Economic Distress and Children’s Mental Health: Evidence from the Brazilian High-Risk Cohort Study for Mental Conditions

Abstract

This paper assesses the effects of adverse economic shocks on children’s mental health. We rely on the Brazilian High-Risk Cohort Study for mental conditions, which provides an unprecedented array of data on psychopathology, life events, family medical history as well as parental behaviour and polygenic scores for mental disorders over a ten-year period. Our empirical strategy exploits parental job loss events over time in a difference-in-differences framework. We document that parental job loss significantly worsens children’s mental health, resulting in increased clinical diagnoses of mental disorders. These results are robust to several specifications and pre-trends. Heterogeneous results and mechanism analysis indicate that psychological distress in the household brought about by job loss events may be a key mechanism affecting children’s mental health.

Mental health conditions account for 14.6% of the years lived with disability globally and can have severe negative economic consequences by impairing cognitive function or distorting beliefs and preferences (Ridley et al., 2020; GBD Collaborators, 2022). Most mental disorders emerge in childhood and adolescence, often amidst an environment of adverse economic conditions and social disadvantage (Patel et al., 2018). Yet, there is scant causal evidence on the extent to which and how socioeconomic disadvantages might lead to childhood adversity and the impairment of children’s mental health. Measuring mental health conditions remains particularly challenging, while household conditions and parents’ behaviour in the face of adverse economic shocks are typically not observed. Notwithstanding these challenges, the characterisation of the potential pathways connecting socioeconomic disadvantages and children’s mental health is crucial for action, and can be instrumental for the prevention of mental health problems and the recovery from mental disorders.

In this paper, we rely on the Brazilian High-Risk Cohort Study (BHRCS) for mental conditions to assess the effects of adverse economic shocks on children’s mental health. The BHRCS assessed 2,510 children at different points in time starting in 2010. These were selected from a pool of 9,937 children aged 6–12 years who were screened at public schools in two major Brazilian cities (São Paulo and Porto Alegre). From the initial pool, researchers selected a sub-sample of children at random (n = 957) and another sub-sample of children under higher risk for mental disorders (n = 1,553).¹ The BHRCS provides an unprecedented array of data on children’s psychopathology, life events, family medical history as well as parental behaviour and psychiatric polygenic scores. Our main indicators of child mental health come from the detailed Development and Well-Being Assessment (DAWBA) instrument, which includes diagnoses of specific disorders confirmed by trained psychiatrists. We focus on parental job losses, recorded in questionnaires, as a marker of a relevant and common adverse economic shock that can lead to negative spillover effects on children.²

Our empirical strategy exploits parental job loss events over time in a difference-in-differences (DiD) framework. Following the recent advances in the DiD literature, we use estimators that are robust to treatment effect heterogeneity across cohorts and time. Exploiting the richness of our dataset, we control for a wide range of covariate-specific trends, including socioeconomic characteristics, parental and child mental health, exposure to life stressors and cognitive development. Our identification strategy thus relies on the parallel-trend assumption conditional on these covariates. We follow Callaway and Sant’Anna (2021) and incorporate these using doubly robust estimators. Given the large number of baseline variables available in our dataset, we additionally test machine learning techniques to select those to be included in the model (Chernozhukov et al., 2018). We provide evidence in support of our research design. Pre-trend differentials are statistically insignificant and close to zero across outcomes, while the timing of the job loss is uncorrelated to a large set of baseline characteristics and their pre-treatment dynamics. Recent evidence on job loss impacts from the Brazilian labour market provides further support to identification.

We find that parental job loss significantly worsens children’s mental health. Parental job loss increases children’s probability of being diagnosed with a mental disorder by 6 percentage points, or 24% relative to the baseline. Similarly, a computer-generated measure capturing the likelihood of mental diagnoses (DAWBA bands) increases by 0.186 SDs. When exploiting granular information on the parental job loss timing, we find that the negative consequences on children’s mental health can last up to five years after the shock, indicating that a large part of their childhood and adolescence years is affected by psychological distress. Our results are stable to several specification checks. They are robust to covariate-specific trends and to how we select and model them within our econometric framework. Furthermore, the magnitude of our estimates remains similar once we restrict the sample to children of parents who eventually lose their job at any point in time and, therefore, solely exploit variation in the timing of the job loss. This is reassuring as treatment and control groups are more likely to be similar in unobservables.

Our findings further indicate that psychological distress in the household brought about by parental job loss may be a key mechanism affecting children’s mental health. In particular, parental job loss leads to a 0.16-SD increase in children’s maltreatment in the first wave after the shock. Additionally, it increases their probability of witnessing constant family fights by 10 percentage points (pps). Auxiliary evidence reveals that temporal variations in these variables strongly predict variations in children’s mental health. We also find suggestive evidence that our main results are more pronounced for children in families with a higher prevalence of mental health issues, potentially more likely to lack the necessary psychological coping resources to deal with adverse contexts. We do not observe heterogeneous results according to children’s genetic endowments, suggesting that the previous result is not driven by inherited risk. While we also find significant adverse impacts of parental job loss on material resources, other pieces of evidence do not support reduced material investments as a relevant mechanism in our setting. Consistent with that, we observe a weak correlation between material resources and child mental health in our data. If anything, our results are weaker among families with fewer material resources.

While there is ample evidence on the effects of parental job loss on children’s educational outcomes (Hilger, 2016; Mörk et al., 2020; Ruiz-Valenzuela, 2021), including a recent study for Brazil (Britto et al., 2022b), evidence on children’s mental health is scarcer. The existing causal studies rely on indirect measures of mental health conditions, such as perceptions reported by household survey respondents or records of health care and prescription drug utilisation (e.g., Schaller and Zerpa, 2019; Mörk et al., 2020; Moghani et al., 2021). However, changes in reported measures may result from variation in the respondent’s perception of a child’s health, or even the respondent’s own mental health, rather than variation in the child’s actual health (Schaller and Zerpa, 2019). Additionally, proxies from health care and prescription drug utilisation may be directly related to changes in the income profile and in the consumption of health care, rather than in actual health status. We contribute novel evidence to the literature in different ways. First, we assess the impact of parental job loss on a wide range of mental health outcomes by using a rich array of psychopathology data, including objective clinical diagnosis of specific disorders.³ Second, we look at parenting practices to illuminate possible mechanisms and document heterogeneous effects by family history of mental disorders and polygenic scores for psychiatric disorders. In this way, we add efforts to the recent literature that investigates whether the interaction between genetic endowments and childhood environment affects human capital development (Houmark et al., 2020; Barcellos et al., 2021). Finally, most of the literature on the mental health impacts of economic distress, as marked by parental job loss, comes from high-income countries. To the best of our knowledge, our work is the first to assess this relationship in a developing country, where social safety nets are weaker and families may be hit the most by economic fluctuations. Our results are therefore particularly informative for policymaking in relatively more deprived contexts.

The remainder of the paper is organised as follows. Section 1 describes the data and the background of our study. Section 2 outlines the empirical strategy. Section 3 describes our main results and their robustness. Section 4 discusses the conceptual background and investigates mechanisms. Section 5 concludes.

1. Data and Background

We use data from the BHRCS for mental conditions (Pan et al., 2017). The cohort is composed of 2,510 children from the cities of São Paulo (SP) and Porto Alegre (POA) who were enrolled in 57 different public schools, aged 6–12 at the baseline (year 2010–1) and who have been followed periodically since then (Salum et al., 2015). SP is Brazil’s most populous city, with a population of 11,253,503 inhabitants according to the 2010 census. POA is the largest city in the Southern region of Brazil, with a population of 1,409,351 inhabitants.

The definition of the BHRCS involved two stages: a screening stage and an assessment stage. A total of 12,500 parents of children enrolled in 22 public schools in POA and 35 public schools in SP were invited to participate. Among them, 8,012 parents (representing 9,937 children) agreed to take part in the study. At that stage, the BHRCS team conducted the family history screen (FHS), a structured interview to screen all family members for mental disorders following the Diagnostic and Statistical Manual of Mental Disorders, fourth edition (DSM-IV). The results of the FHS screening were used to calculate different indexes measuring the percentage of positive family members (adjusting for relatedness to the children) for five disorders: anxiety, attention-deficit hyperactivity disorder (ADHD), obsessive-compulsive disorder (OCD), learning disorder and psychosis. The final sample consisted of a random sub-sample of 957 children and a high-risk sub-sample of 1,553 children, selected based on the highest FHS index scores for specific disorders after random sampling. Only one child per family was included in the final sample. For further details about the screening, sampling and data collection procedures, see Salum et al. (2015).

Our working dataset contains information from the cohort’s baseline (2010–1) and two follow-up waves (2013–4 and 2018–9). All three waves include the DAWBA, a package of interviews, questionnaires and rating techniques generating DSM-IV-based psychiatric diagnosis (Goodman et al., 2000). Based on answers given by parents and children, an algorithm generates diagnostic probabilities, which are then used to compute DAWBA bands (Goodman et al., 2011). These are ordinal and integer scores from 0 to 5 corresponding to probabilities of satisfying the diagnostic: <0.01%, |$0.5\%$|⁠, |$3\%$|⁠, |$15\%$|⁠, |$50\%$| and >70%, respectively. Verbatim responses and structured answers are also carefully evaluated by child psychiatrists, who then provide clinical diagnoses. Our main outcomes for children’s psychopathology are a dummy variable indicating a diagnosis of any mental disorder and an overall score of DAWBA bands (referred to as ‘DAWBA bands’ throughout the text) summing up the bands of all disorders. In supplementary analyses, we assess the individual externalising and internalising DAWBA band scores,⁴ and the diagnosis of four specific disorders—anxiety, depression, conduct and oppositional defiant disorders (CODD) and ADHD.⁵ Summary statistics at baseline for our main outcome variables are provided in Online Appendix Table A1, showing a high prevalence of mental disorders in our sample. Around 26% of the children have a positive clinical diagnosis for mental conditions, with anxiety disorders and ADHD being the most prevalent conditions. This relatively high prevalence is due to the fact that 61.9% of the sample consists of children identified as having a high risk of developing mental health disorders.

For each follow-up wave, our database includes information on exposure to different life events since the previous wave, including parental job losses, which we use to identify children exposed to this shock between any two waves.⁶ Statistics in Online Appendix Table A1 show that 42.9% of children in our sample were affected by parental job loss at least once in the period of our analysis, with levels of exposure being considerably larger between wave 1 and wave 2 (i.e., 2014–8). We also have more granular data on parental job loss timing, gathered from a complementary questionnaire recording when the shock took place between consecutive waves (coded by year |$\times$| semester). Available information does not detail whether the mother or the father lost her/his job.

All three waves additionally contain information about educational attainment, maternal and paternal income, household assets, exposure to parental abuse and neglect⁷ and the cohabitation/marital status of the parents. For the two follow-up waves, we also have information on whether the child witnessed constant family fights, the average time the child and her mother spent together and the number of individuals residing in the household. We use this information to shed light on the impact of job loss on alternative outcomes and possible mechanisms linking job loss to children’s mental health.

In the baseline, our dataset includes a wide range of covariates: basic parental and child demographics, socioeconomic indicators, information on prenatal and perinatal health, exposure to violence or other life stressors, family environment, executive function development and other cognitive skills. Additionally, it includes data on the FHS index measuring the family history of mental disorders, as well as the children's and their parents’ polygenic risk scores (PRSs) for psychiatric diseases, computed from saliva samples.⁸ We use these data to explore the heterogeneous impact of parental job loss according to the family history of mental disorders and the genetic propensity to developmental disorders.

Online Appendix Table A2 presents summary statistics for children and parental employment characteristics in our sample, and the 2010 census (Instituto Brasileiro de Geografia e Estatística, 2016; Instituto Nacional de Estudos e Pesquisas Educacionais Anísio Teixeira, 2020). Children in our sample are mostly white (60.9%) and male (55.2%). They are, on average, 9.8 years old in the baseline. Regarding the employment status of their parents, 71.1% of mothers and 90.7% of fathers in the sample are employed—with 31% of employed mothers and 31.7% of employed fathers in the informal sector. We also present statistics comparing children in our sample to the universe of children aged between 6 and 12 years old from the 2010 census in São Paulo and Porto Alegre (where our sample was drawn from) and to the entire country. We observe that children’s characteristics are relatively similar to those from São Paulo and Porto Alegre, except for a smaller share of females and a slightly higher age-to-grade distortion. Although employment among mothers is higher, particularly because of participation in the informal sector, the overall picture indicates that our study population is not remarkably different from the population at large in São Paulo and Porto Alegre. Yet, we observe that it is not representative of the entire Brazilian population. This is expected as the Northern regions of the country are socioeconomically more vulnerable than Southern Brazil, where São Paulo and Porto Alegre are located.

2. Empirical Strategy

2.1. Econometric Framework

Our goal is to assess the average treatment effect of parental job loss on children’s mental health. While a job loss itself may be transient, its impact may last longer. Hence, we define treatment based on when parents first lose a job. In this case, our treatment is staggered by construction. We then primarily follow the identification, estimation and inference tools proposed by Callaway and Sant’Anna (2021) (C&S) in staggered DiD set-ups.

We start by defining the notation used throughout the section. Denote a particular wave time by |$w$|⁠, where |$w=0,1,2$|⁠, and let |$D_w$| be a binary variable that indicates whether a child’s parent has lost a job up to |$w$|⁠. Also, define |$G_g$| to be a dummy variable that is equal to one if a child is first treated between waves 0 and 1 (let |$g=1$|⁠) or between waves 1 and 2 (⁠|$g=2$|⁠), and define |$C$| as a dummy variable that is equal to one for children who are not treated in any period. Finally, let |$Y_w(1)$| and |$Y_w(0)$| measure potential children’s mental health at time |$w$| with and without parental job loss, respectively. The main building block of our framework is the average treatment effect for children who are members of group |$g$| at a particular wave time |$w$|⁠, denoted by

$$\begin{eqnarray} \textit{ATT}(g,w):=E[Y_w(1)-Y_w(0)|G_g=1]. \end{eqnarray}$$

(1)

Following C&S, we express our parameter of interest in terms of functionals of (1). In particular, we are mostly interested in

$$\begin{eqnarray} \tau _0:=\sum _{j\in \lbrace 1,2\rbrace } P(G_j=1) \textit{ATT}(j,j), \end{eqnarray}$$

which is the average treatment effect of parental job loss on children’s mental health measured in the first wave after this adverse shock among all groups of children whose parents ever lose a job across our panel data. Since waves are multiple years apart and we exploit job displacements taking place at any possible time between waves, |$\tau _0$| averages short-term effects from recent job loss and longer-term effects of job loss that happened years before (up to four in our data). As discussed later, in additional analysis we exploit survey information about how long ago the job loss happened to look at the time profile of effects.

We are also interested in the dynamic treatment effects |$\tau _1:=\textit{ATT}(1,2)$|⁠, which measure whether the impact of an adverse economic shock on children’s mental health is persistent in a follow-up wave. Given that |$\tau _1$| is only identified for group 1, |$\tau _1-\tau _0$| could be the result of not only treatment effect dynamics, but also compositional differences between groups 1 and 2. Since the groups are balanced across a wide range of covariates, this is less likely to be a concern.

Our study design is based on a conditional parallel-trend assumption. That is, we assume that children with the same baseline characteristics |$X$| would follow the same trend in mental health status in the absence of parental job loss. We exploit the richness of our dataset to implement a propensity score–weighting strategy that balances treated and control groups on a wide range of covariates. These cover basic parental and child demographics (city of residence, race, birth year, gender), socioeconomic indicators (socioeconomic strata, mother’s educational attainment and employment status), parental and child mental health, child cognitive development (an index measuring spatial, working memory, reading, writing and math skills) and risk factors (height and weight at birth, and exposure to bullying and violence). By following this route, we deal with potential biases that would arise if the path of mental health outcomes (in the absence of parental job loss) depended on these baseline characteristics. Under the conditional parallel-trend assumption, (1) is identified by

$$\begin{eqnarray} E[\omega _g(X)\times (\Delta Y -E[\Delta Y|X, C=1] ) ], \end{eqnarray}$$

(2)

where

$$\begin{eqnarray} \omega _g(X):=\frac{G_g}{E[G_g]} - {\frac{p_g(X)C}{1-p_g(X)}}\bigg / {E\bigg [\frac{p_g(X)C}{1-p_g(X)}\bigg ]}, \end{eqnarray}$$

with |$p_g(X)$| the probability of being first treated between waves |$g-1$| and |$g$| conditional on covariates, and |$\Delta Y:= Y_w-Y_{g-1}$|⁠.⁹ We estimate (2) and, consequently, |$\lbrace \tau _i\rbrace _{i\in \lbrace 0,1\rbrace }$| using sample analogues and parametric estimators. In particular, we estimate |$p_g(X)$| using a logit model, and |$E[\Delta Y|X, C=1]$| using OLS. Our estimator has the double-robustness property: it only requires one to correctly specify either (but not necessarily both) the outcome evolution for the comparison group or the propensity score model (Sant’Anna and Zhao, 2020). In order to conduct asymptotically valid inference, we use a bootstrap procedure that computes simultaneous confidence bands for the entire path of group-time average treatment effects. Our inference procedure also accounts for the autocorrelation of the data by using clustered bootstrapped SEs at the child level. Our results remain similar when we consider school-level clusters.¹⁰

2.2. Internal Validity and Robustness Checks

Estimating the causal effects of parental job loss can be challenging as workers who lose their jobs may differ from those who remain employed in unobserved ways that can affect child outcomes. An advantage of our setting is that we observe child mental health before and after job loss events, so we can deal with any unobserved reasons for parental job loss that are time invariant in a DiD framework.¹¹ The main challenge for identification is therefore selection into job loss based on time-varying confounders related to child outcomes. To circumvent that, some previous papers restrict the analysis to involuntary job loss stemming from mass layoffs or plant closures. Our data do not contain information on such events. Hence, in principle, the job loss measure we use may reflect variation steaming from both endogenous and exogenous dismissals.

Yet, considering current evidence from Brazil and the institutional background, endogeneity can be considered less of a concern in our setting. Recent papers employing DiD models to assess the impact of job loss on crime, domestic violence and child educational attainment show virtually identical results across several measures of job loss—from general job loss events, supposedly more endogenous, to job loss due to plant closures and through different definitions of mass layoffs (Bhalotra et al., 2021; Britto et al., 2022a,b). These papers use rich Brazilian administrative data that overlap with our period of analysis. The results above come with no surprise for those familiar with the unstable nature of employment in Brazil. For instance, the annual turnover rate among formal workers has historically been high, estimated at 50% in 2013 (Rocha et al., 2019). Moreover, 90% of dismissals in the formal sector occur within less than three years in the job (Britto et al., 2022a). The turnover is even higher in the informal sector (2–4 times), which accounts for over 40% of all jobs in the country (Ulyssea, 2018; Engbom et al., 2022).¹² In such an unstable setting, the scope for selection is expected to be smaller, which might explain why restricting job loss measures to potentially more exogenous events makes less of a difference—this instability, together with the markedly high turnover rate in Brazil, makes it more likely that the treatment is independent of personal characteristics or behavioural traits.

We further back up our identification strategy with additional empirical evidence and robustness checks. We estimate the effects of parental job loss using pre-treatment periods, |$\tau _{-1}:=\textit{ATT}(2,1)$|⁠. We show throughout the paper that |$\widehat{\tau }_{-1} \approx 0$| for several different outcomes. However, since waves are multiple years apart, |$\widehat{\tau }_{-1}$| aggregates pre-treatment effects from children experiencing parental job loss in periods closer to, and further away from the wave the outcomes are measured. This raises concerns about potential time-varying shocks that may only manifest near parental job loss and whose relevance might not be fully captured by our pre-treatment estimate. Bearing this in mind, we leverage more granular data on the timing of the job loss and estimate heterogeneous placebo effects over this dimension. We find that, even for parental job loss events occurring close to the pre-treatment wave, |$\widehat{\tau }_{-1}$| remains close to zero.

We complement the pre-trend analysis by further inspecting the determinants of parental job loss, comparing treated and (weighted) control children (Online Appendix Table C1). In particular, we find that the groups are balanced on a wide range of observable covariates. Notably, we do not see any correlation between parental job loss and the pre-treatment dynamics of relevant outcomes, including the child’s mental health. We also explore granular data on the timing of job loss to flexibly estimate its determinants using survival models (Online Appendix Table C2). The results confirm that baseline characteristics and their pre-treatment dynamics are not associated with the timing of job loss—except for very few and fixed characteristics (e.g., age of the mother).

We also explore several alternative specifications for our DiD model. First, we consider the late-treated group as an alternative control group, which may be more comparable to the treated children.¹³ Second, we test alternative ways of incorporating covariates into the model, such as inverse probability weighting and regression adjustment. Third, we employ double-LASSO regression for variable selection to select covariates for both the outcome regression model and the propensity score model in our primary specification (Chernozhukov et al., 2018). Fourth, we test a specification that neither controls nor balances any covariates. We show in the next section that our results remain remarkably consistent across all these alternative specifications. Taken together, all these tests combined are consistent with unobservable time-varying shocks not playing a major role in our setting. For further details, see Online Appendix Sections C.2–C.4. Finally, we provide evidence that attrition is not a concern. In particular, predicted parental job loss is balanced across attriters and non-attriters, and attrition is independent of a wide range of covariates. Online Appendix C.5 further discusses these results.

3. Main Results

Before turning to the consequences of parental job loss for children’s mental health, we first show that job loss leads to variations in parental employment, as expected. Figure 1 depicts the treatment effects on the probability of either the child’s mother or father being unemployed.¹⁴ In the closest wave after job loss, unemployment increases by 28% for displaced parents relative to the weighted control group. Although smaller, the effects are persistent in the follow-up wave: unemployment is still 18% higher. In the pre-treatment period, the unemployment variation across treated and control groups was virtually the same. To get a more detailed picture of the time profile of the effects, Online Appendix Figure D1 looks at the heterogeneity according to the timing of reported job loss. When measured up to one year after job loss, unemployment is around 40% higher in the treated group relative to the control. This gap closes very slowly over time. These effects are quantitatively and qualitatively similar to the results from previous papers using Brazilian administrative data to track job loss events and employment outcomes (e.g., Britto et al., 2022a).

Fig. 1.

Treatment Effects of Parental Job Loss on Unemployment.

Note: This figure plots simultaneous 95% confidence bands computed with a child-level clustered bootstrap and DiD estimators for the effects of parental job loss on whether the child’s mother or father is unemployed. Post-treatment effect 0 measures the impact of parental job loss at the closest post-treatment wave (⁠|$\widehat{\tau }_0$|⁠). Post-treatment effect 1 measures the impact of parental job loss at the furthest post-treatment wave (⁠|$\widehat{\tau }_1$|⁠). The pre-treatment effect exploits variation across waves that precede parental job losses (⁠|$\widehat{\tau }_{-1}$|⁠). We control for a wide range of baseline covariate-specific trends using doubly robust methods (Callaway and Sant’Anna, 2021). Controls include basic parental and child demographics (city of residence, race, birth year, gender), socioeconomic indicators (socioeconomic strata, mother’s educational attainment and employment status), parental and child mental health, child cognitive development (an index measuring spatial, working memory, reading, writing and math skills) and risk factors (height and weight at birth, and exposure to bullying and violence).

We present our main results in Figure 2 (and Online Appendix Table D1). Figure 2(a) assesses the impacts on the diagnosis of a mental disorder by child psychiatrists. In the first wave following parental job loss, the probability of being diagnosed with a mental health condition increases by 6 pps for treated children compared to the control group. This represents a 24% increase relative to the mean at the baseline.¹⁵ Figure 2(b) presents the effects on the DAWBA bands, a computer-generated measure of psychiatric diagnoses. The results confirm that parental job loss worsens children’s mental health—the index increases by 0.186 SDs among the treated group in the closest wave after the shock.¹⁶ To put our results in perspective, the effects reported in Online Appendix Table D1 (0.186 SDs) correspond to approximately 60% of the difference in mental health scores between high-risk children and those that are not in the high-risk group. In comparison to policy interventions, the effects are greater than the average effects on mental health of traditional anti-poverty programs, including cash transfers specifically (0.07 SDs) and programs that include multiple interventions (0.14 SDs; Ridley et al., 2020).

Fig. 2.

Treatment Effects of Parental Job Loss on Children’s Mental Health.

Note: This figure plots simultaneous 95% confidence bands computed with a child-level clustered bootstrap and DiD estimators for the effects of parental job loss on the probability of being diagnosed with any mental disorder (panel (a)) and DAWBA bands (panel (b)). The DAWBA bands are standardised to a distribution with zero mean and a unit SD. This procedure is applied for each age-bin |$\times$| wave combination. Post-treatment effect 0 measures the impact of parental job loss at the closest post-treatment wave (⁠|$\widehat{\tau }_0$|⁠). Post-treatment effect 1 measures the impact of parental job loss at the furthest post-treatment wave (⁠|$\widehat{\tau }_1$|⁠). The pre-treatment effect exploits variation across waves that precede parental job losses (⁠|$\widehat{\tau }_{-1}$|⁠). We control for a wide range of baseline covariate-specific trends using doubly robust methods (Callaway and Sant’Anna, 2021). Controls include basic parental and child demographics (city of residence, race, birth year, gender), socioeconomic indicators (socioeconomic strata, mother’s educational attainment and employment status), parental and child mental health, child cognitive development (an index measuring spatial, working memory, reading, writing and math skills) and risk factors (height and weight at birth, and exposure to bullying and violence).

Figure 2 also shows the effect of parental job loss on children’s mental health in a follow-up wave, taking place around four years after the previous one. The effects on both clinical diagnosis and DAWBA bands are positive, but small and statistically insignificant. However, as we exploit job loss events that occurred in any period between waves, which are several years apart, the results may mask important heterogeneities in timing. To further inspect the time profile of the effects, Online Appendix Figure D2 presents heterogeneous results according to the timing of parental job loss. The impact peaks immediately for children whose mental health is assessed within one year after the job loss, suggesting that policy interventions should act right upon the shock to mitigate the negative externalities of parental job loss to children. The effect then gradually diminishes, but remains observable for up to approximately five years after treatment.

Compared to the time window of childhood and adolescence, this represents a prolonged period, which can have lasting negative consequences for human capital formation and adult outcomes (Currie et al., 2010). To infer these potential costs, we measure the effects of parental job loss on school attainment (see Online Appendix Table D4). Point estimates are negative and persistent: −0.06 (SE 0.05) and −0.04 (SE 0.09) SDs at event times 0 and 1, respectively. Although imprecisely estimated, these results align with the results of Britto et al. (2022b). They used data from the Brazilian school census and found that parental job loss impacts school enrolment negatively, and age-grade distortion positively.

We show that pre-treatment effects are remarkably close to zero (Figure 2 and Online Appendix Table D1), independent of how close the job loss occurred relative to the pre-treatment wave (Online Appendix Figure D2). These results mitigate concerns about reverse causality and omitted variable bias. We further address selection concerns by showing that our main findings are robust when we restrict the sample to children of parents who eventually lose their job at any point in time and, therefore, solely exploit variation in the timing of the job loss (Online Appendix Figure D3). This is reassuring as in this set-up treatment and control groups are more likely to be similar in unobservables. Online Appendix Figure D3 shows that our results are robust to different specifications. In particular, they remain stable to different ways of controlling for covariate-specific trends and to using double-LASSO regression to select these trends. Moreover, the results from a specification that does not control for any covariates (and, hence, it is based on unconditional parallel trends) fall within the confidence interval of our main specification.

4. Heterogeneity and Mechanisms

4.1. Conceptual Background

Standard economic models include material (e.g., healthcare) and parental time investments as key inputs in the production function of child health (Currie, 2009). Parental job loss may affect these inputs primarily through three channels. A job loss is, in the first place, a financial distress shock that can reduce household income and healthcare investments that are important to child mental health (Currie, 2020). Second, a job loss is also a positive time shock that may lead to increased parental presence at home. However, its consequences on child mental health will ultimately depend on the quality of the time children and parents spend together. Finally, a job loss may also result in psychological distress among adults (e.g., Kuhn et al., 2009; Zhao, 2022), which can affect children by reducing the quantity and quality of parental investments (Baranov et al., 2020; Angelucci and Bennett, 2024). If the shock deteriorates the family’s psychosocial environment, time investments may even have negative returns to child mental health. Exposure to violence, abuse and neglect in the household are established predictors of mental health issues (Kieling et al., 2011; Patel et al., 2018), and past evidence from Brazil indicates that job loss leads to increased levels of domestic violence (Bhalotra et al., 2021).

How job loss affects child mental health depends on the existence of protective personal and family resources for coping with shocks (McKee-Ryan and Maitoza, 2018). For instance, lower-income or single-earner families are likely to have fewer financial resources to cope with adverse income shocks (Schaller and Zerpa, 2019), while shifts in time use may be greater in high-SES families, as displaced parents may choose to remain unemployed longer. Moreover, family mental health is a key protective factor against psychological distress resulting from adverse life events (Rutter, 1985). Past work has shown that individuals who are mentally well off cope with unemployment in a positive and resilient way (Binder and Coad, 2015), and that the adverse impact of job loss on mental health is particularly pronounced for those at the lower end of the mental health distribution (Schiele and Schmitz, 2016). Finally, the genetic psychiatric endowments within families determine the endowments of children, which, in turn, can regulate biological responsiveness to environmental effects (Scarr and McCartney, 1983; Houmark et al., 2020).

4.2. Heterogeneity Analysis

We start by analysing how exposure to parental job loss affects children according to baseline family characteristics. We focus on household composition (single-parent family and the number of household members), household monthly income and history of mental health problems. The family’s history of mental health problems is measured by the FHS, an instrument that computes the percentage of family members screening positively for different mental conditions, used in the screening phase to define at-risk children.¹⁷ Figure 3 shows that the effects of parental job loss on children’s mental health do not significantly vary based on family composition. The same holds for households with above- and below-median income, with results potentially being slightly more pronounced for higher-income families. We also find evidence suggesting that the effects are driven by children within families affected by a mental condition (children whose maximum FHS score is above the median). When breaking down the FHS scores by specific diseases, Online Appendix Figure D4 reveals some heterogeneity, particularly for anxiety and OCD. However, the coefficients are generally statistically similar across the various conditions. So, although point estimates indicate that effects are greater among families affected by mental conditions, we remain cautious when interpreting heterogeneity according to FHS.

Fig. 3.

Heterogeneous Effects of Parental Job Loss on Children’s Mental Health According to Family Characteristics.

Note: This figure plots simultaneous 95% confidence bands computed with a child-level clustered bootstrap and DiD estimators for the effects of parental job loss on the probability of being diagnosed with any mental disorder across different sub-samples defined based on family characteristics. In particular, we present heterogeneous results depending on whether the family is single parent or not (‘Single-parent family’), and for above- versus below-median values regarding the number of people in the children’s household (‘Number of people in the HH’), household monthly income (‘Income’) and the children’s highest FHS score among psychosis, anxiety, learning disorders, OCD or ADHD (‘FHS max score’). The figure plots the post-treatment effect |$\widehat{\tau }_0$|⁠, which measures the impact of parental job loss at the closest post-treatment wave.

Family mental health might confound genetic and environmental factors. That is, children with higher FHS scores may be more susceptible to developing psychopathology either because they live with individuals that lack the necessary psychological coping resources or because of inherited risk, which, in turn, could influence their response to changes in parenting more broadly. To further investigate this, we look at the heterogeneous effects of parental job loss according to children’s PRSs for psychiatric diseases, which measure genetic propensity to develop mental disorders at the individual level.¹⁸ Although PRSs are predetermined, they are not exogenous as they mechanically reflect parental genes, which can influence the environmental factors experienced by children. However, due to Mendelian inheritance, one’s PRS is randomly assigned conditional on the PRS of their parents. Figure 4 plots the interaction between parental job loss and children’s and mothers’ PRSs.¹⁹ Overall, we find no evidence suggesting that our primary results are stronger for children with higher psychiatric PRSs. This is consistent with available evidence showing that the correlation between stressful life events and mental health does not significantly vary across the PRS distribution (Musliner et al., 2015; Østergaard et al., 2020). Therefore, the strong impact we find among children with more family mental health problems might be driven by environmental factors to a greater extent.

Fig. 4.

Heterogeneous Effects of Parental Job Loss on Children’s Mental Health According to Polygenic Risk Scores.

Note: This figure plots the heterogeneous effects of parental job loss on the probability of being diagnosed with any mental disorder according to the mother’s and children’s PRSs for cross-disorder psychiatric illnesses. We estimate these in a DiD specification for the post-treatment effect |$\widehat{\tau }_0$|⁠, which measures the impact of parental job loss at the closest post-treatment wave. All specifications include the interaction between parental job loss and children’s PRSs. ‘Including mother’s PRS × parental job loss’ and ‘Further including FHS × parental job loss’ additionally include the interaction between parental job loss and mother’s PRS. ‘Further including FHS × parental job loss’ further includes the interaction between parental job loss and FHS scores. Following Callaway and Sant’Anna (2021), we estimate the DiD specification for each cohort and then aggregate the results based on the size of each.

4.3. Mechanisms

As discussed earlier, parental job loss can affect material inputs relevant to the production of children’s mental health, as well as the quantity and quality of time investments. We now analyse indicators that map on different dimensions of these inputs. Specifically, we analyse household income, family composition (parental cohabitation and household size), parental time investments (duration of child-mother interaction per day) and proxies for psychological distress within the family (child maltreatment and family conflicts).

First, we estimate the association between these inputs and children’s mental health using a first-difference model with data from the control group. Online Appendix Table D5 shows that temporal variations in indicators of child maltreatment and family conflicts strongly correlate with fluctuations in children’s mental health. This relationship remains robust even after controlling for other inputs and several baseline characteristics. The association between household income and children’s mental health has the expected sign, but is never statistically significant. The same is true for parental cohabitation, household size and hours spent with the mother. We then estimate the impact of job loss on these same variables. Results in Table 1 (and Online Appendix Figure D5) reveal that parental job loss is associated with greater exposure to a conflicting family environment. For instance, parental job loss increases children’s abuse and neglect by 0.16 SDs in the closest wave after the shock. This effect vanishes in the follow-up wave. Similarly, the probability of children witnessing constant family conflicts increases by 10 percentage points. We also observe that job loss induces financial distress, resulting in a decrease of approximately 0.16 and 0.2 SDs in household monthly income during the two subsequent waves. Finally, we do not observe any significant impacts on family composition or the duration of child-mother interaction.

Table 1.

Mechanisms.

	Household monthly income (SD)		Parental cohabitation (pps)		Abuse and neglect (SD)		Child witnessed constant family fights (pps)	Hours spend with the mother	Number of people in the household
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)	(9)
Pre-treatment effect	0.048		−0.015		−0.037
	(0.07)		(0.02)		(0.06)
Post-treatment effect 0	−0.163***	−0.170**	0.006	−0.018	0.159***	0.167**	0.101***	0.499	−0.001
	(0.05)	(0.07)	(0.01)	(0.02)	(0.05)	(0.08)	(0.03)	(0.47)	(0.10)
Post-treatment effect 1	−0.216***		−0.012		0.001
	(0.1)		(0.03)		(0.07)
Mean of the outcome	0	0	0.518	0.47	0	0	0.161	7.609	4.143
Sample	Full	Waves 1 & 2	Full	Waves 1 & 2	Full	Waves 1 & 2	Waves 1 & 2	Waves 1 & 2	Waves 1 & 2

Note: This table reports DiD estimators for the effects of parental job loss on household monthly income (columns (1) and (2)), a dummy indicating parental cohabitation (columns (3) and (4)), an index capturing children’s exposure to abuse and neglect (columns (5) and (6)), a dummy indicating whether the child witnessed constant family fights in the previous three years (column (7)), the average number of hours the child and mother spend together per day (column (8)) and the number of people residing in the children’s household (column (9)). For the outcomes assessed in columns (7)–(9), data are available only for the last two waves. Household monthly income and exposure to abuse and neglect are standardised to a distribution with zero mean and a unit SD. Post-treatment effect 0 measures the impact of parental job loss at the closest post-treatment wave (⁠|$\widehat{\tau }_0$|⁠). Post-treatment effect 1 measures the impact of parental job loss at the furthest post-treatment wave (⁠|$\widehat{\tau }_1$|⁠). The pre-treatment effect exploits variation across waves that precede parental job losses (⁠|$\widehat{\tau }_{-1}$|⁠). We control for a wide range of baseline covariate-specific trends using doubly robust methods (Callaway and Sant’Anna, 2021). Controls include basic parental and child demographics (city of residence, race, birth year, gender), socioeconomic indicators (socioeconomic strata, mother’s educational attainment and employment status), parental and child mental health, child cognitive development (an index measuring spatial, working memory, reading, writing and math skills) and risk factors (height and weight at birth, and exposure to bullying and violence). In parentheses we report SEs computed with a child-level clustered bootstrap.

*** |$p\lt 0.01$|⁠; ** |$p\lt 0.05$|⁠.

We interpret the evidence as indicating that a worsened family environment, resulting from psychological distress in the household, is a relevant mechanism in our setting. This interpretation is supported by the effects of parental job loss on family conflicts and child maltreatment, as well as the strong association between these inputs and children’s mental health. It is also consistent with results that suggest that effects are driven by children in families with a higher prevalence of mental health issues, who are more likely to lack the necessary psychological coping resources. Although we also observe a significant impact of parental job loss on household income, other pieces of evidence cast doubt on whether reduced material investments are a relevant mechanism in our context—in particular, the weak association between material resources and children’s mental health, as well as the lack of stronger effects among families with fewer material resources. Finally, the dynamic effects on child maltreatment and mental health follow a similar pattern, decreasing over time, while the impact on family income is persistent.

5. Conclusion

This paper studies the effects of adverse economic shocks on children’s mental health using detailed data from the Brazilian High-Risk Cohort Study for mental conditions. Our empirical strategy exploits parental job loss events over time in a difference-in-differences framework. We find that parental job loss significantly worsens children’s mental health, as measured by diagnoses confirmed by trained psychiatrists and computer-generated measures of psychopathology. Looking at potential mechanisms, we observe that parental job loss leads to children’s exposure to maltreatment and family strife. These factors have been previously associated with higher susceptibility to developing mental disorders, and in our data, their temporal variation strongly predicts variations in children’s mental health. We interpret the evidence as indicating that a worsened family environment resulting from psychological distress in the household is a relevant mechanism in our setting.

Additional Supporting Information may be found in the online version of this article:

Online Appendix

Replication Package

Notes

The data and codes for this paper are available on the Journal repository. They were checked for their ability to reproduce the results presented in the paper. The authors were granted an exemption to publish parts of their data because access to these data is restricted. However, the authors provided the Journal with temporary access to the data, which enabled the Journal to run their codes. The codes for the parts subject to exemption are also available on the Journal repository. The restricted access data and these codes were also checked for their ability to reproduce the results presented in the paper. The replication package for this paper is available at the following address: https://doi.org/10.5281/zenodo.10235067.

We thank Sule Alan, the co-editor and three anonymous referees for valuable feedback that allowed us to improve this article. We thank the Brazilian High-Risk Cohort Study group for generously sharing the data with us. We are especially grateful to Giovanni Salum, Eurípedes Miguel, Maurício Hoffmann, Luis Rohde, Pedro Pan and Rodrigo Bressan. We would also like to thank Rodrigo Soares, Bruno Ferman, Silvia Helena Barcellos and seminar participants at the LACEA LAMES 2022 and the Workshop on Global health, environment and labour for helpful comments and suggestions.

Luiz Felipe Fontes gratefully acknowledges financial support from São Paulo Research Foundation (FAPESP) Grant 2019/25473-5.

Footnotes

1

High-risk children were selected based on the percentage of family members, adjusted for relatedness, who screened positively for mental health disorders (Salum et al., 2015). The random sub-sample was drawn from a pool of children whose parents agreed to be screened. It is important to note that 60% of the children in the sample face a high risk of mental disorder, thus reducing the total number of children assessed that are not at risk and statistical power in specific heterogeneity analysis by sub-samples.

2

We interpret job loss as an adverse economic-related shock to families that can affect children’s mental health through different pathways, including financial distress directly.

3

Mari and Keizer (2021) used psychopathology data from Ireland to estimate the impacts of parental job loss on children’s health indicators. However, the authors relied on repeated cross sections, conditionally only on observable baseline covariates and subject to omitted variables bias.

4

The externalising scores aggregate the DAWBA bands for ADHD and conduct and oppositional defiant disorders. The internalising scores aggregate the DAWBA bands for several emotional disorders, including depression, anxiety disorders and phobias. These disorders are also included in the clinical ratings by child psychiatrists.

5

In addition to DAWBA, our data include the Child Behavior Checklist (CBCL) and the Strength and Difficulties Questionnaire (SDQ), two popular tools for assessing children’s mental health problems. In additional analyses, we use the CBCL and SDQ scores to show that our results are stable across different measures of children’s psychopathology. For additional details on all these assessments, see Online Appendix A.

6

Note that the data do not allow us to identify children that have experienced parental job loss in the years before the baseline wave.

7

Exposure to parental abuse and neglect was measured by a validated index based on parents’ and children’s answers to questions investigating exposure to physical abuse, neglect, emotional maltreatment, and sexual abuse (Salum et al., 2016). For further details, consult Online Appendix A.

8

PRSs estimate the genetic propensity to a specific outcome at the individual level. They are defined by the sum of risk gene variants that correspond to that outcome of interest in each individual, weighted by the association between those gene variants and the outcome of interest according to a genome-wide association study (Choi et al., 2020).

9

Here we are also implicitly assuming that there are no anticipation effects in the wave prior to parental job loss.

10

Results available upon request. This is expected given that the share of treated children is very similar across schools. In particular, the treatment intracluster correlation is just 0.005.

11

Online Appendix C.1 discusses common methods for estimating the effects of parental job loss on child outcomes, and provides context for our own empirical strategy.

12

For additional details, see Online Appendix B.

13

Individuals who lost their jobs in the first period (2010–4) do not appear to be different from those who did so between 2014–8, based on observable characteristics in our sample—see, for example, column 3 in Online Appendix Table C2. This is corroborated by the Brazilian National Household Survey (PNAD), where individuals with job losses in 2012–4 do not have statistically significant differences in age, race, income or education than those who lose jobs in 2014–8 (results available upon request).

14

Around 23% of children lack information about parental unemployment, primarily due to the underreporting of the father’s employment status, which is missing for 34% of the sample. Nevertheless, the missing share is fairly similar between the treatment and control groups. The difference between groups is not statistically significant (p-value = 0.377), indicating that sample selection is not expected to affect the validity of the estimated effects of parental job loss on parental employment.

15

Online Appendix Table D2 reports the effects on diagnoses of anxiety, depression, ADHD and CODD. The results are particularly strong for anxiety (0.05, SE 0.02), followed by depression (0.035, SE 0.02). We find no statistically significant impact on ADHD and CODD. However, low statistical power limits our capacity of making strong claims. For instance, at a 95% significance level, we cannot reject the fact that the impact on CODD is as high as 2.5 pps or 40% of the mean.

16

Online Appendix Table D3 (first two columns) disaggregates this index by internalising and externalising problems. We find substantial impacts on both scores (0.2 and 0.15 SDs, respectively). The same table presents remarkably similar results when measuring internalising and externalising psychopathology using two alternative instruments, the CBCL and the well-known SDQ. The consistency of our results to different rating techniques alleviates concerns about non-random measurement error.

17

FHS scores are individually computed for each of the following disorders: anxiety, ADHD, OCD, learning disorder and psychosis. Following the protocol used in the screening phase to define at-risk children (those within families severely affected by any of these disorders), we select, for each child, their highest FHS score.

18

Given the shared genetic risks among several psychiatric disorders, we test for heterogeneous effects using cross-disorder PRSs, which capture polygenic risk for ADHD, autism, bipolar disorder, major depressive disorder and schizophrenia (Cross-Disorder Group of the Psychiatric Genomics Consortium, 2013).

19

While we lack data on the fathers’ PRSs, we demonstrate that interacting parental job loss with the mothers’ PRSs does not significantly affect our estimates for the heterogeneous treatment effects based on children’s PRSs. Additionally, conditioning on FHS scores does not alter the results. Therefore, our estimates likely capture the impact of adverse economic shocks on children’s mental health based on their genetic make-up rather than environmental factors correlated with the parental risk of developing psychiatric disorders.