Treatment effect extrapolation in the presence of unmeasured confounding

2025-09-07

AI Summary

Causal extrapolation of conditional average treatment effects (CATE) to target populations is biased under unmeasured confounding. Method: We propose a causal extrapolation framework integrating multiple randomized controlled trials (RCTs) and observational data. Our approach constructs a nonlinear deconfounding function from multiple related RCTs to automatically correct confounding bias in observational data and extend CATE estimation beyond the support region. Contribution/Results: By leveraging auxiliary RCTs, our method overcomes limitations of single-RCT approaches—namely, insufficient sample size and poor covariate coverage—thereby enabling robust extrapolation, especially under nonlinear confounding structures. Empirical results demonstrate that incorporating even one additional RCT significantly improves CATE estimation accuracy, with pronounced gains in low-sample-size and strongly nonlinear settings. The framework provides a scalable, robust solution for synthesizing real-world evidence in causal inference.

Technology Category

• Machine Learning: Causal Learning
• Reasoning under Uncertainty: Causality
• Intelligent Robots: State Estimation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

While randomised controlled trials (RCTs) are the gold standard for estimating causal treatment effects, their limited sample sizes and restrictive criteria make it difficult to extrapolate to a broader population. Observational data, while larger, suffer from unmeasured confounding. Therefore, we can combine the strengths of both data sources for more accurate results. This work extends existing methods that use RCTs to debias conditional average treatment effects (CATEs) estimated in observational data by defining a deconfounding function. Our proposed approach borrows information from RCTs of multiple related treatments to improve the extrapolation of CATEs. Simulation results showed that, for non-linear deconfounding functions, using only one RCT poorly estimates the CATE outside of the support of that RCT. This is emphasised for smaller RCTs. Borrowing information from a second RCT provided more accurate estimates of the CATE outside of the support of both RCTs.

Problem

• Extrapolating treatment effects beyond RCT populations
• Addressing unmeasured confounding in observational data
• Improving CATE estimation using multiple RCTs

Innovation

• Combines RCT and observational data via deconfounding function
• Uses multiple related RCTs to improve CATE extrapolation
• Addresses unmeasured confounding through information borrowing