Spatially Structured Sparseness in Bayesian Spatial Health Modeling


Ranking, Sparseness and Clustering in Disease Mapping

Andrew B. Lawson (Medical University of South Carolina) (Speaker)
Raymond Boaz (Medical University of South Carolina)

Often geospatial disease outcomes are characterized by sparseness when viewed as a count distribution. This is sometimes called zero-inflation. Classically the models for zero inflation are of two types: zero class modelled as ‘structural’ or ‘Poisson’ or as hurdle models where the zeroes are treated completely separately from the truncated Poisson positive counts. Often the idea of structural zeroes is not made specific in that no attempt is made to assess which zeroes are in that class. In this talk I describe an approach which models the spatial structure of the structural zero class and provides estimates of the spatial distribution of the probability of being structural or not. This is applied in a spatio-temporal context. Different assumptions about the spatial prior distribution of the structural probability are compared using a Bayesian formulation. An example of spatio-temporal modeling of sudden infant death in the counties of Georgia USA will be presented.