Training during recessions: recent European evidence

What is the expected effect of an economic slowdown on training participation? While several studies investigate the relationship between education and recessions, less is known about training, and what is known mostly refers to US data before the 2009 recession or to apprenticeships in Europe (see Méndez and Sepúlveda, 2012; Muehlemann et al., 2020, and the references therein).

Why should the business cycle affect training? For firms, the forgone production associated with training declines during recessions, inducing them to hoard temporarily idle employees and train them in the expectation that their productivity will be higher when the economy picks up again.

Labor hoarding is likely to be more pronounced in countries with higher employment protection.

In this paper, we investigate the effects of the business cycle on training participation in Europe by using data from the European Labour Force Survey, which cover 27 European countries and the quarters from 2005Q1 to 2018Q4.

This paper draws on previous unpublished work of ours (Brunello and Bertoni, 2021).

The paper is organized as follows. Section 2 provides a review of the empirical literature on the relationship between training and the business cycle, Section 3 introduces the data and the empirical strategy and Section 4 presents the results. Conclusions follow.

The empirical literature discussing the relationship between training and the business cycle has so far produced mixed results.

Closely related to this literature is the one on wage cyclicality. See, for instance, Martins et al. (2012).

Sepúlveda uses data from the US National Longitudinal Survey of Youth (NLSY) to construct a panel of individuals aged 14–22 years in 1979, which he follows until 1998. Focusing on the incidence and intensity of off-the-job and on-the-job training, measured in hours and net of apprenticeships, he reports that they are weakly countercyclical, lead the cycle, and are highly volatile, with a standard deviation of >10 times that of the output.

Majumdar (2007) also uses NLSY data for the period 1979–1988 but finds that the probability of receiving company training decreases when the local unemployment rate increases, which points to procyclical training. This negative association, however, is only statistically significant for workers who have joined the firm since the last wave. Majumdar explains his results as the outcome of two countervailing effects, with the latter dominating the former: on the one hand, labor market opportunities for trained workers are fewer in a downturn, which reduces their bargaining power with the firm and increases the employer's incentive to train. On the other hand, there are many alternatives to training in a slack labor market, which induces firms to hire rather than train.

Bassanini and Brunello (2008) study the relationship between product market regulation and workplace training, using data for 15 European countries and 8 years – drawn from the European Labour Force Survey. They find that their measure of training incidence – the proportion of employed individuals who received training in the 4 weeks before the reference week – is negatively correlated with their measure of the business cycle, the logarithm of worked hours filtered from trend using the Hodrick–Prescott filter,

The Hodrick–Prescott filter or decomposition is a procedure that decomposes a time series into its trend and business cycle components.

Felstead and Green (1996) report instead that training was procyclical in Britain during the 1970s, 1980s, and 1990s. Felstead et al. (2012) examine the impact of the 2008–2009 recession on training activity in the United Kingdom. Using data from the National Employer Skills Survey 2009, they show that cuts in training expenditures were not as severe as feared. Although a minority of employers did cut expenditure and coverage as a result of the recession, most reported no significant change, and some had even increased their commitment. Training expenditure in real terms fell by only 5% between 2007 and 2009.

In his review of the literature focusing on apprenticeships, Brunello (2009) concludes that the ratio of apprentices to employees tends to be (mildly) procyclical and to decline during a recession, with the notable exception of the Great Depression, when it rose (at least in England). Recent evidence from Switzerland confirms this assessment (see Luthi and Wolter, 2020). When broader measures of training are considered, which exclude apprentices, the weight of the evidence is in favor of countercyclical training incidence.

More recently, Muehlemann et al. (2020) have used German data on apprenticeships from 2007 to 2019 and information on business cycle expectations up to June 2020 and estimate that the coronavirus-related decrease in firms’ expectations about the business cycle is associated with a predicted 8% decrease in firm demand for apprentices and a 6% decrease in the number of new apprenticeship positions in Germany compared to 2019.

Finally, Méndez and Sepúlveda (2012) argue that, in the United States, while aggregate schooling exhibits a countercyclical pattern, the case for countercyclical training is weak at the aggregate level. However, when training episodes are decomposed into independent categories, they highlight two key distinctions: (a) between firm-financed training, which tends to be strongly procyclical, and training financed by the individual, which tends to be countercyclical (see also Alessandrini et al., 2015); (b) between employed and unemployed individuals. Training seems much more procyclical for the former than for the latter.

This paper contributes to this literature in two ways. First, we argue that it is important to distinguish between the employed and the not employed, because firm-sponsored training is mostly undertaken by employed workers and firms are likely to encourage training during recessions. Training by the unemployed and the inactive, instead, may be difficult during recessions, when individuals face liquidity constraints. We estimate the sensitivity of training to cyclical unemployment for the employed and the not employed and explicitly address the endogenous selection of individuals between groups. We show that the cyclical behavior of training participation varies with labor market status. Second, we are the first to investigate the cyclicality of training in a sample that covers all the 27 European member states. We also highlight that the cyclicality of training varies with labor market institutions, such as employment protection, and the importance of public expenditure in training.

We use quarterly data from the European Union Labour Force Survey (EU-LFS), a harmonized household survey that collects information on labor markets across all EU 27 member states. We restrict our sample to the period from 2005 until 2018 to account for the substantial changes in the survey that took place until 2004 and consider individuals aged 25–64 years. Our final estimation sample spans 14 years and 27 countries and counts >43 million observations.

Our key measure of training is training participation (T), a binary variable that is equal to “1” if individuals attended – within the past 4 weeks – courses, seminars, conferences, or private lessons or instructions outside the regular education system and equal to “0” otherwise. We also look at training intensity, measured by training hours during the past 4 weeks. As in Méndez and Sepúlveda (2012), our business cycle indicator is the quarterly country-specific unemployment rate (U), which can be decomposed into a trend component, a cyclical component, seasonal effects, and residual noise.

For each country, we capture seasonality by regressing U on quarter dummies. We then decompose the residuals into a trend (U_trend) and a cyclical component (U_cycle), using the filter proposed by Hodrick and Prescott (1997) and adopting a smoothing parameter of 1,600 (Ravn and Uhlig, 2002). Since training participation T also includes a pronounced seasonal component – its level drops by about 50% during the summer quarter of each year – we filter out this component by retaining the residuals of country-specific regressions of T on quarter dummies. We define these residuals as TR. Descriptive statistics for our final sample are shown in Table 1.

As a preliminary step, we regress both training participation and training hours on gender, age, educational attainment, labor market status, country group, quarter, and year dummies. The results, reported in Table 2, show that both training participation and intensity decline with age and are the highest among the better educated and those who live in Northern Europe. While training participation is the highest for the employed, training intensity is the highest among the not employed. Conditional on training participation, the employed have spent during the past 4 weeks 16.85 hours of training on average, while the unemployed and inactive have spent 47.74 hours and 40.26 hours, respectively. The longer duration of training among the not employed may depend on the fact that those out of work need to master new skills, while those at work often need only to upgrade existing skills.

Consistent with the evidence reviewed in Bassanini et al. (2007), we also find that training participation and intensity are higher for females than for males. One reason for this is that we are considering all types of training. If we restrict our attention to the participation in training that takes place during working hours, which is likely to be paid or organized by the employers, this is higher for males than for females.

The information on whether training occurs during working hours or not is only available with annual data. For example, when we use the 2016 wave and regress the probability of participating in training exclusively or prevalently during working hours on the same variables used in Table 2, we find that the gender dummy (male =1, female =0) attracts a positive and statistically significant coefficient (0.004, standard error: 0.002).

We estimate the following regression model: (1) TRict=∑yαyDy+∑cαcDc+β1 U_cyclect+∑cβ2cDcU_trendct+γXict+uict, {\rm{T}}{{\rm{R}}_{{\rm{ict}}}} = \sum\limits_y {{\alpha _y}{D_y} + } \sum\limits_c {{\alpha _c}{D_c} + {\beta _1}{\kern 1pt} U\_cycl{e_{{\rm{ct}}}} + \sum\limits_c {{\beta _{2c}}{D_c}U\_tren{d_{{\rm{ct}}}} + \gamma {X_{ict}} + {u_{ict}},} } where TR_ict is the net of seasonal effects; D_y and D_c are vectors of year and country dummies; U_cycle_ct and U_trend_ct are the business cycle and trend indicators, and we allow for trend effects that are country specific (β_2c); X_ict is a set of individual-level controls that include age, gender, and whether the individual has a tertiary education degree or not, and u_ict is the error term. We cluster standard errors by country and time period.

We investigate the differential response of training to the business cycle for the employed and not employed by estimating Eq. (1) separately for each group. We address the endogenous selection of individuals into each subsample by augmenting the specification in Eq. (1) with the inverse Mills ratio (IMR), obtained from the estimate of the effects of the explanatory variables in Eq. (1) and the additional variable Z_ct on the probability of individual employment, using a probit specification.

Méndez and Sepúlveda (2012) deal with selection using panel data methods. This approach is precluded here, because our data are repeated cross sections.

The variable Z_ct is a proxy of country-specific demand shocks and serves as the exclusion restriction (i.e., an instrumental variable), guaranteeing that our switching regression model is not identified by functional-form restrictions alone. Following the shift-share logic of Autor et al. (2013), we define Z_ct as follows: (2) Zct=∑k[EktEk,t−1−1]×Eck,τEk,τ, {Z_{ct}} = \sum\limits_k {\left[ {{{{E_{kt}}} \over {{E_{k,t - 1}}}} - 1} \right]} \times {{{E_{ck,\tau }}} \over {{E_{k,\tau }}}}, where [EktEk,t−1−1] \left[ {{{{E_{kt}}} \over {{E_{k,t - 1}}}} - 1} \right] is the aggregate change in EU-27 employment in industry k between quarters t and t – 1 and the weights Eck,τEk,τ {{{E_{ck,\tau }}} \over {{E_{k,\tau }}}} are the ratios of industry k workers in country c to industry k workers in the EU-27 area in year τ. This year is 1995 for all countries except the following: Slovenia (1996); Czech Republic, Estonia, Poland, and Romania (1997); Latvia, Lithuania, and Slovakia (1998); Cyprus (1999); Bulgaria (2000); and Croatia (2002).

By combining the predetermined employment shares by industry and country with the aggregate industry-level employment changes in the whole of EU-27, Z_ct captures the changes in the employment rate that are not purely explained by country- and industry-specific labor supply shocks. Conditional on country-specific trends and aggregate year effects, it is unlikely that overall EU employment changes directly affect country-specific training. Therefore, the validity of the exclusion restriction relies on the assumption that the presample shares Eck,τEk,τ {{{E_{ck,\tau }}} \over {{E_{k,\tau }}}} can be taken as exogenous. Since these shares are computed for most countries between 7 years and 10 years before the start of our sample, we consider this assumption as plausible.

We have investigated the cyclicality of training in Europe, using data from the EU-LFS for the period 2005–2018. Pooling data across all EU-27 countries, as well as across employed and not employed workers, we have estimated that training participation is mildly countercyclical, while training intensity is acyclical.

Considering that firm-sponsored training is mostly undertaken by employed workers and that firms are likely to encourage training during recessions, these average effects may hide heterogeneities by employment status. As a result, we have estimated the response of training to the business cycle separately for employed and not employed workers. We have found that training participation (intensity) is countercyclical (acyclical) for the employed and acyclical (procyclical) for the not employed. Countercyclical training of the employed is consistent with the view of recessions as times of reorganization. Procyclical learning for the not employed is driven instead by the behavior of the inactive and can be explained by the presence of credit constraints preventing investment when the economy is in dire straits.

What are the implications of countercyclical training for the European economy? The available evidence suggests that labor productivity typically increases in economic expansions and declines in economic downturns. If training does not subtract workers’ time from production during recessions, when labor demand is low, and positively affect labor productivity, countercyclical skill accumulation can contribute to attenuation of the procyclical behavior of productivity, in the sense that the slowdown in productivity would have been sharper in the absence of training.

We have shown that in the countries of Europe where training increases during recessions, governments spend more to encourage training and have a higher share of training expenditure on GDP. Government expenditure includes cofinancing schemes directed at firms (levy/grant programs and tax credits) and at individuals (vouchers, individual learning accounts). Although these schemes may induce deadweight losses (by funding training that would have been done anyway), they can increase adult learning by reducing the liquidity constraints faced by workers and firms, especially during recessions (see Costa et al., 2018).

Training investment during recessions can be stimulated by designing countercyclical subsidies, which increase in intensity when the economy is in a recession and the likelihood that liquidity constraint bite is higher. Examples in this direction are the topups to individual learning accounts introduced by France and Singapore to promote training during the coronavirus disease (COVID) recession.

We have also shown that training participation increases more with cyclical unemployment in countries where employment protection is stronger. In these countries, the dismissal of employees when the economy slows down is either costlier or more complicated, which favors training of redundant labor as an alternative viable option. An implication of this is that policies favoring the deregulation of labor markets may have the unpleasant side effect of reducing the incentives that firms have to train labor during recessions.