|
Abstract:
|
The health impacts of many factors are routinely estimated using spatial areal unit data, including air pollution and green space. The disease data are counts of disease cases in each areal unit, and are modelled using Poisson log-linear models with known covariates and random effects (RE). The latter account for any residual spatial autocorrelation in the data, and are typically modelled by globally smooth conditional autoregressive (CAR) priors as part of a Bayesian model. However, such globally spatially smooth RE are likely to be inappropriate, because they may be collinear to any spatially smooth covariate, and the residual autocorrelation is unlikely to be globally spatially smooth after removing such covariate effects. This talk proposes an extension to CAR models to allow for localised spatial smoothing, so that random effects between some pairs of geographically adjacent areal units are correlated while between other pairs conditional independence is assumed. The model is tested by simulation and shown to outperform globally smooth CAR models, before being applied to data on air pollution and health in the UK.
|
