A robust covariate-balancing method for learning optimal individualized treatment regimes

One of the most important problems in precision medicine is to find the optimal individualized treatment rule, which is designed to recommend treatment decisions and maximize overall clinical benefit to patients based on their individual characteristics. Typically, the expected clinical outcome is required to be estimated first, for which an outcome regression model or a propensity score model usually needs to be assumed with most existing statistical methods. However, if either model assumption is invalid, the estimated treatment regime will not be reliable. In this article, we first define a contrast value function, which forms the basis for the study of individualized treatment regimes. Then we construct a hybrid estimator of the contrast value function by combining two types of estimation methods. We further propose a robust covariate-balancing estimator of the contrast value function by combining the inverse probability weighted method and matching method, which is based on the covariate balancing propensity score proposed by Imai & Ratkovic (2014). Theoretical results show that the proposed estimator is doubly robust, ie, it is consistent if either the propensity score model or the matching is correct. Based on a large number of simulation studies, we demonstrate that the proposed estimator outperforms existing methods. Application of the proposed method is illustrated through analysis of the SUPPORT study.