Variable Selection for Optimal Treatment Regime

• Activity Number: 377
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Na Zhang*+ and Howard Bondell and Eric Laber
• Companies: North Carolina State University and NC State University and NC State University
• Keywords: optimal treatment regime ; variable selection ; treatment decision rule ; classification method ; patient's characteristics

Abstract

Personalized medicine aims to tailor treatments to the individual based on measured characteristics. Choosing among the set of possible tailoring variables can be a daunting task. The goal is to determine the subset of variables that will yield the decision rule that results in the optimal treatment regime over the population. Unlike typical variable selection problems, variables that may be highly related to the clinical outcome may not be relevant to distinguish among competing treatments. We propose a two-stage approach by first obtaining a flexible fit to the potential outcomes under each treatment regime to determine an estimate of the optimal treatment for each subject. In the second step, we then perform a sparse classification method to perform variable selection for this optimal decision rule. We show that this approach can not only identify the relevant tailoring variables, but, in the process, the reduction in dimension yields better accuracy in classification of subjects to the optimal treatment.