Spatial modeling with INLA for analysis of unequal care in Skåne

Abstract: The objective of this thesis is to extend on a previous analysis of health care accessibility for patients diagnosed with a chronic disease in Region Skåne. The previous analysis resulted in a logistic mixed effects model having municipality as a random effect and age as a first-degree spline-function. This thesis extends on the random effects from the previous model in order to analyse the spatial dependencies on municipal and postal-code spatial levels. The models being compared are Bayesian structured additive regression models with latent Gaussian Markov Random Fields. The spatial dependencies are modeled using a Conditional Autoregressive model, and a Random Walk is used to approximate a spline-function for age in this framework. To perform approximate Bayesian inference Integrated Nested Laplace Approximation (INLA) is used. It is shown that both on a municipal and postal-code level a Random Walk of order two is preferred for approximating the spline-function. The difference lies in the spatial dependencies, where on municipal level modeling them as i.i.d. is sufficient, which is comparable to the previous analysis. Regarding spatial dependencies with more intricate geographic boarders, such as on the postal-code level, modeling using a Conditional Autoregressive model is preferred.