Model Selection and Combination for Estimating Treatment Effects

• Activity Number: 82
• Type: Contributed
• Date/Time: Sunday, July 29, 2012 : 4:00 PM to 5:50 PM
• Sponsor: Health Policy Statistics Section
• Author(s): Craig Rolling*+ and Yuhong Yang
• Companies: University of Minnesota and University of Minnesota
• Keywords: causal effects ; model selection ; model averaging ; focused information criterion ; cross-validation ; personalized medicine

Abstract

In many applications, it is believed that the effect of a treatment on a response varies as a function of certain baseline covariates. Several methods for estimating conditional treatment effects have recently been proposed; however, little attention has been given to the problem of choosing between estimators of conditional treatment effects. In general, given a set of candidate models, the models that best estimate the conditional mean of the response may not be best for estimating the effect of a treatment. Therefore, traditional model selection methods such as AIC and cross-validation may be unsuitable for the goal of treatment effect estimation. We demonstrate an application of the Focused Information Criterion (FIC) in this setting and propose nearest neighbor-based methods for model selection and combination aimed specifically at minimizing errors of treatment effect estimation.