Flexible Bayesian inference on partially observed epidemics.

Area of Science:

• Epidemiology
• Computational Biology
• Statistical Modeling

Background:

• Individual-based models are crucial for predicting epidemic spread and guiding interventions.
• Contact network data enhances realism by capturing non-random and heterogeneous interactions.
• Bayesian inference on complex contagion models with limited disease status data is challenging.

Purpose of the Study:

• To develop a Bayesian inference method for SIR contagion spreading parameters on known networks.
• To address challenges in sampling posterior distributions for complex or partially observed contagion models.
• To circumvent the need for manual summary statistic selection in Approximate Bayesian Computation (ABC).

Main Methods:

• Utilized Mixture Density Network compressed Approximate Bayesian Computation (ABC).
• Employed a scheme that minimizes expected posterior entropy to learn informative summary statistics.
• Applied to Bayesian inference on spreading parameters of a SIR contagion with partially observed disease status.

Main Results:

• Successfully conducted Bayesian inference on contagion spreading parameters without manual summary statistics.
• Demonstrated an effective method for partially observed contagious processes on static networks.
• The methodology learned informative summary statistics automatically by minimizing posterior entropy.

Conclusions:

• The proposed Mixture Density Network compressed ABC method enables robust Bayesian inference for complex epidemic models.
• This approach overcomes limitations of traditional ABC by eliminating the need for user-defined summary statistics.
• The methodology is extensible to more complex scenarios like behavioral changes or imperfect testing.