Neuroevolutionary representations for learning heterogeneous treatment effects

Michael C. Burkhart, Gabriel Ruiz

Abstract: Highlights•In causal inference, the cate measures the expected net benefit of a potential intervention for persons with a given feature set.•Learned feature representations can improve the performance of black-box estimators for the cate.•Many approaches learn invariant representations that remove information related to treatment assignment.•We use a genetic algorithm to maintain information about treatment assignment if it is also useful for predicting outcome.•Our method can improve the performance of a variety of standard cate estimators.