The impact of father’s job loss on children’s school performance
Table 5 presents different estimates of the effect of father’s job loss on the standardised average grade. Standard errors are clustered at the family level. There are 137 clusters in the main analytic sample. The analysis uses the average grade standardised based on the mean and standard deviation of all students in the sample, with all the models controlling for year and stage of education dummies. For the 2012 cross section (CS in Table 5, see Eq. 4) and value added (VA in Table 5, see Eq. 5) regressions, standardisation of the average grade variables is performed at the year level.Footnote 23
Using the strict definition of involuntary job loss, the results with the main identification strategy used in this paper are shown in Column 1 (see Eq. 1). Fixed effect estimates show that after father’s job loss, students suffer a decrease in average grades of about 15% of a standard deviation. Columns 2 and 3 use cross-sectional data instead, with additional controls including a dummy that indicates whether the father has an education level beyond high school, the gender and the quarter of birth of the student (not shown). Column 2 shows the results when the cross section of 2012 is used, in a fashion similar to that used in the plant closure literature. The magnitude of the coefficient increases considerably compared to that in Column 1. Controlling for the student’s average grade prior to father’s job loss in the 2012 value-added regression shown in Column 3 increases the precision of the estimate. This value-added result indicates that father’s job loss entails a significant decrease in average school performance that almost doubles the impact measured with the fixed effect regression in Column 1. This exercise suggests that, unless the set of controls available is extremely rich, the two latter strategies render estimates that could suffer from potential bias due to (negative) selection of whom is laid off and underlines the usefulness of panel data to address it.
Given that the information to construct kindergarten grades differs from the information in compulsory education grades, Columns 4 to 6 in Table 5 repeat the same exercise but excluding from the analysis and the standardisation process all those grades obtained in kindergarten. The results are very similar to those obtained in Columns 1 to 3. In order to not disregard additional observations, the remaining analysis will continue to use kindergarten grades. All the previous results in Columns 1 to 6 are based on the restricted sample, i.e. the one resulting after applying the sample restriction criteria outlined in Sect. 3. In Column 7, I present results using the full sample of students observed in 2012. The point estimate in Column 7 is smaller (in absolute terms) than the one in Column 1, but the results point in the same direction. This decrease in the magnitude is expected since this regression includes students that are not observed in 2008, but only after father’s job loss has occurred (and therefore, for these students the grade might had already decreased or started decreasing by the time they entered the sample).
Results go in the same direction when using the other two definitions of job loss (see Table 14 in Appendix B). As it could be expected given the more loose definition of job loss used in Panel A, the magnitude of the estimates decreases with respect to those in Table 5. And vice versa for the results in Panel B: Children of fathers suffering job loss due to downsizing or firm closure suffer a decrease in the average grade of almost 22% of a standard deviation (Panel B, Column 1). The remaining analysis will continue to utilise the strict definition of involuntary job loss used in Table 5.
Table 6 Placebo: average effect of father’s job loss on the cross section of 2008 Table 7 Excluding fathers that in 2008 had a lower labour market attachment Tables 6 and 7 suggest that the effects found in Table 5 for the fixed effect regressions are likely to be of a causal nature. In Column 1 in Table 6, I show the results of a placebo test using the strict definition of job loss. Future job losses (i.e. job losses that will happen later in the period) should not have an impact on the average grade of students prior to father’s job loss in 2008 (the first academic year in the sample, where by construction all students have employed fathers). The estimate for the main variable of interest is highly imprecise and very close to zero. This finding provides evidence against the possibility that changes in household’s unobservables simultaneously drive father’s job loss (FJL) and school performance of their offspring, since otherwise we would expect to see significantly worse school performance prior to father’s job loss. I additionally display the coefficients of different control variables in Column 1, in order to show that the average grade measure is sensible even in this rather small sample. The results are in line with facts well established in the economics of education literature: students coming from better socio-economic backgrounds (as measured by the father’s education level) perform better and females obtain better grades as well as those born at the beginning of the year (the oldest in the class). Results in Column 2 confirm that the results are the same when kindergarten grades are excluded from the analysis.
The former placebo test indicates that future paternal job losses do not significantly affect grades prior to father’s job loss. This evidence does not guarantee that the grades of treated students were already suffering a decline prior to father’s job loss. However, the evidence shown in Sect. 4 does, given that Fig. 4 showed that there are no existing negative trends in school performance for treated students prior to treatment.
Another way to partially address this issue is to check whether the impact of father’s job loss is mainly driven by those students whose fathers had a lower labour market attachment prior to job loss. One potential way of defining labour market attachment is to use the information on years of tenure at the firm before losing the job. Workers with lower tenure prior to job loss might have been on a different (negative) trajectory prior to losing their job during the Great Recession, and this could, in turn, have affected the performance of their offspring. In order to verify this, I consider only those students whose fathers in 2008 had at least three (or six) years of tenure in their jobs, respectively. Results are shown in Column 1 (and 2) of Table 7. The estimates show that the impact of FJL remains negative and significant in all cases. That is, the negative effect of FJL does not seem to be driven by those students whose fathers had lower tenure at the firm prior to job loss, but rather, by those students whose fathers had a more stable situation prior to losing the job. This is in line with several papers in the job loss literature that have found that workers with longer tenure prior to losing their jobs suffer more after job loss in terms of income declines and employment probabilities.Footnote 24 In this sense, these children would suffer a larger shock after paternal job loss than those children whose fathers had, prior to job loss, a lower labour market attachment. In Column 3, I use an alternative definition for labour market attachment based on the type of contract the father had prior to job loss: students whose fathers had a temporary contract in 2008 are excluded from the sample. The results are very similar to those shown in Columns 1 and 2. Additionally, suffering multiple job losses during the period under analysis might also indicate a rather unstable attachment to the labour market. However, it is also possible that multiple job losses indicate a better ability to find new employment during the Great Recession. In any case, multiple job losses could be due to unobserved time varying heterogeneity that could bias the estimates. In Column 4, I exclude students whose fathers have experienced more than one job loss in the period. Stevens (1997) studied the effects of multiple job losses on earnings and found that much of the persistence in the earnings losses can be explained by additional job losses in the years following an initial displacement. Initial displacements predict future displacements, and thus, subsequent displacements might not be exogenous (in the sense that they might no longer be attributed to the combination of the Great Recession and fixed effects). By excluding from the sample those students whose fathers experienced multiple job losses during the period under analysis, the estimate remains negative and very similar to the other point estimates in Table 7. All in all, the evidence in Tables 6 and 7, together with Fig. 4, suggests that treated students were not on a different (negative) trend prior to father’s job loss.
Table 8 Robustness check: group-and-year and group-and-stage of education specific effects A further potential concern raised in Sect. 4 is that estimates might be driven by negative trends in school performance for particular groups of students, either because of differential effects by subgroups as the crisis unfolded, or due to differential trends in parental investment as the child advances within the education system. To address these concerns, I show in Table 8 the results of different regressions where the original model is augmented by interacting the year dummies (stage of education dummies) with certain group-specific characteristics measured prior to job loss in Panel A (B). The year (stage of education) dummies are interacted with a variable that is equal to 1 if the father has a high level of education (beyond high school) in Column 1; a variable that equals 1 if the father was classified in the high-income category in 2008 (Column 2) and a dummy that equals 1 if the father owned a business in 2008 (as opposed to working for a firm) in Column 3. The year (stage of education) dummies are also interacted with dummies for the sector the father was employed in 2008 (3 sector categories are used: manufacturing, construction and services) in Column 4, whether the household lives in a house that is fully paid in Column 5 and student’s gender in Column 6.Footnote 25 The last model (Column 7) includes all the group specific trends in Columns 1 to 6 together. The point estimates shown in Table 8 are all very similar to those in Column 1, Table 5. This evidence suggests that the estimates presented so far do not simply reflect differential group-and-year specific effects or group-and-stage of education effects.
Table 9 Impact of mother’s job loss (and labour market status) on school performance The role of mother’s job loss
In Column 1, Table 9, I show the impact of mother’s job loss (MJL) on her children’s school performance. The same analytic sample used so far is employed here, but further excluding those mothers that were unemployed at the beginning of the period. The MJL variable is defined in the same way as FJL, using the preferred definition of strict job loss. It is equal to 1 from the (academic) year that the mother loses her job. The results in Column 1 show that there is a negative but no significant effect of MJL, on average, on her children’s school performance. These results do not seem to be driven by the fact that women in a country like Spain might have a lower labour market attachment. As it was shown in Table 2, 81% of the mothers in the analytic sample were employed in 2008.
This paper differs from almost all other papers in the literature in the sense that job losses happen during a deep economic crisis. If mother and father job losses are correlated (which is indeed the case in this sample), then the effect of FJL could also be capturing the impact of MJL on the average grades of their offspring. The results in Column 2 suggest that this is not the case. The MJL coefficient becomes smaller in magnitude when father’s job losses are introduced, whereas the FJL coefficient remains almost the same (see Column 1, Table 5). Finally, mothers could react to father’s job loss by going back to work (in the case that they were unemployed prior to FJL). The results for the FJL variable shown in Column 3 barely change when augmenting the specification with a dummy variable that equals 1 whenever the mother is employed, and 0 otherwise.
The findings in this section could partly be explained by the results in recent papers in health economics that have found that men suffer more negative health-related consequences after job loss than women. For instance, Kuhn et al. (2009) find that job loss significantly increases expenditures for antidepressants and related drugs, as well as hospitalizations due to mental health problems for men, but not for women. Eliason and Storrie (2009a) find that job loss produces a twofold short-run increase in suicides and alcohol-related mortality for both sexes. However, overall mortality risk among men increased by 44 percent during the first 4 years following job loss while there was no impact in the longer run or on female overall mortality. Eliason and Storrie (2009b) find that job loss significantly increases the risk of hospitalisation due to alcohol-related conditions, among both men and women, and due to traffic accidents and self-harm among men only.
In terms of earnings decline after job loss, both men and women suffer substantial decreases in the probability of being observed in the high-income category (a decline of 31% and 24% for men and women, respectively). The larger contribution of fathers to household income could also be behind these results. Whereas 65% of the fathers were observed in the high-income category in 2008, only 24% of mothers reported to be in the high-income category. Findings reported by social psychologists suggest that there are detrimental effects of job insecurity (something that is likely to be positively related to job loss) on financial anxiety for men but not for women (Lim and Sng 2006).
A further potential explanation could be found in the theories of social roles and identity, as pointed out by Rege et al. (2011) who come to similar conclusions as this paper with regard to the effects of maternal and paternal job loss. These authors highlight that social norms and historical employment patterns have allowed women to develop a greater range of non-employment-related roles. This, in turn, would make women more adaptable and equipped to handle job loss, whereas job loss could be more detrimental for men because a large part of their identity is connected to their specific job.
Alternative treatment definitions and the role of long-term unemployment
So far, given the reasons stated in Sect. 4, the treatment variable has been defined as an absorbing state (i.e. it equals 1 from the moment the father loses the job, irrespective of his employment situation afterwards). However, it is interesting to see what happens if the treatment definition is changed to allow those fathers who find a job after job loss to switch treatment status.
Table 10 Alternative treatment definitions and the role of long-term unemployment I show the results of experimenting with two different treatment definitions in Columns 1 and 2 in Table 10. ‘All FJL’ is a dummy variable that measures the impact of all unemployment spells irrespective of the duration, whereas ‘FJL leading to long-term unemployment’ only equals 1 if (and when) the father stays unemployed at least for an academic year.Footnote 26 In this sense, ‘FJL leading to long-term unemployment’ would be capturing the effect of long-term unemployment spells, whereas ‘All FJL’ would be capturing the impact of father’s job loss and long-term unemployment. The results in Columns 1 and 2 of Table 10 suggest that the negative impact of father’s job loss on the average grade is mainly driven by those fathers that stay unemployed for at least one academic year.
Additional robustness checks
I present a series of additional robustness checks in Table 11. In Column 1, I only use information from academic years 2008, 2010 and 2012. As described in Sect. 3, by restricting the sample to these periods I do not need to make any assumptions with regard to the exact date of job loss. The estimates in Column 1 show that the coefficients of the FJL variable are also negative and significant, and slightly larger in magnitude. In Column 2, I present results when controlling for grade-specific dummies (rather than stage of education dummies). Results are very similar to the preferred estimate. Given the small sample size, it is important to verify that outliers are not the main drivers of the results. In order to address this concern, in Column 3 I show the FJL estimate when observations at the extremes of the grade distribution are dropped. I calculate the average change in the average grade between the academic years 2008 and 2012 and run the main specification excluding observations for which the average change falls in the 5th and 95th percentile. Applying these restrictions has almost no effect on the estimate of FJL.
Table 11 Other robustness checks
Job losses in this article happen during a period where many individuals lost their jobs. In order to understand whether general job losses are affecting the results, I calculate the percent of students whose fathers suffered a job loss in the same grade and year, and also in the same grade, year and class. This controls for potential peer effects of parental job losses of classmates/friends. Introducing these variables in the main specification (see Columns 4 and 5) barely changes the original point estimates of FJL in the preferred specification in Table 5 (Column 1). Moreover, the peer group effect coefficients are not significant. Finally, Column 6 examines the role of job loss of the main earner in the household. As pointed out in Sect. 3, if parental job loss leads to separation or divorce and the father moves out during the period under analysis, the sample will not register these cases. Therefore, the results shown so far are constrained to the sample of students whose parents have stayed together during the period under analysis. In Column 6, JL main earner is defined in the same way as the FJL variable, but taking into account the job losses of the mother when the father is not present in the household. The results barely change.
Table 12 Heterogeneous effects Heterogeneous effects
Following the related literature, I analyse whether the impact of father’s job loss is heterogeneous across different subgroups in Table 12. The results need to be interpreted with caution given the limitation posed by the sample size (standard errors tend to be large), but nevertheless they add new suggestive evidence on likely mechanisms. In Column 1, I reproduce the results of the preferred specification.Footnote 27 The subsequent models in the table interact the FJL variable with one characteristic at a time. The results suggest that the effects of FJL are concentrated on those students whose fathers have low levels of education (in Column 2, Table 12, the P value for the interaction is just under the 10%).Footnote 28,Footnote 29 The results also suggest that those students whose fathers lost their jobs because they closed their businesses suffer a more detrimental effect of FJL (Column 3). However, standard errors for the interaction term are too large to draw any strong conclusions. Additionally, FJL does not seem to have a differential impact for older students in Column 4 (older students are those enrolled in secondary education in 2012). Column 5 shows that the effect of FJL is mainly concentrated on those students whose families either rent or have a mortgage (as opposed to households that fully own their property).Footnote 30 Finally, the results in Column 6 show that the negative effect of FJL seems to be more detrimental for those students whose families moved during the period. The results in the last two columns might be picking up the same mechanism, if families that suffered FJL and were not in possession of a fully paid property, were evicted or had to move as a consequence of FJL.
Table 13 Income reductions across different subgroups Several papers in the literature have documented a considerable reduction in earnings after job loss. For instance, Jacobson et al. (1993) reported that high-tenure workers separating from distressed firms suffer long-term losses averaging 25% per year. Accordingly, Column 1 in Table 13 shows that after job loss, the fathers in the sample are about 32 percentage points less likely to be observed in the high-income category. The remaining models in Table 13 explore whether the heterogeneous effects shown in Table 12 could be due to differential reductions in income by group. This is done by regressing the FJL variable on a dummy variable that equals 1 if the father is observed in the high-income category, controlling for individual fixed effects. The order of the columns is the same as in Table 12. The evidence partly suggests that those groups with a smaller decrease in average grades are also those groups with a smaller decrease in income, although standard errors of the FJL variable are rather large and the coefficients on the interactions are highly insignificant in most cases. The different income reductions across groups could be suggesting that income is one of the mechanisms driving the negative effect of paternal job loss on the school performance of their offspring. However, it is important to note that these results are obtained for students who are enrolled in the same school during the five periods of observation. The observed reduction in income cannot be linked, therefore, to changes in the school attended after job loss. Reductions in income could partly explain the results if fathers decrease hours of extra help with homework (or other extra school activities) after job loss or decrease consumption that could be related to school performance. Additionally, income reductions after father’s job loss could entail higher stress or financial anxiety and uncertainty for affected individuals and households, as some papers in social psychology and health economics have suggested [see, for instance, Lim and Sng (2006) and Kuhn et al. (2009)].