Extending conditional autoregressive models for space time studies of air pollution and health
Alastair Rushworth (University of Glasgow)
Friday 27th June, 2014 15:00-16:00 Maths 204
Ecological-level health and air pollution studies aim to estimate the negative impact of ambient air pollution exposure on related health outcomes such as respiratory deaths. The health data are usually in the form of monthly or annualised total counts of the health outcome and correspond to each of a group of non-overlapping areal units, while air pollution data are typically in the form of fine grids of modelled values which must be first aligned with the set of areal units associated with the health outcome.
It is important to account for those other confounding factors that could contribute to the health outcome and without which, the estimated effect of air pollution may be biased. Typically many of these risk factors are unmeasured and their influence cannot be modelled directly, but they are often correlated in space and as a result, they are often accounted for by introducing a set of spatially smooth random effects often with a smoothness-inducing prior distribution. However, some recent research has shown that in general, the standard choices for smoothing the random effects may not be appropriate, such as those based on conditional autoregressive (CAR) priors. This can be because the CAR prior might imply a global level of smoothness when the unobserved confounding might exhibit more local structure, or because the CAR random effects can be csollinear with air pollution exposure which can result in variance inflation.
We describe the results of fitting a number of spatio-temporal models to long term health and air pollution data in the UK, which have been adapted from the literature designed to tackle these issues. In particular, we discuss the adequacey of some recent approaches that attempt to address the issues of avoiding collinearity issues by using orthogonal random effects, and also 'Wombling' approaches that try to explore localised structure by generalising the usual CAR priors.