On Estimation of Nuisance Working Models in Doubly Robust Estimators

• Activity Number: 658
• Type: Contributed
• Date/Time: Thursday, August 2, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Karel Vermeulen*+ and Stijn Vansteelandt
• Companies: and Ghent University
• Keywords: Causal inference ; Double robustness ; Missing data ; Outcome regression ; Propensity score ; Semiparametric inference

Abstract

We consider estimation of the population mean outcome in studies with ignorable missingness, given completely observed covariates. Doubly robust estimators are asymptotically unbiased when either a working model for the mean outcome given those covariates, or a working model for the propensity of missingness given those covariates is correctly specified, but can be severely biased under misspecification of both working models. This is in particular the case when missingness is strongly dependent on covariates implying that typical doubly robust estimators must cope with highly variable inverse probability weights. In this talk, we will propose simple strategies for estimating the nuisance parameters indexing both working models, thereby targeting minimal asymptotic variance and robustness of the doubly robust estimator. Our work extends that of Cao, Tsiatis and Davidian (2009) who focus on estimation of the outcome regression model, by additionally considering estimation of the missingness model. Our approach guarantees stability of the weights and estimators with simple asymptotic distributions. Simulation studies will be shown to assess finite-sample performance.