A comparison of three empirically based, spatially explicit predictive models of residential soil Pb concentrations in Baltimore, Maryland, USA: understanding the variability within cities

Abstract

In many older US cities, lead (Pb) contamination of residential soil is widespread; however, contamination is not uniform. Empirically based, spatially explicit models can assist city agencies in addressing this important public health concern by identifying areas predicted to exceed public health targets for soil Pb contamination. Sampling of 61 residential properties in Baltimore City using field portable X-ray fluorescence revealed that 53% had soil Pb that exceeded the USEPA reportable limit of 400ppm. These data were used as the input to three different spatially explicit models: a traditional general linear model (GLM), and two machine learning techniques: classification and regression trees (CART) and Random Forests (RF). The GLM revealed that housing age, distance to road, distance to building, and the interactions between variables explained 38% of the variation in the data. The CART model confirmed the importance of these variables, with housing age, distance to building, and distance to major road networks determining the terminal nodes of the CART model. Using the same three predictor variables, the RF model explained 42% of the variation in the data. The overall accuracy, which is a measure of agreement between the model and an independent dataset, was 90% for the GLM, 83% for the CART model, and 72% for the RF model. A range of spatially explicit models that can be adapted to changing soil Pb guidelines allows managers to select the most appropriate model based on public health targets.