Edge effect in disease mapping

In mapping geographical variation of risk or occurrence of disease in tract count data, information from the neighbouring areas is used to improve the estimate for a given area. However such information is incomplete for those areas at the region boundary and ignoring this could lead to a distortion of the estimates (edge-effect). We present a general approach to compensate for edge effects under a hierarchical Bayesian model. In our model the out-of-the-border areas are regarded as having missing values of the number of events but known population at risk. Two algorithms for the imputation of missing data and estimation of relative risks are applied: doubly stochastic EM and Chained Data Augmentation. Results on simulated data as well as a real example on mortality for gastric cancer in Tuscany (Italy) 1981-88 are discussed.