Abstract

Small area estimation has specific challenges and features and Bayesian disease mapping models have been widely used to address some of these. Bayesian models inherently borrow information from larger denser regions to infer prevalence or risk for smaller sparser regions. Most Bayesian models, however, assume second-order stationarity like the well-known ICAR-type model for example the Besag, BYM and LeRoux models. In this work we propose a BYM-type model that accounts for second-order non-stationarity in a principled way, and we enable fast inference through the INLA methodology. We show that accounting for covariance non-stationary is critical for valid inference in small area estimation.