Abstract
BACKGROUND: Pragmatic trials that combine electronic health record data and patient-reported data may be subject to selection bias due to the differential post-randomization exclusion of participants who are randomized in error. Such situations are often caused by inevitable reasons, such as incomplete patient medical records at the pre-randomization stage. This can lead to participants in the intervention arm being identified as ineligible after randomization, while randomized-in-error participants in the usual care are often not discernable. The differential exclusion can present analytic challenges and threaten result validity. METHODS: Under the potential outcomes framework, we developed a Bayesian model that jointly identifies the randomized-in-error status and estimates the average treatment effect among participants not randomized in error. We designed simulation studies with hypothesized proportions of 5 %-15 % randomization in error to evaluate the performance of our model across scenarios where the outcomes of participants randomized in error were either measured or unmeasured. Comparisons were made to intention-to-treat and covariate-adjusted estimators. RESULTS: Simulation results show satisfactory performance of our proposed models, where the estimated average treatment effects among participants not randomized in error have low bias (<1 %) and close to 95 % coverage. Estimates from the alternative approaches can exhibit notable biases and low coverage. CONCLUSIONS: Differential exclusion in pragmatic clinical trials after randomization can lead to selection bias. Under certain assumptions, Bayesian methods provide a feasible solution to jointly identify randomized-in-error status and estimate the average treatment effect among participants not randomized in error, ensuring more reliable and valid inferences about intervention effects.