Oracle, Robust, and Best Linear Model Average Estimation in Randomized Trial

• Activity Number: 471
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Kwun Chuen Chan*+
• Companies: University of Washington
• Keywords: Calibration ; Empirical likelihood ; Moment matching ; Weighted estimation

Abstract

Estimation of average treatment effect in a randomized clinical trial can be done by simple two sample analysis, without the need for extensive parametric modeling. When pretreatment variables are collected, regression based estimate can lead to improvement in efficiency but may not be robust against model misspecification. In this talk, I will talk about semiparametric methods which use regression models to improve efficiency of estimation of average treatment effect without losing validity when model assumptions are incorrect. We will look at methods that can assume multiple non-nested working models instead of a single model considered in the literature. The methodology is shown to yield consistent estimate of average treatment effect even when all working models are wrong, and automatically attains the best asymptotic efficiency of any linear combination of working models. Moreover, when any one of the working models is equivalent to the true model, the estimator attains the same asymptotic efficiency as if the true model is known in advance.