Conditional Average Treatment Effects and Decision Making

This paper develops a decision making empirical method to evaluate welfare programs accounting for heterogeneity of impacts. We find outcome predictive distributions for different subgroups of the population and use a characterization of second order stochastic dominance to give a policy recommendation conditional on covariates with minimal requirements on the social planner's utility function. Further, we can estimate quantile treatment effects within subgroups of the population. We apply this method to the Connecticut's Jobs First program and find subgroups for which the program did not maximize welfare even though some statistics may suggest the opposite and vice-versa.