Abstract
The detection of regions with unusually high risk plays an important role in disease mapping and the analysis of public health data. In particular, the detection of groups of areas (i.e., clusters) where the risk is significantly high is often conducted by public health authorities.
Many methods have been proposed for the detection of these disease clusters, most of them based on moving windows, such as Kulldorff's spatial scan statistic. Here we describe a model-based approach for the detection of disease clusters implemented in the DClusterm package. Our model-based approach is based on representing a large number of possible clusters by dummy variables and then fitting many generalized linear models to the data where these covariates are included one at a time. Cluster detection is done by performing a variable or model selection among all fitted models using different criteria.
Because of our model-based approach, cluster detection can be performed using different types of likelihoods and latent effects. We cover the detection of spatial and spatio-temporal clusters, as well as how to account for covariates, zero-inflated datasets and overdispersion in the data.