Empirical Comparison of Exposure Set Definitions in the Prevalent New‐User Design

ABSTRACT Purpose Prevalent new‐user (PNU) designs aim to provide a wider assessment of treatment effects by incorporating users of a newer study drug who previously received the comparator. Similarity in terms of prior use of the comparator is accounted for via exposure sets based on time in study, prior prescription number, or a hybrid incorporating both plus calendar time. Given a current lack of consensus, we examine choice of exposure set definition using a study investigating upper gastrointestinal bleeding (UGIB) risk between users of non‐steroidal anti‐inflammatory drugs (NSAIDs) and cyclooxygenase‐2 inhibitors (COX‐2is). Methods We identified a cohort of individuals with osteoarthritis initiating NSAIDs or COX‐2is between 2000 and 2004 from the UK Clinical Practice Research Datalink. Considering prescription‐based, time‐based, and hybrid exposure set definitions, we estimated time‐conditional propensity scores (TCPS) using conditional logistic regression and matched COX‐2i users 1:1 to NSAID users. Analyses estimated the hazard ratio (HR) of UGIB bleed comparing COX‐2i and NSAID users overall, in the incident new‐user and prevalent new‐user subgroups. Results We identified 100 185 individuals who received a prescription for either COX‐2is or NSAIDs; 25 742 patients were incident new‐users of COX‐2is and 17 952 were PNUs. Prescription‐ and time‐based exposure sets resulted in the highest proportion of COX‐2i users successfully matched (86% vs. 76% using hybrid definition). We observed variability in the point estimates obtained under the different exposure set definitions; however, conclusions remained consistent. Conclusions Given the potential for differences in the matched cohorts and substantive results obtained under different exposure set definitions, we encourage increased use of sensitivity analyses in PNU studies to explore the robustness of results to this decision.