Statistical relationship between hardness of drinking water and cerebrovascular mortality in Valencia:: a comparison of spatiotemporal models

The statistical detection of environmental risk factors in public health studies is usually difficult due to the weakness of their effects and their confounding with other covariates. Small area geographical data bring the opportunity of observing health response in a wide variety of exposure values. Temporal sequences of these geographical datasets are crucial to gaining statistical power in detecting factors. The spatiotemporal models required to perform the statistical analysis have to allow for spatial and temporal correlations, which are more easily modelled via hierarchical structures of hidden random factors. These models have produced important research activity during the last decade, the most promising line being Bayesian analysis via Markov chain Monte Carlo. Practical issues without a clear answer still remain. Model comparison is one of those with the most practical relevance. In this article we perform the analysis of hierarchical spatiotemporal models to evaluate the protective effect of calcium and magnesium in drinking water against cerebrovascular mortality. Our data register yearly mortality counts from 1990 to 1995 in a municipality scale but we are not sure which temporal aggregation would be adequate. Many small counts might introduce computational artifacts, leading to inaccurate conclusions. We compare hierarchical models of diverse structure for different temporal aggregation settings via pseudo-Bayes factor and deviance information criteria. We carry out a sensitivity analysis of our main conclusions to ascertain their validity independently of the arbitrary degree of data aggregation and model structure. Copyright (C) 2003 John Wiley Sons, Ltd.