FABBS Doctoral Dissertation Research Excellence Award

William E. Pelham III, Arizona State University, Department of Psychology

“Depression in Mothers and Behavior Problems in Children: From Association to Causation”

William E. Pelham III, Arizona State University, Department of Psychology

Abstract

Hundreds of studies have documented an association between maternal depression and child behavior problems, but little attention has been paid to other factors that might confound this
relationship. We used propensity score matching in a sample of 731 low-income families to estimate the causal effect of exposure to maternal depression at child age 2 years on child behavior problems from ages 3 to 14 years. Families with depressed vs. non-depressed mothers were equated via matching on 89 other relevant characteristics at baseline (e.g., living environment, child behavior, marital conflict). We then compared the equated groups on child externalizing and internalizing behavior from ages 3 to 14. Prior to matching, the effect of maternal depression was statistically significant for 34 of 40 different child behavior outcomes, and the magnitude of effects was in the small to medium range (mean d = 0.29). After matching, the effect of maternal depression was statistically significant for 4 of 40 outcomes, and the magnitude of effects was in the very small to small range (mean d = 0.08). Thus, most of the association of maternal depression to child behavioral outcomes was accounted for by confounding variables rather than a causal effect of depression.

Introduction

Over 50 years ago, Rutter (1966) showed that children of depressed mothers had more behavior problems than their peers. Since then, the relation of maternal depression to child behavior has become one of the best-studied topics in developmental psychopathology. A recent meta-analysis of 193 studies found that children of depressed mothers display more externalizing behavior (d = 0.43), internalizing behavior (d = 0.47), and general psychopathology (d = 0.49) in both clinical and community samples (Goodman et al., 2011). The field has replicated the effect in many contexts, delineated the conditions under which the effect is magnified or attenuated, and explored the mechanisms through which the effect occurs (Goodman, 2007).

Despite much progress, this literature has retained a key weakness: the evidence remains correlational rather than causal. Families with depressed vs. non-depressed mothers typically differ on many factors besides depression, including socioeconomic status, living environment, marital conflict, parenting practices, and their children’s pre-existing level of behavior problems. These confounding variables may explain part or all of the observed difference in later child behavior, undermining our understanding of the developmental processes truly at work. For example, greater exposure to concomitant marital conflict, rather than depression per se, may explain why the children of depressed mothers exhibit more aggression. Such a finding would suggest very different priorities for future research and clinical intervention than if depression itself were causing child aggression.

The current study seeks to address this limitation of past work using propensity score matching, an approach first developed in the 1980s (Rosenbaum & Rubin, 1983, 1984) but still receiving limited attention in clinical psychology. Propensity score methods provide machinery for equating exposed and unexposed groups on measured characteristics in observational (i.e., non-randomized) studies. Exposed and unexposed cases are matched on the propensity score, or the estimated probability of the case being exposed given its values on a set of covariates. Statistical theory indicates that after matching, the exposed and unexposed cases will be (in expectation) similar on all covariates that were included in the estimation of the propensity score. In this way, propensity score matching can help us equate families with depressed vs. nondepressed mothers on a large number of other factors that might explain differences in later child behavior.

We used propensity score methods to estimate the causal effect of exposure to maternal depression at child age 2 years on child externalizing and internalizing behavior from ages 3 to 14 years. Data were drawn from a prospective, longitudinal, multisite dataset in which 731 low-income families were recruited at child age 2 and assessed semi-annually until age 14. In Step 1, we created two groups of depressed and non-depressed mothers at child age 2 that were equated on 89 baseline covariates that might potentially confound the relation of maternal depression to child behavior. In Step 2, we compared the equated groups on child behavior outcomes over time in order to estimate the causal effect of exposure to maternal depression during early childhood.

Methods

Sample. Data were drawn from the Early Steps Multisite Trial, a sample of 731 at-risk families recruited from the Women, Infants, and Children (WIC) Nutritional Supplement program (see Dishion et al., 2008). Families were recruited and assessed at child age 2 years, then assessed again at child ages 3, 4, 5, 7, 8, 9, 10, and 14 years.

Measuring mother’s depression. Mothers’ depression at child age 2 was measured via the Center for Epidemiological Studies on Depression Scale (CES-D; Radloff, 1977), a 20-item, self-report measure of depressive symptoms. Following the published literature (Vilagut, Forero, Barbaglia, & Alonso, 2016), a threshold at total CES-D score ≥16 was used to create two groups: mothers who exhibited clinically significant depressive symptoms at child age 2 (45%) and mothers who did not (55%).

Measuring potential confounders. Covariates for equating the depressed and nondepressed groups were drawn from a comprehensive battery of questionnaires and inventories completed at the initial visit, when the child was 2 years old. We selected a total of 89 covariates spanning the domains of demographics (e.g., sex, race, ethnicity, income), areas of family strength (e.g., support from extended family), negative impact factors (e.g., drug use by parent, recent death in the family), child behavior (e.g., aggression, noncompliance, anxiety, sleep), neighborhood factors (e.g., danger, cohesion), parent functioning (e.g., substance use, frequency of contact with friends), and factors related to live-in partners (e.g., relationship satisfaction of
mother, live-in partner’s substance use).

Measuring child outcomes. Child behavior outcomes were measured via broadband rating scales of externalizing and internalizing problems (Achenbach & Rescorla, 2001). Mothers and secondary caregivers completed the Child Behavior Checklist at child ages 3, 4, 5, 7, 8, 9, 10, and 14. Teachers completed the Teacher Report Form at child ages 7, 8, 9, 10.

Analytic step 1. In step 1, we equated families with depressed vs. non-depressed mothers at child age 2 years on all 89 covariates. This required three steps: (a) estimating the propensity score for each family, (b) matching families on the propensity score, and (c) verifying that the matched groups were similar on all covariates. For continuous variables, we verified that the means of the matched groups were within 0.20 SD of each other. For binary variables, we verified that the rates of endorsement in the matched groups were within 5% of each other.

Analytic step 2. In step 2, we compare the matched (i.e., equated) groups on child behavior outcomes from ages 3 to 14 years. For each child outcome variable, we estimated two effects. First, the prima facie (i.e., unadjusted, or correlational) effect was estimated by comparing families with depressed vs. non-depressed mothers at child age 2. Second, the causal
(i.e., adjusted) effect of maternal depression was estimated by comparing only the matched families with depressed vs. non-depressed mothers at child age 2. The latter comparison accounts for baseline differences between depressed and non-depressed mothers on the 89 covariates, whereas the former does not.

Results

Step 1. Prior to matching, families with depressed mothers vs. non-depressed mothers differed significantly on most of the 89 baseline covariates. For example, the depressed group exhibited more home chaos (d = 0.54), more child internalizing problems (d = 0.54), and poorer relationships between mothers and their live-in partners (d = −0.58). Thus, unadjusted comparisons of the child behavior outcomes for depressed vs. non-depressed mothers were confounded by these other baseline differences. Using propensity score matching, 176 families with depressed mothers were matched with 176 families with non-depressed mothers. At child age 2, these two groups were similar on all 89 covariates. Differences on all continuous covariates were less than 0.10 SD and differences in the prevalence of all binary covariates were less than 5%. Thus, after matching, comparisons of the child behavior outcomes for the depressed vs. non-depressed mothers were no longer confounded by any of the 89 baseline covariates.

Step 2. Before matching, the effect of maternal depression at child age 2 was statistically significant for 34 of 40 different child behavior outcomes measured between ages 3 and 14. The magnitude of effects was in the small to medium range (mean d = 0.29), consistent with published literature. Effects were similar for child externalizing (mean d = 0.29) and internalizing behavior (mean d = 0.28). Effects based on mother report (mean d = 0.38) were larger than those based on secondary caregiver report (mean d = 0.23) or teacher report (mean d = 0.21) After matching, the effect of maternal depression at child age 2 was statistically significant for 4 of the 40 outcomes, and the magnitude of effects was in the very small to small range (mean d = 0.08). Thus, the mean effect size shrunk by nearly 75% after adjusting for baseline differences between families with depressed and non-depressed mothers. Effects after matching were similar for child externalizing (mean d = 0.09) and internalizing behavior (mean d = 0.07). Effects based on teacher report (mean d = 0.17) were larger than effects based on mother report (mean d = 0.07) or secondary caregiver report (mean d = 0.05).

Sensitivity analyses. Conclusions remained the same in the following sensitivity analyses: (a) making the adjustment for confounding using inverse probability of treatment weighting instead of propensity score matching; (b) analyzing child outcomes using longitudinal models; (c) measuring child outcomes using DSM symptoms of disorders instead of broadband behavior rating scales; (d) defining the maternal depression groups with more extreme thresholds (i.e., CES-D total scores ≥18 vs. ≤10).

Discussion

We used propensity score matching in data from the Early Steps Multisite Trial to estimate the causal effect of exposure to maternal depression during early childhood on later child behavior problems. First, we equated families of mothers with and without clinically significant symptoms of depression at child age 2 years on 89 other baseline covariates, including mother, child, and family characteristics. Next, we compared the equated groups on child externalizing and internalizing behavior from ages 3 to 14, as reported by mothers, secondary caregivers, and teachers. The mean effect of maternal depression shrunk from d = 0.29 before matching to d = 0.08 after matching. Thus, most (~ 75%) of the association of maternal depression with later child behavior was accounted for by confounding variables rather than a causal effect of maternal depression.

Findings imply that studies seeking to explicate developmental pathways relating maternal depression to child behavior will produce misleading results when they do not account for baseline differences between depressed and non-depressed mothers. Moderator variables that appear to alter the effect of maternal depression on child mental health may in fact be confounding factors that influence both variables in the relationship. Similarly, maternal depression may appear to mediate the effect of another developmental factor on child behavior, when in fact it simply serves as a proxy for other methods of transmission. Finally, despite the apparently strong prima facie effect of exposure to maternal depression, reductions in maternal depression should not be expected to resolve child behavior problems.

Findings also suggest that studies seeking to explore the causal effect of maternal depression on child behavior will need very large samples sizes. The mean causal effect in this study was d = 0.08. If the true causal effect of maternal depression were d = 0.08, more than 4,900 cases would be required to detect this effect with 80% power using a two-sample t-test (Cohen, 1988). Even if the true causal effect were d = 0.20, more than 750 cases would be required. Since these samples sizes are not common in the field of developmental psychopathology, investigators interested in causal pathways that include the link of maternal depression to child behavior should focus on design and analysis techniques that can improve statistical power (e.g., twin designs, integrative data analysis).

Finally, findings underscore the value of causal inference methods (e.g., propensity score matching) for addressing questions in developmental psychopathology, a field in which experimentation is often difficult or impossible. The relation of maternal depression to child behavior is robust and has been observed consistently across hundreds of studies (Goodman et al., 2011). If such a well-established effect is mostly explained by confounding variables, what might we find when applying similar designs to other developmental phenomena? If developmental psychopathologists can increase the correspondence between the estimated effects and the true causal effects, their theories will become more accurate.

Impact Statement

There is great need for effective prevention and intervention programs to help mothers who are depressed and mitigate the impact of their condition on their children. Developing these programs requires an accurate understanding of the way in which maternal depression impacts children, for which children and mothers these impacts are largest, and what modifiable processes might be targeted for intervention. The current study demonstrates that ignoring baseline differences between depressed vs. non-depressed mothers undermines our pursuit of this understanding. By accounting for these differences, future work can produce more accurate knowledge to guide public programs that help depressed mothers and their children.

References

Achenbach, T. M., & Rescorla, L. (2001). Manual for the ASEBA School-age Forms & Profiles: Child Behavior Checklist for Ages 6-18, Teacher’s Report Form, & Youth Self-Report. Burlington, VT: University of Vermont, Research Center for Children, Youth & Families.

Cohen, J. (1988). Statistical Power Analysis for the Behavioral Sciences (2nd ed.). Hillsdale, NJ: Lawrence Erlbaum Associates.

Dishion, T. J., Shaw, D., Connell, A., Gardner, F., Weaver, C., & Wilson, M. (2008). The Family Check-Up with high-risk indigent families: preventing problem behavior by increasing parents’ positive behavior support in early childhood. Child Development, 79, 1395– 1414.

Goodman, S. H. (2007). Depression in mothers. Annual Review of Clinical Psychology, 3, 107– 135.

Goodman, S. H., Rouse, M. H., Connell, A. M., Broth, M. R., Hall, C. M., & Heyward, D. (2011). Maternal depression and child psychopathology: a meta-analytic review. Clinical Child and Family Psychology Review, 14, 1–27.

Radloff, L. S. (1977). The CES-D scale: a self-report depression scale for research in the general population. Applied Psychological Measurement, 1, 385–401.

Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70, 41–55.

Rosenbaum, P. R., & Rubin, D. B. (1984). Reducing bias in observational studies using subclassification on the propensity score. Journal of the American Statistical Association, 79, 516–524.

Rutter, M. (1966). Children of Sick Parents: An Environmental and Psychiatric Study. London: Oxford University Press.

Vilagut, G., Forero, C. G., Barbaglia, G., & Alonso, J. (2016). Screening for depression in the general population with the Center for Epidemiologic Studies Depression (CES-D): a systematic review with meta-analysis. PLOS ONE, 11, e0155431.