Bias-Reduced (BR) Doubly Robust (DR) Estimation

• Activity Number: 67
• Type: Topic Contributed
• Date/Time: Sunday, August 9, 2015 : 4:00 PM to 5:50 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Karel Vermeulen* and Stijn Vansteelandt
• Companies: Ghent University and Ghent University
• Keywords: Bias ; Causal inference ; Double robustness ; Missing data ; Nuisance parameters

Abstract

DR estimators are asymptotically unbiased when one of two working models is correctly specified. While theoretically appealing, DR estimators have been the subject of recent debate because model misspecification is likely to affect all working models. Moreover, their performance can be sensitive to the choice of estimators for the working models under double misspecification. In this talk, I will show that, interestingly, some DR estimators partially retain their robustness properties even under double misspecification. We propose a simple and generic estimation principle for the nuisance parameters indexing each of the working models, which we call BRDR estimation (Vermeulen and Vansteelandt, 2015). It is designed to improve the performance of the DR estimator of interest, relative to the default use of MLE, by locally minimizing the squared first-order asymptotic bias of the DR estimator under double misspecification. I discuss the basic proposal based on parametric models for the nuisance models, as well as extensions that employ machine learning algorithms. Simulation studies confirm the desirable finite-sample performance of the proposed estimators relative to other proposals.