Model Selection for Causal Parameters in Structural Mean Models Based on a Quasi-Likelihood

• Activity Number: 360
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Masataka Taguri*+ and Yutaka Matsuyama and Yasuo Ohashi
• Companies: The University of Tokyo and The University of Tokyo and The University of Tokyo
• Keywords: Akaike's information criterion ; Causal inference ; G-estimation ; Model selection ; Noncompliance ; Structural mean model

Abstract

The structural mean models (SMMs) are proposed for estimating causal parameters in the presence of time-dependent confounders. In addition to the fundamental assumption of the unmeasured confounders, correct specification of the structural model is necessary for obtaining consistent estimates of the causal parameters. There are a few studies to cope with this issue. In this talk, we propose a new approach for model selection of SMMs based on a quasi-likelihood. Our quasi-likelihood is the objective function of certain type of Robins' G-estimators. We provide a likelihood-ratio-type test and a formal model selection criterion which is an extension of Akaike's information criterion. We investigate the finite sample performance of the model selection through simulation studies. The method is illustrated using the analysis of a real clinical trial data with non-random non-compliance.