Outcome-Weighted Learning for Dynamic Treatment Regime Selection

• Activity Number: 155
• Type: Topic Contributed
• Date/Time: Monday, July 30, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Yingqi Zhao*+ and Eric B Laber and Donglin Zeng and Michael Kosorok
• Companies: The University of North Carolina at Chapel Hill and North Carolina State University and The University of North Carolina at Chapel Hill and The University of North Carolina at Chapel Hill
• Keywords: Dynamic Treatment Regimes ; Outcome Weighted Learning

Abstract

Traditional approaches for selecting dynamic treatment regime using data from sequentially multiple randomization trials include Q-learning and A-learning. All these approaches rely on finding the statistical relationship between rewards or regrets given treatment and other predictors, then obtain the optimal treatment assignment by inverting this relation. However, in the presence of many covariates, such methods are likely to overfit models so may lead to suboptimal treatment policy. In this work, we propose a completely new approach by directly maximizing value function associated with each treatment policy. For one time-point decision, we transform this maximization into an outcome-weighted learning and apply a weighted support vector machine for estimation. With multiple state decision, we propose two methods based on outcome weighted learning for estimation. Our procedure can be carried out easily existing software and is shown to lean to optimal treatment decision rule. Finally, we illustrate our approach using a number of numerical studies.