Lifelong learning and employment outcomes: evidence from Sweden

ABSTRACT

We study the relationship between adult education and training (AET) and employment in Sweden. Exploiting rich data from the Programme for the International Assessment of Adult Competencies, and using an inverse-probability weighted regression-adjustment estimator, we find that AET is positively related to the probability of doing paid work. This relationship is driven by non-formal, job-related AET, such as on-the-job training. We also find that the relationship – the strength of which increases with training intensity – is similar across different types of non-formal, job-related AET. The results suggest that policies stimulating relevant AET take-up have promise to secure higher employment.

KEYWORDS:

• adult education and training
• employment outcomes
• PIAAC

JEL CODES:

• I22
• J24
• M53

Acknowledgements

The author thanks Giorgio Brunello, Lorraine Dearden, Henrik Jordahl, Julian Le Grand, the editors Lindsey Macmillan and Colin Green, Olmo Silva, Anders Stenberg, and two anonymous reviewers for comments and discussions.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 These calculations, which do not include students aged 16–24 who are in their first formal cycle of studies, are carried out using micro-level data from the OECD’s (2017) PIAAC database, which we utilise in this paper. The data are derived from questions enquiring whether respondents studied for (1) any formal qualification at primary, secondary, university, or post-secondary level, and (2) whether they participated in non-formal education through ‘Courses conducted through open or distance education, ‘Organised sessions for on-the-job training or training by supervisors or co-workers’, ‘Seminars and workshops’, and/or ‘Other courses or private lessons’. Job-related AET is defined as training that individuals report having undergone for the purposes of improving their employment chances in general. Statistics Sweden (2014) reports very similar figures using other data.

2 While it would also be interesting to study the effect of AET on earnings, we refrain from doing so in this study. This is partly because the PIAAC dataset does not include any measure of earnings history, which is likely key to ensure that the conditional independent assumption holds – just as an indicator of work history appears key for this purpose in this paper. Also, the Swedish PIAAC dataset only includes information of earnings in deciles, which is too coarse to permit a rigorous analysis of how AET affects earnings.

3 However, there is research separating the effects of specifically on-the-job work-related training from off-the-job work-related training on wages. For example, Lynch (1992) studies the impact of on-the-job as well as off-the-job training – with the latter provided through apprenticeships and for-profit institutions – on wages among young, non-university graduates, finding positive effects of both types, while Pischke (2001) finds imprecise evidence suggesting that the wage returns to off-the-job training are larger than the returns to on-the-job training. Another study finds that the wage returns to both on-the-job and off-the-job training are dependent on years of schooling (Ariga and Brunello 2006).

4 In contrast to many international assessments at the school level, PIAAC is not carried out continuously in the same countries. To date, each country has only participated once, with the exception of the United States (which participated in both the 2012 and 2016 rounds).

5 Technically, the final sample was composed of 4,469 respondents, but due to one observation with no values in the database, the number available for analysis is 4,468.

6 However, our preferred inverse-probability weighted regression-adjustment estimator, discussed in Section 4, automatically excludes a few individuals who have/have not undergone AET, but whose values on the covariates do not overlap with any of the individuals who have not/have undergone AET.

7 It is not possible for individuals to simultaneously report that they underwent both job-related and non job-related, formal AET, or both job-related and non job-related, non-formal AET, but it is possible for them to report that they underwent different combinations of formal and non-formal AET. The OECD’s assignment of individuals to the job-related and non job-related categories is based on, firstly, the type of formal AET they underwent, and, secondly, the type of non-formal AET they underwent. For example, individuals who underwent non job-related, formal AET as well as job-related, non-formal AET are assigned to the non-job related AET category. In an unreported robustness test, we instead created a separate indicator for individuals who underwent both job-related, formal AET and non job-related, non-formal AET, or non job-related, formal AET and job-related, non-formal AET (about 2 per cent of the sample), but found little support for interaction effects in this respect.

8 We follow the OECD’s method of assigning individuals to the different categories by, firstly, the type of formal AET they underwent, and, secondly, the type of non-formal AET they underwent (see the previous endnote). In an unreported robustness test, we created a separate indicator for respondents who underwent both formal and non-formal AET (about 9 per cent of the sample), but found little support for interaction effects in this respect.

9 In this analysis, to be able to distinguish differential effects, we exclude individuals who participated in more than one non-formal AET component, for each overall category found to be related to employment. In an unreported robustness test, we further excluded individuals who also underwent some form of formal AET, but the results were essentially identical.

10 Some of the control variables, such as years of schooling, test scores and the number of books at home, may be partly endogenous to AET undertaken in the year prior to the interview. If so, by including these variables as covariates, we may adjust for some of the mechanisms behind the effect of AET. However, since our research strategy hinges on adjusting for rich observable characteristics to make the conditional independence assumption hold, we believe this is a risk worth taking. Furthermore, the variables in question are positively correlated with both AET and employment, meaning that we are less likely to find a positive effect once holding them constant. In other words, we are more likely to err on the side of caution than bias the estimates in favour of the hypothesis that AET has positive effects on employment outcomes. Finally, in robustness tests, we also show that results are essentially identical when only including age, gender, and an indicator of paid work measured in the same period as the AET indicator, showing that potentially endogenous covariates do not drive the findings.

11 However, we still include individuals who did not do paid work in the 12-month period prior to the survey in the analysis. This is because these individuals contribute to the inverse-probability weighted regression-adjustment estimator’s first step, which is used to construct the weights utilised in the second step, as discussed in Section 4. Nevertheless, results are unsurprisingly almost identical if we exclude individuals who did not do paid work in the 12-month period prior to the survey.

12 Individuals who have never done paid work contribute to the inverse-probability weighted regression-adjustment estimator’s first step, as discussed in Section 4. However, as noted in the previous endnote, results are essentially identical if we exclude these individuals.

13 Following previous research, we replace missing values for the covariates with the sample means and include separate indicators for missing values in the regressions (see Hanushek and Woessmann 2011).

14 It is not possible to adjust for such variables, since only those who are currently engaged in paid work answer questions of relevance to the issue of workplace quality.

15 Our outline of the estimator draws on Brunello and Rocco’s (2017, 338–342) discussion.

16 Since the model is dependent on respondents having similar values on the covariates, it is only possible to study AET effects among people in the treatment and control groups for which the covariates overlap. To determine such overlap, we use the default tolerance level in the teffects command in STATA. As discussed in Section 5.2, Figures A1–A4 in the Appendix show graphically that there is generally sufficient overlap in the main models. In robustness tests, to increase overlap with the density of predicted probabilities of not undergoing any AET further, and to ensure that our results are not dependent on a small number of observations, we also trimmed the sample and re-estimated the models. As highlighted in Section 5.2, despite dropping about 60 per cent of the observations, the results are essentially identical.

17 While the assumptions are essentially the same as in propensity score matching, the latter does not allow analyses of multiple treatments simultaneously. In robustness tests, we instead used regular nearest matching to study the average relationship between AET and employment as well as the relationship between employment and each category of training we find to be related to employment. Similarly, we utilised entropy balancing, which achieves complete balance in terms of means, variances, and skewness in single treatment and control groups by using weights constructed through data pre-processing. This estimator has been found to balance treatment and control groups more efficiently than matching estimators (see Hainmuller 2012; Hainmuller and Xu 2013). As discussed in Section 5.2, the results from these alternative estimators are very similar to the equivalent results from the IPWRA estimator.

18 The over-identification test can only be carried out in analyses with one treatment group, and we thus only present these statistics for the main model analysing the average effect of all types of AET, as well as for models separately analysing the types of training that have effects on employment outcomes.

19 While literacy and numeracy scores in PIAAC are estimated from ten ‘plausible values’ derived from multiple imputations, and replicate weights are used to adjust for sampling error (see OECD 2013b), we use the average of all plausible values for each subject and regular robust standard errors. This is to estimate both the inverse-probability weighted regression-adjustment estimator and the covariate balance test correctly. However, the OLS results are identical if we estimate the regressions for each plausible value separately and use replicate weights to adjust the standard errors, suggesting these adjustments matters little – which is supported by prior research analysing similar survey structures (see Jerrim et al. 2017).

20 This may be compared with the coefficient of years of schooling, which in the equivalent OLS model in Table 1 in fact only has a weak, statistically insignificant positive relationship with the probability of doing paid work. However, the coefficient of literacy scores is positive and statistically significant in this model, indicating that a one standard deviation increase in literacy scores is associated with a 4 percentage point increase in the probability of doing paid work.

21 This finding is not that unusual in the field and may reflect the sometimes rather subtle differences between OLS and propensity-score estimators, especially when analysing large-scale datasets and including rich controls (see Goodman and Sianesi 2005; Mendolia and Walker 2015).

22 This also holds true when excluding individuals who did not participate in any form of job-related, non-formal AET and using ‘On-the-job training’ as benchmark category. While the differences between the separate components of job-related, non-formal AET and non job-related, non-formal AET as a whole are not statistically significant in Panel 3, this is not surprising since the disaggregation of the job-related, non-formal AET category naturally decreases the precision of the separate estimates. Also, as highlighted in endnote 9, the number of observations decrease in Panel 3 because, in order to separate the effects of the different job-related, non-formal AET components, we exclude individuals who participated in more than one such component simultaneously.

23 For example, the findings largely rule out one potential source of bias: individuals undergoing on-the-job training as preparation for a new position for which they have already been selected. Such individuals undergo AET as a result of getting the new position rather than vice versa. While we believe adjusting for paid work in the previous year is sufficient to deal with this potential issue, the fact that we find very similar effects across all types of job-related, non-formal AET further indicates it is not an important problem. This conclusion is also supported by unreported analyses in which we found essentially identical effects of job-related, non-formal AET when excluding all individuals who underwent such AET mainly (1) because they were obliged to do so or (2) to do their jobs better and/or improve their career prospects, either of which would presumably apply if AET was undertaken as a result of getting a new position.

24 The only (partial) exception is the probability of undergoing non job-related, formal AET in the untrimmed model. While there is overlap in this case as well, large proportions of the probability mass of the control and alternative treatment densities are close to zero. Yet, as displayed in Table A2, the results are very similar when restricting the sample to observations below the 40th percentile of the predicted probabilities of not undergoing any AET, which decreases this problem significantly. Furthermore, we also re-estimated all relevant models with individuals undergoing non job-related, formal AET excluded, and the estimates for the other categories were essentially identical.

25 These categories roughly correspond to observations in the 25th percentile and below, between the 25th and the 50th percentile, between the 50th and the 75th percentile, and in the 75th percentile and above, among people who participated in at least one job-related, non-formal activity.

26 In order to accurately display the ‘dose-dependent’ relationship, we report coefficients with three decimal places, instead of two, in Table 5 specifically.

27 In unreported analyses, we also studied the relationship between participating in several types of job-related, non-formal AET, among individuals who participated in at least one such AET type, and found evidence of positive effects of participating in three or all four types (about 7 per cent of the sample), and a smaller and statistically insignificant impact of participating in two types (about 14 per cent of the sample), compared with the baseline category of participating in one type only (about 22 per cent of the sample).

28 The average treatment effect in the population is a weighted average of the average treatment effects on the treated and the untreated (Cornelissen et al. 2016).

29 Still, it is worth noting that prior research analysing data from other countries, which have very different labour markets compared with Sweden, does often suggest that training is positive for labour-market outcomes (see, for example, Dearden, Reed, and Van Reenen 2006; Lynch 1992; Pischke 2001). This would support the idea that the relationship found in this paper may well be generalisable to other countries and contexts.

Additional information

Funding

This work was supported by the Economic and Social Research Council [grant number ES/J500070/1] and Jan Wallanders och Tom Hedelius stiftelse.