Primary Submission Category: Heterogeneous Treatment Effects
LOOL: A Simple and Precise Estimator of Heterogeneous Treatment Effects
Authors: Duy Pham, Adam Sales,
Presenting Author: Duy Pham*
Given the ever-present and growing demand for personalized treatment and intervention in healthcare, education, policy, and many other domains, there have been significant developments in methodology to effectively and efficiently measure heterogeneous treatment effects over the last decade. We propose the Leave-One-Out Learner (LOOL) – a new meta-learner approach for the estimation of conditional average treatment effects (CATEs). Given a correctly specified experiment design, we can first obtain unbiased estimates of the individual treatment effects (ITEs) – without requiring any additional assumptions – by applying the Leave-One-Out Potential Outcomes (LOOP) Estimator given by Wu and Gagnon-Bartsch (2018). By regressing these estimates on the covariates in the treatment and control group, we further obtain the expected CATE functions under each condition. Like the X-Learner given by Künzel et al (2019), each observation’s CATE estimate is the sum of the corresponding values of these two conditional functions – weighted by the propensity score. However, ITEs from LOOP are unbiased, and regressing them on the covariates combats the relatively higher variance. Thus, LOOL can potentially obtain more accurate estimates. Compared to existing meta-learners, LOOL’s performance was highly competitive in a wide range of initial simulations.