Nonparametric Variable Selection for Predictive Models and Subpopulations in Clinical Trials.

Most clinical trials have heterogeneous treatment effect among patient individuals. It is desirable to identify a patient subpopulation, which has a stronger treatment effect than the rest of patients, so that researchers will be able to determine who will benefit the most or the least from the treatment and design treatment strategies accordingly. This paper develops a nonparametric method for predicting clinical response and identifying subpopulations. The method first selects predictors using kernel-based local regression and a forward procedure via F-tests. It then defines subpopulations with enhanced treatment effects based on the selected predictors and the nonparametric model of the clinical response. Simulation examples and a pharmacogenomics study of bortezomib in multiple myeloma demonstrate the proposed method and show favorable performances compared to other existing methods. The proposed method provides an alternative way to define subpopulations and is not limited by parametric models and their possible misspecification for the clinical response.