Subgroup analysis using exchangeability probabilities

Area of Science:

• Biostatistics
• Clinical Trial Design
• Statistical Inference

Background:

• Subgroup analyses in clinical trials commonly use p-values from univariate tests or interaction models.
• Existing methods may lack optimal power or control for type I error rates in specific scenarios.

Purpose of the Study:

• To propose and evaluate an alternative method for subgroup analyses in clinical trials using Bayesian posterior weights.
• To demonstrate the superiority of the proposed method over standard approaches in terms of statistical power and type I error control.

Main Methods:

• Developed a Bayesian approach utilizing posterior weights to estimate group exchangeability probabilities.
• Established a significance cutoff based on Bayesian posterior weights.
• Conducted simulations and applied the method to a COVID-19 clinical trial dataset.

Main Results:

• The proposed Bayesian method demonstrated increased statistical power (e.g., 0.9609 vs. 0.6705 in a simulated two-arm trial) compared to traditional likelihood ratio tests (LRT).
• Type I error rates remained stable and comparable to standard methods (e.g., 0.0530 vs. 0.0532).
• The method was successfully applied to real-world COVID-19 trial data for subgroup effect evaluation.

Conclusions:

• The Bayesian posterior weights method offers a more powerful and reliable approach for subgroup analyses in clinical trials.
• This method provides a valuable alternative for researchers seeking to optimize subgroup identification and interpretation.
• The approach is applicable across various trial designs and outcome types, including real-world clinical data.