|
Abstract:
|
We consider the problem of modeling realizations of a semicontinuous response Y(x) at sampling locations x. A semicontinuous response typically has a portion of responses equal to zero and a continuous distribution for the positive values. Following Diggle et al. (1998), conditional on an unobserved stationary zero-mean Gaussian process S(x), Y(x) forms a generalized linear model with S(x) appearing as an offset term in the linear predictor. We break down the response into an incidence part U and a positive part V, then specify a two-part model, U~f{u|S(x)} and V~g{v|Z(x)}, where S(x) and Z(x) are separate but related zero-mean spatial processes with a given cross-covariance structure. We describe a new and improved fully Bayesian inferential approach for this model, focusing in particular on the choice of priors and sampling algorithms.
|
