|
Abstract:
|
Spatial Bayesian models are commonly used in disease mapping literature, especially the case of mapping rare events. However, little has been done to evaluate their properties. Recent work has demonstrated how the informativeness of the conditional autoregressive (CAR) model framework of Besag-York-Mollie (BYM) can be quantified relative to the number of observed events using the spatial and non-spatial variance parameters. Here, we build off that work by reparameterizing the BYM CAR model by allowing users to place priors on the model’s informativeness and the proportion of the total variability associated with the spatial dependence structure. Using county-level heart disease death data, we illustrate how users can specify priors that discourage oversmoothing in an intuitive way while still preserving the benefits of explicitly accounting for spatial structure.
|
