Subgroup Analysis Using Doubly Robust Semiparametric Procedures

Subgroup analysis is commonly performed in clinical trials and is an important means for precision medicine as the treatment improvement could well be different between two subgroups of patients. One goal is then to identify if there exists such a subgroup in the patient population. Since the treatment assignment in the study population may not be covariate balanced, and the naive estimate of treatment effect such as a group mean may be biased. To infer the true effect of a new treatment against a current one, and classify patients into the new treatment favorable subgroup if it exists, causal inference methods are needed. The doubly robust (DR) procedure for causal inference represents an important methodological advancement, where a propensity score model and a regression model need to be specified. As long as one of the two models is specified correctly, the DR causal estimate of the treatment will be unbiased asymptotically. In practice, DRE may fail and have large bias and variance, even when the propensity and/or outcome models only depart slightly from the true model. Recently we developed an enhanced DRE by utilizing more flexible semiparametric propensity and regression models and applied it to subgroup analysis. The key idea is to enhance and construct the DRE to achieve robustness and numerical stability, and then identify treatment favorable subgroup if it exists. In this chapter, we will explore an alternative construct of the DRE after a brief review of the DRE and enhanced DRE. Then we will use the enhanced doubly robust procedure in subgroups analysis, which is to identify subgroup if there exists groups of patients who benefit significantly more from the treatment over the control, and compare it with the previously developed enhanced procedure and the naive estimator in both simulation and in the analysis of a smoking cessation study. We found that the two enhanced DR estimators are both robust and overall give consistent results.