A Dynamic Power Prior for Bayesian Non-inferiority Trials

Non-inferiority trials compare new experimental treatments to active controls. Previous information on the control treatments is often available and, as long as the past and the current experiments are sufficiently homogeneous, historical data may be useful to reserve resources to the new therapy’s arm and to improve accuracy of inference. In this article we propose a Bayesian method for exploiting historical information based on a dynamic power prior for the parameter of the control arm. The degree of information-borrowing is tuned by a quantity based on the Hellinger distance between the two posterior distributions of the control arm’s parameter, obtained respectively from the current and the historical experiments. Pre-posterior analysis for type-I error/power assessment and for sample size determination is also discussed.