At ISPOR Europe 2025, Marie Génin and Antoine Movschin presented Quinten Health’s large-scale simulation study assessing Bayesian borrowing methods for extrapolating treatment effects from a source population to a target population. These methods are particularly relevant in paediatrics, rare diseases, and small-population contexts where direct evidence generation is difficult.
Why Borrowing Matters
Indirect evidence from former studies or from different populations can be complementary to scarce available data, thus helping inform decision-making. Bayesian approaches offer structured ways to incorporate external information, thereby increasing precision and statistical power, but this may come at the cost of type I error inflation, especially if treatment effects differ strongly between the source and target populations. Currently, there are no practical guidelines as to which Bayesian borrowing method to use in a specific context.
Methods
The simulation framework recreated realistic clinical scenarios inspired by six clinical studies across six different therapeutic areas and multiple outcome types. Variations included treatment effect drift between the source and target populations, sample size and variance differences.
Methods evaluated included Conditional Power Prior (CPP), Robust Mixture Priors (RMP), Normalized Power Prior (NPP), Commensurate Power Prior (Com.PP), Empirical Bayes Power Prior (EBPP), p-value based Power Prior (p-PP), and Test-then-Pool (TtP) approaches.
Metrics assessed were Type I Error (TIE), power, bias, Mean Squared Error (MSE), coverage of the Credibility Interval (CI), precision, and prior Effective Sample Size (ESS).
Key Findings
- All borrowing methods can increase power but at the cost of type I error inflation.
- Overall, CPP and RMP seem to provide the best balance between bias, precision and TIE.
- p-PP and TtP tend to yield suboptimal results, with higher MSE and poorer coverage.
Implications
The study offers practical guidance for HTA assessments, small-population clinical trials, and evidence synthesis using external data.
Simulation-based calibration is essential before applying borrowing methods in confirmatory settings.
Conclusion
Bayesian borrowing methods can strengthen evidence generation but require careful scenario-specific evaluation. CPP and RMP emerge as the most reliable options, provided that their performance is validated through tailored simulation studies.
Official ISPOR Poster
You can view the official ISPOR Europe 2025 poster here: Evaluating Bayesian Borrowing Methods for Treatment Effect Extrapolation
.
