Identifying Lead Service Lines with Field Tap Water Sampling

Abstract

Cross-sectional associations between tap water lead levels and service line materials are not fully understood. As a result, stakeholders are unable to make use of routinely collected tap water samples for myriad decisions. This analysis leverages a novel data set associating field tap water observations with service line materials under varying seasonal and sampling conditions. The analysis demonstrates a staged approach that leverages interpretable model results to strategically filter the sample to boost material diagnoses. Predicted precisions guided by interpretable model results easily exceed those from an incumbent black box, Random Forest model. For example, precisions of 94% were achieved for samples collected between July and October, when lead more naturally leaches. Precisions for the 7% of households with the highest predicted probabilities exceed 96%, which could help municipalities locate lead service lines. Models were also used to inform customers about their risks, predicting lead with a 95% probability when either a single sample exceeds 20 ppb or multiple samples exceed detection thresholds. These results should help stakeholders make better use of tap water samples for risk mitigation and regulatory compliance. Similarly, the staged approach, using interpretable model results to guide classification, can support other water research domains.