Comparison of Model Averaging and Model Selection in Dose Finding Trials Analyzed by Nonlinear Mixed Effect Models

Area of Science:

• Pharmacometrics and Drug Development
• Clinical Trial Simulation
• Statistical Modeling

Background:

• Pharmacometric approaches in drug development often use model selection (MS) to identify the best data-describing model structure.
• Using a single selected model for predictions neglects model structure uncertainty, potentially reducing predictive performance.
• Model averaging (MA) addresses this by incorporating uncertainty across multiple candidate models, weighted by information criteria.

Purpose of the Study:

• To compare the predictive performance of model selection (MS) versus model averaging (MA) in dose-finding clinical trials using clinical trial simulations (CTSs).
• To evaluate the performance of MA and MS using the Akaike Information Criterion (AIC) and four other information criteria.
• To assess the impact of these approaches on characterizing dose-response relationships and identifying minimum effective doses.

Main Methods:

• Clinical trial simulations (CTSs) were employed to compare MS and MA.
• A nonlinear mixed-effects model was used to simulate the visual acuity time course in wet age-related macular degeneration patients.
• Predictive performance was assessed using criteria relevant to Phase II clinical trial objectives.

Main Results:

• Model averaging (MA) demonstrated superior predictive performance compared to model selection (MS).
• MA improved the accuracy of dose-response relationship characterization and minimum effective dose identification.
• The Akaike Information Criterion (AIC) was associated with the best predictive performance, irrespective of the modeling approach (MA or MS).

Conclusions:

• Model averaging (MA) is a more robust approach than model selection (MS) for predictive modeling in dose-finding clinical trials.
• MA enhances the reliability of characterizing dose-response relationships and determining optimal drug dosages.
• AIC remains a valuable criterion for model evaluation, showing strong performance in predictive tasks within this simulation framework.