Approaches to mitigate bias in the design and analysis of pRCTs

Abstract Unlike traditional randomized clinical trials (RCTs), pragmatic randomized trials (pRCTs) enroll a broader population of patients typically encountered in routine clinical practice, generating results that can usually be better generalized to routine clinical practice. Randomization in pRCTs mitigates potential confounding at baseline that may be present in observational studies, providing support for valid statistical inference. Despite the advantages of randomization, however, pRCTs can be prone to a number of biases. Appropriate attention to the potential for these biases at the design and analytic stage can significantly mitigate such biases. While masking to treatment assignment is not typical of real-world pRCTs, certain aspects of pRCTs such as outcome evaluation and ascertainment can be masked without masking the full study. Prior validation of outcomes and removing subjectivity in outcome evaluation is useful. The broader population of pRCTs may yield issues with competing risks and informative loss-to-follow-up, which should be addressed in analyses to avoid underestimating or overestimating treatment effects. As with RCTs, accounting for post-randomization intercurrent events, such as discontinuation or non-adherence, may be critical to derive appropriate conclusions. Evaluation of treatment effect heterogeneity, while appropriate for the broader, typically larger pRCTs, should be approached thoughtfully as individual subgroups can be highly imbalanced even when with randomization, especially if at the cluster level. Further evaluation during the interpretation phase through quantitative bias analysis is essential. Replication of results in similar studies with comparable populations and outcome risks increases credibility. Finally, interpretation of pRCTs should be made in light of the advantages and limitations and in the context of all available evidence.