ABSTRACT
The Bayesian framework facilitates use of historical control data in the analysis of a clinical trial, and hence to reduce the number of subjects randomized to control. This decreases costs, makes the trial shorter, facilitates recruitment, and may be more ethical. We focus on the meta-analytic-predictive (MAP) approach to historical data, but also describe alternatives such as bias models, commensurate and power priors. The prospective MAP approach derives a prior distribution from historical data based on a hierarchical model. We discuss mixture approximations of MAP priors, robustness to prior-data conflict, and prior effective sample sizes. A proof-of-concept study with placebo data from eight historical studies is used to illustrate the methodology. The MAP approach can be extended in various directions. We discuss two such extensions: over-dispersed count data in multiple sclerosis, where the historical placebo data consist of individual patient data and aggregate data; and noninferiority trials, which depend critically on a proper integration of historical data.
