Labor Market Policy and Subjective Well-Being During the Great Recession

Abstract

While subjective well-being generally decreased in Europe during the Great Recession, some countries fared better than others. We assess how this experience varied across population subgroups and countries according to their labor market policies, specifically two types of unemployment support policies and employment protection legislation. We find both types of unemployment support, income replacement and active labor market policy (which assists the unemployed to find jobs), reduced the negative effects for most of the population (except youth); however, income replacement performed better, reducing the impacts of the Recession to a greater degree. In contrast, stricter employment protection legislation exacerbated the negative effects. This finding may be explained with suggestive evidence that indicates: legislation limiting the dismissal of employees curbed increases in unemployment but this benefit was more than offset, plausibly by perceptions of increased employment insecurity; and legislation limiting the use of temporary contracts may have exacerbated increases in unemployment. Our analysis is based on two-stage least squares regressions using individual subjective well-being data from Eurobarometer surveys and variation in labor market policy across 23 European countries.

Appendix

1.1 Assessing Instrument Validity Using Overidentification Tests

The overidentification tests are conducted based on the main analysis using additional instruments, which were generated using the Lewbel (2012) method. The Lewbel (2012) method uses heteroskedasticity in the data and higher order restrictions to generate instruments without introducing external data. While somewhat new, it has been used numerous times now: Lewbel (2012) documents papers as early as 2007, and more recently by (Arampatzi et al. 2018; Denny & Oppedisano, 2013; Le Moglie et al. 2015; O’Connor, 2020; O’Connor & Graham, 2019; Sarracino & Fumarco, 2018).

Using the Lewbel (2012) method, we generate the instruments as follows: (1) run a regression of our endogenous variable, ${group}_{g} X {trough}_{t} X {policy}_{j}$, on the other covariates from Eq. 1 and store the residuals (${\mu }_{igjt}$), (2) de-mean the covariates and multiply them by the stored residuals. For example, instrument ${Z}_{igjt}^{*}=\left({Z}_{igjt}-\overline{{Z }_{gjt}}\right)* {\mu }_{igjt}$, where ${Z}_{igjt}$ is any subset of the covariates and ${\mu }_{igjt}$ are the stored residuals. Beyond the standard assumption (all of the covariates are exogenous), the method relies on two key conditions. First, heteroskedasticity, which can be tested using the standard Breusch-Pagan test. The second condition is untestable and relies on an assumption. Specifically, the residual from the first step above (${\mu }_{igjt}$) multiplied by the second stage residual of life satisfaction (${\varepsilon }_{igjt}$) must be unrelated to the covariates used to generate instruments, formally: $cov\left({Z}_{igjt},{\varepsilon }_{igjt}*{\mu }_{igjt}\right)=0$. For Z, we only use gender. Gender is chosen as the only additional variable because we only need one additional variable per endogenous variable to test the overidentification restrictions and gender is exogenously determined. Indeed, introducing additional excludable instruments would weaken any findings that suggest peak policy is excludable. Overidentification tests apply to the full set of excluded instruments, and peak policy forms a smaller proportion of the set with more excluded instruments. We use gender instead of age because age is used as one of the groups of interest. In this way, we generate three additional excluded instruments for each 2SLS regression, one for each endogenous variable (policy by group). For a further description of the approach, see Baum et al. (2013) and Lewbel (2012). The user written command ivreg2h (Baum & Schaffer, 2012) can be used to generate the instruments in STATA.

We had to make two further adjustments to the main analysis in order to run the Hansen J overidentification test. First, because the number of clusters (countries) is too small for the Hansen test, we clustered at the country-period level to double the number of clusters. However, the clustered standard errors were quite similar in both cases, providing some reassurance that the change does not affect the overidentification test. Second, we partialed out the country fixed effects (dummies) to reduce the number of variables.

Table 6 presents the results using the Lewbel generated instruments. They read the same as in Tables 1 and 2, with the addition of the Hansen J p-value, which reports the overidentification test results. As noted in the main text, the overidentification tests support the validity of our instruments. We fail to reject that the instruments are valid across each column. What is more, the coefficient estimates are nearly identical to the main results, consistent with expectations. We expected the coefficients to be similar because in the main analysis our instrument, peak policy, strongly predicts trough policy. This strength means that adding additional instruments should not greatly affect the first and second stage results. Across Tables 1 and 2, the lowest F-stat was nearly 24, more than double the often-used cut off value of 10 for weak instruments.

Table 5 Peak and trough policy variable values, by country

Table 6 Effects of policy variables on life satisfaction by level of education and cohort with additional Lewbel generated instruments. Dependent variable: life satisfaction (1–4)

1.2 Ordinal Treatment of Life Satisfaction

Life satisfaction is inherently ordinal, yet the main analysis treats it cardinally. More generally, there are a few papers which critique inference based on subjective well-being, stating limitations based on the way it is reported or assumptions necessary to operationalize it (e.g., Bond & Lang, 2019; Schröder & Yitzhaki, 2017). Relatively early work has addressed the ordinal versus cardinal treatment of subjective well-being. For instance, Ng (1997) argues subjective well-being should conceptually be treated as cardinal, and Ferrer-i-Carbonell and Frijters (2004) find little difference between estimates when using cardinal versus ordinal models. However, the debate continues today, recently with responses to Bond and Lang (2019) and Schröder and Yitzhaki (2017) such as Chen et al. (2019) and Kaiser and Vendrik (2019). Cardinal treatment is often preferred because ordinal treatment has some limitations, for instance individual fixed effects are not possible in standard ordered choice models and instrumental variable methods would require using different models for each stage (linear and nonlinear). To address these issues, practitioners sometimes use both methods.

As mentioned in Sect. 5.3, we use two additional analyses that treat life satisfaction ordinally, ordered probit and linear probability. For the ordered probit model, we do not use an instrumental variable approach because we prefer to maintain the same treatment, linear or nonlinear, for each stage. To conduct the linear probability model, we compress life satisfaction into a binary outcome, coded: 1 if the respondent responds in the top two categories, very or fairly satisfied; and 0 if they respond not very or not at all satisfied. This variable has an intuitive description, the probability that someone is more satisfied than not. See Sechel (2019) for a description of how this approach addresses measurement issues in subjective well-being (e.g., Bond & Lang, 2019). Given both stages are linear, we use instrumental variables like the main analysis.

The results are presented in Table 7. For each policy variable there are two columns, one for each method. Two panels present the results separately by education and cohort. The estimates are consistent in significance and direction (magnitudes are not directly comparable) with the main results, with two exceptions. For ALMP, two estimates according to the linear probability model are inconsistent; the most important change is for those with at least college education. Indeed, the effect changed direction. While the ordered probit and the results by cohort using both models are consistent, this change should be considered when interpreting the impacts of ALMP. For EPL, the estimate for the high school group was statistically significant but is no longer when using the linear probability model. This change does not pose much of a limitation however. It may be due to lower variability in the binary measure of life satisfaction, and the other estimates are consistent with the main results.

Table 7 Ordinal treatment of life satisfaction, two sets of regressions, by education and cohort

1.3 Nonlinear Specification of Job Situation Worse

According to theory, Job Situation Worse should be specified using a nonlinear model because it is a binary variable. To test the sensitivity of our two-stage least squares estimates, performed in Table 4, we replicate the analysis instead using a probit model. However, once again the nonlinear model precludes the use of two-stage least squares, meaning the relations should not be interpreted as causal. Nonetheless, the results are solely intended to supplement Table 4, and the analysis in which is free from many concerns; it is only limited in the linear treatment of Job Situation Worse.

The results presented in Table 8 are qualitatively similar to those of Table 4. With stricter employment protection, the at least college and eldest age groups perceive statistically significantly worse job situations, while the other groups do not. There are some differences between EPL and EPL-T however. In Table 4, EPL affects both groups (at least college and eldest) while EPL-T does not affect either. Here EPL does not affect the eldest age group and EPL-T affects both groups.

Table 8 Probit regressions of Job Situation Worse (Binary, Worse = 1) on EPL and EPL-T by level of education and cohort