CMStatistics 2016: Start Registration
View Submission - CMStatistics
B1451
Title: The covariate balancing generalized propensity score for continuous treatment regimes in the presence of censoring Authors:  Samantha Noreen - Emory University (United States) [presenting]
Qi Long - University of Pennsylvania (United States)
Abstract: The propensity score is widely used for causal inference in observational studies. The covariate balancing propensity score (CBPS) methodology for binary treatment assignments has been proposed to address potential misspecification of the propensity score by exploiting the covariate balancing property of the propensity score. Extending the CBPS, the covariate balancing generalized propensity score (CBGPS) considers general treatment regimes. While the CBGPS has several appealing features, our preliminary numerical studies showed that the CBGPS tends to be numerically unstable. We investigated refinements to the CBGPS approach (iCBGPS), which demonstrated superior performance over the CBGPS in our empirical studies. Extending this approach further, our goal was to develop the iCBGPS in the presence of censoring. Specifically, many observational studies include information on patients censored by death or dropout, and standard propensity score methods including the CBGPS use a complete-case analysis. The iCBGPS methodology in the presence of censoring takes advantage of this extra information previously unused in such analyses. In our subsequent empirical studies, the inclusion of censoring information improved performance over the iCBGPS, as well as standard generalized propensity score and CBGPS methods, in the absence of censoring.