Potential Bias Models With Bayesian Shrinkage Priors for Dynamic Borrowing of Multiple Historical Control Data

Area of Science:

• Biostatistics
• Clinical Trial Design
• Statistical Modeling

Background:

• Integrating historical control data in clinical trials presents challenges due to potential conflicts and relationships among controls.
• The 'Potential biases' assumption defines differences between current and historical control parameters as 'potential bias parameters.'

Purpose of the Study:

• To extend the 'potential biases model' by incorporating various shrinkage priors for analyzing historical control data.
• To compare the performance of proposed methods (spike-and-slab, Dirichlet-Laplace, spike-and-slab lasso) against existing methods like the horseshoe prior.

Main Methods:

• Development of a 'potential biases model' class encompassing existing methods.
• Application of spike-and-slab, Dirichlet-Laplace, and spike-and-slab lasso priors to the potential biases model.
• Simulation studies and analysis of clinical trial examples to evaluate method performance.

Main Results:

• The horseshoe prior and the proposed priors effectively utilize historical controls when they are homogeneous.
• These priors mitigate the influence of heterogeneous historical controls when only a few are present.
• The spike-and-slab prior exhibited the best performance in scenarios with heterogeneous historical controls.

Conclusions:

• The proposed shrinkage priors offer robust methods for incorporating historical controls in clinical trial analysis.
• The spike-and-slab prior is particularly effective for managing heterogeneity in historical control data.
• The potential biases model provides a flexible framework for leveraging historical control information.