Bayesian Dynamic Borrowing to Enhance Evidence for New Therapies
Author(s)
Sean Yiu, PhD1, Katya Galactionova2, Steven Yuen, MD3, Chris Skedgel, MA, PhD4, Isabelle DURAND ZALESKI, MD5, Maarten Jacobus Postma, PhD6, Mark Sculpher, PhD7, Keith R. Abrams, BSc, MSc, PhD8.
1Roche Products Ltd, Welwyn Garden City, United Kingdom, 2Evidence Lead, Roche, Basel, Switzerland, 3Genentech, Inc, South San Francisco, CA, USA, 4Office of Health Economics, London, United Kingdom, 5Hôpital de l’Hôtel Dieu, Université de Paris, CRESS, INSERM, INRA, URCEco, AP-HP, CRETEIL, France, 6University of Groningen, Groningen, Netherlands, 7University of York, York, United Kingdom, 8University of Liverpool, Liverpool, United Kingdom.
1Roche Products Ltd, Welwyn Garden City, United Kingdom, 2Evidence Lead, Roche, Basel, Switzerland, 3Genentech, Inc, South San Francisco, CA, USA, 4Office of Health Economics, London, United Kingdom, 5Hôpital de l’Hôtel Dieu, Université de Paris, CRESS, INSERM, INRA, URCEco, AP-HP, CRETEIL, France, 6University of Groningen, Groningen, Netherlands, 7University of York, York, United Kingdom, 8University of Liverpool, Liverpool, United Kingdom.
OBJECTIVES: Historical trial data can enhance treatment effect (TE) estimation in trials of rare/slowly progressing diseases. Bayesian Dynamic Borrowing (BDB) is an attractive approach to integrating historical data as it dynamically adjusts the degree of borrowing of historical data based on the agreement between historical and new trial data. We illustrate BDB for improving precision of TE estimators on disability outcomes in Multiple Sclerosis (MS).
METHODS: We simulated MS trials that were adequately powered to detect TEs on confirmed disability progression on the Timed Twenty-Five Foot Walk Test (CDP-T25FWT, intermediate outcome), but not the Expanded Disability Status Scale (CDP-EDSS, final outcome). BDB TE estimates on CDP-EDSS were obtained by (1) fitting a model to historical data to predict TE on CDP-EDSS based on TE on CDP-T25FWT in the new trials; (2) then using this prediction to construct a prior distribution; and (3) updating the prior distribution with the TE on CDP-EDSS from the new trial to form a posterior distribution summarizing all available evidence. We investigated the performance of BDB relating to bias and Type I error using cross-validation and simulation.
RESULTS: BDB always preserved the TE estimate from new trials and resulted in reasonably large reductions in uncertainty (SD of posterior 10.1-19.3% smaller than SE) when observed and predicted TEs were well aligned or when agreement was moderate, but SE was relatively large. Simulations confirmed that BDB exhibits minimal bias and Type I error inflation (<4 % points increase).
CONCLUSIONS: We demonstrated that BDB can effectively integrate historical data and improve precision without introducing noticeable bias or Type I error inflation. This is particularly useful when data for decision-making is limited, such as from important, but under-powered, components of composite endpoints. BDB facilitates timely regulatory and health technology assessments, thereby supporting the advancement of effective therapies.
METHODS: We simulated MS trials that were adequately powered to detect TEs on confirmed disability progression on the Timed Twenty-Five Foot Walk Test (CDP-T25FWT, intermediate outcome), but not the Expanded Disability Status Scale (CDP-EDSS, final outcome). BDB TE estimates on CDP-EDSS were obtained by (1) fitting a model to historical data to predict TE on CDP-EDSS based on TE on CDP-T25FWT in the new trials; (2) then using this prediction to construct a prior distribution; and (3) updating the prior distribution with the TE on CDP-EDSS from the new trial to form a posterior distribution summarizing all available evidence. We investigated the performance of BDB relating to bias and Type I error using cross-validation and simulation.
RESULTS: BDB always preserved the TE estimate from new trials and resulted in reasonably large reductions in uncertainty (SD of posterior 10.1-19.3% smaller than SE) when observed and predicted TEs were well aligned or when agreement was moderate, but SE was relatively large. Simulations confirmed that BDB exhibits minimal bias and Type I error inflation (<4 % points increase).
CONCLUSIONS: We demonstrated that BDB can effectively integrate historical data and improve precision without introducing noticeable bias or Type I error inflation. This is particularly useful when data for decision-making is limited, such as from important, but under-powered, components of composite endpoints. BDB facilitates timely regulatory and health technology assessments, thereby supporting the advancement of effective therapies.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
CO26
Topic
Clinical Outcomes, Methodological & Statistical Research
Topic Subcategory
Clinical Outcomes Assessment, Comparative Effectiveness or Efficacy
Disease
Neurological Disorders