Estimating Spatial Intensity and Variation in Risk from Locations Subject to Geocoding Errors

The accurate assignment of geocodes to the residences of subjects in a study population is an important component of the data acquisition/assimilation stage of a spatial epidemiological investigation. Unfortunately, however, it is not a simple matter to obtain accurate point-level geocodes. Recent investigations have demonstrated that when residential address geocoding is performed by the most common method of street-segment matching to a georeferenced road file and subsequent interpolation, positional errors of hundreds of meters are commonplace, especially in rural locations. Ignoring these errors in a statistical analysis may lead to biased estimators, a reduction in power, and incorrect conclusions. This article modifies some existing likelihood-based procedures for estimating the intensity and relative risk of Poisson spatial point processes from locations ascertained without error, so as to permit valid inferences to be made from locations observed with error. The superior performance of the modified methods compared to methods that ignore positional errors is demonstrated by simulation.