Extending randomized trial estimates to different target populations: An application to a nested trial comparing anticoagulant regimens for percutaneous coronary intervention

Background: Two key reasons populations in randomized trials differ from target populations of interest are: 1) not all eligible individuals are invited, or consent, to enroll in trials, and 2) trials often have stringent eligibility criteria, excluding individuals who are candidates for treatment. These key reasons coincide with two tasks when extending inferences from trials: 1) generalizability, when the target population is limited to individuals who were eligible for the trial, and 2) transportability, when the target population includes individuals who do not satisfy trial eligibility criteria.

Objectives: To extend inferences from VALIDATE, a randomized trial nested in the SWEDEHEART registry comparing bivalirudin to heparin for percutaneous coronary intervention (PCI), to two target populations: 1) generalizing to the “trial-eligible” population, consisting of all patients eligible for the trial, including those who were not invited or did not enroll and 2) transporting to the “treatment-eligible” population, consisting of all patients undergoing PCI who are candidates for both treatments.

Methods: We identified both target populations using SWEDEHEART, which collected data on all PCI procedures in Sweden during the VALIDATE study period. The “trial-eligible” population is identified by applying the eligibility criteria from VALIDATE to the registry data (with exceptions such as Swedish language requirement and ability to give consent), and the “treatment-eligible” population relaxes eligibility criteria to exclude only individuals with contraindications. We then estimate separate weights for each target population by combining 1) the inverse-odds of trial participation (estimated in each target population) and 2) the inverse-probability of treatment (estimated in the trial population). We use a weighted Cox model to extend inferences from the original trial to each target population for the effect of bivalirudin versus heparin on the 30-day risk of major bleeding.

Results: VALIDATE enrolled 6,006 individuals, constituting 51% of the trial-eligible target population (N=11,763) during center-specific recruitment periods, and 25% of the treatment-eligible target population (N=23,614). The trial population was healthier on average than the target populations. The hazard ratio (HR) for major bleeding in the trial was 0.89 (95% CI: 0.71, 1.11). After weighting, the HR was 0.93 (95% CI: 0.71, 1.22) for the trial-eligible population, and 0.82 (95% CI: 0.60, 1.12) for the treatment-eligible population.

Conclusions: Highlighting the importance of defining explicit target populations, we demonstrate the application of generalizability and transportability to extend inferences to target populations relevant for decision makers.