Advanced Statistical Methods for Infectious Disease Epidemiology with a focus on Human Social Contact Dynamics

Human social contact data are crucial for understanding the transmission dynamics of airborne and sexually transmitted infections within populations. However, current modeling frameworks often limit their scope to variables such as age and sex, overlooking other influential factors, including socioeconomic status, geographic location, and individuals' attitudes toward disease transmission. This project seeks to develop a more comprehensive model that incorporates these additional dimensions, enabling more accurate predictions of disease spread and facilitating targeted public health interventions.
The first aim of this project is to enhance the Bayesian hierarchical modeling framework, previously developed by the student , with capabilities to stratify across diverse population subgroups. This enhancement will leverage the symmetrical nature of social contacts (e.g., if Harry meets George, then George also meets Harry) to establish mathematical constraints that improve the model's consistency with real-world contact dynamics. Additionally, the project will explore advanced computational methods in Bayesian inference, including Hilbert space approximate Gaussian process priors, deep-learning-based meta-learning techniques (e.g., pi-VAE, prior-VAE), GPU-accelerated Markov Chain Monte Carlo (MCMC) simulations, and variational inference, to optimize the model's accuracy and efficiency.
On the theoretical front, this project will investigate robust statistical methods to account for the heterogeneity in contact dynamics. This includes utilizing heavy-tailed distributions, such as the Beta-Negative Binomial, and generalized Bayesian inference to better capture variability and mitigate model misspecification. Furthermore, a key challenge in pandemic preparedness is the scarcity of social contact data, particularly in underdeveloped regions. To address this, the project aims to develop Bayesian and deep learning-based transfer learning techniques to infer social contact patterns with limited data, using models trained on data from other countries or regions.
We will apply these methods to analyze social contact data from the POLYMOD study (pre-COVID-19 data from eight major European countries), the COVIMOD study (COVID-19 era data from Germany), and the CoMix study (COVID-19 era data from the United Kingdom and Belgium).
This project will make significant contributions to methodologies for analyzing human social contact data, ultimately enriching the existing toolkit for infectious disease modeling. The insights gained will support improved management and preparedness strategies for future pandemics.

1. Dan et al. (2023), Estimating fine age structure and time trends in human contact patterns from coarse contact data: The Bayesian rate consistency model, PLOS Computational Biology