Confounding Adjustment in Comparative Studies of Newly Marketed Medications

Abstract

Observational studies of newly marketed medications can add important safety and effectiveness information to what is known from pre-approval randomized controlled trials. As new medications are often prescribed to select group of patients, confounding can be particularly large in this setting. Multivariable modeling of the exposure or the outcome is commonly used to reduce confounding, but the modeling can be challenged by rapidly evolving prescribing patterns and the limited number of patients exposed to the new medication. Furthermore, while adjustment for empirically identified potential confounders, and proxies thereof, has been shown to reduce confounding, it is difficult to empirically identify potential confounders in settings with small number of outcomes. In this thesis, we explored two new methods to improve confounding control in the setting of newly marketed medications with few exposures and outcomes and many potential confounders. The first method is the high dimensional disease risk score (DRS) developed using an historical cohort of comparator drug initiators. We developed prediction models for the outcomes of interest in an historical cohort and applied the models to the concurrent study cohort of new and comparator drug initiators. We then used individual patient’s predicted risk score to balance the baseline risk between the two groups. In chapter 1, we compared combinations of shrinkage and dimension reduction approaches to reduce model over-fitting when developing DRSs from large numbers of potential covariates. In chapter 2, we compared the performance of the high dimensional DRSs to the standard high dimensional propensity score (hdPS) approach. In chapter 3, we developed a new method that augments the hdPS variable selection process with data from an historical cohort and compared this approach to standard hdPS. All evaluations were conducted in example comparative studies of newly marketed medications using large US claims database. The hdPS with variable selection augmented by historical data showed good performance in confounding adjustment even in small outcome settings. Future studies should evaluate the use of this method in other settings and should explore improvements in the use of high dimensional DRSs.