Title: Cluster detection of disease mapping data based on latent Gaussian Markov random field models
Authors: Wataru Sakamoto - Okayama University (Japan) [presenting]
Abstract: Detecting clusters of higher prevalence in spatial data is of primary interest. Most of the existing methods use spatial scan statistics based on the likelihood-ratio test. The echelon scan based on the echelon analysis is useful in detecting significant clusters of non-circular shape effectively. Bayesian analysis methods for spatial data have been also studied. The latent Gaussian models, in which a Gaussian Markov random field prior is assumed on the spatial effect, provide very flexible tools. A method of detecting clusters was proposed using Poisson models with the latent Gaussian Markov random field. The clusters are scanned on the echelons constructed from the posterior means of the spatial effect, and the clusters giving maximum marginal likelihood were detected. It can be easily extended to adjustment for covariates and the random effect. An example of applying to disease mapping data illustrated that the proposed method constructed more aggregated echelons and clusters than the echelon scan based on empirical Bayes estimates of relative risk, and that detected clusters provided smallest deviance information criterion values.