Nonparametric Bayesian analysis of Korean elderly suicide rates from small areas

Abstract

Disease mapping is an area of epidemiology concerned with estimating the spatial pattern in disease risk over a geographical region. Typically, the study area is partitioned into spatial districts (small areas), with the disease map displaying the disease incidence associated with each district. We consider Bayesian nonparametric small-area models with a Dirichlet process and a simultaneous conditional autoregressive model for the spatial correlation between neighboring areas to estimate the standardized mortality ratio and mapping. The observed mortality rates are unstable for regions with rare events such as cancer. Bayesian methods, similar to the proposed small-area analysis method, provide mechanisms for smoothing the estimates of rates by region. We compare nonparametric Bayesian models using a Dirichlet process prior with the parametric Bayesian models using a gamma prior for small-area estimation.