Bayesian adaptive randomization design incorporating propensity score-matched historical controls

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Health Research Methods

Background:

• Randomized controlled trials (RCTs) can be enhanced by incorporating historical control data to improve efficiency and feasibility.
• Existing Bayesian adaptive randomization designs using historical data do not differentiate between measured and unmeasured covariate imbalances.
• Prior-data conflict, arising from imbalances between historical and concurrent control groups, can affect the reliability of trial results.

Purpose of the Study:

• To extend Bayesian adaptive randomization designs by incorporating propensity score-matched historical controls.
• To address both measured and unmeasured covariate imbalances in the control arm of RCTs.
• To reduce sample size and bias in treatment effect estimates while maintaining statistical power.

Main Methods:

• Propensity score matching is used at interim assessments to select historical controls similar to concurrent controls based on measured covariates.
• The extent of borrowing information from matched historical controls is quantified by an effective historical sample size.
• The conditional power prior and commensurate prior approaches are employed to design priors and manage prior-data conflict, particularly due to unmeasured covariates.

Main Results:

• The proposed method demonstrated reduced bias in treatment effect estimates compared to existing approaches.
• Simulations confirmed that the method maintains type I error rates at the nominal level.
• The approach successfully reduced the required sample size for the control arm while preserving statistical power.

Conclusions:

• The extended Bayesian adaptive randomization design effectively incorporates propensity score-matched historical controls, enhancing RCT efficiency.
• The method robustly handles imbalances in both measured and unmeasured covariates, allowing for greater information borrowing.
• This approach offers significant implications for facilitating the conduct of adequate RCTs by optimizing the use of historical control data.