Mixtures of spatial and unstructured effects for spatially discontinuous health outcomes

This paper proposes mixture models for spatially adaptive smoothing of health event data (e.g. mortality or illness totals). Such models allow for spatial pooling of strength but adopt a mixture strategy that also reflects health risks that are discordant with those of surrounding areas. Mixing is either discrete or based on beta densities. A fully Bayesian estimation and specification strategy is applied with fit based on DIC and BIC criteria. Illustrative applications are to long term illness in 133 London small areas, where event counts are large, and to lip cancer in Scottish counties where the majority of event totals are under 10.