Summary Score-Based Confounding Control in Studies of New Drugs

Abstract

Epidemiological studies of newly marketed drugs are often undertaken to assess questions of safety and effectiveness that did not arise during a drug’s pre-approval studies. However, unlike in randomized controlled trials, such observational studies are prone to confounding., which can bias estimates of a drug’s effect on the study outcome. In this dissertation, we explore three challenges to confounding control in the study of new drugs. In Chapter 1 we propose a family of privacy-preserving, propensity scored-based methods that allow for adjustment for many confounders when not all confounders are physically stored in the same location. We evaluate the performance of the proposed methods in five empirical studies and in simulation. In Chapter 2 we consider the use of historical data for the estimation of disease risk scores and attempt to determine the impact of different follow-up approaches on achieved confounding control. Finally, in Chapter 3 we address the conditions under which adjustment for calendar time will aid in confounding control in studies of new drugs, and assess the impact of such adjustment in four previously published empirical comparisons of new drugs vs. their established comparators. In 203 (75%) of the 270 empirical scenarios considered in Chapter 1, the proposed methods for the analysis of vertically distributed data had a bias of less than 5% on the relative risk scale. Plasmode simulation confirmed that these methods performed nearly as well as would be possible if data pooling was allowed. In Chapter 2 we found evidence that, in an empirical comparison of the Dabigatran vs. Warfarin for the risk of major bleeding events, developing disease risk scores under an “as-treated” follow-up principle often led to overfitting and consequentially reduced confounding control. Additional work is needed to determine if developing disease risk scores under an “intention-to-treat” principle can alleviate this issue in a diverse set of examples. Finally, in Chapter 3 we found that, across four previously-published drug comparisons, adjustment for calendar time did not improve confounder balance or lead to appreciably different inferences regarding drug effects. Further research is needed to determine if calendar time adjustment may amplify unmeasured bias or reduce precision in other settings.