Comparative Effectiveness/Safety Research With Multiple Treatment Groups

Abstract

Thanks to the continued development of new medications in many therapeutic areas, patients and clinicians often are faced with the need to choose from multiple treatment options. All treatment decisions should ideally be informed with randomized controlled trials. However, randomized controlled trials with multiple active medications are rare. In the absence of trial evidence, comparative effectiveness/safety research utilizing observational data can play important roles. Although propensity score methods have become a standard tool in comparative effectiveness/safety research, they are less frequently used in questions involving three or more treatment options. This is in part due to the lack of familiarity and methods. In this dissertation, we extended several existing methods that had originally been developed in the two-group setting to the multi-group setting to overcome this. In Chapter 1, we extended the matching weights, an alternative propensity score weighting method, to the general multi-group setting. We showed its asymptotic equivalence to multi-group simultaneous propensity score matching and confirmed its similarity to three-way simultaneous matching in a simulation. In Chapter 2, we applied the multi-group matching weights method to an applied question on the bone safety of analgesics. The analysis based on the initial treatment assignment showed similar changes in bone mineral density although the on-treatment analysis suggested a potentially detrimental effect of opioids. In Chapter 3, we developed an empirical equipoise tool for the multi-group setting to address the question familiar to pharmacoepidemiologists: Are the treatment groups similar enough? We examined the settings in which the tool helped identify the danger of residual confounding due to dissimilar patient characteristics. In Chapter 4, we proposed extensions of three existing propensity trimming methods into the multi-group setting. We examined their ability to reduce confounding due to unmeasured variables more common in the tails of the multinomial propensity score distribution. In conclusion, we extended several existing propensity score methods to the multi-group setting. We hope these methods promote and improve comparative effectiveness/safety research with multiple treatment groups.