Propensity Score Approach to Treatment Effect Evaluation

Propensity score has been increasingly used to control for confounding in observational studies. The inverse probability weighting (IPW) is widely used to estimate the average treatment effect (ATE). The excessively large IPW weights caused by lack of overlap in the propensity score distributions between groups can cause numeric instability in estimation. In Chapter 2, we studied a class of modified IPW estimators that can be used to avoid this problem. These weights cause the estimand to deviate from ATE. We provided justification for the deviation from the perspective of treatment effect discovery. We showed that in the presence of lack of overlap, the modified IPWs may achieve substantial gain in statistical power compared with IPW and other relevant propensity score methods. We developed analytical variance estimation that properly adjusts for the sampling variability of estimated propensity scores, and augment the modified IPW estimator with outcome models for improved efficiency. The proposed methodology was illustrated with a blood cell transfusion study. We provided open source software to implement the proposed methodology. The modified IPWs in Chapter 2 has not been studied in the context of survival data, and we investigated this problem in Chapter 3. We considered estimands of average survival time, restricted average survival time, survival probability, survival quantile effect, and the marginal hazard ratio in the two-sample context. We proposed a unified analytic framework to obtain the point and variance estimators. Simulations showed that the point and variance estimators possess desired finite sample properties, and demonstrate better numerical performance than some existing weighting and matching methods commonly used in the literature. The proposed methodology was illustrated with data from a breast cancer study. Like weighting methods studied in previous two chapters, regression adjustment is frequently used for treatment effect estimation and it can be highly efficient when the relationship between outcome and propensity score is correctly specified, but can lead to biased estimation when model assumption is violated. To address this issue, we proposed a flexible regression approach that allows for nonlinear association between outcome and propensity score in Chapter 4. The flexible approach has connections to other commonly used propensity score methods. We developed a penalized spline based method to obtain point and variance estimators that properly adjust for the estimated propensity scores. Simulations showed that the proposed approach demonstrates satisfactory numerical properties, and are at least comparable with other propensity score methods. The proposed methodology was further illustrated with a right heart catheterization study. In summary, this dissertation added to current propensity score literature by considering modified IPW weights for treatment effect discovery and in studies with survival outcome, addressing limitations in regression adjustment, and providing software for convenient implementation. These advances would promote its application in public health research where observational data is commonly encountered for intervention effect evaluation and subsequent decision making