Keywords

1 Introduction

Throughout recent decades, technological change, international trade, and economic turbulences have accelerated in the globalized world, making labour markets more and more dynamic. As a result, workers have to handle steadily changing technologies and compete with highly competitive workers in an international labour market. In addition, many workers have to switch jobs more frequently as a consequence of cyclical fluctuations. In such dynamic labour markets, workers continuously have to invest in human capital to maintain and increase their productivity. Otherwise, they are likely to experience long-lasting negative consequences for their careers. Thus, it is not surprising that further training and lifelong learning, such as on-the-job training and retraining, have gained increasing attention in the public media and political debate. However, existing research provides only little evidence about the influence of dynamic labour markets on workers’ further training participation.

To understand the importance of employment-related further education in dynamic labour markets, this study concentrates on technological change and further training. A large body of literature documents that computers and industry robots have had a substantial impact on labour markets (e.g., Acemoglu, 2015; Acemoglu & Restrepo, 2017a, b, 2018), because these technologies can perform many routine tasks previously performed by low- and medium-skilled workers.Footnote 1 As emerging technologies, such as artificial intelligence (AI), can even perform many non-routine tasks of high-skilled workers,Footnote 2 many policymakers and researchers fear that technological change might disrupt labour markets even more substantially in the future.Footnote 3 Therefore, many scholars and practitioners emphasize the growing importance of on-the-job training and lifelong learning to prevent disruptive effects of technological change; and governments across the world are investing heavily in training policies.

However, no study has analysed whether and to what extent workers whose jobs are likely to be replaced by new technologies invest in on-the-job training.Footnote 4 Obviously, workers have strong individual incentives to update their skills to avoid wage loss or unemployment in response to technological changes.Footnote 5 Yet, firms—not workers—initiate and finance most on-the-job training. According to the results of the Adult Education Survey (AES), 72% of all further training activities in 2018 were financed by firms or took place during working hours (Bundesministerium für Bildung und Forschung, 2019). And firms have different incentives than workers when it comes to training. On one hand, they may invest in training to retain firm-specific human capital, even if new technologies can substitute certain tasks that workers currently perform. On the other hand, they may refrain from investing in training for workers whom they eventually plan to automate. Without understanding whether and how technological change affects workers’ training participation, designing efficient training policies that are tailored towards the most affected workers remains difficult.

This study sheds light on the relationship between technological change and training by analysing whether and why workers whose jobs are at a high risk of being changed or automated by technologies invest either more or less in training. Therefore, we use a novel and unique data source: the adult survey of the German National Educational Panel Study NEPS (Allmendinger et al., 2019) combined with administrative register data from the German Federal Employment Agency (BA). Thus, our data combines detailed survey information about workers’ training participation with precise register data about individual careers and firms.

Our first goal is to document the training gap between workers in routine jobs who are at a high risk of being substituted by technology and workers in non-routine jobs who are less exposed to the negative consequences of technological change. Results reveal a substantial gap in training participation of approximately 14 percentage points between workers in routine and non-routine jobs. This gap is almost as large as the training gap between high-educated workers (i.e., workers with a university degree) and low educated workers (i.e., workers without a university or apprenticeship degree). The gap persists within educational groups and barely changes after accounting for variables that are strongly related to the workers’ unobserved productivity differences, such as information about their health status and results from competence tests. Thus, although we cannot present causal inferences from these results, they do not suggest that unobserved ability differences are able to explain entirely the training gap between routine and non-routine workers.

Our second goal is to exploit the detailed information within our data source to identify the most important factors that are able to account for the training gap between routine and non-routine workers. Detailed decompositions of the training gap reveal that education and other variables strongly related to the employees’ unobserved ability account for only a small amount of the training gap (i.e., 6.3%). Similarly, firm characteristics, such as firm size and industry, account for only a small share of the training gap between routine and non-routine workers.

Instead, firms’ training support accounts for the lion’s share of the training gap. In more detail, we find that firms’ active human resource activities (e.g., individual offers to train during working time and individual financing offers) account for 29.7% of the gap, and firms’ passive human resource strategies (e.g., their general financing strategy) for about 12.1%. Thus, overall, our results suggest that the firms’ individual and financial support are the most important determinants of the low training participation of workers in routine jobs. This result is in line with the argument that firms may have low incentives to support and finance the training participation of employees who are likely to be replaced by modern technologies in the near future.

Our results contribute to at least two strands of the literature: first, the literature on routine-biased technological change (Autor et al., 2003). Whereas previous studies have analysed the wage and employment effects of routine-biased technological change, our study shows that routine-biased technological change not only has induced a polarization of the wage and employment structure, but is also related to a polarization of workers’ training participation. In this sense, our results provide evidence for a channel through which routine-biased technological change translates into a polarization of wages and employment. Moreover, if routine workers who are most likely to be exposed to the negative consequences of technological change invest less in training and re-training than non-routine workers, the polarization of workers’ training participation may reinforce the polarization of the wage and employment structure in the long run.

Second, we contribute to the training literature by showing that routine-biased technological change may have a substantial influence on workers’ human capital accumulation. Previous studies have focused on the education level and the enhancement of the training participation rate for low-educated workers (see Fouarge et al., 2013; Hidalgo et al., 2014). We now show that, even within educational groups, workers’ training participation is very heterogeneous and depends on their tasks and their firms’ support.

2 Data

This section describes our data set that combines administrative records of workers and firms from the Integrated Employment Biographies with detailed survey data for workers’ training participation from the adult cohort of the National Educational Panel Study (NEPS).

2.1 The National Educational Panel Study (NEPS)

This study uses Starting Cohort 6 of NEPS, a longitudinal data set surveying the educational trajectories and labour market careers of about 10,000 adults between 2009 and 2017 (Allmendinger et al., 2019). The data contain detailed information about the individuals’ entire labour market careers, education histories, and training activities. Moreover, we can merge the survey data to register information from social security contributions that allow us to obtain precise individual worker and firm characteristics.

The 2011/2012 wave of NEPS is ideally suited for our investigation, because it contains detailed information on workers’ training participation and information about their firms’ human resources practices, e.g., information about financial and structural company support for on-the-job training. Moreover, the data contain many individual and job characteristics that are likely to influence the workers’ training participations but are commonly not observed by researchers. For example, we have information on the workers’ health status and their competencies in math and reading. As a result, the lion’s share of our analysis relies on the 2011/2012 wave of the NEPS. However, whenever possible, we include information from all waves in our analyses.

2.2 Administrative Data from the German Federal Employment Agency

The administrative records from the German Federal Employment Agency (BA) contain individual-level data on labour market activities that are relevant for calculating the amount of social security contributions. The BA uses this information to provide two data sources: first, the Integrated Employment Biographies (Integrierte Erwerbsbiografien, IEB) containing the entire population of German social security records from 1975 to 2018. The data cover all employees subject to social security contributions except civil servants and the self-employed—that is, all dependently employed workers who make up to 80% of the workforce in the German labour market. Unique person and establishment identifiers identify all individuals and establishments such that we can merge the IEB data to the NEPS data. Second, the Establishment History Panel (Betriebs-Historik-Panel, BHP) that contains information about all German firms with at least one worker subject to social security contributions (Schmucker et al., 2018).

Both data sources allow us to rely on precisely measured individual labour market and firm characteristics instead of relying exclusively on survey measures that are commonly prone to substantial measurement error. From the IEB, we obtain precise individual information about each person’s occupation, gender, age (in years), exact tenure and labour market experience, and education. From the BHP, we obtain exact information about the firms’ size, share of full-time or (marginal) part-time workers, qualification structure, median wage, and location.

3 Sample Restrictions and Main Analysis Variables

3.1 Sample

Our main analysis sample is composed of individuals who are working at the time of the interview and give consent to link their survey data to the administrative data from social security records (93% consent rate, see Antoni et al., 2018, for more information). We further restrict our sample to individuals who are not undergoing a vocational training and are between 25 and 60 years old. Finally, we delete all observations with missing values in our main analysis variables. These restrictions leave us with 3246 individuals.

3.2 Dependent Variables

The purpose of our study is to analyse the training participation of workers whose jobs are likely to be changed or automated. Thus, our dependent variables describe the workers’ participation in further training activities. More specifically, we analyse labour-related non-formal training courses that relate to employment spells. However, we exclude training that is not employment-related, informal training that is not organized in courses or seminars, formal training, such as apprenticeship training, and training courses that occur during periods of non-employment.

Our main dependent variable is a dummy variable indicating whether a worker participated in at least one training course throughout the last 12 months before the interview. However, a training dummy measures only the extensive margin of the workers’ training participation. Therefore, we also analyse the frequency and duration of training to uncover effects at the intensive margin.

3.3 Independent Variables

Our main independent variable measures whether a job is likely to be changed or automated by modern technologies in the near future. To create such as measure, we follow the literature on routine-biased technological change (e.g., Autor et al., 2003). The core idea of this literature is that job tasks are more likely to be replaced by modern technologies if they follow routines that are easy to automate or program. In contrast, tasks that do not follow routines cannot be programmed and automated so easily.Footnote 6 Moreover, non-routine tasks often even complement modern technologies.

More specifically, we rely on a set of six items from the NEPS data to measure the routine-task intensity of an individual’s job. Matthes et al. (2014) designed these items explicitly to infer the extent to which a worker performs routine tasks. Table 14.A1 in the appendix reports the content of each item we use. Workers responding to each item indicate the routine intensity of their jobs by rating the frequency of repetitive tasks on a 5-point Likert scale ranging from 1 (not at all) to 5 (always or very often). For the largest part of our analysis, we define a worker as being highly exposed to routine-biased technological changes if the sum score of items lies at or above the 75th percentile. However, in certain robustness checks, we also use a continuous version of this measure.

3.4 Control Variables

Because workers do not choose their jobs randomly, the routineness of a worker’s job correlates with a large number of observable and unobservable variables that are likely to influence training participation (selection problem). For example, workers’ training participation is likely to correlate with firm size, i.e., larger firms tend to invest more in the training of their workers than smaller ones. Because it is very difficult, if not impossible, to find experimental or quasi-experimental variation that assigns workers randomly to their jobs, we have to rely on a huge set of control variables to capture as many confounding factors as possible.

Our data set contains a very rich set of control variables from the administrative register data and the survey data. Whenever possible, we rely on the administrative data to measure our control variables, because register data is less prone to measurement error than the subjective information from survey data. In detail, we obtain information about the workers’ gender, part-time work, vocational education, marginal employment, work experience, tenure, and wage from the register data. Moreover, we obtain the following information about the workers’ firms from the register data: median wage of full-time workers, firm size, share of full-time workers, and shares of high- and medium-qualified workers.

However, like most register data sources, our register data is limited to a few core labour market and firm characteristics that might not capture all potential confounders. Therefore, we also include important variables from the survey data. Namely, we have information about the workers’ migration background, their self-assessed health status, their disability status, and their math and reading competencies. Moreover, the 2010/11 wave of the NEPS provides information about whether a worker’s firm implemented human resource practices to support that worker’s training participation. For example, we know whether a worker’s firm has an official training agreement, an official unit responsible for on-the-job training, or whether the firm generally offers financial training support.

4 Descriptive Statistics

Figure 14.1 shows the distribution of our measure for the self-reported routineness of individuals’ jobs. The figure reveals that our measure for routineness ranges between 5 and 30 and almost resembles a normal distribution with a mean of approximately 15. The vertical line indicates the 75th percentile of the distribution. Throughout the study, we refer to workers with a routineness below the 75th percentile as non-routine workers and to workers with a value above the 75th percentile as routine workers.

Fig. 14.1
A line graph of self-reported routineness of job tasks versus density of Kernel density estimate. The highest point on the wave graph is 17, 0.12. A note at the bottom reads Kernel = epanechnikov, bandwidth = 0.5812.

Kernel density estimates of self-reported routineness. (Source: 2011/2012 wave of the NEPS Adult Starting Cohort)

Full size image

Table 14.1 presents descriptive statistics for the core variables of our analysis. The first column presents the results for routine workers the second for non-routine workers. The third column shows the differences between routine and non-routine workers along with significance levels from t tests.

Table 14.1 Descriptive statistics of core analysis variables
Full size table

With a training participation of approximately 27% routine workers train substantially less than non-routine workers of whom approximately 41% reported having participated in at least one training throughout the year. The difference in training participation is highly significant on the 1% level. However routine and non-routine workers do not just differ in their training participation, they also differ with respect to many individual characteristics. For example, they are more likely to be female, work part-time, and have a migration background. Moreover, they appear to work in firms with different management and training policies. For example, they are substantially less likely to work in firms that have a specific training agreement or a designated person who is responsible for on-the-job training. Finally, they are substantially less likely to receive individual training support from their firms, such as opportunities to train during working time and financial training support.

5 Empirical Strategy

We apply an Oaxaca–Blinder decomposition to estimate and explain the difference in training participation between routine and non-routine workers. The Oaxaca–Blinder decomposition explains differences in outcomes for individual members of two groups in which random assignment is impossible—for example, gender or race. For the purpose of this study, we use this method to explain training participation differences between routine and non-routine workers—that is, between workers whose jobs are likely to be changed or automated by modern technologies and all other workers. To decompose the differences in training participation, we estimate the following equation:

$$ E\left({T}_{NR}\right)-E\left({T}_R\right)={\left(E\left({X}_{NR}\right)-E\left({X}_R\right)\right)}^{\prime }{\beta}_{NR}+E{\left({X}_R\right)}^{\prime}\left({\beta}_{NR}-{\beta}_R\right) $$
(14.1)

The left-hand side of Equation 14.1 shows the difference in the expected training participation between non-routine (NR) and routine (R) workers. The right-hand side shows two terms. The first term (E(XNR) − E(XR))′βNR denotes the cumulated mean difference of all explanatory variables between the two groups weighted by the slope of the non-discriminated group. This term is usually referred to as the explained part of the decomposition. Thus, in our case, the first term indicates the part of the difference in the training participation between routine and non-routine workers that is related to observable individual and firm characteristics in our data set. The second term on the right-hand side denotes the cumulated average of the explanatory variables of the reference group of routine workers weighted by the differences of the slopes between non-routine and routine workers. This term is usually referred to as the unexplained part of the decomposition. In our case, this term can be interpreted as the difference in the training participation of routine and non-routine workers that is not related to observable worker and firm characteristics.

6 Results

6.1 Determinants of the Training Gap Between Routine and Non-routine Workers

Figure 14.2 shows the training gap between routine and non-routine workers for different subgroups of our population. The first bar shows that the raw gap in training participation between routine and non-routine workers amounts to 14 percentage points— i.e., a gap of approximately 40% relative to the average participation rate of 35%. As the second and the third bar reveal, the training gap between routine and non-routine workers remains large within the groups of women and men—although the routine training gap is somewhat larger among men than among women. The training gap also remains large and persistent within the groups of migrants and non-migrants, and within different age groups. Interestingly, the routine training gap is larger among non-migrants than among migrants, and at 22 percentage points, it is particularly large for individuals aged between 30 and 34. In contrast, the routine training gap is smaller for both younger and older workers.

Fig. 14.2
A bar chart of differences in training participation between routine and non-routine workers. The highest point is at age 30-34, 0.22, and the lowest participation is at age 25-29, 0.057.

Differences in training participation between routine and non-routine workers. (Source: 2011/2012 wave of the NEPS Adult Starting Cohort)

Full size image

Table 14.2 shows the results of simple ordinary least squares regressions (OLS) of a dummy indicating whether workers have participated in at least one training throughout the year on a dummy variable for a non-routine job and different sets of observable worker characteristics. The first column replicates the raw difference in training participation between routine and non-routine workers without further controls. It shows a coefficient estimate of 14 percentage points that is highly significant on the 1% level. The second column adds control variables for the workers’ highest level of education. Adding these controls decreases the coefficient estimate only slightly to a value of approximately 12 percentage points. This result suggests that the workers’ education and related differences in unobserved ability do not substantially influence the training gap between routine and non-routine workers

Table 14.2 OLS regression of training participation on different individual characteristics
Full size table

The third column adds a dummy indicating the workers’ disability status and a dummy indicating whether the worker is of poor health or not. Adding these health variables barely changes the result and suggests that the workers’ health is unlikely to explain the training gap. Column four adds further individual controls: migration status, gender, age, and labour market experience. Adding these controls slightly inflates the coefficient estimate to a value of approximately 13 percentage points suggesting that measurement error in our main explanatory variable for routine jobs biases the coefficient estimate towards zero. Column five adds job characteristics, such as tenure or a dummy indicating whether a worker is marginally employed. The results barely change relative to column four. The sixth column adds a substantial amount of firm characteristics: firm size, industry shares of high- and medium-qualified workers, share of full-time employees, share of marginally employed workers, location of the company, and the imputed median wage of full-time workers. Adding all these variables decreases the coefficient to approximately 11 percentage points. Finally, the seventh column adds controls for different training policies of the workers’ company, such as dummy variables indicating whether a worker’s firm has a training agreement has an official unit responsible for on-the-job training or whether it generally offers financial training support. Adding these controls further decreases the estimated training gap between routine and non-routine workers to approximately 9 percentage points

Overall, the results of Table 14.2 reveal that observable individual worker and job characteristics barely influence the estimated training gap between routine and non-routine workers. Moreover, the results suggest that unobserved ability differences that are commonly related to observed factors, such as education, migration status, health, and other individual and job characteristics, are unlikely to explain large parts of the training gap between routine and non-routine workers. However, differences in firm characteristics and, specifically, differences in the firms’ training policies have a quite substantial impact on the training gap between routine and non-routine workers.

6.2 Decomposition Results

To put our results into perspective, this section presents results from our Oaxaca–Blinder decomposition as described above. Fig. 14.3 presents the results.

Fig. 14.3
A stacked bar chart of Oaxaca-Blinder decompositions of training gap. The two are explained gaps and unexplained gaps. The four characteristics are raw gap, individual worker characteristics, job, and firm characteristics.

Oaxaca-Blinder decompositions of training gap. (Source: 2011/2012 wave of the NEPS Adult Starting Cohort)

Full size image

The grey bars describe the entire training gap between routine and non-routine workers, the green bars describe the share that can be explained by observable characteristics.

The figure reveals that individual worker characteristics can explain only approximately 7% of the entire training gap between routine and non-routine workers. Moreover, the workers’ job characteristics are also unable to explain a substantial part of the training gap. However, when we add firm characteristics and the measures for the firms’ training policies, the explained share of the gap increases to approximately 40%. Thus, firm characteristics and company policies explain almost one half of the training gap between routine and non-routine workers.

7 Conclusion and Outlook

This study investigates whether automation and digitalization of work processes across different occupations have affected workers’ training participation. The analyses show that workers whose jobs are likely to be substituted by technology in the near future participate substantially less in on-the-job training than workers whose jobs are not likely to be substituted. In addition, the study explores the mechanisms that explain the training gap between routine and non-routine workers. We find that common observable individual and labour market characteristics, such as education or migration status, cannot explain the training gap, whereas differences in firm characteristics and training policies can explain a large share of the training gap. This result is in line with the argument that firms may have low incentives to support and finance the training participation of employees who are likely to be replaced by modern technologies in the near future. If workers who are potentially exposed to the negative consequences of technological change invest less in training and retraining than workers who are not exposed, the polarization of training participation may reinforce the polarization of the wage and employment structure in the future. Our results strongly support the idea of training subsidies for employees.Footnote 7 Against this background, these programmes should have a special focus on the disadvantaged group of routine workers whose jobs are particularly likely to be substituted by technology in the near future.

A relevant question for future research will ask how shocks caused by labour market dynamics—in particular, job displacements following economic turbulences—influence workers’ training participation. During economic downturns, many workers lose their jobs, become unemployed, or have to search for new jobs. A great number of prominent studies from the U.S. and Europe have analysed the consequences of job displacement for the careers of workers while focusing on wage profiles. Although evidence is somewhat mixed, most studies found that worker displacement—usually defined as unexpected and involuntary job loss due to firm closures or mass layoffs—has large and long-lasting negative consequences for the careers of individual workers with earnings losses of up to 40% per quarter of a year (e.g., Davis & von Wachter, 2011).Footnote 8

Although researchers widely agree that displaced workers’ earning losses are large and long lasting, it remains unclear why displaced workers’ large earnings losses persist over the long run. In particular, standard economic theory fails to explain the magnitude and persistence of these long-term displacement losses. To date, existing research has not considered workers’ foregone investments in human capital that may help standard theory to explain empirical results of displaced workers’ long-term earnings losses. In other words, displaced workers may not only lose specific human capital in response to a job loss (Becker, 1964; Neal, 1995; Topel, 1991), but also forgo access to (employer provided) on-the-job training for long periods after the job loss—because they either remain unemployed, move to lower-paid jobs, or move to firms that provide less training. This would imply that displaced workers experience long periods during which they fail to accumulate human capital through occupation-related further training and, therefore, experience long-lasting negative consequences for their careers. Future work based on NEPS data aims to analyse the relationship between worker displacement and occupation-related further training to fill this research gap.

Another question that should be addressed in the future refers to the returns to employment-related further training in dynamic labour markets. The effects of further training have been analysed for various types of returns, such as income, productivity, career and occupational mobility, prevention of unemployment, and re-integration of non-working individuals into the workforce.Footnote 9 As a consequence of technological change, international labour markets, and cyclical fluctuations, workers have to remain flexible and be prepared to switch their jobs and tasks more frequently. Hence, the relevant question is whether job-related further training increases workers’ stock of human capital and enables them to conduct different tasks and deal with future changes in job contents. However, to date, it is unclear whether potential effects of further training on labour market outcomes are driven by training-induced skill increases and subsequent changes in job contents or by other mechanisms such as signalling higher productivity (Spence, 1973) or gift exchange (Becker et al., 2013). Future research based on repeated measures of human capital in the NEPS adult cohort aims to explore the returns to further training in the form of worker competencies and job contents to fill this gap.

Findings will complement the new knowledge about the firms’ and workers’ role in employment-related further training and may help policymakers to target further training investments.