Keywords: generalizability, transportability, preference trials, missing data, causal inference
The need to generalize trial findings from randomized patients to all eligible patients arises naturally in singly randomized preference trials. In such studies, methods that use baseline covariates from all eligible patients but only use treatment and outcome information from the randomized patients are appealing because they avoid confounding of the treatment effect among eligible patients who refuse randomization. Leveraging connections with missing data theory, we show that the data from preference trials can be used to identify the average treatment effect in the superpopulation of all eligible patients. We examine three classes of estimators of the average treatment effect: (1) outcome model-based; (2) probability of trial participation-based; and (3) doubly robust estimators. We assess the finite-sample performance of different estimators in simulation studies. Lastly, we demonstrate the implementation of the methods using data from the Coronary Artery Surgery Study of 2099 eligible patients with coronary artery disease, of whom 780 were randomized into coronary revascularization surgery or medical therapy and 1319 refused randomization and self-selected into treatment.