With any such parametric specification, the amount of smoothing performed (e.g., controlled by the parameters σ2 and τ2) is affected globally by all the areas and is not adaptive. Concerns that such parametric models could oversmooth have led several authors to develop semiparametric spatial models that replace the continuously varying spatial distribution for {θi} by discrete allocation or partition models. Such models allow discontinuities in the risk surface and make fewer distributional assumptions. Partition models that allow a variable number of clusters have been proposed by Denison and Holmes (2001) and Knorr-Held and Rasser (2000).