The spatial structure of chronic morbidity: evidence from UK census returns

BACKGROUND Disease prevalence models have been widely used to estimate health, lifestyle and disability characteristics for small geographical units when other data are not available. Yet, knowledge is often lacking about how to make informed decisions around the specification of such models, especially regarding spatial assumptions placed on their covariance structure. This paper is concerned with understanding processes of spatial dependency in unexplained variation in chronic morbidity. METHODS 2011 UK census data on limiting long-term illness (LLTI) is used to look at the spatial structure in chronic morbidity across England and Wales. The variance and spatial clustering of the odds of LLTI across local authority districts (LADs) and middle layer super output areas are measured across 40 demographic cross-classifications. A series of adjacency matrices based on distance, contiguity and migration flows are tested to examine the spatial structure in LLTI. Odds are then modelled using a logistic mixed model to examine the association with district-level covariates and their predictive power. RESULTS The odds of chronic illness are more dispersed than local age characteristics, mortality, hospitalisation rates and chance alone would suggest. Of all adjacency matrices, the three-nearest neighbour method is identified as the best fitting. Migration flows can also be used to construct spatial weights matrices which uncover non-negligible autocorrelation. Once the most important characteristics observable at the LAD-level are taken into account, substantial spatial autocorrelation remains which can be modelled explicitly to improve disease prevalence predictions. CONCLUSIONS Systematic investigation of spatial structures and dependency is important to develop model-based estimation tools in chronic disease mapping. Spatial structures reflecting migration interactions are easy to develop and capture autocorrelation in LLTI. Patterns of spatial dependency in the geographical distribution of LLTI are not comparable across ethnic groups. Ethnic stratification of local health information is needed and there is potential to further address complexity in prevalence models by improving access to disaggregated data.