In southeastern Michigan, several spatial models of groundwater arsenic were developed using samples from 6050 private wells, and validated [19] using samples from 371 private wells in a case-control study. A geographic model that secondarily took into account geologic formations and geographic boundaries of bedrock performed best (ρ = 0.46). Models more similar to our basic GIS linkage methods yielded precision closer in magnitude to what we observed for our small subgroup of homes on private wells. One model used mean arsenic within a township (typically 6 × 6 miles; 4.24 miles maximum from the centre), and when we focused on homes served by private wells and used a radius of 4 miles our results were nearly identical (ρ = 0.35 vs. ρ = 0.36). The authors repeated this model using a township section (1 × 1 mile; 0.7 miles maximum from the centre). As in our analysis, precision improved (ρ = 0.42) with a shorter "radius," but data were unavailable for half of homes. Also as we observed, there was little difference between this method and using only samples from the closest well (ρ = 0.35). The similarities between our results are especially interesting given that arsenic levels were greater in that study (median 2.30 μg/L; 90th percentile 22.73 μg/L).