*3.1. Main Results*

Table 3 reports the main results of the paper. Columns 1–3 present results for Ordinary Least Squares (OLS) with PDS-selected variables and full regressor set. Each column shows the results for a measure of mental health, as defined above. Columns 4–6 show results for the IV with PDS-selected variables and full regressor set as depicted in Equation (2).

First-stage estimates for the exposure to an external shock on maternal mental health are presented in Panel B. Selected instrument by the LASSO regression was suffering crop or livestock loss during pregnancy or within the first year of the index child. The outcome variable was a measure of maternal mental health in 2002. The coefficient indicated changes in maternal mental health after experiencing a shock of crop or livestock loss during pregnancy or after giving birth. Across columns, the precision of the estimate does not change, but the size of the coefficients is sensible to the measure of mental health used.

As presented in Panel A, our IV estimations indicate that poor maternal mental health has a negative impact on child cognition. An increase by one standard deviation in maternal mental health problems when children were 1-year-old or younger was associated with a reduction of 0.5–0.54 standard deviations in vocabulary Z-scores when children were 5 years old. This effect corresponds to a reduction of 31 percent of the mean PPVT raw score. These large estimated effects were consistent with existing evidence. For example, Aizer et al. [3] found that exposure to stress hormones in utero negatively affects cognition (verbal IQ at age 7), behavior, and motor development. Specifically, the authors found that exposure to cortisol in the top quintile of the distribution was associated with a 43 percent of a standard deviation reduction in verbal IQ.

The LASSO regression selected the following controls: Mother's age, wealth index, living in an urban area, child's age, consumption of durable goods index, household size, number of children younger than 5 years in the household, and height for age Z-score. The effects on child cognition of demographic controls (not shown in the table) are in the expected direction. Z-scores of children living in urban areas were higher than those of children living in an urban area. In addition, children's nutritional status also affected performance in the PPVT. The coefficients for wealth were positive, statistically significant, and among the highest, which was in line with research that points to socioeconomic status gradients of cognition as measured by vocabulary [16,33–35].



Panel A in Table 3 presents the estimated impacts of maternal mental health on children vocabulary at the age of 5 years obtained from Equation (2), using Ordinary Least Squares (OLS) and Instrumental Variables (IV) as estimation approaches. The dependent variable was measured using the standardized value of the PPVT test. Mental health indexes are standardized values of the SRQ-20 items using three different estimation approaches. Columns (1–3) present estimated coefficients using OLS and columns (4–6) show coefficients using IV. Both approaches were implemented using the option PDS-selected variables and full regressor available in the LASSO command. The selected instrument was suffering crop or livestock loss during pregnancy or within the first year of the index child. Selected controls are mother's age, wealth index, living in an urban area, child's age, consumption of durable goods index, household size, number of children younger than 5 years in the household, and height for age Z-score. The sample is restricted to children with available information on maternal mental health in 2002 and PPVT scores in 2006 and 2009. \*, \*\* and \*\*\* indicate statistical significance at 10%, 5% and 1%, respectively. Confidence intervals at the 95% confidence are presented in Appendix C. Robust standard errors in parentheses.

We then explored if these negative impacts were held three years after the first measure of vocabulary. Using data of child vocabulary at the age of 8, we estimated the model presented in Equations (2) and (3). The main results are presented in Table 4. Our estimations showed that maternal depression had no effect on the child's vocabulary at the age of 8. Not only were the estimated coefficients not statistically significant, but also their sizes were very small—that is, the vocabulary of children whose mothers' experienced mental health problems when they were 1-year-old caught up with the vocabulary of children whose mothers did not suffer mental health problems. As we present in the table in Appendix D, the effects of a shock on maternal mental health are not statistically significant a year after the woman experienced the shock. These results suggest that the effect of exposure to maternal depression during early childhood need not undermine language development permanently, and exposure to rich vocabulary environments later on during childhood can compensate for earlier developmental gaps. For our sample, it is possible that the convergence in vocabulary development is explained by the fact that by the time they reached age 8, all children had had exposure to formal education opportunities (99.9% of children in our sample), which may have a compensatory effect on children's vocabulary development. Still, given that early vocabulary constitutes a foundational skill that facilitates the development of other cognitive skills, based on our results, we cannot rule out the possibility that exposure to maternal depression during early life does not undermine cognitive development and academic achievement.
