Propensity Score Modeling to Reduce Channeling Bias When Exposure Is Rare
Background: Channeling bias in post-market surveillance of treatments introduces challenges in contextualizing rates of adverse events, particularly with rare treatments indicated for a common disease or when treatment is first-in-class. New methodology is needed to quantify the baseline risks of adverse events in the indicated population.
Objectives: To generate a counterfactual group consisting of patients who were clinically-similar to the treatment initiators, but did not receive treatment.
Methods: Women with endometriosis (EM) were identified within the IBM MarketScan insurance claims database between 01Jul18 and 01Jul2019 and followed until the earliest of disenrollment, end of data (30Sep2020), death, exposure termination, or outcome. Treatment exposure was elagolix, a product to treat EM-associated pain. Univariable regression models quantified the association between demographics, comorbidities, medication and symptoms and initiation of elagolix. Variables with a p-value<0.20 were included in the propensity score (PS). Women with EM unexposed to elagolix were matched 1:1 with a 1% caliper to exposed patients, to generate an unexposed counterfactual group. Kernel density plots of and standardized mean differences between the pre- and post-matched populations assessed PS distribution and matching performance. Chi-square tests of independence and Fisher's exact test compared baseline characteristics between the exposed and unexposed counterfactual group using the Aetion Evidence Platform.
Results: During the study, 136,027 women with EM were identified; 1,207 (0.9%%) had a prescription claim for elagolix. Among 121 baseline covariates considered, 83 were included in the PS model. Due to the low frequency of elagolix, the PS were low (mean 0.04 in the exposed group); yet, matching was robust due to a large referent group size and wide range of baseline covariate distribution. There were no significant differences in baseline covariate distribution between the two groups after matching. The exposure and counterfactual groups demonstrated significant heterogeneity from the general EM population.
Conclusions: Propensity score matching generated a counterfactual drug group that may be a suitable comparator to quantify baseline adverse event rates in the indicated population for post-market surveillance. These methods can inform the benefit-risk profile of the treated population and can account for differences in comorbidities and disease severity in patients with the same disease who initiate treatment compared to those who do not.