw data to sequential data. The SIAB is organised in spells. Each spell has a commencing date and an end date, which are precise to the day. The dates mark the beginning and the end of an individual episode. The term ‘precise to the day’ implies that a spell may start and end on any given day in any given month in any given year between January 1999 and December 2014. In the raw version, spells can overlap with each other. Hence, spells for the same individual may cover the same period or overlap in part. Therefore, creating sequential data means organising the spells so that there will (a) be no congruency or partial overlapping and (b) that each spell starts exactly one day after the end date of the previous spell if the same individual is considered. We took the STATA code to generate sequential data from the data report for the SIAB‐7514. 27 Once the sequential data structure was established, we created a variable reporting the single year of the individual spell using the year specified in the variable that marked the beginning of an episode and erased all spells that started and ended before 1999. Next, we pay attention to those spells that overlapped years. That is, spells that started in 1 year and ended in the following year or the year after that. We focus solely on spells that cover a maximum 2‐year span based on the year reported in the column for the start date of the episode. All spells exceeding this limit were dropped from the panel. This is because the actual number of these spells is very small. Less than 3% of the entire panel is affected. Since these spells almost exclusively report attributes in the variable depicting the employment status we do not use for our analysis, almost all those spells are dropped later. The remaining spells, which overlapped 1 year or more, are split into several episodes to attach a unique value for the year variable to each episode. We then drop those spell episodes that started before 1999. For example, if a spell started in 1998 and ended in 2000, we duplicated the original spell two times, ending with three identical spells. Each spell is assigned to a year, that is, 1998, 1999, and 2000. We kept the two latter spells and dropped the first. Once the spells are in order, we split the panel into individual years, effectively giving up the panel structure. In the next step, we develop a code that enables us to brand an individual as employed or unemployed in any given month from January 1999 to December 2014. We employ a syntax that operates in a way that an individual is assigned to the status of being employed or unemployed for a given month if the individual status exceeds half of a given month. For example, a person reports an episode of employment in a given year that starts on January 1st and ends no later than March 15th. This person is assigned to be employed throughout January and February of that year. For March, however, the individual could be assigned either status. In March, an individual will be employed if the consecutive episode starts the day after the previous episode and experiences a change in any variable, but the variable indicates the employment status. On the other hand, an individual is unemployed in March if the variable indicates a change in employment status from 101 to any number in 1,2,31,32,41,51. We create our variables once we have transformed the individual episodes into monthly episodes. We recode the original industry branches in the panel by categorising them into 13 branches of economic activity using the standards provided in the FDZ data report of the Sample‐of‐ Integrated‐Labour‐Market‐Biographies Regional‐File 1975–2010 (SIAB‐ R 7510). 28 We computed the respective shares of employment and unemployment for each combination of cross‐section and time unit for all variables. Finally, we aggregate the individual data at the administrative district level and receive a strongly balanced panel, which is the basis of our empirical analysis. Computation of the False Positive Risk The false positive risk (FPR) was introduced by Colquhoun (2019,2017) and measures the probability that the result occurred by chance P(H 0 | data). The approach is based on the Bayes theorem that we express in odds: posterior odds on H1=Bayes factor ×prior odds This is equal to P(H1|data) P(H0|data)=P(data |H1) P(data |H0)×P(H1) P(H0) Following Colquhoun, the Bayes factor becomes a likelihood ratio (LR), and the prior odds can be expressed using the probability that there is a real effect, P(H 1 ): P(H 1 )/(1 −P(H 1 )). Among others, Sellke, Bayarri, and Berger (2001) provide an approach to calculate the LR based on the p‐value: LR =1/(−ep log (p)). However, this measure can be considered only as long as p<1/e, with eas Euler's number. Taking things together and considering P(H 0 | data) =1−P(H 1 | data) gives us the FPR: FPR =1 1+1 −ep log(p) P(H1) 1−P(H1) Applying the FPR approach requires to specify P(H 1 ) first. However, specifying the prior probability in regression analysis is difficult, and we should always be careful when defining this unknown number. We use P(H 1 )/(1 −P(H 1 )) =0.5/(1 −0.5) =1, which means that both probabilities have the same weight. This is equal to a 50:50 chance for a real effect specified before the data are analysed. This seems reasonable when we do not know what to choose or are open to the results. When the prior probability of a real effect is 0.5, the FPR is much larger than the corresponding p‐value, and, for example, p=0.05 is equal to a FPR of 0.2893. 15 of 20 FIGURE A1 |Unemployment rates by age groups in Germany. FIGURE A2 |Unemployment rates by education groups in Germany. 16 of 20 Growth and Change, 2025 FIGURE A3 |First difference of shares and lagged first difference of unemployment rate. 17 of 20 TABLE A1 jFurther results using age, education or schooling. Age Reference group 40–64 50–64 50–64 Apprenticeship Secondary education Considered groups (1) (2) (3) (4) (5) Log youth 0.041 b 0.039 b 0.038 b (0.003) (0.003) (0.003) W(log youth) 0.001 0.001 −0.006 (0.006) (0.005) (0.006) Log 25–39 0.016 0.013 (0.010) (0.010) W(log 25–39) 0.034 0.053 (0.020) (0.021) Log 25–49 0.023 (0.018) W(log 25–49) 0.192 b (0.034) Log 40–49 0.004 (0.011) W(log 40–49) 0.083 b (0.019) Log no apprenticeship 0.082 b (0.005) W(log no apprenticeship) −0.022 (0.008) Log academics −0.006 (0.006) W(log academics) 0.059 b (0.009) Log no graduation 0.001 (0.001) W(log no graduation) 0.001 (0.001) Log high school −0.012 (0.006) W(log high school) 0.069 b (0.009) Spatial lag (λ) 0.241 b 0.239 b 0.239 b 0.240 b 0.238 b Time lag (γ) 0.809 b 0.807 b 0.807 b 0.799 b 0.811 b Spatial‐time lag (π)−0.116 b −0.120 b −0.121 b −0.116 b −0.119 b Within R 2 0.919 0.919 0.919 0.905 0.894 Note: Dependent variable: log of unemployment rate; spatial lag, time lag, and spatial‐time lag refer to the dependent; all regressions include fixed and time effects; county‐cluster robust standard errors are in parentheses; period: monthly data for January 1999 to December 2014; balanced county level panel. a FPR ≤0.05. b FPR ≤0.01. 18 of 20 Growth and Change, 2025 TABLE A2 jShort‐run and long‐run elasticities: age and schooling. Short‐run elasticities Long‐run elasticities Considered groups Direct Indirect Total Direct Indirect Total Dependent variable: log unemployment rate Table 5: Regression (2): reference: secondary education and 40–64 years Youth 0.043 −0.008 0.035 0.234 0.157 0.391 (0.003) (0.008) (0.009) (0.020) (0.149) (0.161) 25–39 0.016 0.006 0.022 0.091 0.168 0.259 (0.010) (0.025) (0.027) (0.057) (0.363) (0.393) No Graduation 0.001 0.001 0.001 0.003 0.007 0.010 (0.001) (0.001) (0.001) (0.004) (0.013) (0.015) High School −0.011 0.076 0.065 −0.020 0.765 0.745 (0.006) (0.011) (0.011) (0.037) (0.309) (0.327) Table 5: Regression (4): reference: secondary education and 50–64 years Youth 0.040 −0.014 0.026 0.214 0.051 0.265 (0.003) (0.008) (0.009) (0.019) (0.105) (0.117) 25–39 0.018 0.035 0.053 0.115 0.429 0.544 (0.010) (0.027) (0.028) (0.059) (0.323) (0.351) 40–49 0.004 0.072 0.076 0.055 0.713 0.768 (0.011) (0.023) (0.022) (0.057) (0.287) (0.304) No Graduation 0.001 0.001 0.001 0.003 0.006 0.008 (0.001) (0.001) (0.001) (0.004) (0.012) (0.015) High School −0.016 0.060 0.044 −0.061 0.504 0.443 (0.006) (0.011) (0.012) (0.036) (0.172) (0.186) Note: Direct effects come from the local region, and the indirect effects come from the neighbouring regions. Long‐run effects cumulate feedback over the period considered. Robust standard errors are in parentheses; period: monthly data for January 1999 to December 2014; balanced county‐level panel; observations: 76,782. 19 of 20 TABLE A3 jResults for age and education with second order neighbourmatrix. Dependent variable: log unemployment rate Reference age group 25–64 40–64 50–64 Considered groups (1) (2) (3) (4) Log youth 0.026 b 0.025 b 0.024 b 0.024 b (0.003) (0.003) (0.003) (0.003) W(log youth) −0.012 −0.011 −0.008 −0.012 (0.009) (0.011) (0.010) (0.011) Log 25–39 −0.007 −0.013 (0.010) (0.012) W(log 25–39) 0.011 0.041 (0.033) (0.036) Log 25–49 −0.035 (0.020) W(log 25–49) 0.137 (0.053) Log 40–49 −0.012 (0.011) W(log 40–49) 0.070 (0.025) Log no apprenticeship 0.079 b 0.080 b 0.080 b 0.080 b (0.006) (0.006) (0.006) (0.006) W(log no apprenticeship) −0.059 b −0.060 b −0.058 b −0.058 b (0.010) (0.010) (0.010) (0.010) Log academics −0.013 −0.012 −0.011 −0.011 (0.006) (0.006) (0.006) (0.006) W(log academics) 0.054 b 0.054 b 0.045 a 0.040 (0.014) (0.014) (0.014) (0.015) Spatial lag (λ) 0.474 b 0.474 b 0.473 b 0.473 b Time lag (γ) 0.788 b 0.788 b 0.788 b 0.788 b Spatial‐time lag (π)−0.304 b −0.304 b −0.306 b −0.307 b Within R 2 0.919 0.919 0.920 0.920 Observations 76,782 76,782 76,782 76,782 Note: Spatial lag, time lag, and spatial‐time lag refer to the dependent; all regressions include fixed and time effects; county‐cluster robust standard errors are in parentheses; period: monthly data for January 1999 to December 2014; balanced county‐level panel. a FPR ≤0.05. b FPR ≤0.01. 20 of 20 Growth and Change, 2025