Self-adapting Mixture Prior to Dynamically Borrow Information from Historical Data in Clinical Trials

Mixture priors, such as robust meta-analytic predictive (MAP) prior, provide an intuitive way to borrow information from historical data, while acknowledging the possibility of prior-data conflict, by mixing an informative prior and a non-informative prior. The key question when applying mixture priors is how to pre-specify the mixing weight for each mixture component. Ideally, the mixture weight should be chosen based on the degree of prior-data conflict, which unfortunately is often unknown a priori. This has been a major barrier to the application and acceptance of mixture priors. To address this issue, we propose self-adapting mixture (SAM) priors where the mixture weight is data-driven and self-adapting --- it favors the informative (non-informative) prior component when there is little (substantial) evidence of prior-data conflict. As a result, SAM priors achieve dynamic information borrowing. We show that SAM priors possess desirable finite-sample and large-sample properties and outperform existing methods, such as robust MAP prior. SAM priors are simple to calculate, data-driven and calibration-free, avoiding the risk of data dredging and lowering the hurdle for the acceptance of the method.

Real-world data

Historical data

Dynamic information borrowing