Evaluating Bayesian Borrowing Methods for Treatment Effect Extrapolation: A Simulation-Based Study
Author(s)
Marie Génin, MSc, Tristan Fauvel, PhD, Antoine Movschin, MSc, Billy Amzal, MBA, MPH, MSc, PhD.
Quinten Health, Paris, France.
Quinten Health, Paris, France.
OBJECTIVES: Bayesian borrowing methods are increasingly used to extrapolate treatment effects when direct evidence generation is challenging, such as for pediatrics or rare diseases. This study aimed to evaluate and compare the performance of Bayesian methods for treatment effect extrapolation using realistic trial scenarios.
METHODS: A comprehensive simulation study was performed inspired on real-world use cases, exploring design factors such as sample size, size of treatment effect, and dissimilarity (“drift”) between source and target populations. Frequentist and Bayesian methods - including test-then-pool (TTP) strategies, Robust Mixture Priors (RMP), Conditional Power Priors (CPP), and other Power Prior variants - were evaluated. Performances were assessed through frequentist operating characteristics (OCs) (probability of success (POS), type I error (TIE), bias, mean squared error, and estimation precision) and Bayesian OCs.
RESULTS: No method showed systematic better performance across all simulation scenarios. CPP and RMP methods showed stronger performance across diverse scenarios. They achieved higher POS than other methods at similar TIE levels, with accurate estimates when source and target data were similar. Under substantial drift, RMP downweighted source data, preserving robust performances. In contrast, test-then-pool and p-value-based power priors performed less reliably, particularly in the presence of drift. The study highlights the importance of selecting methods that balance informativeness with statistical operating characteristics across diverse design contexts.
CONCLUSIONS: Bayesian borrowing methods, when carefully selected and supported by thorough sensitivity analyses, enable extrapolation of treatment effect estimates to small populations and increase the probability of identifying true effects, at the cost of TIE inflation that depends on the method. While regulators generally require strict TIE control, in the context of extrapolation they may accept the risk of T1E inflation if the overall evidence package is robust and well justified.
This study was funded by the EMA. This abstract expresses only the opinion of the authors and not EMA’s.
METHODS: A comprehensive simulation study was performed inspired on real-world use cases, exploring design factors such as sample size, size of treatment effect, and dissimilarity (“drift”) between source and target populations. Frequentist and Bayesian methods - including test-then-pool (TTP) strategies, Robust Mixture Priors (RMP), Conditional Power Priors (CPP), and other Power Prior variants - were evaluated. Performances were assessed through frequentist operating characteristics (OCs) (probability of success (POS), type I error (TIE), bias, mean squared error, and estimation precision) and Bayesian OCs.
RESULTS: No method showed systematic better performance across all simulation scenarios. CPP and RMP methods showed stronger performance across diverse scenarios. They achieved higher POS than other methods at similar TIE levels, with accurate estimates when source and target data were similar. Under substantial drift, RMP downweighted source data, preserving robust performances. In contrast, test-then-pool and p-value-based power priors performed less reliably, particularly in the presence of drift. The study highlights the importance of selecting methods that balance informativeness with statistical operating characteristics across diverse design contexts.
CONCLUSIONS: Bayesian borrowing methods, when carefully selected and supported by thorough sensitivity analyses, enable extrapolation of treatment effect estimates to small populations and increase the probability of identifying true effects, at the cost of TIE inflation that depends on the method. While regulators generally require strict TIE control, in the context of extrapolation they may accept the risk of T1E inflation if the overall evidence package is robust and well justified.
This study was funded by the EMA. This abstract expresses only the opinion of the authors and not EMA’s.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR91
Topic
Methodological & Statistical Research
Disease
Oncology, Pediatrics, Rare & Orphan Diseases