Examining the effect of residual spatial autocorrelation on fixed effect estimation

Estimation of fixed effects in spatial data sets can be challenging, as spatial autocorrelation can occur in the residuals as well as the covariates. The residual spatial autocorrelation can be caused by spatially autocorrelated risk factors for the response data that are unknown or unmeasured, and leads to unmeasured confounding. Spatial regression models have been developed to allow fixed effect estimation whilst accounting for residual spatial autocorrelation, and three of these methods have been compared here through a simulation study along with a method which ignores the spatial autocorrelation. The aim of this thesis is thus to determine if accounting for the spatial autocorrelation produces better results in terms of fixed effect estimation, and if so which method is the best. These aims are first examined through simulation studies, and then the methods are applied to a study of air pollution and respiratory illness hospital admissions in the central belt of Scotland in 2010. The analysis shows that higher concentrations of particulate matter air pollution result in an increased risk of hospital admission due to respiratory illness