Large-scale empirical identification of candidate comparators for pharmacoepidemiological studies

Background: The active comparator new user (ACNU) cohort design has emerged as a best practice for the estimation of drug effects from observational data. However, the process of identifying comparators and determining their adequacy can be challenging, particularly in the context of many potential comparators.

Objectives: We introduce an efficient empirical approach to rank candidate comparators in terms of their similarity to a target drug in high-dimensional covariate space.

Methods: We generated new user cohorts for each RxNorm ingredient observed in five administrative claims databases mapped to the OMOP Common Data Model: Merative™ MarketScan®’s Commercial Claims and Encounters, Multi-State Medicaid, and Medicare Supplemental databases; Japan Medical Data Center; and Optum© De-Identified Clinformatics© Data Mart. We then extracted aggregated data on thousands of pre-treatment covariates for each cohort across five clinically oriented domains related to historical condition occurrence, medication use, healthcare utilization, and demographics. We formed all pairs of cohorts with ≥ 1,000 patients and computed a scalar similarity score, termed the cohort similarity score (CSS), defined as the average of cosine similarities computed within each covariate domain, for each pair. Ranked lists of candidate comparators were then generated for each cohort. We present focused results for six example drugs (warfarin, apixaban, glatiramer acetate, ocrelizumab, methotrexate, and adalimumab) and provide an interactive application for exploration of the full result set at data.ohdsi.org/ComparatorSelectionExplorer/.

Results: Across up to 1,350 cohorts forming 922,761 comparisons observed across data sources, drugs that were more closely related in the ATC hierarchy tended to have higher cohort similarity scores. The most similar candidate comparators for each of six example drugs consistently corresponded to alternative treatments for the target drug’s indication(s), as identified in literature or publicly registered studies. For example, the top five comparators for ocrelizumab by average rank across data sources were: teriflunomide (avg. CSS: 0.980), dimethyl fumarate (0.969), natalizumab (0.967), fingolimod (0.965) and dalfampridine (0.967). Across data sources, 80%-95% of cohorts had at least one comparator with a cohort similarity score ≥ 0.95.

Conclusions: Empirical comparator recommendations may serve as a useful aid to investigators and could ultimately enable the automated generation of evidence from ACNU designs, a process that has previously been limited to self-controlled designs and temporal scans.