Efficient estimation of the target population average treatment effect from multi-source data
🤖 AI Summary
This paper addresses the challenge of estimating the target-average treatment effect (TATE) when outcomes are unobserved in the target population. We propose a cross-population causal inference framework leveraging multi-source data. Our key innovation is introducing the “conditional average treatment effect (CATE) transportability” assumption—a weaker condition than conventional distributional transportability—which enables derivation of the semiparametric efficiency bound for TATE. Building upon this, we develop a doubly robust, optimally weighted estimator that simultaneously supports TATE estimation and low-dimensional characterization of treatment effect heterogeneity. The method integrates causal inference, transportability modeling, semiparametric efficiency theory, and meta-analytic principles, combining inverse probability weighting with outcome regression adjustment. Applied to a multicenter semaglutide weight-loss trial, our approach successfully transports evidence from non-U.S. sites to accurately estimate TATE for U.S. subgroups, markedly improving estimation precision and robustness.
Application Category
📝 Abstract
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.
Problem
Research questions and friction points this paper is trying to address.
Estimating target population treatment effects without outcome data
Combining multi-source treatment effects under weaker transportability
Developing efficient estimators for heterogeneous treatment effects
Innovation
Methods, ideas, or system contributions that make the work stand out.
Assumes transportability of conditional average treatment effects
Derives semiparametric efficiency bound for TATE
Proposes doubly robust estimators with optimal weights
🔎 Similar Papers
No similar papers found.
