A Bayesian hierarchical modeling approach for studying the factors affecting the stage at diagnosis of prostate cancer.

We extend the baseline-category logits model for categorical response data to accommodate two distinct kinds of clustering. Our extension introduces random effects that have one component exhibiting spatial dependence and a second component that is distributed independently. We use this enhanced categorical logits model for investigating the factors that affect the geographical distribution of the diagnostic stage of prostate cancer (PrCA) in South Carolina (SC). Using incidence data from the SC registry, we fit three types of models: the baseline-category logits model, the proportional odds model, and the adjacent-categories logits model, each incorporating our two-component random effects. The deviance information criterion (DIC) is used for selecting the best-fitting model. The results from the best model are presented and interpreted. The county-specific random effects are mapped to characterize the spatial distribution pattern of diagnostic stage of PrCA in the study region. In terms of spatial distribution of the diagnostic stage of PrCA, an area of excess (unexplained) risk was found in the north-west area, and an area of low excess risk in the north-east area for regional-stage cancer in SC was identified through the analysis of the cancer registry data. 2008 John Wiley & Sons, Ltd