Title: Bayesian clustering-based adjacency modelling in disease mapping
Authors: Xueqing Yin - University of Glasgow (United Kingdom) [presenting]
Duncan Lee - University of Glasgow (United Kingdom)
Gary Napier - University of Glasgow (United Kingdom)
Craig Anderson - University of Glasgow (United Kingdom)
Abstract: Conditional autoregressive (CAR) models are the most common modelling approaches in disease mapping to quantify the spatial pattern in disease risk across $n$ areal units. In these models the spatial autocorrelation is typically induced by a $(n\times n)$ binary neighbourhood matrix based on the sharing a common border specification, such that spatial correlation is always enforced between geographical neighbours. However, geographically adjacent areas may sometimes exhibit step changes in risk due to factors such as population behaviours and socio-economic deprivation. Therefore, we propose a novel methodology to account for these step changes via a two-stage modelling approach. In stage one we produce a set of candidate neighbourhood matrices via a variety of common clustering methods. In the second stage, an appropriate spatial autocorrelation structure is estimated by estimating the neighbourhood matrix as part of a hierarchical Bayesian spatio model. The proposed model yields improved risk estimation and simultaneously identifies clusters of areas exhibiting elevated or reduced risks. The effectiveness of the methodology is evidenced by a simulation study, and the methodology is motivated by a study of respiratory disease risk in Greater Glasgow, Scotland, in 2016.