Estimation and Inference for Individualized Treatment Rules Using Efficient Augmentation and Relaxation Learning

Individualized treatment rules aim to identify the best strategy—if, when, which, and to whom treatment should be applied, to optimize the health outcomes. One type of approaches to learn such rules estimates individualized treatment rules through a weighted classification framework. Particularly, in this chapter, we introduce doubly-robust learning methods called efficient augmentation and relaxation learning (EARL) and its extensions to high-dimensional data such as genetic profiles. These methods can consistently estimate the optimal individualized treatment rule as long as model specifications of either a propensity or outcome models are correct. More importantly, we also provide an inferential tool for EARL and its extensions under a linear decision rule. These inferential tools provide p-values and confidence intervals to characterize how predictors affect the best treatment decision. With these inferential tools, EARL and its extension can inform the practical use of the treatment and facilitate scientific discovery.