In this paper, we implement a Bayesian potential outcomes model to evaluate the impact of program interventions using non-randomized data. The approach jointly addresses selection bias in program placement, heterogeneous treatment intensity among the treated, and heterogeneity in treatment effects. Using data from a non-randomized household survey, we evaluate the impact of Ethiopia’s Health Extension Program on fertility and child mortality outcomes. We find that there is significant selection bias in both program placement and intensity of exposure to the program among the treated. On average, the program has significant impact on reducing fertility and child mortality. However, there is notable heterogeneity in the treatment effects ranging from negative impacts for some individuals to positive impacts for the majority in the sample. We recover individual-level treatment effects and present the distributions graphically.
Supplementary notes can be added here, including code and math.