Leveraging Artificial Intelligence and Field Experiments to Explore Novel Features of Parental Speech and Foster Child Development

Parents play a critical role in shaping children's skills during the first years of life. Yet, identifying the contributors to richer learning environments remains difficult due to various unobservable factors. In this paper, we combine field experiments with AI to explore new acoustic features of parental speech. Specifically, we develop a signal processing model that uses more than 600 hours of recorded parent-child interactions combined with assessment data from two home-visiting experiments conducted in the Chicagoland area to identify features of parental speech that map into children's skills. Our two experiments consist of the same intervention helping parents provide nurturing interactions to their child. We exploit the experimental and natural variation in our data to explore two causal channels and one potential moderator. First, our intervention improves parental speech consistently across the two studies, as measured by acoustic features that are predictive of higher socioemotional skills and adult-child conversational turns. Further, we find that it also increases children's language skills in both experiments, as well as socioemotional skills in the second experiment. Interestingly, our heterogeneity analyses reveal that some of the interventions' impacts vary by socioeconomic groups, with patterns across the two experiments suggesting that the mechanisms are context-dependent.