Intergenerational Scars: The Impact of Parental Unemployment on Individual Health Later in Life

3.1 Data, Sample Selection Criteria and Treatment Definition

We obtained our sample by combining two datasets: (i) the German Socio-Economic Panel (SOEP-Core v36, EU edition), a household panel survey representative of the adult German population (Wagner, Frick, and Schupp 2007);^[1] (ii) the Federal Employment Agency statistics. The SOEP represents the main building block. Started in 1984, the survey is carried out annually. In 2021, more than 30,000 adult individuals participated in the panel.^[2] The SOEP presents several useful features. First, once a household is selected, all members are interviewed and they may remain in the sample even if they separate from the original household. If an individual gets married or starts a cohabitation, also the partner and the future kids join the sample.^[3] This longitudinal structure allows to track individuals over time, regardless of complex family dynamics. Individuals exit the sample only if they die, move abroad or decide to quit.^[4] Second, the SOEP collects retrospective information. By combining both retrospective and currently acquired information, a researcher can recreate the full history of each SOEP member. Third, the SOEP covers a vast range of topics, such as health and consumption habits, well-being, labor conditions, political affiliations, income, etc.

We started with 149,565 individuals for a total of 1,164,296 observations. We kept only those individuals with at least one parent interviewed by the SOEP when the individual was between 0 and 15 years old. Hence, these parents could have been either married, cohabiting or singles when the individual was a child. We restricted on individuals born from 1984 to 2000 and aged between 18 and 31 years old when interviewed. In addition, we required individuals to have lived with their parents until they turned 15. Then, we dropped individuals who had missing observations in the outcome variables.^[5] Finally, we removed those individuals with at least one parent who experienced unemployment not due to plant closure. Moreover, for the sake of having similar control and treated units, we also required that the parents of both groups had the same initial labor market condition: at the time when an individual was born, her/his father had to be a private employee, whilst her/his mother had to be a either private employee or not employed.^[6] We ended up with 1,859 individuals for a total of 4,822 observations. Table 1 reports how the sample size shrank through our selection process.

We applied these sample selection criteria for different reasons. First, by requiring parents to be in the sample, we are able to collect information on parental unemployment directly from their files, with their children providing information about themselves when adult. By doing so, we reduced the risk of reporting bias, i.e. the measurement error due to an individual misreporting information, in this case, about someone else. Second, by focusing on children born between 1984 and 2000, we can observe the full history of each individual and his/her family and we avoided losing potentially relevant information. Third, by requiring individuals to have spent their whole childhood living with their parents, we are sure that the kid was fully exposed to parental unemployment when it occurred. Finally, with the last sample selection criterion, we leave in our sample only individuals who either did not experience parental unemployment during their childhood (the control units) or experienced parental unemployment due to plant closure during their childhood (the treated units). The latter sample selection criterion is the pillar of our identification strategy, which is discussed in Subsection 3.4.

We defined the treatment using information from two SOEP questions. The first question is about the reason for the job termination, which is part of the survey since 1985 and since 1991 it includes “due to shutdown” among the possible answers. The second question is on the (un-)employment status. We defined as treated those individuals who, when they were between 0 and 15 years old, had at least one parent that was fired once or more times due to a plant shutdown and registered the unemployment event at the Federal Employment Agency. Individuals whose parents became unemployed for reasons different from a plant closure were excluded from the sample.^[7] We also excluded individuals whose parents were in other labor market status, like workers for the public sector, retired or self-employed.^[8]

Given our sample selection criteria as summarized by Table 1 and our treatment definition, we identify the effect of interest by comparing health outcomes of individuals who were born between 1984 and 2000, whose fathers worked as private employees and for whom at least one parent experienced at least an unemployment event due to plant closure when they were 0–15 years old to health outcomes of similar individuals whose parents never experienced any type of unemployment. Therefore, one may wonder if our findings may have enough external validity. Some of the adopted sample selection criteria are not likely to have an impact on the external validity, like the exclusion of persons with fathers who were in the public sector, self-employed or retired. This is justified by the fact that we are interested in a population that is homogeneous in terms of previous labor market history, health insurance status, or probability of job dismissal. Similarly, the selection of individuals who lived either in single-parent or two-parent households until they were 15 years old should very likely represent the vast majority of the German population born between 1984 and 2000. However, individuals born in single-parent households may fare differently against stressful events such as unemployment compared to peers born in two-parent households. Typically, two-parent households can benefit from the insurance of the ‘added worker effect’, that is the labor supply response of spouses to the loss of job of the partner (see e.g. Halla, Schmieder, and Weber 2020; Lundberg 1985; Stephens 2002). As such, individuals born in two-parent households may be more resilient toward parental unemployment, as one parent may have compensated the income loss of the other. In Subsection 4.3, we tested this hypothesis considering only individuals born in two-parent households. Finally, the most important restriction to the external validity of our analysis is probably the choice of limiting the focus on job displacements due to plant shutdowns (Suppa 2021). This was done to reduce endogeneity concerns about selection into treatment. It comes nonetheless with a loss of external validity, because the effect of a parental unemployment due to other causes may differ from the one due to a plant shutdown. As a sensitivity check, in Subsection 4.3, we provide estimates when the treatment is enlarged to those who experienced at least one parental unemployment event, independently on the reason for the job displacement.

3.2 Health Outcome Variables

Our health outcome variables are built on the items available in the SOEP version of the Short Form Health Survey (SF-12v2) (Ware, Kosinski, and Keller 1996). The SOEP introduced this questionnaire for the first time in 2002 and ever since every two years. This survey covers several health dimensions like bodily pain, emotional draining, and social functioning, with each item referring to the last 4 weeks period of the life of the individual.^[9]

By a factor analysis and as suggested by Andersen et al. (2007),^[10] we calculated two indicators of general health, the Mental Component Summary (MCS) and the Physical Component Summary (PCS) scales (Ware, Kosinki, and Dewey 2000). They take values from 0 to 100, with mean 50 and standard deviation 10 in the 2004 SOEP sample. Higher scores indicate better health. The two variables were proved to be valid and reliable measures of the current general health status compared to other scales (Gill et al. 2007) and are widely used in the economic and psychological literature (Marcus 2013; Neubert et al. 2019; Schmitz 2011; Stauder 2019]).^[11]

Table 2 reports their summary statistics and Figure A.1 in the Online Appendix plots the density distributions of the outcome variables distinguishing between treated and control units. On average, the treated individuals reported worse mental and physical health conditions. Whereas the difference in mental health is very tiny, it is substantial in terms of physical health, with the treated having a lower physical health of almost one third of the standard deviation.

Our health variables and the life satisfaction used by Nikolova and Nikolaev (2021) are positively correlated.^[12] This and the figures in Table 2 suggest that previous findings by Nikolova and Nikolaev (2021) on life satisfaction may be the consequence of both a physical and a psychological scar.^[13] On top of the speculation in Nikolova and Nikolaev (2021) for their findings, which was based on parental unemployment generating a psychological trauma, it may be that the treated individuals engaged in risky behaviors later in life, like smoking or drinking, which in turn impaired their physical health, leading to a decreased level of life satisfaction. One of the contributions of our analysis is that it sheds light into the mediating channels possibly explaining the findings about life satisfaction in Nikolova and Nikolaev (2021).

3.3 Control Variables

Following previous studies on parental unemployment effects on individuals’ outcomes (see e.g. Bratberg, Nilsen, and Vaage 2008; Ekhaugen 2009; Oreopoulos, Page, and Stevens 2008; Siedler 2011), we included a large set of individual and family socio-economic characteristics. The individual characteristics were gender, age, migration background and whether the individual was the first-born child in his/her household. We also included number of siblings when the individual was born and three sets of dummy variables for the survey in which the individual was interviewed, the year of birth and the region of Germany where the individual was born. This set of variables allowed to control for time and spatial heterogeneity. The family characteristics were the age that the parent had when the child was born, the parental migration background, the highest educational level achieved by parents, the cumulative unemployment experience of parents and the income quintile in which the household found itself (Bratberg, Nilsen, and Vaage 2008), measured when the individual was born. We included the household income variable in order to separate the potential poverty component from the parental unemployment effect (Siedler 2011). Furthermore, we controlled for local labor market conditions at the time of the eventual treatment by using the average unemployment rate at länder level.^[14] Finally, we included missing indicators for each of the aforementioned variables in order to save observations and avoid losing statistical power. Tables containing summary statistics of the control variables are in Online Appendix A.

We did not control for contemporaneous characteristics because their realization may be the consequence of the treatment and be therefore themselves outcome variables, i.e. “bad controls” (Angrist and Pischke 2009, p. 64). For example, the educational level that an individual achieves may be causally impacted by past parental unemployment. Education is linked with earnings and income which, in turn, are used to buy health goods and improve the health condition (Grossman 1972, 2000]; Mincer 1974).

3.4 Identification Strategy

Identifying the causal effect of parental unemployment on children’s health later in life may be tricky due to omitted variables (Brand 2015; Farré, Fasani, and Mueller 2018; Kassenboehmer and Haisken-DeNew 2009). In our framework, omitted variables are those unobservables which are jointly correlated with the likelihood of parental unemployment and children’s health later in life, for example family and social background, predetermined and innate offspring’s health conditions and household wealth. We tackled this endogeneity concern by using plant closures as the reason for unemployment entry. Plant closures may be regarded as a quasi-experimental variation of the individual labor market condition (Brand 2015). However, this assumption has been questioned (see e.g. Hilger 2016). A plant closure rarely happens overnight. Typically, a public announcement is made in advance. The time interval between the announcement and the shutdown may represent a ‘breeding ground’ for a selectivity bias to arise (Suppa 2021). Those workers who are more likely to find a new job elsewhere may anticipate the plant closure, look for a new job before the plant closure and avoid the transition into unemployment. Those who do not anticipate the plant closure and/or are not able to find a new job before the layoff experience transit to unemployment. Therefore, they may not be a random sample from the workforce and they may be endowed with labor market characteristics that are systematically different from those of individuals at work. We tried to limit these concerns about the exogeneity of plant closures by including in our empirical model a rich set of control variables predetermined with respect to the treatment (e.g. parental migration background, parental education, parental cumulative unemployment and lagged parental job loss expectations)^[15] As the question was asked biennially, missing indicators were used when this information was unavailable to mitigate potential non-random attrition bias due to non-response (Nikolova and Ayhan 2019). And by checking whether the estimated parental unemployment effect changed across the various model specifications (Marcus 2013; Nikolova and Ayhan 2019; Nikolova and Nikolaev 2021).

We observed the treatment and the outcomes in two different moments. The outcome variables were observed multiple times when the individuals were adults, between 18 and 31 years of age, while the treatment was observed when they were children, between 0 and 15 years old.^[16] For individual i at time t, the health outcome variable y _it is specified as follows:

where U _i is the binary indicator for parental unemployment, X _it is a vector of individual and family controls and α _i and ɛ _it are the error term components.^[17] Because parental unemployment is time invariant with respect to the outcome, we could not include fixed-effects; α _i is therefore assumed to be a random effect, uncorrelated with the covariates and with the treatment.^[18]

Equation (1) imposes a particular parametric structure to the conditional mean of the dependent variable. The estimates of the treatment effect using Equation (1), for instance by Generalized Least Squares (GLS), may be biased if the functional relationship with the control variables is misspecified or if there are areas outside of common support which require extrapolation (Ho et al. 2007; Peter 2016). In order to reduce biases due to model dependence, we followed Ho et al. (2007); Hainmueller (2012); Gambaro, Marcus, and Peter (2019) and combined matching methods and standard outcome regression. More in detail, it consists in preprocessing the data over a set of observed confounders using matching methods so that the relationship between U _i and X _it is eliminated or reduced. Thereafter, standard outcome regression techniques for modeling the conditional mean of the dependent variable are applied on the preprocessed sample. The matching part, by covariate balancing, is meant to reduce the risk of model dependence in the subsequent parametric estimation step, while the outcome regression to model the conditional mean of the dependent variable is meant to address the potentially leftover sample imbalance (Gambaro, Marcus, and Peter 2019; Ho et al. 2007).

In the matching step, we used entropy balancing (EB) (Hainmueller 2012). EB increases the balancing properties of the sample by reweighting the control units in a such way that the joint distribution satisfies some pre-specified moments conditions. The selected weights minimize an entropy distance metric on a r-dimensional set of constraints. We followed Marcus (2013) and imposed the conditions on the first and second moments of the full set of pre-treatment variables.^[19] Table 3 reports the gains in balance for the mean of each covariate. Table B.1 in the Online Appendix instead shows the gains in balance for the variance. The balancing properties improved remarkably. Before matching, significant differences were observed between treated and control individuals in terms of the means of family socio-economic characteristics at the time of birth. Treated individuals were more likely than their control counterparts to be born into households with younger, less educated parents and limited financial resources. These differences were statistically significant at the 1 % level. However, after matching, the socio-economic profiles of the treated and control groups became very comparable. The biases, i.e. the standardized differences of the mean and the variance between treated and controls, were importantly reduced for every covariate. Further, each difference in the means of family socio-economic characteristics became statistically insignificant after matching.

Finally, we computed the Average Treatment Effect on the Treated (ATT) by estimating Equation (1) using the EB weights. To account for within-individual correlation, we estimated standard errors using Liang and Zeger’s (1986) cluster-robust variance estimator.

4.1 Main Findings

The main findings are reported in Table 4. Its column (1) reports the results for the baseline specification. It includes the set of pre-treatment characteristics used in the EB matching plus the controls for the local unemployment rates and the sets of dummy variables for the year when the individual was interviewed and born. Local unemployment rates are meant to control for the precariousness in the labor market at the period of when the treatment occurred. The survey year and the year of birth are meant to control for potential unobserved heterogeneity at time level, at both the childhood and the adulthood periods. Column (2) extends the baseline specification by adding other predetermined family characteristics, measured when the child was born, in order to reduce the risk of possible omitted variables bias: household size and house size in squared meters. Finally, column (3) includes the lagged parental job loss expectation. Being the richest specification in terms of predetermined covariates, column (3) is our preferred specification.

We found very similar results across the three different specifications, with parental unemployment affecting differently mental and physical health. On the one hand, mental health is positively but not significantly affected by parental unemployment. On the other hand, physical health is negatively and significantly reduced by treatment: those individuals who experienced parental unemployment in their childhood have a lower physical health of approximately 26 % of their standard deviation.

Our findings may be the result of different mechanisms that occurred during individuals’ lives. The small and insignificant effect on mental health may be the consequence of a ‘normalization’ process that the laid-off parent operated over his/her unemployment condition (Georgellis et al. 2008; Thill, Houssemand, and Pignault 2019). Individuals often use coping strategies to overcome stressful situations (Ashforth and Kreiner 2002). A laid-off parent might have justified his/her condition by pointing to the impossibility of preventing an event that was out of his/her control, like a plant closure (Pignault and Houssemand 2017, 2018]). Moreover, when generalized precarious conditions occur, people suffer less from their own misery because it is becoming the prevailing social norm (Clark 2003, 2006]; Clark, Knabe, and Rätzel 2009; Stavrova, Schlösser, and Fetchenhauer 2011). Normalizing his/her own unemployment might prevent displaced parents from developing anxiety or depression disorders, which are likely to affect offspring’s well-being. The negative effect on physical health may be instead explained by a longer exposure of the child to parental risky behaviors. Laid-off workers are more likely to engage in unhealthy behaviors, like smoking or drinking (Mörk, Sjögren, and Svaleryd 2020; Reine, Novo, and Hammarström 2013; Schunck and Rogge 2010). Children with parents who smoke or consume alcohol are at increased risk of developing health problems while growing (Kuppens et al. 2020; Richter and Richter 2001). If displaced workers are more likely to develop unhealthy habits, then a far-reaching effect of parental unemployment may be the capability of triggering unhealthy behaviors on children, impairing offspring’s physical health.

We empirically tested whether parental unemployment may have consequences on children’s physical health later in life by increasing their probability of engaging in risky behaviors or incurring in unhealthy physical behaviors or diet. More in detail, we examined the impact of parental unemployment on child alcohol and tobacco consumption, as proxies of risky behaviors, and obesity when adult, as potentially correlated to physical activity and diet. We coded alcohol consumption as a dummy indicator equal to 1 if the individual had a moderate/severe level of alcohol intakes.^[20] Tobacco consumption was defined by using a binary indicator which took value 1 if the individual reported to be a smoker at the time of the interview and 0 otherwise. Finally, obesity was defined using a binary indicator which took the value 1 if the individual reported a BMI score equal to or greater than 30, and 0 otherwise. We estimated three linear probability random-effect models for these three binary variables. Table 5 reports the results.

We found that parental unemployment increases the probability of engaging in risky behaviors later in life. Treated individuals are more likely to consume alcohol and tobacco. Individuals who experienced parental unemployment due to plant closure are, compared to their control peers, 6.6 and 12.3 percentage points more likely to frequently consume alcohol and to smoke, respectively. These findings are in line with those in Ermisch, Francesconi, and Pevalin (2004), who found that parental joblessness increases the likelihood of both smoking and experiencing greater distress later in life. Furthermore, the treated are also more prone to obesity (+5 percentage points). These estimates are substantial when compared to the baseline prevalence of alcohol consumption, smoking, and obesity among the treated group, which are 25.4 %, 41.3 %, and 6.7 %, respectively.^[21] However, these estimates should be interpreted with caution, as they may be subject to omitted variable bias due to our limited ability to control for parental risk factors.

4.2 Heterogeneity Analysis

Cunha and Heckman (2007) argued that parental investments are not equally effective at any point in childhood for the human capital formation process. Childhood is a complex and heterogeneous moment in the lives of individuals. It is made up of many different stages, with different impacts in human development. Because parental inputs are transformed in human capital and since human capital tends to both accumulate over time and increase in efficiency, investments made in the early periods of childhood may generate higher returns compared to those made later. In the same spirit, setbacks suffered in the early periods of the childhood may generate greater losses in the future. Following these ideas, we checked whether parental unemployment exerts different effects on the human capital of individuals depending on the timing of its occurrence. We interacted the treatment indicator with the indicator which divides the sample into those who experienced parental unemployment during the pre-pubescent period (0–8 years of age) and those who suffered it subsequently and we re-estimated the model. Table 6 reports the results. On the one hand, we found that there is a positive effect on mental health if parental unemployment occurred in earlier years. The positive effect may originate from the positive elements of the unemployed condition, such as the increase in leisure time that parents can use in a more intensive parenting activity with their own children (Doepke, Sorrenti, and Zilibotti 2019; Knabe et al. 2010). On the other hand, we found that parental unemployment due to plant closure remains negatively and significantly related to physical health regardless of the childhood period when it occurred. However, the effect is not evenly distributed across periods. It is significantly stronger if parental unemployment occurred in the later years (9–15 years old) compared to the earlier years (0–8 years old). This piece of evidence is in contrast to the idea proposed by Cunha and Heckman (2007), for which setbacks in early childhood periods may generate greater losses later in life. Nonetheless, this may be attributable to the fact that parental unemployment experienced between the ages of 9 and 15 occurred closer in time to the observation of health outcomes, allowing less time for its effects to fade away.

Parental unemployment may exert different effects according to the gender of the parent or the child. We followed the previous empirical literature and tested this hypothesis in our sample (Briody 2021; Haisken-DeNew and Kind 2012; Nikolova and Nikolaev 2021; Pieters and Rawlings 2020; Powdthavee and Vernoit 2013). First, we decomposed the treatment distinguishing between maternal and paternal unemployment. Second, we did it according to the gender of the child. Tables 7 and 8 report the findings. The estimates for both mental health in model (1) and physical health in model (2) of Table 7 suggest that the effect does not dependent on whether it is the mother or the father to experience the dismissal due to plant closure. By splitting the sample according to the gender of child, we found that daughters suffered from parental unemployment in terms of physical health, whereas sons remain generally unaffected. This piece of evidence is in contrast to previous empirical literature, which instead found that sons suffered the most from the negative dynamics of the parents’ labor market (Haisken-DeNew and Kind 2012; Nikolova and Nikolaev 2021). However, this literature focused on life satisfaction, while our article focuses on mental and physical health. In addition, we speculate that the differential effect of parental job loss on daughters may be due to sons and daughters having different mechanisms to cope with family stressors. In fact, women can internalize and ruminate more (Johnson and Whisman 2013) and can take on more care-giving roles, leading to larger physical strain, fatigue and somatic complaints.

Next, we assessed if the effects of parental unemployment depended on the status of the labor market in the year of job loss. More in detail, we estimated the treatment effect by distinguishing those who were affected by parental unemployment before and after 2003. The choice of the 2003 cutoff is related to important changes in the German labor market occurring between 2003 and 2005, when employment services were restructured, temporary employment was made more attractive, the duration of unemployment benefit for the elderly was lowered and activation schemes were planned for all welfare recipients considered able to work (Rinne and Schneider 2017). After these reforms, the unemployment rate steadily decreased. Given the institutional changes in the labor market and its improved shape after these reforms, also the consequences of an unemployment event may have mutated, becoming less important after the reforms because the labor market has become more flexible, with shorter unemployment duration and higher tightness. Table 9 shows the results: the negative effect of parental unemployment on the physical health of their offspring is due to those parental unemployment events experienced in a difficult time for the German labor market. However, given that we observe only 39 individuals who experienced a parental unemployment event after 2003, the heterogeneity results in Table 9 must be taken with caution as affected by low statistical power.

Finally, we checked if parental unemployment differently affected individuals depending on whether they were from East or West Germany. Regional differences between East and West Germany due to different historical, economic, institutional and labor market backgrounds may translate into different reactions to and consequences of a job displacement, with differential impacts on children. For example, women in East Germany might be more hurt by unemployment than women in West Germany because historically they were more attached to employment (Kassenboehmer and Haisken-DeNew 2009). We distinguished the treatment indicator between parents who were living in West Germany and those who were residing in East Germany at the time in which the individual was born^[22] and we redid the main analysis. Table 10 shows the estimated parental unemployment for East and West Germany. Although the negative effect of parental unemployment on physical health does not depend on the region of origin, the impact on mental health was nil in East Germany and significantly positive in West Germany. The positive effect on mental health in West Germany may be due to positive factors of parental unemployment, such as the increase in the amount of leisure time to be spent with children (Knabe et al. 2010) and the development of soft skills, like resilience (Masten and Narayan 2012; Masten and Palmer 2019), dominating negative ones, such as the higher chances of incurring economic and emotional deprivation (Jahoda 1982; Janlert and Hammarström 2009) and emotional distress leading to family breaks (Di Nallo et al. 2022; Eliason 2012; Jensen and Smith 1990; Marcus 2013; Mörk, Sjögren, and Svaleryd 2020). This dominance is more likely in West Germany than in East Germany, because in West Germany the attachment to employment was historically lower and the wealth of the household was higher than in East Germany.

4.3 Sensitivity Analysis

We performed a series of sensitivity checks to verify the robustness of our baseline findings. The corresponding results are reported in Online Appendix C.

First, we tested the robustness of the EB weighting scheme by re-estimating the model by inverse probability weighting (IPW). IPW assigns to each control unit a normalized weight equal to P ( X i ) 1 − P ( X i ) / π ̄ 0 , where P(X _i ) is the predicted probability of being treated and π ̄ 0 is the average of P ( X i ) 1 − P ( X i ) for the controls. Treated units are instead assigned weights equal to 1. Then, the outcome equation is estimated by weighted least squares using the square root of the normalized weights. Table C.1 reports regression adjusted estimated effects for IPW. Using IPW instead of EB weighting did not lead to relevant changes. The estimated effects are indeed very similar to the benchmark ones. The only difference is in the precision, which is reduced when using IPW.

Second, we tested whether our results were sensitive in magnitude and significance to the sample size of the treated group. Table 2 showed that the number of treated units is not very large. We replicated the analysis nine times by removing each time one of the survey waves from the sample. In a particular wave, answers given by respondents about their health may be affected by something exceptional that took place in that year. By excluding one wave from the dataset at a time and re-estimating the model with the remaining data and by doing so for each wave, we can evaluate how much the estimated effects change when each wave is excluded. We can therefore verify that our findings are not driven by specific peculiarities in some waves and that they are not too sensitive to observations of any single wave. Table C.2 reports the results, which are in line with the baseline findings.

Third, in our sample, the health outcomes are observed multiple times for each individual and each individual enters the dataset as many times as the times we observe her/his health between 18 and 31 years of age. We reestimated the parental unemployment effect using as outcome variable the within-individual average of the health outcome. We therefore collapsed the sample to one observation per individual. Table C.3 reports the results and shows that they are very similar to the benchmark estimates.

Fourth, we assessed whether the effects of parental unemployment depend on the cohorts of birth. If these effects are heterogeneous across cohorts or over years and in each subsample we have a different composition of other observables, we may get biased estimates of the ATT. In fact, this imbalance may affect both the outcomes and the impact that the treatment generates on them. Hence, we estimated the treatment effect by distinguishing it between those born before and after 1991. The choice of the 1991 threshold is suggested by two reasons: first, the German reunification after the fall of the Berlin wall in November 1989; second, the fact that plant closure as a reason for unemployment was collected in the SOEP starting from 1991, which results in an under-representation in our sample of individuals born before 1991 and treated in the first years since childbirth. Table C.4 in the Online Appendix reports these results, which are very similar to the benchmark ones: there is no effect on mental health and there is a negative effect on physical health, independently on being born before or after 1991. The negative effect on physical health is larger in magnitude for those born after 1991. However, the test for equality of coefficients rejected the null hypothesis, suggesting that there may be possible imbalances across cohorts or over years.

Fifth, we tested whether our results may be driven by the presence of individuals born in single-parent households. Typically, single-parent and two-parent households possess different earning capacities. In particular, in case of unemployment, two-parent households may rely on the insurance of the ‘added worker effect’, that is the increase in the labor supply response of one partner to the joblessness condition of the other (Lundberg 1985). This compensation mechanism may make individuals born and grew in two-parent households more resilient to episodes of parental unemployment compared to their single-parent household peers. If the effect of parental unemployment is heterogeneous across the household type, this unbalancedness may have generated bias in our previous results. We re-estimated the parental unemployment effect considering only individuals born in two-parent households.^[23] Table C.5 reports the estimated coefficients showing that they are in line with the benchmark estimates.

Finally, we assessed the parental unemployment effect by changing the definition of the treatment and enlarging it to include unemployment events due to other reasons for job displacement. By doing so, we increase the external validity in terms of the population to which our results can speak, but we are likely to generate biases because the chances that the treatment indicator is endogenous due to omitted variables are larger. The results are presented in the Online Appendix in Table C.6. On the one hand, the negative effect on physical health is confirmed and it is in line with the one generated by dismissals due to plant shutdown. On the other hand, the relation between parental unemployment and mental health becomes significantly negative.