Nonparametric Bayesian approach for dynamic borrowing of historical control data.

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Bayesian Statistics

Background:

• Incorporating historical control data in randomized controlled trials (RCTs) requires accounting for dataset differences.
• Unmeasured factors can cause heterogeneity, making simple covariate adjustment insufficient.
• Dynamic borrowing methods are needed to mitigate the impact of heterogeneous historical controls.

Purpose of the Study:

• To propose a nonparametric Bayesian approach for analyzing current RCT data with historical controls.
• To address between-trial heterogeneity and enable borrowing from homogeneous historical controls.
• To introduce a dependent Dirichlet process (DP) mixture method for conflict resolution between historical and current controls.

Main Methods:

• Developed a nonparametric Bayesian framework adaptable for both aggregated and individual participant data.
• Introduced a dependent Dirichlet process (DP) mixture model for enhanced borrowing and conflict resolution.
• Created a novel similarity index based on the posterior distribution to compare historical and current control data.

Main Results:

• The dependent DP mixture method accurately borrows from homogeneous historical controls.
• It effectively reduces the impact of heterogeneous historical controls compared to standard DP mixtures.
• Proposed methods outperform existing approaches, particularly in heterogeneous historical control scenarios where meta-analysis fails.

Conclusions:

• The proposed dependent DP mixture offers a robust method for integrating historical controls in RCTs.
• This approach improves the reliability of trial results by selectively utilizing relevant historical data.
• The methods provide a valuable tool for biostatisticians and clinical researchers facing data heterogeneity challenges.