Disease cluster detection: a critique and a Bayesian proposal.

This paper reviews issues in the analysis of non-focussed clustering, and proposes a novel approach to cluster modelling that can be used in a surveillance context. The novel approach involves the use of local likelihood models for the analysis of clustering in small area health data. Local likelihood is used when interdependence between data events at locations is modelled directly, as opposed to the modelling of a hidden process of cluster centres. This approach allows the use of conventional posterior sampling. It also allows a less parameterized approach to the form of clusters detected. The idea of a spatially dependent lasso which provides the local maxima for the aggregation of locations is considered as an approximation. The methods are applied to a well known data set and compared with Satscan, and a conditional logistic Bayesian model.