Cluster detection diagnostics for small area health data: with reference to evaluation of local likelihood models

Abstract

The focus of this paper is the development of a range of cluster detection diagnostics that can be used to assess the degree to which a clustering method recovers the true clustering behaviour of small area data. The diagnostics proposed range from individual region specific diagnostics to neighbourhood diagnostics, and assume either individual region risk as focus, or concern areas of maps defined to be clustered and the recovery ability of methods. A simulation-based comparison is made between a small set of count data models: local likelihood, BYM and Lawson and Clark. It is found that local likelihood has good performance across a range of criteria when a CAR prior is assumed for the lasso parameter.