B0971
Title: Computationally efficient localized spatial smoothing of disease rates using anisotropic basis functions
Authors: Duncan Lee - University of Glasgow (United Kingdom) [presenting]
Abstract: The data used to quantify the spatial variation in disease rates relate to a set of areal units, and conditional autoregressive priors are applied to a set of random effects to model this spatial variation. These priors force all pairs of neighbouring areal units to exhibit correlated disease rates. However, disease rate surfaces are likely to contain boundaries, which are locations where neighbouring areal units exhibit a step-change in disease rates. A small body of work has extended CAR-type models to facilitate localised smoothing that accounts for these boundaries, but they would be computationally prohibitive for big spatial data. Therefore, motivated by a new study of mental ill health across the $N=32,754$ lower super output areas in mainland England, a computationally efficient approach is proposed for localised spatial smoothing for big spatial data. A set of anisotropic spatial basis functions is first created on the geodesic distances between all pairs of areal units and a second set of ancillary data. These basis functions are then included in a generalised linear model framework with a ridge regression shrinkage penalty to prevent overfitting. The efficacy of the approach is evidenced by simulation, before using it to identify the highest risk areas and the magnitude of the health inequalities in four measures of mental ill health, namely antidepressant usage, benefit claims, depression diagnoses and hospitalisations.
