General adult education of displaced workers in a recession: the effects on university enrollment and graduation

ABSTRACT

This article studies a sample of displaced workers during the deep 1990s recession in Sweden and estimates the effect of secondary-level adult education on tertiary-level educational attainment. Plant closures and mass layoffs are used to identify job separations unrelated to individual productivity. Results indicate a large positive effect of general adult education on displaced workers’ further investment in human capital by continuing on to university education. These findings offer some explanation of findings in previous research, i.e. slow recovery of post-treatment earnings among workers enrolled in adult education, with positive treatment effects on earnings emerging in the long run.

KEYWORDS:

• Adult education
• displaced workers
• university education
• human capital

1. Introduction

Public funding of general adult education is often targeted toward persons with low educational attainment, a vulnerable position in the labor market, or both – e.g. low-income workers and those at risk of being unemployed. Economic research on adult education and policy has, for good reasons, focused on its effects in terms of income and employment. Although the official goals of general adult education (internationally and in Sweden) include preparation for eventual investment in further education (e.g. Collins 2014, Congressional Research Service 2014), very little is known regarding the effects in this context. In this study, I use nationwide population register data from Sweden and estimate the effect of general secondary-level adult education of displaced workers on tertiary-level educational attainment.

Facilitating additional investment in human capital by increased capacity for continuous learning and enhanced formal qualifications (OECD 2021) is an intermediate goal that may be perceived by evaluators and policymakers as a spillover effect.¹ The lack of studies on this subject in light of the stated official ambitions, including targeting vulnerable groups, decreasing generational gaps in educational attainment, and providing an educational system with no ‘dead ends’ (e.g. OECD 2000, 31, 56), is a major motivation for the present study. Moreover, although income effects lie outside the scope of this study, investments in education have implications for the estimation of earnings over time. Specifically, investments in longer educational programs have negative effects on wage earnings in the short run and positive effects in the long run, eventually stretching beyond the follow-up period in available data. Economic evaluations of general education for displaced or unemployed workers have either omitted individuals enrolling in higher education from the sample or estimated the long-term income effects of secondary-level education without examining post-treatment transitions into university-level education (Jacobson, LaLonde, and Sullivan 2005a, 2005b; Stenberg and Westerlund 2008, 2015).

Research on the economic effects of general adult education at secondary and tertiary levels indicates positive effects on income; see, e.g. Stenberg (2022) on Swedish data and for further references. Similar findings apply to samples of displaced (Jacobson, LaLonde, and Sullivan 2005b) and long-term unemployed (Stenberg and Westerlund 2008) workers enrolling in primary or secondary-level general adult education. Job loss may encourage an individual to invest in education with the goal of positive effects on future career and income. Displaced workers are more likely than the average worker to have skills that are linked to industries and sectors that are waning or disappearing (Kletzer 1998). This suggests that retraining and further education can result in both public and private benefits. Displaced workers can suffer from reduced wages for several years after displacement (Couch and Placzek 2010), which can be attributed to a multitude of factors surrounding the displacement. e.g. loss of specific human capital, loss of health, poor matching, and peer effects (Carrington and Fallick 2017). Therefore, displacement may reduce the opportunity cost of investment in education. Enrolment in adult education can also be motivated by the goal of becoming eligible for tertiary education. In a Swedish context, those not qualified to apply to the university can enroll in general adult education and complete the required courses or an entire high-school program to become eligible for university studies.²

Two to seven percent of all employees in OECD countries are affected by displacement yearly (Quintini and Venn 2013), with higher displacement numbers typically seen during periods of economic downturn.³ The current recession in many economies and the aftermath of the Covid-19 crisis may spur substantial global investment in adult education to ease the adverse labor market effects. Sweden is no stranger to using adult education as a labor market policy measure. In the early to mid-1990s, Sweden went through a period of sustained economic hardship, with negative GDP growth and many displaced workers. Investment in adult education (secondary-level general education) for displaced and unemployed workers was expanded during this period (e.g. Stenberg and Westerlund 2008). This was partially intended as a labor market policy measure to activate the unemployed in the short run and partially aimed at increasing employment and occupational mobility with a longer perspective, i.e. under more normal macroeconomic conditions. Another motive for increased investment in general adult education was that workers in the private sector in Sweden had a relatively low educational level at the time compared with other OECD countries (Rubenson, Tuijnman, and Wahlgren 2000).

The longitudinal Swedish population register data with linked employee-employer information offer good opportunities to identify displacement from employment, analyze selection into secondary-level education (treatment), and estimate the treatment effect on the treated in terms of completion of a university degree corresponding to a three-year bachelor's degree or higher. The sample is restricted to displaced workers who were separated from their workplaces due to plant closures or mass layoffs during the deep recession in Sweden in the 1990s. This sample is designed to reduce potential bias due to selection into job loss based on unobserved individual characteristics. The underlying assumption is that job separation due to plant closure and mass layoffs are unrelated to individual ability/productivity and are not systematically correlated with unobserved characteristics affecting the outcome of interest. I estimate the effect of enrolment in general adult education at the secondary level, in conjunction with job loss, and on the probability of earning a bachelor's degree within 10 years after job displacement. Bachelor's degree completion is the measured outcome in this study because earning a college degree is widely accepted as a distinct increase in an individual's educational attainment.

The results indicate significant and large positive treatment effects. For the full sample of males, enrollment in secondary-level adult education in conjunction with job loss is associated with an 11.2 percentage point increase in completion of at least a three-year university degree within ten years after job loss. The corresponding estimate for females is 10.0 percentage points. Moreover, the results confirm expected patterns of systematic self-selection on observed characteristics into education.

The following section provides a literature review on enrollment in higher studies and graduation. Section three presents the macroeconomic context and institutional features of general adult education and the educational system in Sweden. Sections four and five describe the data and the method, section six presents the results, and section seven provides a summary and discussion.

2. Previous findings

The literature on the effects of adult education on displaced or unemployed workers is sparse and is mainly focused on economic returns and labor market outcomes. To the best of my knowledge, there are no previous studies on the effects of general adult education on displaced or unemployed workers’ investment in tertiary/university-level education. Although previous research does not focus on displaced workers and adult students, there are findings relevant to this study.

Enrollment in higher studies is affected by a multitude of factors. Even if the exact motivations and incentives differ between individuals, some common factors are parents’ education, unemployment rate, academic ability, income level, cost of education, and wages (e.g. Corazzini, Dugan, and Grabowski 1972; Juszkiewicz 2017; Leslie and Brinkman 1988). The interaction between these and other determinants of enrollment can be complex, e.g. Casarico, Profeta, and Pronzato (2016) on EU data.

Regarding the role of unemployment, studies suggest (but not conclusively) that higher rates of unemployment may increase enrollment in higher education. On US data, Bozick (2009) reported higher enrollment in higher-level education in local job markets with relatively higher unemployment and fewer jobs that did not require a bachelor's degree. Casquel and Uriel (2009) found in Spain that lower unemployment has a significant negative effect on enrollment, except for young individuals from well-off families with an income above the 75th percentile. Based on UK data, Meschi, Swaffield, and Vignoles (2019) found that local labor market conditions were not a key driver of university enrollment but cautioned that there were some indications that higher local unemployment could increase males’ willingness to invest in education.

Turning to factors that may affect degree completion, Baum, Ma, and Payea (2010), using US data, found that college students who have parents with a bachelor's degree and students from better-off families were more likely to graduate within six years. Hughes (2012), also using US data, found that the earlier a student applied to (and enrolled in) college after high school, the more likely the student was to complete a bachelor's degree. Findings in Nelson, Froehner, and Gault (2013) indicate a negative influence of childbearing and child-rearing on the completion of higher studies. Based on Swedish data, Dryler (2013) found that students in four longer degree programs (engineering, law, medicine, and psychology) who had parents with lower education started their studies when they were older and had a lower completion rate.

Studies on the economic return to general adult education at the primary or secondary level for displaced or unemployed workers (e.g. Jacobson, LaLonde, and Sullivan (2005b); Stenberg and Westerlund (2008)) indicate positive returns in terms of wage earnings, although it often takes more than 5–10 years for this positive effect to materialize and often longer to recoup the forgone earnings. Generally, evaluation of the effects of general adult education at the primary or secondary level is associated with various methodological problems, one of which is how to treat observations of students continuing to higher education. One strategy is to include these observations and interpret positive effects on earnings of adult education at the secondary level as reduced form effects that include positive effects of an (unmeasured) increased propensity to enroll in higher education (e.g. Stenberg and Westerlund (2008)). The inclusion of students who enroll in higher studies may offer an explanation for the absence of strong positive short-term effects on earnings and explain some of the long-term positive effects of adult education. A study that shows large positive short-run returns is Jacobson, LaLonde, and Sullivan (2005b). In this study, individuals continuing to higher education were excluded from the sample. This exclusion was likely motivated by an interest in isolating the effect of completion of schooling on earnings in a relatively short follow-up period in available data. However, very little is known about the transition rates from adult education to university enrollment. For the OECD, EU, and Sweden, no studies or statistics are available that indicate the transition rates from general adult education to university studies. Cejda (1997), using US data, reported that around 80% of community college students in his study expressed a wish to transfer to a university. While a ‘wish’ is not the same as enrolling in university studies, it may indicate that actual transition rates could be substantial but are largely unknown.

3. Macroeconomic, educational, and institutional context

Sweden experienced an increase in unemployment from 3 percent in 1990–10 percent in 1993. Although the economy started to recover in 1994–1995 with positive GDP growth, unemployment did not decrease until 1998. The sample in this study experienced job loss in 1994 or 1995, and the treated enrolled in general adult education in the year of displacement or the first semester of the following year (i.e. an effective enrollment window of three semesters). Plant closures and mass layoffs are overrepresented in mature industries and low-skilled occupations. This overrepresentation should provide extra incentives for skill upgrading. Moreover, the high unemployment and bleak labor market prospects most likely strengthened the incentives for participation in education due to a relatively low opportunity cost. On the other hand, what may deter enrollment under such conditions is that the returns on investment in education are likely to be below normal returns in the short run. Also, the labor laws in Sweden, with ‘last in-first out’ principles (with some exceptions), contribute to an increase in the value of tenure, which may encourage displaced workers to engage in job-seeking activities rather than to enroll in long educational programs.

The school system in Sweden mandates compulsory education for children aged 6–16. Upper secondary and tertiary education are optional. In 1994-95, the upper secondary programs were one to three years in length. The three-year programs were theoretical and preparatory for tertiary-level education, while the shorter programs were mainly vocational. Completion of a three-year upper secondary schooling program was (and still is) the basic requirement for university entrance. A policy change in 1994 made three years of upper secondary school the new norm. For some university programs and courses, eligibility also hinges on whether a student has passed specific courses at the high school level (e.g. math, physics, chemistry, biology).

The level of education in Sweden has increased over time, as in most other developed economies. In 1992, 46% of the Swedish population aged 25–64 years had upper secondary education as the highest level of education, which put Sweden far above the OECD average of 36% (OECD 1995). Over time, this figure has been relatively static; in 2005, it was 48%, and in 2018, 43%. The proportion of the population with tertiary level (≥ 3 years) education has increased from 11% in 1992–26% in 2018. During the same time, the proportion with only basic schooling (less than nine years) dropped from 21% to 3%. The current average age in Sweden for completing a bachelor's degree (a three-year degree) is 28.3 years of age (OECD 2018), which is well above the OCED average of 26 years of age. As in many other countries, more females than males enroll in university education and graduate with a degree. In Sweden, this change has come gradually and was not visible until around the year 2000. In 1990, the proportions for tertiary education (≥ 3 years) were the same for both genders, at 11%, but in 2018 they were 32% for females and 21% for males (SCB 2020).

Swedish law requires municipalities to offer adult education at both the primary and upper secondary school level.⁴ The municipalities administer adult education through education centers called ‘Komvux’ (Kommunal vuxenutbildning), mainly financed by grants from the Swedish state. At the time of enrollment for the sample in this study, the adult courses at the compulsory level were mainly Swedish and mathematics. In contrast, the adult upper secondary-level courses included both academic high school programs and subject-specific courses in information technology and health care. Both part-time and full-time students have the right to financial aid. At the time of displacement in 1994-1995, the monthly financial aid for a full-time student consisted of a €700 low-interest loan from the state and a €300 tax-free grant (both in 2020 prices). Depending on the individual's age (maximum 56) and family situation, the amounts could differ. Depending on an individual's previous employment sector, additional financial support was available from other sources. Employees have the right by law to take an unpaid leave of absence for studies. In addition, the Swedish social security system and other transfers may reduce the private risk and costs for participation in adult education. University studies in Sweden are free of charge and do not have any upper age limit for applicants. Admission is solely based on high school GPA or the Swedish Scholastic Aptitude Test (SweSAT) score, given that the basic requirements are fulfilled. Financial aid has the same rules for students at the university as students enrolled in Komvux, which amounts to around €1000 in grants and loans per month at current prices.

4. Data

4.1. Sources and sample

The data comes from different longitudinal population registers administered by Statistics Sweden (SCB) and from the military draft records administered by the Swedish Defense Recruitment Agency (Rekryteringsmyndigheten). The sample in this study is composed of displaced workers who were between the ages of 20 and 50 in 1994 or 1995 ($t_0$) who had a maximum registered educational attainment corresponding to three years of upper secondary school (high school) in $t_{-1}$.⁵ All were employed at $t_{-1}$, experienced job loss at $t_0$, and were not enrolled in Komvux the autumn semester prior to $t_0$ (i.e. autumn of $t_{-1}$). All workers are linked to workplaces and firms, and data include yearly observations between $t_{-3}$ and $t_{10}$, i.e. the follow-up period is 10 years. All individuals in the sample are Swedish citizens (but not necessarily born in Sweden) with a minimum of one year of tenure in a workplace with ten or more employees at $t_{-1}$. Furthermore, their workplace either shut down or reduced its workforce by 30% or more in conjunction with the individual's experience of job loss. Displacement took place between observations in November $t_{-1}$ and November $t_0$. Some plant closures and mass layoffs could have resulted from relocation or production restructuring. Therefore, employees who were separated from their workplace but kept working in other units of the same company are excluded from the sample.

4.2. Treatment and outcome of interest

Treatment is defined as enrolling in Komvux and completing at least one course (partial credits are not given) in any of the three consecutive semesters, starting with the spring semester $t_0$.⁶ The potential comparison group adheres to the same sampling restrictions as the treated and is composed of the workers who were displaced but did not register in Komvux during the same three consecutive semesters starting with the spring semester $t_0$. From this group of potential comparisons, a control group is constructed through propensity score matching as described in Section 5. Data include yearly information on individuals’ participation in education and educational attainment from 1990 to 2005. Variables used for matching and variables indicating outcomes are measured equally for treated and untreated. The data is not self-reported and has low attrition.⁷

Given the definition of treatment and the sampling criteria, the total underlying sample contains 48,101 individuals (26,831 males and 21,270 females), of whom 3.17% (1,524) are treated (Appendix Table A.1). Females are underrepresented in the total sample (44.22%) but overrepresented among the treated (63.91%). Among females, the proportion of the treated is more than twice that of males, 4.58% versus 2.05%. As expected, the proportion of treated individuals is higher among the young.

However, missing observations for some variables reduce the samples used in estimations. The inclusion of the mother's educational level (both genders) and the test score for cognitive ability (males)⁸ reduces the full sample observations from 26,714–17,910 (−33%) for males and from 21,162–17,702 (−16%) for females. Missing observations have only small effects on the estimated treatment effects and have no implications for the conclusions of this study (see section 6.2).

The outcome of interest is the attainment of at least a three-year university degree. A three-year university degree in Sweden is typically referred to as ‘kandidatexamen’ and is equivalent to a bachelor's degree. The treatment effect can be perceived as the effect of adult education as a short-term policy measure in direct conjunction with displacement, compared with no treatment, treatment later, or all other treatments. The estimate of the treatment effect is not conditioned on future events. For example, migration or participation in labor market programs during the follow-up period are considered to be potential effects of treatment or no treatment. Figure 1 shows the time trajectories of the unconditional cumulative proportion of individuals who have achieved at least a three-year university degree among the treated and the potential control group.

Figure 1. Sample means of graduation from university education, males (left) and females (right).

4.3. Variables

The data include variables measuring individual characteristics, such as wage earnings, various other labor market outcomes, income transfers related to the social insurance system, and parents’ socioeconomic status from 1990 onward. The data also provide information on high school GPA for individuals aged 20–27 and cognitive test scores from the mandatory military conscription tests for males. Additionally, the data include transcripts on course registration and credits from Komvux adult education (1986 onward). Since the data includes information on the mother's education, the individual's GPA, and cognitive test scores, it is possible to adjust for heterogeneity in nurture and nature factors affecting taste, motivation, and aptitude for education. These attributes are also not only associated with the costs of education, in terms of efforts required to obtain a university degree, but are also associated with increased returns to investment in human capital. (e.g. Roth et al. 2015; Coyle et al. 2018; Carneiro, Crawford, and Goodman 2007; Heckman, Stixrud, and Urzua 2006)

Table 1 presents means by gender and treatment status for the variables included in the full sample main specification for estimation of the propensity score matching model.

Table 1. Sample means for treated and potential control group, and p-values for the difference in means between treated and potential control group, and between treated and matched controls.

		Males					Females
	AE	No AE	p-val.	Matched p-val		AE	No AE	p-val.	Matched p-val
Age group 25 −29 (at $t_0$)	0.320	0.277	0.049	0.869		0.309	0.243	0.000	0.362
Age group 30+ (at $t_0$)	0.396	0.571	0.000	0.889		0.420	0.605	0.000	0.622
Born in Sweden	0.871	0.901	0.046	0.797		0.888	0.895	0.520	0.775
No. Children at home age 0–3	0.221	0.333	0.000	0.946		0.309	0.314	0.785	0.719
No. Children at home age 4–15	0.193	0.271	0.000	0.983		0.291	0.319	0.086	0.690
Sick benefits	0.112	0.108	0.589	0.830		0.226	0.212	0.725	0.645
Proportion of sick benefit of total income	0.019	0.012	0.018	0.603		0.038	0.036	0.656	0.984
Unemployment benefit	0.373	0.289	0.000	0.563		0.375	0.240	0.000	0.870
Difference in income ($t_{-2}-t_{-1}$)	36.0	9.4	0.405	0.772		49.1	15.6	0.118	0.958
Average income ($\left(t_{-3}+t_{-2}\right)/2$)	1930.3	2226.8	0.000	0.953		1360.0	1549.9	0.000	0.991
Employed at the same place from $t_{-3}$	0.492	0.595	0.000	0.905		0.512	0.608	0.670	0.636
Employed in $t_{-1}$ at a company with >250 workers	0.373	0.303	0.002	0.710		0.285	0.294	0.609	0.789
Workplace was shut down	0.427	0.405	0.371	0.809		0.308	0.315	0.641	0.927
Workplace was private	0.876	0.916	0.004	0.917		0.557	0.578	0.218	0.769
3-year high school education	0.214	0.215	0.961	0.901		0.281	0.268	0.422	0.903
GPA *	318.4	304.0	0.001	0.722		316.4	317.5	0.723	0.709
Military draft IQ test	0.534	0.489	0.067	0.932		no obs.
Education level mother ($\ge$ 2 years highschool)	0.625	0.498	0.000	0.847		0.590	0.475	0.000	0.692
Access to jobs	230000	200000	0.021	0.792		260000	240000	0.000	0.478
Nuts region 2 (NUTS1 level)	0.368	0.361	0.749	0.723		0.363	0.335	0.088	0.669
Nuts region 3 (NUTS1 level)	0.161	0.188	0.156	0.945		0.134	0.180	0.001	0.872
Cohort 1995	0.452	0.403	0.040	0.959		0.485	0.468	0.348	0.771
N (main sample)	429	17481				848	16854
N* (main sample plus GPA, age 20-27)	195	4655				340	4286

Although plant closures and mass layoffs reduce the problem of self-selection into job loss, there is likely selectivity into adult education on observed and unobserved individual characteristics associated with the outcome of interest. It is reasonable to believe that factors such as age, ability, unemployment, and parental background could affect the selection of enrollment into treatment (Becker 1962; Ben-Porath 1967; Card 1994). The significant differences in pre-matching means in Table 1 are generally in line with a priori expectations. Selection on unobserved factors is discussed in Section 5.

5. Method

The parameter of primary interest is the average treatment effect on the treated (ATT) of secondary-level general adult education on attaining at least a three-year university degree. In this study, ATT is estimated using propensity score matching (PSM) (Rosenbaum and Rubin 1983). The propensity score is estimated with logistic regression using the variables presented in Table 1. The measured outcome for each of the treated is then compared with the weighted average outcome of the four nearest neighbors among the non-treated in terms of the propensity score. The matching quality depends on the data quality and the balance of covariates for the treated group and the matched comparison group Table 1 shows significant differences in sample means between treated and non-treated before matching. However, after matching, the t-tests of differences in means between the treated and the matched comparison group do not indicate significant differences in means for any of the covariates (columns 4 and 8 in Table 1). Another threat to the identification of ATT is insufficient overlap in the propensity score of treated and untreated (Caliendo and Kopeinig 2008). As shown in Appendix Tables A2 and A3 and Figures A2 and A3, the distributions of the propensity score for the matched sample of treated and untreated indicate sufficient overlap.

For a causal interpretation of the ATT estimate, treatment assignment must be ignorable. As in all evaluations without random assignment into treatment, it is not possible to completely rule out influence of unobserved characteristics that could affect both participation in adult education and the probability of attaining a university degree. This study's advantage is access to high-quality data that enable control for a variety of characteristics potentially affecting selection into treatment and the outcome of interest.⁹ While this does not guarantee unconfoundedness, it makes the assumption of unconfoundedness more plausible. The biggest concern for biased estimates of ATT in this study lies with the potential influence of an individual's unobserved ability and its association with educational outcomes. This is addressed by using data that allow for extended control for heterogeneity in ability for certain subsamples in terms of the test score for cognitive ability and high-school grades (GPA).

While matching allows control for common support in data, there is still variation in treatment after matching, i.e. some individuals chose to enroll in education (the treated), and others chose not to enroll (the untreated). Some factors that may influence an individual's choice are difficult or impossible to measure, e.g. the assignment process involving interactions with an individual caseworker, peers and family opinions and attitudes, individual preferences, and motivation. This study uses two different methods to assess the plausibility of the unconfoundedness assumption (see section 6.2 for the results).

One method is a simulation-based sensitivity analysis (Nannicini 2007). This is used to test the sensitivity of the ATT estimates to unobserved confounding that would violate the conditional independence assumption (CIA). There are two ways to simulate this, and this study uses both. One is to simulate the effect of a ‘calibrated’ confounder, which is to test the effect of a confounder with the same empirical distribution as one of the binary variables used in the propensity score specification. This test makes it possible to simulate the effect on the ATT if an important variable were unmeasured, hence a possible confounder. The other way is to simulate a search for a ‘killer confounder,’ which is a confounder that would drive the ATT to zero or at least make the ATT insignificant. The ‘search’ for a killer confounder is done by iteratively testing different distributions of a possible confounder and evaluating the effect of such a confounder on the ATT. Both approaches have in common that the simulation gives the selection and outcome effect of the confounder as an odds ratio for the given ATT point estimate. While it is not possible to give a clear-cut answer as to whether it is reasonable to expect a given confounder, the simulated odds ratios can at least serve as indicators of whether such a confounder is likely.

The other method to assess the credibility of the CIA assumption is Imbens’s (2015) method of estimating the ATT on ‘pseudo-outcomes.’ This method relies on using pre-treatment variables as outcomes, which is crucial since we know that these outcomes should not be affected by the treatment and therefore should generate a (pseudo) ATT of zero (or close to zero) when matching on the same variables that are used to estimate the ATT on earning a university degree. Imbens argue that if the ATT estimates of the pseudo-outcomes are small and statistically insignificant, this can be seen as evidence supporting the assumption of unconfoundedness; if the contrary is true, it will cast doubt on the assumption of unconfoundedness. Four pseudo outcomes are used for the assessment: Wage earnings ${\mathrm{t}}_{-4}$, Wage earnings below the 10th percentile for two or more years ${\mathrm{t}}_{-4}$ to ${\mathrm{t}}_{-1}$, Wage earnings below the 10th percentile for three or more years ${\mathrm{t}}_{-4}$ to ${\mathrm{t}}_{-1}$, Average wage earnings ${\mathrm{t}}_{-4}$ to ${\mathrm{t}}_{-1}$.

6. Results

The individuals were displaced in 1994 or 1995, which gives a sample comprised of two cohorts of treated and non-treated. To exemplify this for the 1994 cohort, the baseline year ${\mathrm{t}}_0=1994$. The follow-up period $\mathrm{t}+$ until $T={\mathrm{t}}_{10}$ is then 1995, 1996, … , 2004. The pre-treatment observations are indexed $\mathrm{t}-$. The treated were enrolled in general adult education and completed at least one course at some point during a treatment period of three consecutive semesters (spring and autumn semester ${\mathrm{t}}_0$ plus the following spring semester ${\mathrm{t}}_1$). The outcome variable is dichotomous; it is equal to 1 if the individual has graduated from university education at time t, and 0 otherwise. Hence, the estimated ATT can be interpreted as a percentage point difference between the treated and matched comparison groups in the proportion of university degree attainment corresponding to at least a bachelor's degree.

The results are presented in the following order. First, estimates of ATT are shown for full samples of males and females using the main specification for the propensity score matching.¹⁰ Second, robustness tests that focus on missing observations are presented for the full sample estimates, as well as the full sample sensitivity analysis of the CIA and unconfoundedness assessment. Third, heterogeneity in effects by pre-treatment education is shown. Fourth is an examination of heterogeneity in effects by age and robustness checks by extended controls for ability. The ATT estimates are reported graphically for each year with a 95% confidence interval. The propensity score for selection into treatment (secondary-level adult education) is estimated for each subsample of displaced workers. Estimates for the full samples of men and women are presented in Appendix A3 and A4.

6.1. Full sample estimates of ATT

Figure 2 shows the results for both males and females using the full sample main specification of the propensity score. For the treated males (left), there is a small negative difference in the years $t_1$ to $t_3$, which indicates that the matched comparison group had a slightly larger proportion that hurried through university or could complete a previously unfinished degree. From $t_4$ to $t_{10}$, the point estimates are positive and statistically significant from $t_5$ onward. The ATT estimates indicate that a significantly higher cumulative proportion among the treated attained at least a three-year degree. The point estimate at $t_{10}$ shows a difference of 11.2 percentage points between the treated and the potential comparison group. The relative difference is +373% (14.2% compared to 3.0%).

Figure 2. Estimates of ATT, full sample of males (left) and females (right) aged 20-50, 95% CI:s.

For females (Figure 2, right panel), the negative estimates in the period $t_1$ to $t_3$ are slightly larger in magnitude than those for males, but from $t_4$ to $t_{10}$ the positive trend in ATT estimates is as clear as it is for males. At $t_{10}$, the point estimate indicates a difference of 10.0 percentage points between the treated and matched comparisons. The relative difference is +145% (16.9% compared to 6.9%).

Appendix Tables A2 and A3 and Figures A2 and A3 show treated and matched untreated balance in terms of the covariates’ means, overlap in the propensity score, and a variance ratio well within the rule-of-thumb range of 0.5-2.0 (Rubin 2001). Ex-ante assumptions of self-selection on observables into treatment are supported by the differences in pre-matching means in Table 1. The propensity score estimates (Appendix Tables A4 and A.5) indicate that factors such as age (-), parental education (+), cognitive ability (+), and prior unemployment (+) have statistically significant associations with investment in general adult education. These associations are in line with theory (e.g. Becker 1962; Ben-Porath 1967; Card 1994).

To test the sensitivity of the full sample results, the sample was trimmed based on the estimated propensity scores. Four levels of trimming were used – 5, 8, 10, and 12% (e.g. removing the observations with the 2.5% lowest and the 2.5% highest propensity scores). The ATT estimates at $t_{10}$ using the four different levels of trimming are shown in Table A.8 (see appendix). For males, the maximum deviation from the untrimmed result is at the trimming level of 10%, where the ATT is 1.9 percentage points lower (11.4 compared to 9.5). The maximum deviation from the untrimmed result for females is at the trimming level of 12%, where the ATT is 0.7 percentage points higher (10.7 compared to 10.0). While there is some deviation for both males and females from the untrimmed results, the estimates indicate positive treatment effects on the treated – estimates that are substantial in magnitude.

In summary, the full sample results indicate a significant positive effect of secondary-level general adult education on university graduation with at least a three-year degree. Due to the time needed to complete a degree, there is a time lag until positive treatment effects start to show. The main difference between males and females is the relative difference between the treated and the matched comparison at the end of the follow-up period. While the ATT point estimates for both genders at $t_{10}$ is of similar magnitude, the estimated effect in relative terms is much larger for males (+373%) than for females (+145%).

6.2. Robustness checks

Section 4.2 showed that there are missing observations, mainly due to two variables in the main specification of the propensity score model. To test the robustness of the full sample results with respect to the sample restrictions, the covariates indicating IQ test score (males) and mother's education level (both genders) were removed from the propensity score specification. This increases the total number of observations for males from 17,910–26,714 and for females from 17,702–21,162. Despite the relatively large increase in observations and less control for potential confounding, the results are stable. For males, the point estimate at $t_{10}$ indicates a treatment effect of 10.7 percentage points, which is 0.5 percentage points lower than the main specification point estimate. For females, the ATT estimate at $t_{10}$ is 9.7 percentage points, which is 0.3 percentage points lower than the main specification estimate.

Next, let us turn to the evaluation of the unconfoundedness assumption with the two methods described in Section 5. With the usual caveat that unconfoundedness is impossible to test, the two methods provide plausibility tests. Tables A.6 and A.7 (see appendix) show the result of the sensitivity test of the CIA assumption using simulation-based sensitivity analysis (Nannicini 2007). None of the calibrated confounders gave rise to any significant change of the $t_{10}$ ATT estimate for either males or females. Many calibrated confounders were tested, including the mother's education level, previous unemployment, and the cognitive test score (males), but the $t_{10}$ ATT estimates remained large and highly significant. The results were somewhat different for males and females when simulating a killer confounder (not shown in table A.6). For the males, the ATT estimate remained stable and significant even when using a simulated confounder with very large selection and outcome effects (odds ratio, OR > 70). For the females, it was possible to simulate a confounder that would render the positive result insignificant with smaller but still large selection and outcome effects (odds ratio, OR > 10). While the ATT estimate at $t_{10}$ seems less robust for the males, the simulations still indicate that it would take a rather extreme confounder to eliminate the positive and significant ATT for the females. The existence of such an extreme confounder among possible unobservables is not very likely. While these tests cannot rule out bias due to violation of the CIA (Ichino, Mealli, and Nannicini 2008), they indicate that the results for both males and females in the main sample are stable, at least with respect to moderate unobserved confounding.

Using the alternative method to assess the credibility of unconfoundedness (Imbens 2015) gives similar results. The method involves the estimation of the ATT for ‘pseudo-outcomes.’ Table A.9-A.12 (see appendix) shows the estimated full sample ATT at ${\mathrm{t}}_{10}$ on four pseudo outcomes: Wage earnings ${\mathrm{t}}_{-4}$, Wage earnings below the 10th percentile for two or more years ${\mathrm{t}}_{-4}$ to ${\mathrm{t}}_{-1}$, Wage earnings below the 10th percentile for three or more years ${\mathrm{t}}_{-4}$ to ${\mathrm{t}}_{-1}$, and Average wage earnings ${\mathrm{t}}_{-4}$ to ${\mathrm{t}}_{-1.}$ Estimates using trimming levels of 5, 8, 10, and 12% are reported for all four pseudo outcomes. All ATT estimates for the pseudo outcomes for males and females are statistically insignificant. Thus, the results do not indicate a violation of the assumption of unconfoundedness.

6.3. Heterogeneous estimates by education and age

Some displaced workers had 3-year high school education before job loss, making them eligible for some university programs without additional studies. I tested if the results varied for those with 3-year high school education compared to those without. Using the main specification of the propensity score model gives 3,857 males with three-year high school education and 14,053 males with less education. For females, there are 4,756 individuals with three-year high school education and 12,946 with lower educational attainment.

The ATT estimates for males at $t_{10}$ amount to 10.9 percentage points for both educational groups, while for females, they were 9.1 and 10.4 percentage points. All ATT estimates are significant at the 5 percent level. The difference in percentage points for both genders is relatively small between the two groups. However, displaced workers with a three-year secondary-level education have a higher propensity to enroll in university education after job loss and also a higher propensity to graduate in general. The relative difference between the treated and the potential control is much larger for the group with lower educational attainment than the sample with a three-year high school education. For males, the relative differences are +523% (13.1% compared to 2.1%) vs. + 143% (18.5% compared to 7.6%), and for females, + 247% (14.6% compared to 4.2) vs. + 67% (22.7% compared to 13.6%).

The following results focuses on heterogeneity in effects by age, which is examined using two subsamples conditioned on age. The first sample is restricted to individuals aged 20–27 at $t_0$. The age interval is based on available data on school grades (GPA), which allows for robustness checks by extended control for ability. The second sample is restricted to individuals aged 28–50 at $t_0$.

Figure 3 (left panel) shows results for males aged 20–27 using the main specification of the propensity score model. At $t_{10}$, the estimated ATT is 18.7 percentage points with a relative difference of +479% (22.6% compared to 3.9%). Adding the GPA covariate (right panel) shows a non-negligible difference of 2.5 percentage points lower at $t_{10}$, with a point estimate of 16.2 percentage points and a relative difference of +253% (22.6% compared to 6.4%). The disparity suggests positive ability bias using the main specification. Nevertheless, both specifications show a significant positive treatment effect on the treated.

Figure 3. Estimates of ATT, males aged 20–27 using the main specification of the propensity score model (left) and main specification plus GPA (right), 95% CI:s.

Turning to the females aged 20-27, the left panel of Figure 4 shows the results using the propensity score model. The results for females aged 20–27 have a different time pattern compared with males aged 20-27, with negative ATT estimates for the females in the short run. This difference may be caused by how childbearing and child-rearing affect females’ human capital investments and labor supply differently than males. The ATT estimate at $t_{10}$ is 10.4 percentage points, with a relative difference of +93% (21,8% compared to 11,3%). Adding the GPA to the propensity score specification does not change the results much (Figure 4, right panel). The point estimate at $t_{10}$ shows the same 10.4 percentage point difference as in Figure 4, left panel. Suppose GPA reflects abilities that affect the outcome of interest. In that case, it seems that the main specification of the propensity score model for females already accounts for it through other covariates.

Figure 4. Estimates of ATT, females aged 20–27 using the main specification of the propensity score model (left) and main specification plus GPA (right), 95% CI:s.

For the younger samples, the estimates indicate significant positive treatment effects. The extended control for ability suggests a positive ability bias in the estimation results for males using the main specification of the propensity score model. The most significant difference between males and females is the magnitude of the estimated treatment effect, both in terms of percentage points and relative size. However, there are no gender differences in outcomes among the treated. Instead, the difference between males and females among the matched comparisons, where a much smaller portion of the untreated males than the females attain a three-year degree, leads to a significant difference in the estimated treatment effect.

Changing the focus to the older cohorts, Figure 5 (left panel) shows the estimation results for males aged 28–50 using the main specification of the propensity score model. The ATT at $t_{10}$ is 5.7 percentage points, which is relatively small, but the relative difference between the treated and the matched comparisons is still large at 356% (7.3% compared to 1.6%).

Figure 5. Estimates of ATT, males (left) and females (right) aged 28–50 using the main specification of the propensity score model, 95% CI:s.

Figure 5 (right panel) shows the results for females aged 28–50 using the main specification. At $t_{10}$, the difference is 9.2 percentage points, whereas the relative difference between the treated and the matched comparisons is +270% (12.6% compared to 3.4%).

Although there is a significant and positive treatment effect for both older cohorts, it varies compared to the results observed in other samples. Specifically, the sample of older males is the only group in this study that exhibited a peak in the time trajectory prior to $t_{10}$. For the older males, the positive estimate of ATT in percentage points is reduced to half (5.7) at $t_{10}$, compared with the full sample (11.2) and less than a third compared to the age 20–27 sample (18.7). For the older females, the 9.1 percentage point ATT at $t_{10}$ is close to both the ATT for the full sample (10.0) and the sample aged 20–27 (9.3). The relative difference between the treated and the matched comparisons among males aged 28–50 was the lowest among all the male samples. On the contrary, for females aged 28-50, the relative difference is the highest among all female samples.

7. Discussion

This study investigates the effect of secondary-level general adult education of displaced workers on graduation from university-level education. The results indicate strong positive long-run treatment effects. For the full sample of males, enrollment in secondary-level adult education in conjunction with job loss is associated with an 11.2 percentage point increase in completion of at least a three-year university degree within ten years after job loss. The corresponding estimate for females is 10.0 percentage points.

For a policy measure applied so broadly, surprisingly little is known about educational outcomes after participating in general adult education at the primary or secondary level. This type of education is mainly targeted at non-typical students, such as displaced or unemployed workers, who are often the object of public investment in human capital. This study indicates positive treatment effects of secondary-level general adult education on enrolling in and graduating from a university within ten years after job loss. To my knowledge, there are no comparable previous studies. The results in this study indicate that general adult education, in conjunction with job loss, can serve as a stepping stone for displaced workers to boost human capital and skill by earning a university degree. This study has a relatively long follow-up period, and the results show that it takes four to five years before the positive treatment effect on educational attainment starts to show. The results suggest that a substantial proportion of the treated enrolled in higher education in the years following general adult education. When individuals who enroll in general adult education continue to higher education and pursue a degree, it will take time to complete their studies and reap potentially positive labor market effects. This finding may explain why studies that restrict the sample to individuals who do not continue to higher studies show positive short-run returns of general adult education on wage earnings (e.g. Jacobson, LaLonde, and Sullivan (2005b)), while studies with less restrictive samples show lackluster results in the short run but positive effects on wage earnings in the longer run (e.g. Stenberg, de Luna, and Westerlund (2014)).

In the case of university studies, the individual might be away from the job market for an extended period, during which she generates little or no tax revenue and may receive social benefits and subsidies. From a fiscal perspective, this raises a serious concern about recommending long stints of education for adults who could instead be working. However, there are potentially positive long-term effects that come with higher education, e.g. better health and higher productivity. These could have positive effects on government expenditure and tax revenue. Further studies are needed to investigate the total costs and benefits from a societal perspective over a longer period.

Finally, while the results indicate a strong positive effect of general adult education in conjunction with job loss on the attainment of a university degree, more studies are needed to examine the generality of the results for other samples, periods, and countries.

Disclosure statement

No potential conflict of interest was reported by the author.

Additional information

Funding

This work was supported by Vetenskapsrådet: [Grant Number 2015-01706].

Notes

1 There are longer-term ultimate goals such as increased productivity, closing gaps in educational attainment over generations, social mobility, and reduced inequality (e.g., OECD 2000, 56; UNESCO 2010; Skolverket 2022).

2 During the Covid-19 pandemic recession, the Swedish government has expanded the number of slots in adult education and introduced new regulations for secondary-level adult education. The latter are intended to help students with incomplete secondary education become eligible for university education (Swedish Ministry of Education 2021).

3 See, e.g., Jacobson, LaLonde, and Sullivan (1992) and Eliason and Storrie (2006) for similar sampling strategies.

4 There is no clear-cut distinction or definition in the academic literature about when youth education turns into adult education. In the Swedish educational system, an individual is eligible for publicly financed adult education from the age of 20, which is the definition used in this study.

5 At the time of displacement, 3,857 males and 4,756 females already had a three-year high school education, which meets the basic requirement for university eligibility. Still, individuals with three-year high school education may need to study additional topic-specific secondary-level courses to fulfill the requirements for some university courses or programs. Heterogeneity in effects by pre-treatment educational attainment is examined in Section 6.3.

6 The possibility of a ‘showing-up’ effect of education and a credit completion restriction are discussed in, e.g., Jacobson, LaLonde, and Sullivan (2005b).

7 These circumstances indicate that many potential sources that can lead to a biased estimation of the treatment effect are not present or are effectively addressed (e.g., Heckman, Ichimura, and Todd 1998).

8 This information is only available for males because it is based on mandatory military conscription tests. These are generally done at age 18-19.

9 Caliendo, Mahlstedt, and Mitnik (2017) found that normally unobserved characteristics, such as personal traits and attitudes, have little effect on the outcome of labor market programs if one uses comprehensive control variables that include labor market history. The relevance to evaluation of educational outcomes is not known. However, it is possible that some important unobserved heterogeneity can be correlated with observed pre-treatment labor outcomes or other observed characteristics.

10 The nearest neighbur (4) propensity score matching estimator in the t-effects package in Stata version 14 is used for estimation. The standard errors implemented in t-effects are calculated based on the formulas in Abadie and Imbens (2016).