Augmenting randomized trials
We present novel methods for augmenting randomized clinical trials with external data – historical experimental data or observational data – to improve trial efficiency without endangering the guaranty of unbiasedness afforded by randomization. Specifically, we characterize a class of randomization-aware estimators that integrate external data via data-adaptive models (such as machine learning or generative models), thereby increasing trial efficiency and statistical power compared to estimators that utilize trial data alone. We demonstrate how members of this class exploit randomization to remain unbiased even if the external data are not well aligned with the trial data or are subject to unmeasured confounding. We illustrate that several commonly used estimators, including the efficient trial-only estimator, belong to this class. Finally, we show how two or more randomization-aware estimators can be combined to construct estimators with two key properties: (1) robustness to data misalignment and unmeasured confounding in the external data, and (2) efficiency that is at least as high as, and typically higher than, that of their component estimators. We study the finite-sample behavior of the proposed methods in simulations and illustrate their use in an empirical analysis of the Coronary Artery Surgery Study (CASS).
LOCATION:CGIS Knafel Building, Room K354 STATUS:CONFIRMED DTSTART:20260311T160000Z DTEND:20260311T173000Z END:VEVENT END:VCALENDAR