How much information should a prior contain? Investigating informative priors through dynamic borrowing in Bayesian trials
- Authors:
- Center:
-
Contributing OfficeCenter for Drug Evaluation and Research
Abstract
Bayesian analyses are particularly useful in the rare disease context as they can increase efficiency of trial design with increased precision of treatment effect estimates and decreased necessary sample size through informative priors, which effectively borrow information from previous studies. Yet previous and subsequent studies might have divergent results for many reasons such as the previous study results were spurious, there might be treatment drift, or the target population in the subsequent study has somehow changed, resulting in previous studies that are no longer similar enough to be leveraged for the subsequent study. In this context, using an informative prior that fully borrows information from previous studies could introduce substantial bias, resulting in inflated type 1 error rates and erroneous results. In defining informative priors, we need to account for how similar the previous and subsequent study results are to mitigate bias. Most approaches to borrowing propose discounting the previous study information to avoid the introduction of bias. There are two general approaches to borrowing: 1) static and 2) dynamic. For static borrowing, the degree of discounting is defined a priori, while dynamic borrowing uses the subsequent and previous study data to determine an appropriate degree of borrowing based on the degree of similarity between the data. Especially in the rare disease context where the patient population is extremely limited, research is needed to determine how much discounting should occur to achieve a balance between gathering valuable information and minimizing bias. This work investigates some approaches to dynamic borrowing with prior specification in the rare disease context. To do this, we will use a simulation study investigating the operating characteristics for different approaches to prior specification. The approaches we focus on are: 1) fully informative – or borrowing without discounting previous study information, 2) uninformative – or no borrowing, 3) Robust Meta-Analytic-Predictive Priors – a prior specified as weighted average of an informative and uninformative prior , 4) power prior approach – consistency of previous and subsequent study is tested and used to inform a discounting parameter, and 5) calibrated Bayesian hierarchical model – a dynamic borrowing approach that uses Bayesian hierarchical modeling and simulations to pre-specify the degree of borrowing, and 5) Multisource Exchangeability Modeling (MEMs) – a dynamic borrowing approach that uses Bayesian model averaging to average over all exchangeability assumptions. From there, we will explore the operating characteristics for each approach and determine the necessary sample size for a desired power and type 1 error rate that mitigates the degree of borrowing.