Utilizing Propensity Score Methods for Ordinal Treatments and Prehospital Trauma Studies
Propensity score stratification, matching, and weighting methodologies are commonly used when estimating causal effects of treatment in observational studies. While PS methods for binary treatments are well established, methods for studies with more than two treatments (e.g. multinomial, ordinal, or even continuous) are less wide-spread. Specifically, the most popular PS method for studies with ordinal treatments requires the proportional odds (PO) assumption to be satisfied. If this assumption is violated, as is often the case in real data, the estimated propensity scores may be incorrect, and the estimation of average treatment effect may be biased. In Aim 1, we investigated the implications of varying levels of proportional odds violations through simulation, and used the most robust methods to analyze a relevant trauma dataset of patients who underwent an emergency laparotomy. We found that treatment effect estimates became more biased, as the severity of the PO assumption increased. Furthermore non-parametric methods, and parametric methods which used the balancing score from the PO model more coarsely (such as stratification), performed the best. In Aim 2, we developed the novel GPS-CDF propensity score method for propensity score stratification and matching with ordinal treatment. This method does not rely on the proportional odds assumption and is flexible enough to be used with any parametric or non-parametric propensity model. When applied to the Mexican-American Tobacco Use in Children (MATCh) study, we found a significant treatment effect between increased exposure to smoking imagery in movies and cigarette experimentation among previously smoking-naive adolescents. Finally, in Aim 3 we conducted a systematic review of current propensity score literature for prehospital blood transfusion (PHT) studies and present recommendations for best practice. In this review, we uncovered numerous issues in current practice and made clear recommendations for future researchers who wish to study the effect of PHT on mortality using propensity score methodology.
Greene, Thomas J, "Utilizing Propensity Score Methods for Ordinal Treatments and Prehospital Trauma Studies" (2017). Texas Medical Center Dissertations (via ProQuest). AAI10681743.