Hierarchical Bayesian spatio-temporal point level modelling with an application to estimating exposure to air pollution exposure

Professor Sujit Sahu, University of Southampton. Part of the Statistics seminar series.

Spatio-temporal modelling is an essential data science tool to estimate quantities of interest which vary in both space and time. Performed hierarchically in the Bayesian paradigm such modelling enables us to incorporate information from disparate sources into the modelling which results in more accurate inference and prediction. Implemented using computational Bayesian methods such as MCMC, these models also deliver inference and prediction at any coarser geographical and temporal scale than the scale at which data has been observed.

In this talk we demonstrate the use of the Bayesian modelling and computation for estimation of long term exposure to air pollution levels over all of England and Wales in the UK. Using observed data from a very sparse network of monitoring sites and output of an atmospheric air quality dispersion model developed recently especially for the UK we obtain empirically verified accurate maps of air pollution aggregated upto coarser level geographies such as postcode areas and local authority areas. Land use information, incorporated as a predictor in the model, further enhances the accuracy of the space-time model. 

These estimates for aggregated administrative areas can readily be used for many purposes such as modelling of aggregated health outcome data. This talk illustrates the use of the air pollution estimates in assessing their health effects for respiratory diseases where hospitalisation were necessary for the five year period 2007-2011 in England.