The Dynamic Coregionalization Model in Air Quality Modeling and Risk Assessment
Francesco Finazzi1, Marian E. Scott2, Alessandro Fassò1
1Department of Information Technology and Mathematical Methods, University of Bergamo, Dalmine, Bergamo, Italy; 2Department of Statistics, University of Glasgow, Glasgow, United Kingdom

One major role of environment agencies is to provide concise indicators about a country's air quality and its impact on population health. On the one hand, such environmental indicators must be easily understood by the public and on the other, they should be useful to compare air quality over different years. Indeed, it is important to assess whether any actions undertaken to improve air quality have been successful or not (Scott, 2007).

When the sets of airborne pollutants measured at different monitoring sites are different, it is not always clear how to define daily and yearly indicators for the whole country or how to evaluate their uncertainty. Moreover, it is not straightforward to compare across years when in each year, the structure of the monitoring network (sites included), the quantity of missing data and meteorological conditions differ.

In order to define a sound statistical framework within which to address the above issues, the dynamic coregionalization model introduced by Fassò and Finazzi (2010 and 2011) is considered. Based on temporal and spatial latent variables, the model is a multivariate hierarchical space-time model able to cope with missing data and with data coming from heterogeneous monitoring networks, where different pollutants may be measured at different sites.

When only the temporal latent variable is considered, the model can be used to derive a daily air quality indicator for the entire region of interest. If different pollutants have different impacts on population health, then a weighted indicator can be evaluated. When a set of covariates and the latent spatial variable are also included, the model can provide either daily or yearly average concentration maps for the pollutants considered.

By applying non-parametric bootstrap techniques, quantities such as the probability that a pollutant exceeds a threshold level L over a period of time at a particular point in space, the expected number of days that L has been exceeded during the year and the probability that L has been exceeded for at least N days can be evaluated. These results can then be related to population figures in order to assess population exposure and risk. This approach is illustrated using Scottish air quality data for 6 pollutants for the year 2009.


Fassò A. and Finazzi F. (2010) The dynamic coregionalization model with application to air quality remote sensing. A. Bowman (Ed.). Proceedings of International Workshop on Statistical Modelling 2010. Glasgow, July 5-9th 2010, pp 195-200. On-line version: Graspa-Working paper n.42, on

Fassò A. and Finazzi F. (2011) Maximum likelihood estimation of the dynamic coregionalization model with heterotopic data. Accepted by Environmetrics.

Scott E M (2007). Setting and evaluating the effectiveness of environmental policy. Environmetrics 18(3), 333-343.

Keywords: Air quality indicators; Multivariate space-time models; Exposure and risk assessment; Expectation-maximization algorithm

Biography: Francesco Finazzi is research fellow at the department of Information Technology and Mathematical Methods of the University of Bergamo, Italy. His main research topics are multivariate space-time models, remote sensing data analysis, air quality assessment and parallel computation for large datasets.