Socioeconomic disparities in cognitive and language skills develop in the first years of life, jeopardizing long-term educational opportunities and life chances of low-income children (Duncan & Murnane, 2011; Hart & Risley, 1995). This growing income-achievement gap may result from increasingly focused investment in children’s cognitive development, among higher-SES parents (Reardon, 2011). To counter harmful consequences of poverty on children’s development, Head Start (HS), as part of its theory of change, aims to improve parenting practices that contribute to children’s early cognitive and language development. Indeed, evidence suggests that low-income children acquire larger vocabularies more rapidly when their parents talk with them and engage them in learning activities (Fernald & Weisleder, 2013). In examining this theory of change of Head Start, we investigate whether the randomized offer to enroll in Head Start improved the frequency of parent-child literacy activities, and if that experimentally-induced improvement in parent-child literacy interaction contributed to program impact on children’s vocabulary and early reading skills during the first follow-up year.

We extend prior work in several ways. First, while Head Start also includes services that promote positive parenting practices, the Head Start Impact Study (HSIS) did not report the program impact on overall parenting behaviors that characterize cognitive stimulation and are theorized to promote early language development. In estimating the causal effect of HS on parent-child language-and-literacy activities, we constructed and included a psychometrically-sound measure of parent-child language-and-literacy activities for each child built from parent responses to six indicators in the HSIS. Additionally, HSIS did not investigate explicitly whether changes in parenting practices could be a key source of variation in explaining program impacts on children’s learning. Therefore, we unpacked the mechanism behind the average treatment effects on children’s early literacy detected in the HSIS. Third, we applied the principal stratification framework and Bayesian estimation techniques to explore the question of causal mediation.  

            We estimated program impact on parent-child literacy activities by using a multi-level model specification, with random slopes with fixed effects for center groups. Then we used a principal stratification framework with a hierarchical random effects model and Bayesian estimation strategies to stratify children into latent subgroups by the joint distribution of each child’s intermediate values: (1) the observed frequency of parent-child literacy activities under the observed experimental condition, and (2) the unobserved frequency of parent-child literacy activities they would have experienced under the counterfactual experimental condition. By examining these treatment-control differences in early language outcomes for those children whose parents would report high parent-child literacy activities, but who otherwise would provide low parent-child literacy activities, we were able to estimate the causal impact of Head Start for the key group that represents the mediating mechanism (those whose literacy activities increased as a result of being assigned to the treatment condition). Our preliminary findings suggest that the randomized Head Start offer has larger positive impacts on parent’s cognitive stimulation of their children, and operates as an important causal mechanism through which HS improves early language development.