A multi-method study evaluating the inference of compartmental model parameters from a generative agent-based model.

Area of Science:

• Epidemiology
• Computational Biology
• Mathematical Modeling

Background:

• Calibrating compartmental models like SIR to real-world data is challenging due to unknowns such as under-reporting.
• Synthetic data offers a controlled environment to assess calibration method performance.
• Agent-based models can generate realistic synthetic epidemic data reflecting complex contact structures.

Purpose of the Study:

• To evaluate and compare the performance of optimization (Nelder-Mead) and Bayesian (HMC) calibration techniques for SIR models.
• To investigate how varying agent contact structures in synthetic data affect SIR model parameter estimation.
• To determine the impact of contact patterns and population susceptibility on the effective infectious period.

Main Methods:

• Generating synthetic epidemic data using an agent-based model with diverse contact structures.
• Calibrating SIR models to these synthetic datasets using Nelder-Mead (optimization) and HMC (Bayesian) methods.
• Comparing calibration accuracy using Mean Absolute Error, Mean Absolute Scaled Error, and Relative Root Mean Squared Error.

Main Results:

• Both Nelder-Mead and HMC demonstrated comparable accuracy in overall model fitting.
• HMC significantly outperformed Nelder-Mead in accurately recovering the ground truth SIR model parameters.
• The effective infectious period was found to be sensitive to contact patterns and the proportion of susceptible individuals.

Conclusions:

• For general accuracy, both optimization and Bayesian methods are suitable for SIR model calibration.
• HMC is preferred when the goal is to accurately estimate underlying epidemiological parameters.
• Understanding parameter sensitivity to contact patterns and vaccination is crucial for interpreting real-world epidemic data.