Sébastien Coube: "Bayesian hierarchical space-time models with nonseparability and nonstationarity, for air pollution exposure assessment"

Data i lloc: Dimarts 9 de maig de 2023 · 12:00 h · Saló de Graus de la Facultat de Ciències Matemàtiques

Resum: The starting point of this talk is a project at the BCAM aiming to assess exposure to air pollution in Euskadi, both during peaks (leading to admissions in hospitals) and in the long term (leading to chronic diseases). Several pollutants (fine particles, Nitrogen Oxides, Sulfur Dioxide) are monitored. The method I chose is to use a Bayesian space-time model for prediction purposes.
This work aims to combine 3 ingredients which should allow for better predictions and
interpretations: Nonseparability, Multivariablity and Nonstationarity with moderation.
However, this project is a computational Pandora's box :
1. A data set with modest marginal dimensions can have a mean total number of observations.
2. It is needed to assess the strengths of space, time and inter-variate links, the model will be dense in the Markovian sense.
3. Auto-correlation within the model is a problem.
4. The parallelizability of the algorithms is a critical point to provide scalable methods.
This talk will present the model and outline the methods used to deal with the computational difficulties I met until now.