Causal machine learning for understanding heterogeneous effects of childhood obesity prevention

Abstract Community-based interventions (CBIs) are widely used to prevent childhood obesity, yet evaluating their causal effects is challenging due to multi-level interactions and variation across individuals. Standard statistical approaches often have difficulty addressing these complexities, which can produce biased estimates. To address this, we propose a causal machine learning framework that integrates causal discovery, causal inference, and tree-based models to estimate individual treatment effects and examine heterogeneity in program outcomes. Using data from seven CBI programs involving more than 7000 participants, we first applied causal discovery to identify relationships between children’s demographic and behavioral characteristics, finding links between body mass index (BMI), age, sex, and dietary behaviors such as takeaway food consumption. Individual treatment effect estimation suggested notable differences in intervention outcomes across participants, with age emerging as an important moderator of effectiveness. Tree-based models further highlighted subpopulation characteristics associated with these variations. Our findings suggest that CBI effectiveness is not uniform across children, and that identifying sources of heterogeneity can support the design of more targeted and age-sensitive prevention strategies. While demonstrated here in the context of childhood obesity, the framework provides a generalizable approach for analyzing heterogeneous treatment effects in complex, community-based interventions.