Assessing spatial patterns in disease rates.

We describe the empirical performance of three indices of spatial autocorrelation (Moran's I, Geary's c and a rank adjacency statistic D) in the analysis of regional cancer incidence data. Heterogeneity in regional population sizes and age structure leads to variable precision in estimated rates; the usual methods for assessing I, c and D, which ignore such heterogeneity, are shown to be liberally biased, especially for c and D. The power of these indices to detect likely disease patterns is estimated by stimulation; the power is quite variable, depending on the exact pattern assumed, although I tends to have the highest power. The null distributions appear quite robust in small samples, even when several regions have no observed case. Preliminary work on the Ontario cancer registry showed generally unimportant effects on the spatial analysis of variation in case registration rates or missing residence data.