OBJECTIVES : Clinical trials often select participants on the basis of covariates that are also determinants of the outcome and modifiers of the treatment effect. Consequently, the average treatment effect estimated in a trial is not the same as the average effect in other populations in which the trial’s interventions may be applied. We consider methods for transporting the results from a completed trial to a new target population in which no experimental data are available.

METHODS : We show how a composite dataset, formed by appending data from a completed trial to a sample from the target population, can be used to estimate the average treatment effect in the target population. We examine methods that use the treatment, outcome, and baseline covariate information from the trial, but only baseline covariate data from the sample of the target population. We assess the finite-sample performance of these methods in simulation studies.

RESULTS : We show that, under certain assumptions, estimators based on modeling the expectation of the outcome in the trial, the probability of participation, or both (doubly robust) consistently estimate the average treatment effect in the target population. In simulation studies, when all models are correctly specified, all estimators are approximately unbiased. Estimators that only rely on modeling the probability of trial participation have substantially larger variance compared to estimators that rely on modeling the outcome or doubly robust estimators. To illustrate their use, we apply all these methods in a composite dataset formed by a clinical trial of anti-hypertensive agents and a sample from a target population represented by electronic medical record data.

CONCLUSIONS : We propose methods for transporting the results of a completed clinical trial to a new target population. These methods may be particularly attractive in the drug approval setting.