NOVEL PROPENSITY SCORE METHODS FOR MULTIPLE AND CONTINUOUS TREATMENTS: APPLICATIONS TO EHR DATA

DEREK W. BROWN, UTHealth School of Public Health

Abstract

Propensity scoring is often utilized to overcome the challenges posed by covariate imbalance to make causal inferences within observational studies. While methods for utilizing propensity scoring in a binary treatment case are well studied and established, generalizations to multiple unordered (multinomial) and continuous treatments are more complicated. In Aim 1, we developed and tested a novel multinomial treatment propensity score method, the GPS-CDF method, which derives a single scalar balancing score that can match and stratify subjects. Simulation results showed superior performance of the new methodology compared to standard multinomial propensity score methods. The proposed GPS-CDF method was also applied to an electronic health records study to determine the causal relationship between vasopressor choice and mortality in patients with non-traumatic aneurysmal subarachnoid hemorrhage (SAH). The GPS-CDF method indicated that phenylephrine may be the superior vasopressor choice for patients that present with non traumatic SAH. We further applied the GPS-CDF method to the Emergency Truncal Hemorrhage Control Study to determine whether emerging hemorrhage control interventions influence patient mortality. Based on the GPS-CDF method, patients receiving resuscitative

endovascular balloon occlusion of the aorta (REBOA) had similar morality as patients who received Laparotomy. In Aim 2, we extended the GPS-CDF method to the continuous treatment setting and further introduced the npGPS-CDF method. Both novel methods use empirical cumulative distribution functions (CDF) in order to stratify subjects based on pretreatment confounders to produce causal estimates. A detailed simulation study showed superiority of the novel methods based on the empirical CDF when compared to standard weighting techniques. The proposed methods were applied to the “Mexican American Tobacco use in Children” (MATCh) study and found a significant association between exposure to smoking imagery in movies and smoking initiation among Mexican-American adolescents. Finally in Aim 3, we developed an R package for researchers to implement the proposed GPS-CDF method in practice. Overall this research provides investigators with new options for implementing multinomial and continuous treatment propensity scoring.