Frequentist performance of Bayesian inference with response-adaptive designs

Area of Science:

• Clinical Trials and Biostatistics
• Statistical Inference in Healthcare

Background:

• Response-adaptive designs offer an alternative to 1:1 randomization in clinical trials, adjusting treatment assignments based on prior outcomes.
• Minimizing treatment failures is a critical objective in controlled clinical trials.

Purpose of the Study:

• To evaluate the frequentist performance of Bayesian inference for response-adaptive designs in comparing two treatments.
• To assess Bayesian inference using metrics like coverage probabilities, power, and failure count minimization.

Main Methods:

• Considered several response-adaptive designs: Play-The-Winner (PW), Randomized Play-The-Winner (RPW), Drop-The-Loser, Generalized Drop-the-Loser, and Doubly adaptive Biased Coin Designs.
• Employed Bayesian inference with two independent Beta prior distributions.
• Assessed performance using frequentist criteria: coverage probabilities, statistical power, and minimization of treatment failures.

Main Results:

• Bayesian inference demonstrated favorable frequentist performance, comparable to existing frequentist procedures.
• The power of response-adaptive designs closely matched that of 1:1 randomized designs.
• Failure count reductions were generally modest, with notable exceptions for PW and Doubly adaptive Biased Coin designs under specific success rate conditions.

Conclusions:

• Bayesian inference provides a viable and effective frequentist approach for response-adaptive clinical trial designs.
• While power is comparable to 1:1 randomization, significant failure reduction is design- and condition-specific.
• Play-The-Winner, Generalized Drop-the-Loser, and Doubly adaptive Biased Coin Designs show potential for superior performance in certain scenarios, whereas RPW was least effective.