A Modification of Location Commensurate Power Prior in Clinical Trials

Abstract

A statistical approach to incorporate historical data efficiently while maintaining statistical integrity is needed when the use of historical data along with current data is justified. Bayesian statistics enables the utilization of historical data in its priors for inference. In clinical trial analysis, if historical data support current data, incorporating priors can help improve statistical power as well as reduce sample sizes. However, if historical and current data are not compatible, incorporating Bayesian priors can result in estimation bias, hence inflated Type I error. A more recent Bayesian approach, location commensurate power prior (LCPP) helps resolve inflated Type I error issue and gain power in certain scenarios. Our simulation results show that in the presence of large conflicts (a conflict is defined as an average of historical data minus current control data), all Bayesian approaches do not perform well especially in the presence of negative large conflicts, all Bayesian approaches have much lower power compared to frequentist. To address this issue, we propose a modification of location commensurate power prior (mLCPP) that leverages frequentist, traditional Bayesian and LCPP. The proposed approach results in better performance across all conflict profiles. We also apply mLCPP to an adaptive clinical trial design and show that if large conflicts present, mLCPP signals sample size re-estimation in a control arm to avoid under-powered studies and if moderate to no conflicts, it signals partial to full use of historical information, thus improving power while allowing an acceptable inflation of Type I error and offering opportunity for sample size reduction.

KEYWORD:

• Bayesian adaptive clinical trial design
• Bayesian dynamic borrowing
• Commensurate power prior
• Interim analysis
• Sample size re-estimation

Acknowledgments

I thank my previous supervisor, Zorayr Manukyan for his expertise and great insight that greatly assisted and guided the research. I also would like to thank Steve Gilbert for his experience in using Bayesian in clinical trial designs.

Supplementary Material

Supplementary material provides the posterior distributions of the power parameter α from MCMC.