Feature Selection for Discovering Distributional Treatment Effect Modifiers

Yoichi Chikahara, Makoto Yamada, Hisashi Kashima

Keywords: heterogeneous treatment effects, potential outcomes, causal inference

TL;DR: As an approach to elucidating why treatment effect heterogeneity exists, we propose a feature selection framework for discovering the features relevant to the distributional difference in treatment effects.

Abstract: Finding the features relevant to the difference in treatment effects is essential to unveil underlying causal mechanisms. Existing methods seek such features by measuring how greatly the feature attributes affect the degree of the {\it conditional average treatment effect} (CATE). However, these methods may overlook important features because CATE, a measure of an average treatment effect, cannot detect the difference of other distribution parameters than the mean (e.g., variance). In this paper, we propose a feature selection framework for discovering {\it distributional treatment effect modifiers}. To resolve the weakness of the existing methods, we formulate a feature importance measure that quantifies how strongly the feature attributes influence the discrepancy between potential outcome distributions. We derive its computationally efficient estimator and develop a feature selection algorithm that can control the type I error rate at some desired level. Experimental results show that our framework successfully discovers important features and outperforms the existing mean-based method.