Improved Outcome Weighted Learning for Dynamic Treatment Regimes

• Activity Number: 473
• Type: Invited
• Date/Time: Wednesday, August 6, 2014 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Donglin Zeng*+ and Yuanjia Wang and Ying Liu and Michael Kosorok
• Companies: University of North Carolina at Chapel Hill and Columbia University and Columbia University and University of North Carolina at Chapel Hill
• Keywords: Dynamic treatment regimes ; outcome weighted learning ; robustness ; Fisher consistency ; support vector machines

Abstract

Outcome weighted learning (O-learning) has been proposed to estimate personalized treatment regimes in single-stage or multiple-stage SMART (Zhao et al. 2012, 2013). This learning has been demonstrated to be superior to Q-learning in small sample studies. However, O-learning may be sensitive to outcomes of large variability and only relies on subjects who follow optimal treatment regimes in future stages. The latter will particularly result in efficiency loss in estimation. In this work, we propose an improved weighted learning. The new approach can handle non-positive weights and highly variable outcomes. More importantly, the new approach will make use of all data but still be able to learn optimal regimes consistently without specifying any correct models. The numerical performance demonstrates these advantages of the new approach. We will use one real study from mental health to illustrate this approach.