Joost van Rosmalen, PhD, Erasmus University Medical Center; Rotterdam, Netherland

Title: Calculation of power and type I error rate of methods for borrowing data from historical controls: Should we sample the historical data?

Abstract: Joost van Rosmalen is an assistant professor at the Department of Biostatistics of Erasmus MC in Rotterdam, the Netherlands. His research mainly focuses on methods for including historical and real-world data in the analysis of clinical trials and the analysis of repeated measurements. Link to website: https://www.erasmusmc.nl/en/research/researchers/rosmalen-joost-van

Many methods for borrowing information from historical controls in the analysis of a randomized trial have been proposed, such as the (modified) power prior and the meta-analytic predictive prior. These methods aim to increase the power and reduce sample size requirements while providing a safeguard against prior-data conflict (i.e. substantial differences between data sources). However some methods can increase the type I error rate, which may not be acceptable to regulatory authorities. Power and type I error rate are calculated using simulation, which can be done using a conditional (taking the historical data as fixed) or an unconditional approach (sampling also the historical data in the simulation).

In this talk we give a brief introduction to these borrowing methods and we discuss the ideas behind the unconditional and the conditional approach. We propose a framework for the unconditional approach that can incorporate a systematic bias of the historical data as well as between-study heterogeneity based on an assumption of exchangeability. In this framework, prior information on the similarity of the current trial and the historical data can be included. The framework is illustrated using a simulation study. The results of the unconditional approach show that if we assume exchangeability and no systematic bias of historical and current controls, a gain in power while keeping the type I error rate fixed is possible. Using the conditional approach it is difficult to incorporate knowledge on the similarity of the historical controls.

Via Zoom: Please visit our website for the Zoom Link: https://www.mcgill.ca/epi-biostat-occh/seminars-events/seminars/biostati...

web : https://www.mcgill.ca/epi-biostat-occh/seminars-events/seminars/biostatistics