Myra Samuels Memorial Lecture
More Robust Doubly Robust Estimation
Marie Davidian
Department of Statistics, North Carolina State University
Start Date and Time: Fri, 20 Apr 2012, 10:30 AM
End Date and Time: Fri, 20 Apr 2012, 11:30 AM
Venue: SC 239
Abstract:
Considerable recent interest has focused on so-called doubly robust estimators for parameters in a model for full, intended data when some data may be missing. In the simplest case of where the full data consist of an outcome Y and covariates X and interest focuses on the population mean of Y, these estimators involve models for both the propensity score, the probability that Y is observed given X, and the regression of outcome on covariates. These estimators have the appealing property that they are consistent for the true population mean even if one of the outcome regression or propensity score models, but not both, is misspecified. However, despite this appealing property, the "usual" doubly robust estimator may yield severely biased inferences if neither of these models is correctly specified and can exhibit nonnegligible bias if the estimated propensity score is close to zero for some observations, and hence has been criticized as not suitable for practical use. We review doubly robust estimation and propose alternative doubly robust estimators that achieve comparable or improved performance relative to competing methods, which we motivate in this simple setting. We then describe how the approach may be extended to more complicated situations, including causal inference on a treatment effect from observational data and to analysis of a longitudinal study with dropout.