Observational Data for Heterogeneous Treatment Effects with Application to Recommender Systems

Abstract

Decision makers in health, public policy, technology, and social science are increasingly interested in going beyond ‘one-size-fits-all’ policies to personalized ones. Thus, they are faced with the problem of estimating heterogeneous causal effects. Unfortunately, estimating heterogeneous effects from randomized data requires large amounts of statistical power and while observational data is often available in much larger quantities the presence of unobserved confounders can make using estimates derived from it highly suspect. We show that under some assumptions estimated heterogeneous treatment effects from observational data can preserve the rank ordering of the true heterogeneous causal effects. Such an approach is useful when observational data is large, the set of features is high-dimensional, and our priors about feature importance are weak. We probe the
effectiveness of our approach in simulations and show a real-world example in a large-scale recommendations problem.

Keywords: Heterogeneous treatment effects, Machine learning, Online platforms, Personalization, Recommender systems, Social networks

JEL Classification: C5, C53, M31, C9