"Smooth-CAR mixed models for spatial count data"

Monday, September 28, 2010, 12.00 - 16.00 PM
Erasmus Medical Center 21st floor, room Ee 21-69

by Dae-Jin Lee, Universidad Carlos III de Madrid

Abstract
Areal data are very common in disease mapping applications. Usually this type of data are units such as counties, states or provinces, where the number of disease counts are aggregated. Disease counts are assumed to be Poisson distributed. We propose the use of Penalized Likelihood splines (P-splines) and individual random effects for the analysis of spatial count data. P-splines are represented as mixed models to give a unified approach to the model estimation procedure. First, we model the spatial variation by two-dimensional P-splines at the centroids of the areas or regions. In addition, individual area-effects are incorporated as random effects to account for individual variation among regions. Finally, we extend the model by considering a conditional autoregressive (CAR) structure for the random effects, these are the so-called Smooth-CAR models, with the aim to separate the large-scale spatial trend and the local area-level spatial correlation. We apply the methodology proposed to the analysis of lip cancer incidence rates in Scotland.