Robust Estimation of Causal Effects for Comparative Effectiveness Research
*Linbo Wang, University of Washington
Keywords: Comparative effectiveness research; Inverse probability weighting; Propensity score; Robust estimation; Subclassification
The propensity score plays a central role in addressing comparative effectiveness questions with observational data. In particular, weighting and stratification methods based on propensity scores are widely used to adjust for observed covariates. Weighting methods, such as inverse probability weighting (IPW)--based methods, are appealing theoretically as they can be used to produce consistent estimates of treatment effect estimates. However, in practice, these estimates are often highly unstable and very sensitive to misspecification of propensity score models. On the other hand, subclassification methods are widely used in practice to produce stable and robust estimates, but the estimates are not consistent due to residual confounding. In this presentation, we first reveal the intrinsic connection between weighting methods and subclassification methods, and then propose a novel subclassified weighting method that is 1) asymptotically unbiased and consistent so it is theoretically appealing and 2) more stable and robust to model misspecification than subclassification methods so it is practically useful. We illustrate our method with a real-world observational data set.