Model Averaging in Viral Dynamic Models

Area of Science:

• Mathematical modeling
• Virology
• Statistical inference

Background:

• Paucity of experimental data complicates viral dynamic model inference and prediction.
• Model selection (MS) presents results from the 'best' model, ignoring uncertainty.
• Model averaging (MA) weights predictions from multiple models based on data consistency.

Purpose of the Study:

• Evaluate the performance of MS versus MA in viral dynamic models.
• Assess methods under conditions of poorly identifiable parameters and immune response uncertainty.
• Compare parameter uncertainty estimation and prediction accuracy between MS and MA.

Main Methods:

• Simulations using a nonlinear mixed-effect model framework.
• Two realistic acute viral infection scenarios were simulated.
• Performance metrics included false selection rates and coverage rates for parameter uncertainty.

Main Results:

• MS exhibited a high rate of false selection in some scenarios.
• MS resulted in coverage rates below the nominal 0.95, sometimes below 0.50.
• MA demonstrated improved parameter uncertainty estimation with coverage rates mostly within the nominal range (0.72-0.98).
• MA yielded predictions comparable to those from MS.

Conclusions:

• Parameter estimates from MS should be used cautiously, especially when multiple models fit the data well.
• MA offers superior performance in accounting for model uncertainty.
• MA is recommended for more reliable inference and prediction in viral dynamics modeling.