Abstract:
|
Propensity scoring methodology is commonly used when estimating causal effects of treatment in non-randomized observational studies. Historically, observational studies have focused on comparing binary treatments. However, recent developments in theory and application have allowed for accurate causal inference in studies with more than two treatments. These treatments can be multinomial, ordinal, or even continuous. For ordinal dosing schemes, the only technique established to conduct propensity score analysis involves the assumption of proportional odds for the treatment and matching based on the linear predictor given by McCullagh's ordinal logit model. If this assumption is violated, using the linear predictor could provide an inaccurate propensity score and biased causal effect. This analysis conducts a simulation study under various violations of the proportional odds assumption to investigate the implications of propensity score model misspecification for ordinal doses. Potential methods to address this violation are presented and applied to a relevant data set.
|