Posterior maximization and averaging for Bayesian working model choice in the continual reassessment method

Area of Science:

• Biostatistics
• Clinical Trial Design
• Pharmacometrics

Background:

• The Continual Reassessment Method (CRM) is standard for estimating maximum tolerated doses in dose-finding studies.
• Traditional CRM relies on a single, pre-selected working model for dose-toxicity, which can be arbitrary.
• Model averaging has been proposed to address the limitations of a single working model.

Purpose of the Study:

• To propose and evaluate alternative Bayesian model selection criteria for adaptive CRM.
• To introduce three novel adaptive model-selecting CRM approaches.
• To compare these new methods against model averaging via simulation.

Main Methods:

• Developed three adaptive CRM designs utilizing Bayesian model selection criteria.
• Candidate working models for dose-toxicity relationships were pre-specified.
• Adaptive selection of the optimal working model was performed during the trial using posterior model probability, posterior predictive loss, or deviance information criteria.
• Compared proposed methods with model averaging using simulation studies.

Main Results:

• The proposed adaptive model-selecting CRMs demonstrated a viable alternative to model averaging.
• Simulation results indicated the performance characteristics of the novel Bayesian approaches.
• The study provides a framework for more robust dose-finding through adaptive model selection.

Conclusions:

• Adaptive model selection using Bayesian criteria offers an improvement over traditional single-model CRM and model averaging.
• These novel CRM approaches enhance the flexibility and robustness of dose-finding studies.
• The findings support the use of adaptive Bayesian model selection in clinical trials for personalized dose escalation.