Revealing the Complexity of Health Determinants in Observational Systems Epidemiology Datasets using Additive Bayesian Networks

03.12.2021 11:15 – 12:15

RESEARCH CENTER FOR STATISTICS SEMINAR / ABSTRACT

Observational studies in systems epidemiology are situated at the intersection of classical epidemiology and machine learning that aims at studying the determinants of health and holistic thinking, which acknowledges that the biological processes, generating the data, are incredibly complex, and possibly not captured by classical statistical approaches. In a data-rich environment, such as modern veterinary epidemiology, there is a need for appropriate analytical tools to fill the gaps in the understanding of the complexity of animal-based disease dynamics. Additive Bayesian Networks (ABNs) are types of graphical models that are useful to model high-dimensional and strongly correlated data in an acausal perspective. They have an impressive track record of successful applications to real-world data. ABN methodology is a score-based approach which extends the classical Bayesian or a frequentist generalized linear model (GLM) to multiple dependent variables through the factorization of their joint probability distribution. We will first present the ABN methodology in a general context and highlight some of the limitations using real-world examples. Then we will use a re-analysis of a dataset about disease complex infection among cats in Switzerland to showcase the practical challenge of ABN modelling. Moreover, we will emphasize on the practical need to account for uncertainty in the final reported structure through Bayesian Model Averaging (BMA).

Lieu

Uni Mail & Online

Boulevard du Pont-d'Arve 40
1205 Geneva

Room M 5220, 5th floor

entrée libre