Efficient estimation of the target population average treatment effect from multi-source data

We consider estimation of the target population average treatment effect (TATE) when outcome information is unavailable. Instead, we observe the outcome in multiple source populations and wish to combine the treatment effects therein to make inference on the TATE. In contrast to existing works that assume transportability on the conditional distribution of potential outcomes or conditional treatment-specific means, we work under a weaker form of effect transportability. Following the framework for causally interpretable meta-analysis, we assume transportability of conditional average treatment effects across multiple populations, which may hold with fewer standardization variables. Under this assumption, we derive the semiparametric efficiency bound of the TATE and characterize a class of doubly robust and asymptotically linear estimators. Within this class, an efficient estimator assigns optimal weights to observations from different data sources. Additionally, we suggest estimators of a low-dimensional summary of effect heterogeneity in the target population. We illustrate the use of the proposed estimators on a multicentre weight management clinical trial for semaglutide, a glucagon-like peptide-1 receptor agonist, on overweight or obese patients. Using outcome information from other regions, we estimate the weight loss effect of semaglutide in the United States subgroup.