Disease mapping using empirical Bayes and Bayes methods on mortality statistics in The Netherlands.

Data from the 64 public health service districts in the Netherlands, describing the health status of the Dutch population, were used to construct maps for several causes of deaths. The choice and estimation of the relative risk measure is described. The prior expected number of deaths was modelled using a Poisson regression approach based on a model with main effects of district and age. Desirable properties of risk parameters for disease mapping are that they both reflect the level of risk and cope with the instability in the observed measure caused by the numbers at risk in each district. Different techniques for estimation of parameters were applied: empirical Bayes estimation (EB) using a nonparametric prior and a gamma prior, and a Bayesian approach (B) with a uniform prior. For the parametric EB also a constrained estimator was used. The EB techniques studied in this paper shift or smooth the values of the risk parameter towards a global mean. In the Bayesian method applied here, spatial dependence among districts can be modelled, that is the estimates are smoothed towards a local mean. The three EB estimates gave by and large similar results, although the constrained EB estimate smoothed less, as was expected. The Bayesian estimates smoothed the estimates more or less similarly to the constrained EB.