A Policy Search Method for Estimating Treatment Policies

• Activity Number: 232
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Xi Lu*+ and Susan A. Murphy
• Companies: University of Michigan and University of Michigan
• Keywords: dynamic treatment regime

Abstract

A treatment policy or dynamic treatment regime is a sequence of decision rules. At each stage a decision rule inputs patient history and outputs a treatment. The value of a treatment policy is the expected outcome when the policy is used to assign treatment. Data from sequential, multiple assignment, randomized trials can be used to estimate an optimal treatment policy. One approach is to parameterize the policy value (Robins et al. 2008); this may result in bias if the model is misspecified. Alternately the value of any specific policy can be estimated nonparametrically; however this method may have high variance. We propose a new method in which each stage's treatment effect or "blip" is parameterized. These treatment effects are easily interpretable to scientists and thus more meaningfully parameterized than the policy value. To estimate the parameters we utilize a telescoping sum representation of the policy value and uses ideas from missing data theory. We illustrate the proposed method with data from the ExTENd trial, a recently completed alcohol dependence study.