Estimating Exposure Using Kriging: a Simulation Study

Retrospective studies of disease often are limited by the resolution of the exposure measurements. For example, in a typical study of adverse health effects from contaminated groundwater, the number of wells sampled may range from only a few to as many as several dozen, while the number of cases and controls may be in the hundreds or more. To derive individual estimates of exposure for wells that were not sampled, investigators must extrapolate. In this study, we compare three methods of extrapolating from a limited number of observations to estimate individual exposures. Using two naive models of groundwater contamination, we compare nearest neighbor interpolation, inverse distance squared weighting, and kriging for estimating exposure based on a limited number of measurements. Our results show that although kriging is a statistically optimal method, it is not markedly better than simpler interpolation algorithms, though it is considerably more complex to use. Aberrant well measurements and discontinuities are problematic for all methods. We provide some guidance in interpolating data and outline a more comprehensive comparison of methodology.