Beyond Conditional Averages: Robust and Agnostic Learning of Conditional Distributional Treatment Effects
Miruna Oprescu, Cornell University
The conditional average treatment effect (CATE) is the difference in average outcomes between treatment and control units in a study, conditional on individual features. Estimating the CATE is a crucial task across many fields, helping policy and decision-makers take better actions. However, the CATE only captures the (conditional) average, and can overlook risks and tail events, which are important to treatment choice. In this work, we go beyond conditional averages to provide an algorithm for learning the conditional distributional treatment effect (CDTE) given by differences in (conditional) distributional measures between treatment and control groups. Examples of CDTEs include conditional quantile and super-quantile treatment effects, as well as treatment effects based on coherent risk measures given by f-divergences. Our algorithm is based on constructing a special pseudo-outcome and regressing it on features using any regression learner. Our method is robust and agnostic in that even if we learn the distributional measures nonparametrically, at very slow rates, we can still learn CDTEs at rates that depend on the CDTE class complexity and even conduct valid inference on linear projections of CDTEs. We investigate the performance of our method in simulations, and we demonstrate its use in a real-world case study of 401(k) eligibility effects on wealth.
Abstract Author(s): Miruna Oprescu, Nathan Kallus