Bayesian clustering prior with overlapping indices for effective use of multisource external data

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Bayesian Inference

Background:

• External data in clinical trials offers benefits like reduced enrollment and increased power.
• Synthesizing information from multisource external data for Bayesian inference can be distorted by heterogeneity.
• Clustering can identify heterogeneity but optimal clustering faces a trade-off between congruence and robustness.

Purpose of the Study:

• To develop a robust prior synthesis framework for borrowing information from multisource external data.
• To introduce novel Bayesian clustering priors that address data heterogeneity in clinical trials.
• To provide methods for identifying optimal clustering that balances congruence and robustness.

Main Methods:

• Introduction of two overlapping indices: the overlapping clustering index and the overlapping evidence index.
• Application of K-means algorithm with these indices to identify optimal clustering.
• Development of (robust) Bayesian clustering meta-analytic predictive (MAP) priors by incorporating the MAP prior.

Main Results:

• The proposed indices and K-means algorithm effectively balance the trade-off for optimal clustering.
• Bayesian clustering MAP priors demonstrate advantages over commonly used priors in heterogeneous data.
• Simulation studies and real-data analysis validate the effectiveness of the developed priors.

Conclusions:

• The developed Bayesian clustering priors offer an effective framework for prior synthesis from multisource external data.
• These priors enhance information borrowing while mitigating issues caused by data heterogeneity.
• The priors are applicable to both clinical trial design and data analysis, without requiring prospective study data.