Abstract
We investigate the impacts of a perinatal psychosocial intervention on trajectories of maternal mental health and child skills, from birth to age 3. We find improved maternal mental health and functioning (0.17–0.29 SD), modest but imprecisely estimated improvements in parenting (0.07–0.11 SD), and transitory improvements in child socio-emotional development (0.06–0.39 SD). The intervention had negligible influence on physical health and cognition. Estimates of a skill production function reveal the intervention attenuated the negative association between maternal depression and child outcomes, and it narrowed outcome gaps between mothers who were and were not depressed in pregnancy.
I. Introduction
Social and emotional skills are an integral component of human capital. Children living in disadvantaged families, with mothers more likely to be suffering from poor mental health or depression, tend to show greater socio-emotional difficulties (Rahman et al. 2013; Hollins 2007; Bennett et al. 2016; Halfon et al. 2014; Attanasio, de Paula, and Toppeta 2022). Socio-emotional difficulties become apparent early in life and are prone to get ingrained and intensify over time, in a cascading cycle of disadvantage (Feil, Walker, and Severson 1995; Sprague and Walker 2000).1 For instance, there is some evidence that socio-emotional problems at ages 1–3 predict socio-emotional difficulties in elementary school (Briggs-Gowan and Carter 2008), which in turn reduce school performance (Fletcher 2010; Ding et al. 2009; Busch, Golberstein, and Meara 2014; Bhalotra et al. 2021) and predict mental health issues in early adulthood (Class et al. 2019; National Scientific Council on the Developing Child 2012). Despite these patterns, causal evidence of the consequences of maternal depression, or its treatment, for child socio-emotional skills is relatively scarce.
We analyze the impact of a perinatal psychosocial intervention targeted at depressed mothers on the joint evolution of maternal mental health and functioning, parental investment, and child skills from birth through three years of age. The skills that we analyze encompass not only socio-emotional skills, but also cognitive skills and physical health. This inclusion is important because these skills tend to evolve jointly. Our focus is on the mother since she is the primary caregiver in our setting and, in general, the child interacts with her more than with anyone else. As a result, her mental health and functioning have a potentially strong influence on the child. Depression and stress often manifest in low energy, impaired functionality, insomnia, poor concentration, pessimism, and a lack of interest in one’s environment (de Quidt and Haushofer 2018). It is thus plausible that depression modifies the mother’s parenting behaviors and investment in the child (Herba et al. 2016; Baranov et al. 2020; Angelucci and Bennett 2021). A rich literature in developmental psychology posits that improving maternal depression can lead to more responsive mother–child interactions and support secure infant attachment (Erickson, Julian, and Muzik 2019; Tsivos et al. 2015).
The intervention we study, the Thinking Healthy Programme Peer-Delivered Plus (THPP+), was targeted at perinatally depressed women in rural Pakistan. To improve scalability, it was delivered by trained peer volunteers through a combination of home visits and group sessions. In total, between the third trimester of pregnancy and the child turning 3 years of age, a mother in the program received 32 sessions. The intervention provided cognitive behavioral therapy with a focus on behavioral activation, self-care, and the child’s health and development. Thus, the content of the program was such that it could influence child outcomes directly or through modifying the mother’s mental health and her parenting behaviors.
Rich longitudinal data on mother–child pairs were collected in multiple waves throughout the intervention period. Socio-emotional skills are measured using the Ages and Stages Questionnaire: Social-Emotional (ASQ-SE), which contains validated psychometric indexes of competencies in self-regulation, adaptive functioning, emotional balance, communication, and prosociality. Cognitive skills are tested using the Bayley Scales of Infant Development (Bayley-III), which includes cognitive, language, and motor skills. Maternal mental health and functioning are assessed using established scales for measurement of depression (a clinical assessment and symptom severity), stress, and functional disability. Parental investment in children is measured using the HOME score, assessing the quality of cognitive stimulation and emotional support in the household.
In order to link the observed variables in the data set to the underlying developmental trajectories of children, we use a latent variable approach, common in psychometrics and economics (Spearman 1904; Joreskog and Goldberger 1975; Carneiro, Hansen, and Heckman 2003). We generate factor scores and use these to estimate treatment effects on six indicators: (i) cognition of the child, (ii) physical development of the child, (iii) socio-emotional skills of the child, (iv) parental investment, (v) maternal mental health, and (vi) maternal functioning. The reduced form treatment effects tell us how the intervention modified inputs to child development (note that child skills at a younger age are inputs to child skills at an older age). So as to synthesize these reduced form results and describe how the intervention modified the returns to the inputs, we estimate the production functions for child skills in the first three years of life, using, as a point of departure, the formulation in Cunha and Heckman (2008) and Attanasio, Meghir, and Nix (2020).
We make two main contributions. First, we identify the impact of the intervention on trajectories of maternal mental health, parenting, and child development. In contrast to much of the literature, we use longitudinal data on children from birth to age 3, and we allow intervention impacts to vary with age, identified at 6, 12, and 36 months after birth.2 We find that the intervention improves maternal mental health (ranging from 0.17 to 0.27 standard deviations, SD) and daily functioning (0.18–0.29 SD) immediately (at 6 months) and persistently up until the end of the study window, 36 months after birth. The intervention results in weakly identified increases in parental investment at 12 and 36 months (0.07–0.11 SD) and a short-term positive effect on the child’s socio-emotional skills at 6 and 12 months (0.19–0.39 SD), without any discernible impacts on the other domains of child development (physical health impacts ranged from −0.17 to 0.02 SD and cognitive development impacts ranged from −0.08 to 0.06 SD). All of these results are stronger for mothers of boys.
Our second contribution is to estimate a model of child skill formation in which, in a departure from related studies (Cunha and Heckman 2008; Cunha, Heckman, and Schennach 2010; Attanasio et al. 2020), we include dynamic latent factors measuring maternal mental health (including depression and stress) and functioning. We conceptualize maternal mental health as a capital input in the production function, a stock that can depreciate over time or that can be invested in, for instance with therapy.
Similar to other studies estimating the child skills production function, we do not have multiple instruments, and we cannot identify causal effects of (endogenous) inputs. However, we allow the model parameters to vary by treatment arm. The model allows the intervention to influence both the levels and the productivity of the inputs, similar in spirit to Kitagawa (1955), Oaxaca (1973), and Blinder (1973).
Our findings suggest that maternal mental health is an important input in the technology of skill formation, especially for moderately or severely depressed mothers, and that the intervention changes the shape of the production function, especially in the first year of life. In the control group, all child skills are increasing in maternal mental health, but at a decreasing rate. We see a larger improvement in children’s skills when moving from severe to moderate depression than we do when moving from moderate to mild depression. The intervention shifts the production function up (an increase in total factor productivity, TFP) and also attenuates the slope of the production function with respect to maternal mental health, leading to improvements in skills, particularly for children whose mothers did not recover from depression.3 The intervention modifies the production function parameters in a manner that brings outcomes for treated mothers closer to outcomes for mothers who were not depressed at baseline.
These findings help reconcile the pattern of results identified in the reduced form models of intervention effects, namely, the absence of durable effects of the intervention on child skills alongside long-lasting improvements in maternal mental health. The intervention helps women recover from depression and stress, and it also improves children’s socio-emotional skills in the short run. We might have surmised from this evidence that the improvement in maternal mental health at 6 and 12 months was the cause of the improvement in socio-emotional skills at 12 months. However, the production function estimates suggest this is not the case, as the intervention improves child socio-emotional skills particularly in the sample of mothers who do not recover from depression. This is consistent with the fact that the intervention did not just provide therapy for the mother’s depression, it also provided information and support for child development. Our estimates indicate that the latter shielded children from the negative consequences of poor maternal mental health. As time progresses, additional increases in the stock of maternal mental health shift the treated group into a flatter part of the production function, where variation in the underlying measure of mental health is less predictive of child skills. Thus, it makes sense that, although on average maternal mental health in the treated group remained persistently better than in the control group, other aspects of the intervention enabled the control group children to catch up with the treated children, resulting in convergence by 36 months in their socio-emotional skills. The production function estimates also contribute to still scarce evidence on self- and cross-productivity of skills across domains at very early ages.
Understanding how maternal depression at birth may influence the formation of skills in the early years is important given the high prevalence of maternal depression. It is estimated that between 10 and 30 percent of children worldwide are exposed to maternal depression at birth, and this share is higher in developing countries (O’Hara and Swain 1996; Parsons et al. 2012). Maternal depression is often undiagnosed and untreated, and between one-third and one-half of all women who are depressed during pregnancy remain depressed a year later, which implies a significant duration of exposure for many children.
Our finding that maternal mental health (in the control group of women diagnosed as clinically depressed at baseline and not treated) is linked to the child’s socio-emotional development has important consequences. A number of studies have documented that socio-emotional skills in childhood are predictive of adult outcomes, including mental health, educational attainment, and earnings (Currie and Stabile 2006; Bennett et al. 2016; Halfon et al. 2014). Another strand of the literature demonstrates that socio-emotional skills have an even longer-lasting impact, influencing the next generation. In particular, a number of studies show a positive intergenerational correlation in socio-emotional skills (Loehlin 2005; Groves 2005; Anger 2012; Dohmen et al. 2012; Grönqvist, Öckert, and Vlachos 2017; Attanasio, de Paula, and Toppeta 2022). Most of the cited studies measure socio-emotional outcomes in adolescence or adulthood. One study that, like ours, measures socio-emotional outcomes in childhood is Attanasio, de Paula, and Toppeta (2022). However, they associate the child’s outcome with the mother’s socio-emotional skills when she was a child, whereas we are primarily interested in the mother’s socio-emotional skills when she is parenting the newborn child. A second difference in our study from the cited literature is that it is set in a developing country, and we know much less about socio-emotional developmental paths in these settings. Third, none of the cited studies use experimental variation in the mother’s socio-emotional skills.
The paper proceeds as follows. Section II provides the details of the intervention, discusses baseline balance and attrition over time, describes the data set and the outcomes, and discusses the methodology used to reduce the dimensionality of the outcome space and estimate the treatment effects. Section III presents the analytical framework, and Section IV presents the empirical results; Section V discusses the mechanisms through the lens of a simple structural model. Section VI concludes.
II. Study Design and Data
A. The Intervention
We use longitudinal data on a pregnancy cohort, established in the context of a clustered randomized controlled trial (RCT) in rural Punjab, Pakistan, a low-resource context characterized by a high prevalence of maternal depression and limited access to clinical mental health care. The trial recruited women who were depressed during pregnancy and provided them with a three-year-long, peer-delivered psychosocial intervention (Thinking Healthy Programme Peer-Delivered Plus, THPP+) consisting of cognitive behavioral therapy with a focus on behavioral activation, self-care, and attention to the infant’s health and development.
1. Depression screening
Between October 2014 and February 2016, all pregnant women who were eligible for the study—married, resident in Kallar Syedan, a subdistrict of Rawalpindi in Pakistan, and not in need of immediate medical attention, were approached and screened for depression using the Patient Health Questionnaire (PHQ9). The PHQ-9 is a standard instrument for screening and monitoring the severity of depression; it includes questions about the frequency of depressive symptoms in the last two weeks, such as lack of interest or ability to concentrate, feelings of sadness or hopelessness, sleeping or eating problems, restlessness, and suicidal thoughts. Pregnant women who scored 10 or more on the PHQ-9 were invited to participate in the trial.
Among 1,731 women who were screened for depression, 572 (33 percent) were identified as depressed according to the PHQ-9 criteria. Of these mothers, 287 were in the clusters randomized to the intervention, 283 in the control clusters, and two mothers refused to participate before the baseline assessment. Of the 1,159 pregnant women who were screened as not depressed, 584 were randomly selected to constitute the nondepressed arm of the study. They represent a natural reference group to understand the evolution of maternal and child outcomes and to benchmark the potential effectiveness of the intervention.
2. Randomization
The trial was randomized across 40 village clusters. These clusters were geographically separate to minimize the risk of spillover. Twenty clusters were randomized into receiving the intervention and 20 to the control arm. Each village cluster contributed approximately 14 perinatally depressed women. Research teams responsible for identifying, obtaining consent, allocating, and interviewing study participants were not apprised of the participants’ original depression status and their allocation across the study arms.
3. THPP+ intervention
Thinking Healthy Programme Peer-Delivered Plus (THPP+) is a low-intensity scaleable psychosocial intervention delivered by volunteer peer women from the same community as the mother. Peers received prior classroom training in accordance with the intervention content, which builds on a previous intervention that proved very successful in a similar context (Rahman et al. 2008). They were provided supervision throughout the trial period. The intervention strategy includes behavioral activation to overcome unhealthy thinking, with a focus on self-care and infant development.
The timeline of the THPP+ intervention and all follow-up surveys is summarized in Figure 1. In the intervention group, depressed women received a combination of individual and group sessions. Starting from the third trimester of pregnancy until six months postpartum, participants attended ten individual and four group-based sessions, with a primary focus on modifying maladaptive thoughts and behaviors frequently observed among individuals experiencing depression. From seven to 36 months postnatal, another 18 group sessions were delivered; the first six sessions were delivered monthly, the rest every two months. The content of these lower-intensity booster sessions was a continuation of the behavioral activation strategy, with a special focus on contributing to the mother–child interaction and child development by providing examples of age-appropriate activities, as well as information about childcare. Since a large part of the intervention was delivered in group sessions, the social component of meeting with other mothers, alongside the behavioral activation content discussed during the sessions, might have contributed to any intervention effects.
Timeline of THPP+ Intervention and Follow-Ups
Perinatally depressed women in the treatment arm received the THPP+ intervention throughout the trial, while women in the control arm received Enhanced Usual Care (EUC), which is the routine healthcare provided in the region. It is enhanced in the sense that the participants were informed of their depression status and offered guidance about how to seek help. Women who were not diagnosed as perinatally depressed (nondepressed group) did not receive any treatment.4
4. Sample and longitudinal follow-up
Our study sample consists of the experimental group of depressed mothers who were randomized into treatment and control arms and the group of mothers who were not depressed at baseline. Data collection on the mother–child dyads was done six times: at the third trimester of pregnancy and 3, 6, 12, 24, and 36 months postpartum. Figure 1 provides the compositions of the follow-up samples and the respective loss-to-follow-up rates (LTFU). A longitudinal comparison requires a similar measurement system over time, but we have no measure of cognition at 3, 6, and 24 months, and we have a different measure of parental investment at 24 months (see Online Appendix Table A1). For consistency, we only analyze data from the waves at 6, 12, and 36 months.5
B. Measurement and Outcomes
The data contain multiple validated and widely used scales of maternal mental health and functioning and of the cognition, socio-emotional, and physical health of children. Online Appendix Table A1 presents the full list of measures.
To measure maternal mental health across all of the waves, we use the Patient Health Questionnaire (PHQ-9) and the Structured Clinical Interview for DSM (SCID), a 13-item semistructured interview for making the major DSM-5 diagnoses. We also include the Cohen Perceived Stress Scale (PSS), a ten-item instrument among the most widely used in the psychological literature to measure self-reported stress. To measure her functioning, we use the WHO Disability Assessment Schedule (WHO-DAS), a 17-item assessment instrument developed by the World Health Organization (WHO) to evaluate, across cultures and domains, a person’s ability to perform various activities of daily living.
To assess the child’s cognitive development at 12 and 36 months of age, we use five scales from the Bayley Scales of Infant Development (Bayley-III). These scales measure various aspects of infant and toddler development in the following domains: cognitive, language (receptive and expressive), and motor (gross and fine).
To measure the child’s socio-emotional skills, we use the social-emotional subscale of the Ages and Stages Questionnaire (ASQ-SE), a validated screening tool for assessing social-emotional development in children aged 1 month to 6 years (Lamsal, Dutton, and Zwicker 2018).6 The ASQ-SE uses parent-reported questions to identify potential difficulties or delays in the areas of self-regulation, compliance, communication, adaptive functioning, autonomy, interaction with people, and affect (the child’s ability or willingness to demonstrate their own feelings and empathy for others). When the child is 36 months, we also include the Strengths and Difficulties Questionnaire (SDQ), a brief behavioral screening questionnaire used to assess children’s mental health. It has subscales to detect emotional symptoms, conduct problems, hyperactivity and inattention, peer relationship problems, and prosocial behavior.
The child’s physical health was assessed by measuring their weight, height, and head circumference from 3 to 36 months. These measurements were converted to age-adjusted z-scores, serving as proxies for the child’s anthropometrics and indicating their physical growth and development.
To measure parental investment at 12 and 36 months, we used the HOME inventory, a well-established observational tool that evaluates the quality of cognitive stimulation and emotional support offered by parents to their child. It is as a widely used measure to examine the level of parental investment in a child’s development.
Given the richness of the data for both mothers and children, we aggregate outcomes into indexes to overcome measurement error problems, improve statistical power, reduce the dimensionality of the data, and mitigate the issue of multiple hypothesis testing. We present the main results using latent factor scores, described below, although patterns are similar using inverse covariance weighted (ICW) indexes.7
C. Balance and Attrition
1. Balance
The experimental sample was slightly imbalanced at baseline, as shown by the summary statistics in Table 1. For instance, pregnant women in the treatment arm were on average 1 cm taller and lived in households with 0.3 more people per room than women in the control clusters. Treated women also suffered from slightly—albeit not significantly—worse mental health, scoring 0.4 higher on the PHQ-9 (depression), 0.6 on the WHODAS (functioning), and 0.9 on the PSS (stress). A joint F-test rejects balance of baseline characteristics (p-value = 0.01).8 Splitting by gender of the index child shows that the sample of mothers of boys is more balanced than that of girls: treated mothers of girls scored 1.6 higher on the PSS and had 0.5 higher number of people per room, lower socioeconomic status, and less educated husbands (Online Appendix Table A8). However, a joint test of balance for covariates within each gender group does not indicate any statistically significant imbalance, possibly due to lower statistical power (p-values of 0.41 for mothers of boys and 0.12 for mothers of girls). Overall, this slight imbalance seems to be driven by small differences in participants’ baseline characteristics, not by systematic differences between treatment and control village clusters. We confirm balance across village clusters by using the mothers who were not depressed at baseline and lived in the same villages as treated and control mothers (Online Appendix Table A9). A joint test of balance using the baseline characteristics of mothers who were not depressed in pregnancy (baseline) shows balance across treatment and control clusters (p-value 0.456). Similarly, a joint test of balance using the whole sample (nondepressed and depressed mothers pooled) is not rejected (p-value = 0.317).
Baseline Balance
2. Attrition
Lost to follow-up (LTFU) rates range between 18.5 percent and 23 percent in the study period. These attrition rates compare favorably with attrition rates in pregnancy cohort data. The main reason for being lost to follow-up was the death of the index child (constituting around 40 percent of the attritors),9 and this was balanced across study arms. Attrition did not differ by treatment status, despite some small imbalance in attritor characteristics (Online Appendix Tables A2–A5).10 Attritors generally had more crowded households and higher baseline PHQ-9 total scores. Attritors at 6 months additionally differ by having higher blood pressure and lower socioeconomic status and were more likely to be pregnant for the first time. Mothers who were lost to 36-month follow-up had higher weight and were more likely to coreside with their mother or mother-in-law.
In the analysis to follow, we address baseline balance concerns by including covariates in the model, demeaned and interacted with the treatment indicator (Goldsmith-Pinkham, Hull, and Kolesár 2022). Although attrition is not differential by treatment status, we show that our estimates are robust to using inverse probability weights to adjust for attrition.
III. Analytical Framework
To study the impact of THPP+ on the developmental trajectory of maternal mental health and child skills, we use latent factor scores, following a long history in psychometrics (Spearman 1904; Jöreskog and Goldberger 1975) and a more recent one in economics (Cunha and Heckman 2008; Cunha, Heckman, and Schennach 2010; Attanasio et al. 2020; Attanasio, Meghir, and Nix 2020). Latent factor analysis is a model-based approach that facilitates the study of maternal and child developmental trajectories by reducing measurement error and the dimensionality of the outcomes.
We construct the factor scores by assuming a separate measurement system for each domain and then employ Exploratory Factor Analysis (EFA) to select a concise set of measures. This approach helps us identify key factors that best represent the underlying constructs within each domain, while maintaining simplicity and efficiency in the measurement process. Following Agostinelli and Wiswall (2016), the scaling of each factor is standardized by normalizing the measure with the highest factor loading to one, while maintaining the same measure at all time points. The location is fixed by normalizing the means of the latent factors to zero for the control group at the initial time point (6 months). This approach ensures consistent and comparable scaling across the factors over the different time points in the analysis, allowing us to capture the growth of the latent factors over time.11
To close the model, we connect factor scores over time and capture the dynamic evolution of the child’s latent human capital. We follow Cunha and Heckman (2008) and Attanasio et al. (2020) and specify the production function for child development as:
1
where 
and 
are vectors for child skills in treatment arm d—where d = 0 indicates the control group, d = 1 indicates the treatment group, and d = 2 the baseline nondepressed—at time t and t + 1 respectively. 
stands for parental investment, which occurs between the realizations of 
and 
.12 
is maternal mental health and functioning at time t, which we conceptualize as a capital input, X contains baseline covariates measured before the treatment assignment, and η is the vector of random shocks to child development. We allow the distribution of the latent factors and the parameters of the production function 
to vary by the child’s age t + 1 (we estimate one production function at age 12 months and another one at age 36) and by intervention arm d.
Allowing the factors and the parameters to vary by intervention arm, this model can be seen as a generalization of a Kitagawa–Oaxaca–Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973). The intervention can act through two main mechanisms: (i) a change in the level of inputs, parental investments, and maternal health and (ii) a change in the returns to inputs, that is, the efficiency with which the inputs translate into child outcomes. This change in efficiency can happen through a combination of changes in TFP, that is, a different intercept Ad and a shift of the whole production function upwards or downwards, and changes in input-specific returns, that is, a different slope of the production function and shift in the derivative between child development and a particular input 
or 
.13
We present illustrative examples of the pathways through which the intervention may impact the outcomes. First, it may increase the level of inputs, including maternal mental health and parental investments. For instance, the intervention could alleviate maternal depression by facilitating mothers’ engagement in activities that provide them with a sense of achievement and enjoyment through behavioral activation. Alternatively or in addition, it might encourage mothers to dedicate more quality time to their children and invest in enriching resources like new toys and educational materials, thereby enhancing the home environment (parental investment).14
The intervention may also alter the production function, influencing the way that inputs translate into outputs. For example, the intervention may increase the productivity of each unit of maternal time spent with the child by improving maternal focus and empathy or by inducing a more age-appropriate use of time and physical resources. Conversely, it could potentially diminish the productivity of maternal mental health if there are decreasing returns. In particular, cases of maternal depression shifting from moderate to mild may exert a weaker influence on child development than cases where the shift is from severe to moderate depression.
The analysis is conducted in two steps. First, we employ maximum likelihood to estimate the factor model and extract the predicted factor scores, as described in detail in Online Appendix Section D.1. The factor scores are then used to assess the causal impact of the intervention on maternal mental health and child outcomes at each time point. The results of this reduced form analysis are discussed in Section IV.
In the second step, we estimate the parameters of Equation 1, aggregating the reduced form results of the first step in two systems of equations—one at age 12 and the other at age 36 months. Since we lack instrumental variables that might induce quasi-exogenous variation in the inputs of the production function, this analysis is descriptive. Yet, this synthesis helps us explore the reasons why intervention effects on maternal mental health did not spillover to child development. The results of the production function estimates are presented in Section V.
IV. Treatment Effects
We evaluate the impact of the perinatal psychosocial intervention on maternal mental health, daily functioning, and child skills during the first three years of life, leveraging the cluster-randomized nature of the intervention and using ordinary least squares. We estimate intention-to-treat (ITT) effects on the latent factor scores for the domains of maternal mental health, maternal functioning, child cognition, physical health, socio-emotional skills, and parental investment (Table 2 and Figure 2). Table 2 reports the estimated ITT on the latent factors normalized to mean zero and standard deviation one for the control group at the initial time point (6 months) only, to understand the evolution of the latent factors over time, while Figure 2 and Online Appendix Table A22 report ITT on latent factors that are normalized to mean zero and standard deviation one at each time point, to allow for comparison of effect sizes in standard deviation units. For completeness, we also report ITT effects on each individual measure in Online Appendix Tables A18–A21.
Coefficient Plots of Factors
Notes: Plot of the adjusted beta coefficients reported in Online Appendix Table A22 and their 90 percent and 95 percent confidence intervals. Latent factors are normalized to mean zero and SD one in the control group at each time point to allow comparability of effect sizes in standard deviations. Coefficients are obtained from the regressions of items on the treatment indicator and its interactions with the (demeaned) baseline covariates including baseline PHQ Total, baseline WHODAS Total, baseline PSS Total, mother’s baseline age, weight, height, waist circumference and blood pressure, family structure, grandmother being resident, total adults in the household, people per room, number of living children (split by gender), whether the index child is the first child, parental education levels, asset-based socio economic status (SES index), life events checklist score, interviewer fixed effect, union council fixed effect, and days from baseline. All estimations control for child gender and age (in days). Standard errors clustered at the village cluster level. Factor scores are coded so that a higher score always indicates a better outcome and standardized to have mean zero and SD one in the control group at each time point.
Trajectory of Summary Indexes
As our baseline and follow-up samples were not completely balanced along baseline characteristics, the regressions control not only for child age in days, interviewer fixed effects, and union council fixed effects (stratification unit), but also for the full set of baseline characteristics (demeaned) and their interactions with the treatment indicator (adjusted beta).15 Note that we can only identify the overall causal effect of the THPP+ intervention and not the causal effect of recovering from depression or of any individual component of the intervention, such as behavioral activation or group-based aspects of the treatment. Including interactions with treatment allows us to control for possible heterogeneity in the impacts of baseline characteristics on outcomes. Reported standard errors are clustered at the village cluster level (that is, the randomization unit). We also compute p-values using randomization inference based on Young (2019), with the randomization permuted at the cluster level. We observe minimal changes in the p-values due to the randomization inference, as shown in Online Appendix Tables A27–A28.
We additionally present results of the intervention on the distribution of outcomes. Online Appendix Figure A6 presents the estimated densities of the latent factors for the control and treatment clusters. To compare the CDFs of the two groups, we perform a Kolmogorov–Smirnov test with bootstrap.16 Quantile treatment effects are reported in Online Appendix Figure A7.
A. Maternal Mental Health and Functioning
The intervention is effective in improving the mother’s condition at 6, 12, and 36 months postpartum. The upper panel of Table 2 and the first panel of Figure 2 present the adjusted beta coefficient plots of latent factor scores. Improvements range between 0.17 and 0.27 standard deviations in maternal mental health and between 0.18 and 0.29 SD in maternal functioning, with the largest effect sizes observed at 36 months.
Plots of the outcome distributions show a rightward shift in the latent factor score for maternal mental health throughout the trial period. These effects are bigger in the lower half of the distribution, although this difference in quantile treatment effects is not always statistically significant.
Treatment effects on individual maternal outcomes are reported in Online Appendix Table A18. Treated women experienced a significant reduction in depression scores (PHQ-9) at 6 and 36 months postpartum relative to women in the control clusters (p-values 0.014 and 0.001, respectively). Splitting the PHQ score into different categories, the greatest reduction is concentrated in the moderate severity category (15 ≤ PHQ-score ≤ 19), with an increase in the women in the minimal category (PHQ-score ≤ 4). Treated women were less likely to have a major depression episode at 6, 12, and 36 months, with a reduction of likelihood ranging between 7 and 12 percentage points (p-values 0.011, 0.011, and 0.001, respectively). Their stress score is significantly lower and their daily functioning significantly better than in the control group. Overall, we observe positive and significant treatment effects across multiple indicators of maternal depression, stress, and functioning in the three waves analyzed.
B. Parental Investment and Behavior
The adjusted beta coefficients related to the parental investment factor score are all positive (0.08–0.11 SD), but not statistically different from zero. These treatment effects, even if they were to be more precisely estimated, would suggest only modest improvements when compared to other global studies focusing on at-risk parents (Rayce et al. 2017; Jeong et al. 2021).
Analyzing the different measures of parental investment in Online Appendix Table A21, we find the intervention improved most subscales of the HOME inventory indicating maternal responsivity, avoidance of restrictions and punishment, organization of the child’s environment, and provision of appropriate learning materials at 12 months postpartum. At 36 months, the intervention had positive effects on the total HOME score, acceptance, and learning materials, albeit imprecisely estimated, but only small positive and sometimes negative effects on other subscales.
C. Child Outcomes
The estimated treatment effects on child outcomes are generally noisier than on mothers. The intervention seems to have no clear effect on cognition—with estimated ITT coefficients smaller than 10 percent of a standard deviation and hovering around zero, or on physical health, which displays both slightly positive and mildly negative adjusted beta coefficients. Notably, the intervention has a sizeable, albeit transitory effect on socio-emotional skills; the estimated ITT at 6 and 12 months are 0.19 and 0.39 SD respectively, indicating considerable improvements. However, these treatment–control differences fade out by the 36-month mark, when the estimated ITT effect is only 0.06, and it is neither economically meaningful nor statistically different from zero. The transitory effect might have persistent consequences, even if it does not itself persist. For instance, socio-emotional skills in infancy might fuel self-regulation, interaction, and curiosity (and possibly other domains that are hard to measure, especially at an early age), which in turn might improve school achievement and later life outcomes. In line with this, in the next section, we report small but positive estimates of cross-skill productivity between socio-emotional ability and cognition up until age 3.
Looking at the individual indexes in Online Appendix Tables A19–A20, we observe significant improvements only in certain socio-emotional and cognitive domains. The total ASQ-SE score is generally lower (indicating better socio-emotional skills) in the treatment group. Looking at the subcomponents of ASQ-SE shows that, at 12 months, the improved ASQ-SE in the treatment group is driven by significant improvements in self-regulation (measuring the child’s ability to regulate their emotions and adjust to new environments). These effects are mainly driven by male children. At 36 months, the intervention impacts are once again on self-regulation and now also on autonomy.
In terms of cognitive outcomes, the estimated treatment effect on the Bayley receptive domain score (one of the two components of Bayley-III) is significantly positive at 36 months, with a score increase of 0.39 (p-value 0.06) in the treatment group, which brings the mean scores of the treatment group close to the scores of the nondepressed group. However, treatment effects on the aggregate cognition index and factor score are small (0.09 and 0.07 SD, respectively) and imprecisely estimated.
Looking at the distribution of outcomes, there is a shift to the right in the distribution of children’s socio-emotional skills in the treatment group in the first 12 months of the trial. However, at 36 months, the two densities overlap again, suggesting a short-term effect. Quantile treatment effect analysis yields larger effects in the lower half of the distribution in the first two years, which become insignificant at 36 months postpartum (Online Appendix Figure A7).
The distribution of the child cognition factor shows a scale shift at 12 months and a small location shift at 36 months postpartum. For children’s physical health, the densities for the control and treatment groups overlap, and the Kolmogorov–Smirnov test cannot reject that they are equal. Quantile treatment effects are also not generally different from zero in any part of these distributions.
D. Heterogeneity
Exploring treatment effect heterogeneity on maternal outcomes by gender of the index child reveals that the estimated benefits are larger for the mothers of boys (Figures 3–4 and Online Appendix Tables A29–A31). As discussed earlier, intervention effects on investment and child skills also show a tendency to be stronger for boys. There is a well-documented son preference in South Asia, and some evidence that women who have sons are treated better by the family than women who have daughters (Sathar et al. 2015; Milazzo 2018; Bhalotra, Chakravarty, and Gulesci 2020). It seems plausible that women who are in a generally more supportive environment are more responsive to treatment, and this would explain our finding. However, we can imagine the reverse—that is, that treatment effects are larger where the environment is harsher. Indeed, in Baranov et al. (2020), we found that a similar intervention (THP) run on a different sample of new mothers in rural Pakistan was more effective for mothers of girls in a seven-year follow-up. The length of the follow-up aside, the intervention analyzed in this study (THPP+) differs in duration and in intervention modality (see Section I for details). THPP+ was peer-delivered, while THP was delivered by trained community health workers. One possible explanation is that peers (other mothers in the community) might implicitly reinforce gender norms, whereas community health workers might act to empower mothers of girls. We have no hard evidence of this potential channel, but it is a relevant consideration to highlight when considering task-shifting to peers in an attempt to scale up interventions.
Coefficient Plots of Factors—Boys
Notes: Plot of the 90 percent and 95 percent confidence intervals and the adjusted beta coefficients obtained from the regressions, using only the sample of families where the index child is a boy, of items on the treatment indicator and its interactions with the (demeaned) baseline covariates including baseline PHQ Total, baseline WHODAS Total, baseline PSS Total, mother’s baseline age, weight, height, waist circumference and blood pressure, family structure, grandmother being resident, total adults in the household, people per room, number of living children (split by gender), whether the index child is the first child, parental education levels, asset-based SES index, life events checklist score, interviewer fixed effect, union council fixed effect, and days from baseline. All estimations control for age (in days). Standard errors clustered at the village cluster level. Factor scores are coded so that a higher score always indicates a better outcome and standardized to have mean zero and SD one in the control group at each time point.
Coefficient Plots of Factors—Girls
Notes: Plot of the 90 percent and 95 percent confidence intervals and the adjusted beta coefficients obtained from the regressions, using only the sample of families where the index child is a girl, of items on the treatment indicator and its interactions with the (demeaned) baseline covariates including baseline PHQ Total, baseline WHODAS Total, baseline PSS Total, mother’s baseline age, weight, height, waist circumference and blood pressure, family structure, grandmother being resident, total adults in the household, people per room, number of living children (split by gender), whether the index child is the first child, parental education levels, asset-based SES index, life events checklist score, interviewer fixed effect, union council fixed effect, and days from baseline. All estimations control for age (in days). Standard errors clustered at the village cluster level. Factor scores are coded so that a higher score always indicates a better outcome and standardized to have mean zero and SD one in the control group at each time point.
We investigated heterogeneity by birth order of the index child, an asset-based index of socioeconomic status of the family, education of the mother, and baseline depression severity (PHQ-9 total score). We find no systematic patterns here.
E. Discussion
The group-based, peer-delivered psychosocial intervention was effective at achieving one of its targets, which was improving maternal mental health and daily functioning. These improvements in well-being are complemented by smaller and imprecisely estimated increases in parenting behavior of 8–11 percent of a standard deviation and by a sizeable but transitory change in children’s socio-emotional development. This improvement in child skills at 12 months appears to be a direct effect of the intervention, which included training and support for child development.
Our results do not appear to be driven by attrition. Attrition-adjusted estimates using inverse probability weighting (IPW) and Lee bounds (Lee 2009) are shown in Online Appendix Table A6, with gender-specific results in Online Appendix Table A7. Our results are robust to the IPW correction, which only marginally changes the estimated coefficients and their precision. The attrition-corrected Lee bounds are wide, as is typically the case, but in the sample of mothers of boys (in which baseline characteristics are more balanced) show positive and significant effects on maternal mental health and on child socio-emotional skills both at 6 and 12 months.
To provide a benchmark for the effectiveness of the intervention and to put the magnitude of the treatment effects in perspective, we compare the adjusted beta coefficients with the mean level of the summary indexes for the mothers who were not depressed at baseline. Standardizing the summary index for the control group to have a mean of zero, the average outcome for the nondepressed mothers represents the association between prenatal depression and outcomes. We call this descriptive statistic the “depression gap” and display this in Online Appendix Table A33, Columns 5–7.
The intervention acted to narrow depression gaps, tending to bring the medium-term outcomes of perinatally depressed women closer to the outcomes of women who were in the same pregnancy cohort but not depressed at baseline. This is the case for child socio-emotional skills and parental investment.17 The depression gap in child health and cognitive skills is often small and imprecisely estimated. As such, there was limited leeway for the intervention to improve these domains.
The results in this paper build on our findings in Maselko et al. (2020). We extend that analysis in the following ways. We investigate dynamics, exploring multiple indicators and their evolution throughout the study period. As child development is not a linear process, a more granular approach is of substantive importance. At each age, we estimate treatment effects by gender of the child and on the distribution of outcomes rather than only at the mean. We provide treatment effects on a broader set of outcomes (including, for instance, the ASQ-SE for socio-emotional development). We use aggregate summary indexes and factor scores to provide summary measures of maternal well-being and child development and to improve statistical power. We also adopt a less restrictive statistical specification.18 A final and key differentiation is that we now impose some structure on the dynamic evolution of children skills, accounting for the trajectory of maternal mental health, functioning, and parenting, and estimate the production function for skills at ages 12 and 36 months. We discuss this next.
V. The Technology of Skill Formation
The results above indicate that the intervention improved maternal mental health, but these enhancements did not consistently transfer into lasting improvements in child skills. This discrepancy is at odds with some of the descriptive literature comparing the socio-emotional outcomes of children of depressed and nondepressed mothers (Herba et al. 2016; Leung and Kaplan 2009; Gaynes et al. 2005), but it is consistent with other literature that finds that moderate levels of maternal depression are not systematically associated with impaired child development (Laplante et al. 2008; DiPietro et al. 2006). To reconcile the suite of reduced form findings and understand the mechanisms by which the intervention might have influenced the outcomes, we impose the simplifying structure discussed in Section III on the dynamic evolution of the child’s latent human capital.
We contribute to the literature on mental health and child development in two related ways. First, we include in the model two dynamic latent factors measuring maternal mental health and functioning 
. Their measurement is consistent over time and uses state-of-the-art measurements for the screening and assessment of three relevant dimensions of maternal mental health—depression, stress, and daily functioning. We estimate their contribution to the production function of the child’s cognitive, socio-emotional, and physical health. Earlier related studies at best include a time-invariant measure of maternal characteristics, such as cognitive skills, physical health, or noncognitive skills (Cunha and Heckman 2008; Cunha, Heckman, and Schennach 2010; Attanasio, Cattan, and Meghir 2022). To distinguish maternal mental health from parental investments, we conceptualize it as capital in the production function, similar in principle to the conceptualization of physical health as capital (Grossman 1972).
Second, this study is the first to estimate how a psychosocial intervention targeting the mother might influence the production function of children’s skills, allowing some parameters of the production function to vary with the intervention. Similar to a Kitagawa–Oaxaca–Blinder decomposition (Kitagawa 1955; Oaxaca 1973; Blinder 1973), we allow the intervention to act through two potential mechanisms: (i) a change in the level of parental inputs and (ii) a change in the productivity of these inputs, that is, the slope of the production function.
These two channels are embedded into our specification of the dynamic model of skill formation. For ease of interpretation and estimation, we assume that the production functions for child socio-emotional skills, physical health, cognition, parental investment, and maternal mental health described in Equation 1 are log-linear (Cobb Douglas).19
2
3
4
where H, S, and C stand for physical health, socio-emotional skills, and cognition of the child, respectively. Equation 2 reflects that children’s socio-emotional skills, health, and cognition in period t + 1 
are functions of the previous period stock of human capital 
, investments made by parents up to that point {Iit+1}, parental education, and maternal mental health and functioning 
. Xi denotes the same baseline covariates used in the treatment effect estimation in Section IV, notably the mother’s baseline age, weight, height, waist circumference and blood pressure, family structure, grandmother being resident, total adults in the household, people per room, number of children (split by gender), whether the index child is the first child, asset-based socioeconomic status index, and child gender.20 
stands for total factor productivity (TFP) and 
represents unobserved shocks to child development. Equation 3 and 4 model the evolution of the main inputs: the stock of maternal mental health and the flow of parental investment. The same control variables are included as in Equation 2.
We estimate the production and investment functions in Equations 2–4 in two stages: at 12 months and at 36 months. To do so, we use the factor scores resulting from the measurement system discussed above and in Online Appendix Section D.1. We exclude lagged cognition in the estimations for 12 months, as we did not measure cognition at 6 months.
While all of the distributions of latent factors are allowed to be different across treatment, control, and baseline nondepressed mothers—capturing potential changes in the level of inputs—we only allow the coefficients of 
, and Iit+1 to vary with treatment status (d)—capturing potential changes in slope and therefore productivity. We do this by including an indicator for the treatment group (treat) and an indicator for the group of mothers who were nondepressed at baseline (nondep) and interacting them with parental investment and maternal mental health (the two main inputs of interest). This simplifying assumption focuses the estimation on the two main channels that were targeted by the intervention: maternal mental health and investments. It allows us to study how the productivity of maternal mental health and investments changes as a function of the intervention.
Theoretically, we may observe the productivity of mental health increase or decrease as a result of the intervention. On the one hand, if the true relationship between the input (lagged maternal mental health) and the output (child skill development) is subject to diminishing returns, then an improvement in maternal mental health due to the intervention could move the treatment group further up and to the right along the curve, where the slope is flatter (see this notional curve in Figure 5A). The estimated relationship between the input and output in the control group would then exhibit a lower constant and a steeper slope than the treatment group. This would manifest as a positive parameter on the interaction between TFP and treatment (TFP × treat) and a negative interaction between maternal mental health and treatment (motherMH × treat). Plotting the observed nonparametric relationship between maternal mental health at 6 months, 
, and child skills at 12 months, 
, using the control group data does in fact indicate a nonlinear, concave relationship (Figure 5B), similar to Figure 5A and in line with the descriptive findings of Laplante et al. (2008) and DiPietro et al. (2006).21
Relationship between Maternal Mental Health (6 Months) and Child Skills (12 Months)
Notes: (A) Theoretical representation of a concave relationship between maternal mental health (input) and children skills (output) and the consequence of log-linearization at different average levels of the input. (B) Kernel-weighted local polynomial smoothing of the relationship in the control group between maternal mental health at 6 months and child socio-emotional skill factor (solid dark line), child cognition factor (dashed gray line), and child physical health factor (dotted line) at 12 months. (C) Left y-axis is kernel-weighted local polynomial smoothing and 95 percent confidence interval of the relationship between maternal mental health at 6 months and child socio-emotional skill factor at 12 months in the control group (solid dark line) and the treatment group (dotted line). Right y-axis is kernel density estimation of the distribution of maternal mental health at 6 months in the control group (dash-dotted dark line) and the treatment group (dotted gray line).
Alternatively, the intervention could change the shape and the location of the production function. For example, intervention components not specifically targeting maternal mental health (for example, improving mother–child bonding or seeking social support) may change the relative productivity of mental health, parental investments, or both. These non-mental-health components of the intervention might reinforce and complement the intervention-lead effects on maternal mental health, for example, allowing mothers who have recovered from depression to engage in more fruitful parental interaction with the children. This would lead to a positive coefficient on the interaction between maternal mental health and treatment (motherMH × treat). But the non-mental-health component of the intervention could also act as a substitute, shielding children from maternal depression and improving particularly the outcomes of children whose mothers did not recover from depression even after therapy. This would lead to a negative coefficient for motherMH × treat. Plotting the observed nonparametric relationship between maternal mental health at 6 months, 
, and child socio-emotional skills at 12 months, 
, separately for treated and control group indicates a potential substitution effect, with greater intervention effects for children whose mothers did not recover from depression (Figure 5C).
We now turn to the discussion of the empirical estimates of Equations 2–4. It is important to remember that we have one instrument (the intervention) and multiple endogenous inputs, for which it is difficult to find a plausible source of exogenous variation. The results of this analysis should therefore be considered as descriptive, similar to the existing literature estimating child skill production functions (see, for instance, the summary of the literature in Table 1 of Attanasio et al. 2022).
A. Estimates of the Technology
Tables 3–4 report estimates for the outcomes at 12 months and 36 months, respectively. The estimates reveal that children’s skills and maternal mental health are persistent over time, indicating “self-productivity” in skills. Socio-emotional skills, physical health, and maternal mental health exhibit persistence through from 6 to 36 months. For cognitive and socio-emotional skills, self-productivity is larger earlier in childhood. Consistent with estimates of skill formation in other settings (Attanasio, de Paula, and Toppeta 2022; Bufferd et al. 2012), skills are less predictive across domains—the “cross-productivity” of skills is at least a degree of magnitude smaller than self-productivity, often nonstatistically different from zero, except for the predictive power of physical health on cognitive skill development at both 12 and 36 months. Evidence on self- and cross-productivity of skills across domains at very early ages, 0–3, is relatively scarce, therefore providing an important contribution to the literature.
Estimates of the Production Function and Input Equations I
Estimates of the Production Function and Input Equations II
We now discuss the role of maternal mental health and parental investment across the three groups of women (control, treatment, and baseline nondepressed), initially for outcomes at 12 months of age, Table 3, and then for outcomes at 36 months, Table 4.
First, consider Column 1 of Table 3 for child socio-emotional (SE) skills, as this is where we find intervention effects at 12 months (see Figure 2). In the control group of women who were depressed in pregnancy but received no intervention (top panel), maternal mental health at 6 months is a significant predictor of socio-emotional skills at 12 months, consistent with Figure 5B.22 Parental investment at 12 months is positive but imprecisely determined.
The intervention modifies the shape of the production function in two significant dimensions (second panel, interactions). We see a positive coefficient on the interaction of TFP with treat and a negative coefficient on the interaction of maternal mental health with treat. A higher TFP in the treatment group indicated that the intercept and the whole production function have shifted up—the outcome is higher for each level of input. The negative interaction with maternal mental health tells us that the slope of the curve describing how the outcome varies with maternal mental health is flatter in the treated group than in the control group. This is consistent with decreasing returns to improvements in mental health (Figure 5B), as well as a larger effect of the intervention on the socio-emotional skills of children whose mothers did not recover from depression (Figure 5C).
Both the positive TFP shift and the shallowing of the slope of the relationship with maternal mental health that we see in the treated group are also evident in the group of mothers who were nondepressed at baseline. Thus, in line with expectations, the intervention moved the outcomes of children and mothers with prenatal depression closer to the outcomes of children and mothers who were not depressed during pregnancy. Put differently, the intervention bridges the “depression gap” in the production function, morphing the technology of skill formation for depressed mothers to look more like that for women who did not suffer depression during pregnancy.
We now summarize the main results for other outcomes at 12 months of age, in Columns 2–5 of Table 3. In the control group, maternal mental health at 6 months is predictive not only of socio-emotional skills but also of cognitive skills and physical health; it is significantly associated with child development across domains. Maternal functioning at 6 months has no direct relationship with child development above and beyond other inputs, such as maternal mental health and parental investments, but it raises parental investment. Parental investments at this early age are not predictive of any domain of child development, but they are related to maternal mental health. It is also notable that parental education and assets have a significant positive impact on parental investments but, conditional on investment, no direct impact on child outcomes.
Intervention effects are reported in the second and third panel of Table 3. The intervention raises TFP in the production of physical (but not cognitive) development. It attenuates the relationship between maternal mental health and cognition, and it strengthens the return to investments when the outcome is cognition. Once again, the direction of effects in the intervention arm is the same as the direction of effects among the group of mothers not depressed in pregnancy.
Now consider estimates for the production function for child skills at 36 months (Table 4). The estimates for the control group show that maternal mental health at 12 months has only small and statistically insignificant associations with child skills at 36 months, but is predictive of higher parental investments. In turn, parental investment at 36 months predicts higher socio-emotional and cognitive skills at 36 months, conditional on maternal mental health. Intervention effects at 36 months are also most evident for the investment outcome. The pattern is similar to that observed for skills outcomes at 12 months: TFP is higher, and there is an attenuation of the relationship between maternal mental health and parental investment. The only significant intervention effect in the production functions for child skills indicates lower TFP in the production of the physical health of the child, for which we have no clear explanation.
Mirroring our reduced form analysis of treatment effects and following recent trends in the literature focusing on socio-emotional skills and mental health (Moroni, Nicoletti, and Tominey 2019), we split the sample by gender and estimate the technology of skill formation separately for boys and girls. Online Appendix Tables A36–A37 suggest that the overall pattern of production function results is similar across child genders. If anything, maternal mental health seems to be more predictive of parenting for mothers of girls, although statistical power is limited for this comparison.
B. Discussion
Taking stock, the intervention changes both the level of the inputs and their associations with the outcomes (returns). First, it improves maternal mental health (at 6, 12, and 36 months). This is an input to the production function, being directly associated with an improvement in child skills at 12 months, while at 36 months, it is associated with improved child skills through increasing parental investment. Second, the intervention changes the shape of the production function, changing both its intercept (TFP) and its slope (the productivity of specific inputs—maternal health at 12 months and parental investment at 36 months).
These results are well summarized by Figure 5C, plotting the distribution of maternal mental health at 6 months in the treatment and control group, and their nonparametric relationship with child socio-emotional skills at 12 months. The intervention improves the whole distribution of maternal mental health, bringing more mothers into a flatter part of the production function, as modeled in Figure 5A. It seems plausible that, in the sample of women who have largely recovered from depression, marginal improvements in mental health have smaller impacts. Consistent with this, the productivity of maternal mental health is weaker in both the intervention group and the nondepressed group relative to the control group.
At the same time, our results are not consistent with a simple shift along a single production technology curve: treatment induced an upward shift of the whole curve, with an overall improvement in productivity (a positive TFP × treat coefficient). The upward shift is largely driven by mothers at the low end of the mental health distribution (who did not respond to the intervention), and this induces a greater flattening of the curve (a negative motherMH × treat coefficient). The evidence here indicates that the intervention had a direct impact on the improvement in socio-emotional skills of children at 12 months.
VI. Conclusion
We estimate the impacts of a peer-led psychosocial intervention delivered to women diagnosed as depressed in pregnancy, starting in the third trimester of pregnancy and continuing until the child was 36 months of age. Our findings reveal that the intervention resulted in significant and lasting improvements in maternal mental health and functioning. There was also a moderate increase in parental investment, although the estimate is not precisely estimated. However, despite these positive changes, we did not observe any noticeable improvements in overall indicators of child development in the long term.
To understand the associations of the multiple endogenous variables and the dynamics more clearly, we estimated a production function for child skills. Among women diagnosed as depressed in pregnancy but untreated (the control group), mental health is strongly related to child outcomes in early childhood and to investments in children in later childhood. These relations are economically significant; for example, they tend to be larger in magnitude than the associations between socioeconomic status and child skill development.
This suggests that an intervention targeting maternal mental health and parenting behaviors might improve children’s future skills. However, this does not seem to be the case. The potential reason for this is that the intervention mutes the relationship between maternal mental health and children’s outcomes. Just as in the sample of nondepressed mothers, the rate of return of mental health in the production function is close to zero for the treatment group. Therefore, any potential impact of the increase in mental health induced by the intervention is offset by a reduction in its efficiency in producing children’s skills.
Overall, both the reduced form and the production function estimates suggest that the intervention is effective and tends to move outcomes for perinatally depressed mothers towards outcomes for those who were not depressed during pregnancy.
Acknowledgments
Sponsored by the NOMIS Foundation and the Center for Health and Wellbeing at Princeton University.
Atif Rahman and Siham Sikander designed the intervention. The authors are grateful to Siham Sikander and his colleagues at the Human Development Research Foundation in Pakistan for implementing the trial and the follow-up and the Bachpan Study Team for interdisciplinary guidance and support. For their comments and suggestions, the authors thank Tania Barham, Davide Dragone, Alex Frankel, Allison Frost, Cheti Nicoletti, Jonathan de Quidt, Denni Tommasi, Joe Vecci, Matt Wiswall, and participants at the 13th Workshop on the Economics of Health and Wellbeing, the University of Milan Bicocca Seminar, the Warwick-CAGE conference on mental health, and the Causes and Consequences of Child Mental Health conference hosted by the Center for Health and Wellbeing at Princeton University. Sonia Bhalotra acknowledges partial funding for her time from ESRC award ES/S003681/1 to the Centre for Microsocial Change and from Warwick-CAGE. The data collection was partially funded by NICHD award R01HD075875 and NIMH U19MH95687. The study procedures were approved by ethics committees in the United States (Duke University, UNC) and Pakistan (Human Development Research Foundation’s Ethics Committee). This paper uses confidential data from the Bachpan Cohort Study https://www.bachpanstudy.com/. The data can be obtained upon request to the last author. At the end of the grant period, the anonymized data will be published in an open-access NIH repository. The code to replicate the results found in the paper is publicly available on github here https://github.com/pietrobiroli/tech_depression_BACHPAN.
Footnotes
↵1. These skills are malleable in childhood, but there is limited causal evidence of how malleable they are in the early years of life (Heckman, Stixrud, and Urzua 2006; Roberts et al. 2007; Almlund et al. 2011; Lundberg 2017; Nangle, Erdley, and Schwartz-Mette 2020; Abrahams et al. 2019).
↵2. We analyzed a similar but distinct intervention in Baranov et al. (2020). The intervention analyzed in Baranov et al. (2020) was delivered by salaried community health workers rather than by peer volunteers. It ran for 10 months rather than the 36-month intervention we analyze here ran for 36 months, and the two interventions were run in different locations in Pakistan. Also, there was no follow-up between 12 months and 8 years in the previous study, and no measurement of child socio-emotional or cognitive skills prior to 8 years. A key feature of this study is that we have frequent follow-up between birth and 36 months. Baranov et al. (2020) report intervention effects on skills at age 8, but do not estimate the production function for skills.
↵3. This is the case for the production of skills at 12 months of age. At 36 months, we see both an increase in TFP and diminishing returns to parental investment, which, in turn, is shaped by maternal mental health. Details are in Section V.
↵4. More information about the trial is in Sikander et al. (2015, 2019a,b) and Turner et al. (2016).
↵5. Further discussion of the 24-month wave is presented in Online Appendix Section C.
↵6. Longitudinal analyses using the ASQ-SE as a screening tool have been performed in several countries, when involving the general population (Marks, Sjö, and Wilson 2019), at-risk groups (Keenan et al. 2019; Cho, Chien, and Holditch-Davis 2021), and as an evaluation of a randomized controlled trial (Salisbury et al. 2022; Nores, Bernal, and Barnett 2019).
↵7. In Online Appendix Tables A18–A21 and Online Appendix Figures A4–A5, we show reduced form results with ICW indexes constructed by weighting the mean vector of outcomes by the row-sum of the inverse of their covariance matrix, following Kling, Liebman, and Katz (2007) and Anderson (2008). ICW indexes are useful to minimize the noise resulting from random errors that are uncorrelated across indicators and provide an efficient estimation of the treatment effect by allowing single hypothesis testing, which increases statistical power. They also offer flexibility to aggregate information from the observed measures that are not highly correlated or from different domains. The ICW index puts more weight on measures that are less correlated and thus capture new information. That is why, apart from estimating an index for each domain of child and maternal outcomes, we also construct an overall ICW index (for example, child index) to capture a comprehensive effect of treatment on mothers and their children. Each index for each domain at each time point is normalized to have a mean of zero and a standard deviation of one in the control group.
↵8. We regress a treatment dummy on all the baseline controls and report the p-value of the F-test of overall significance.
↵9. The high level of mortality in our sample is sadly in line with the region. Pakistan in 2020 had an infant mortality rate of 65.2 per 1000 deaths. Our mortality rate is higher, at 89/1000 by age 3, likely because our sample is rural and more disadvantaged, and it also includes stillbirths, whereas under-five mortality is reported with reference to live births.
↵10. An exception is the 24-month follow-up, when a joint test for balance is rejected (p-value = 0.046, see Online Appendix Table A4). We do not use the 24-month wave in the analysis. See Online Appendix Section C for additional information.
↵11. Additional details on the construction of latent factor scores are provided in Online Appendix Section D.1.
↵12. We use
instead of as an input in the production function to capture investments that accumulated up until t + 1. As parental investment is a flow variable and our indicators for investment mostly measure material investment (for example, whether the index child has certain toys), is more relevant in the production of .↵13. Put differently, the specification is flexible enough to capture several ways in which the intervention can boost children’s skills. It could improve the inputs (including maternal mental health or parental investment), leaving the production function parameters constant. It could shift the production function upwards (increasing total factor productivity, TFP). It could differentially boost the skills of the children whose mothers did not recover from depression, reducing the slope of the production function (that is, the marginal productivity) with respect to maternal mental health.
↵14. This channel is found in two related studies (Baranov et al. 2020; Angelucci and Bennett 2021) and appears even more likely in this intervention, which directly encouraged mothers to engage with and stimulate the child.
↵15. Adjusted beta coefficients are obtained from the regressions on the treatment indicator and its interactions with the (demeaned) baseline covariates including baseline PHQ Total, baseline WHODAS Total, baseline PSS Total, mother’s baseline age, weight, height, waist circumference and blood pressure, family structure, grandmother being resident, total adults in the household, people per room, number of living children (split by gender), whether the index child is the first child, parental education levels, asset-based SES index, life events checklist score, interviewer fixed effect, union council fixed effect, and days from baseline. All estimations control for child gender and age (in days). Considering the baseline imbalance in some key characteristics, we always focus our discussion on the adjusted treatment effect coefficients and the respective p-values in the text below.
↵16. The null hypothesis is that two CDFs are the same. Bootstrapped p-values are reported at the upper left corner of each plot.
↵17. For these outcomes, the depression gap is positive, favoring the healthy comparison group. It is statistically significant only for socio-emotional skills at 12 months and for parental investment at 12 and 36 months. The treatment effects on socio-emotional skills are similar in magnitude to the depression gap, but for parental investment, the effects are only about a quarter of the depression gap.
↵18. In Maselko et al. (2020) we controlled for only those variables that were statistically significantly imbalanced by treatment arm at baseline or predicted missingness at 36 months at the p < 0.10 level, following common practice in the public health literature. In this paper, we include a broader set of covariates and their interaction with the treatment indicator. Our controls include baseline PHQ-9 (depression), PSS (stress), and WHODAS (functionality) scores, which are significantly imbalanced at baseline when considered jointly, but not individually. The inclusion of baseline mental health measures drives the differences in point estimates between the findings in Maselko et al. (2020) and this paper. Another difference is that in Maselko et al. (2020) we report the impacts of the THPP+ intervention only on a pre-registered set of maternal and child outcomes at 36 months postpartum. For instance, Maselko et al. (2020) focus on clinical measures of depression (PHQ-9 score, depression remission, and major depressive episode), while we construct a broader measure of maternal mental health. Focusing on a narrow set of prespecified outcomes increases transparency and replicability, but might hinder our ability to learn systematically from the data (Coffman and Niederle 2015).
↵19. Freyberger (2020) shows that an erroneous normalization or misspecification of the latent factor structure might lead to biased estimates, especially in the case of nonlinear production functions such as CES.
↵20. As a robustness check, we also include in the controls the baseline level of mental health during pregnancy,
, to capture the idea that pregnancy might be a critical developmental window in terms of exposure to depression and to test for potential departures from the simple Markov dynamics as suggested by Attanasio, de Paula, and Toppeta (2020). Results in Online Appendix Tables A34–A35 show that the estimates do not change sizeably once baseline depression is added as a control.↵21. To mitigate the imposition of functional form assumptions and focus solely on localized averages, in Figure 5B we employ a Kernel-weighted local polynomial smoothing estimator. Results for 36 months display a similar pattern, but with a lower degree of concavity. In the estimation of the production function, we impose a linear-in-log relationship, which might be an approximation of the true underlying production function.
↵22. Note that the coefficient on maternal mental health reflects its direct association with child skills, conditional upon maternal functioning, maternal investment in children, and lagged child skills, all of which are potentially a function of maternal mental health.
- Received December 2023.
- Accepted September 2023.
This open access article is distributed under the terms of the CC-BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0) and is freely available online at: https://jhr.uwpress.org.
References
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.