University of Washington - Department of Statistics
We consider the problem of detecting clusters in space of non-infectious and rare diseases. Cluster detection is the routine surveillance over a large expanse of small administrative regions to identify individual "hot-spots" of elevated residual spatial risk of disease without any preconceptions about their locations. It is vital to assess the sensitivity and specificity of any proposed method as there are consequences associated with too many false alarms and missed clusters.
A class of cluster detection methods known as moving-window methods superimpose a large number of circular regions onto the study region. For each of these circular regions, we determine the significance of any evidence of elevated risk, with those that are deemed significant being flagged as potentially containing a cluster. However, many of the established methods suffer from two drawbacks. First, they rely on Neyman-Pearson tests of the same pre-specified size alpha (e.g. 0.05) for all circular regions regardless of the population inside, which has repercussions on the balance of sensitivity and specificity. Second, given the large number of circular regions which often overlap, multiple testing issues arise.
As a solution to these drawbacks, we propose a Bayesian decision theory approach to cluster detection, which incorporates a loss function and prior knowledge, where regions will be flagged based on Bayes rules. In this presentation, we outline the framework for identifying elevated risk of disease in a single region and address the issues involved with conducting surveillance over multiple regions at once. Further research is needed to extend the methodology to incorporate longitudinal data and consider clusters of multiple cancers. We will eventually apply the methodology to the Surveillance, Epidemiology and End Results (SEER) program database of cancers for the 13 counties in Western Washington for the years 1995-2005.