Overview
This study applied random forest machine learning to publicly available school drinking water lead testing datasets from California, Massachusetts, New Hampshire, and New York to predict which untested schools are likely to have elevated lead concentrations. The study did not collect new water samples but analyzed existing state-mandated testing data. The principal predictor of a school having elevated lead (at or above 15 µg/L) was geographic distance to the nearest school with a known positive result, reflecting the documented clustering of elevated lead schools within municipalities with shared infrastructure age and plumbing material profiles. The model achieved receiver operating characteristic AUC above 0.7 with training sets as small as 10 to 50 percent of schools, depending on state.
Key numbers
Positive threshold used throughout: 15 µg/L (US EPA LCR action level for community water systems, applied to schools).
Alternative thresholds noted for context: 5 µg/L (FDA bottled water allowable), 10 µg/L (WHO provisional guideline value), 1 ppb (American Academy of Pediatrics recommended remediation level for schools).
Model performance:
- Massachusetts: ROC AUC 0.88, precision-recall AUC 0.91
- California: ROC AUC 0.77, precision-recall AUC 0.88
- New Hampshire and New York: ROC AUC above 0.7 achieved with as few as 10 percent training set
No new occurrence concentration data were generated by this study; it uses existing state-collected datasets.
Methods (brief)
Random forest classification trained on each state’s existing school lead testing dataset with leave-out cross-validation. SMOTE-Tomek resampling used for the California dataset (highly imbalanced due to predominantly compliant results post-AB 746 implementation). Key predictors: Euclidean distance to nearest school with known elevated lead, school building age, type of water system, geographic variables. Positive class defined as a school having at least one first-draw sample at or above 15 µg/L in the state dataset.
Evidence fitness
Supports: establishing that geographic clustering of elevated lead schools is a replicable finding across four states and is predictive enough to prioritize testing; demonstrating that random forest modeling can achieve useful predictive performance even with sparse state-level testing data; informing public health prioritization of school water lead testing programs.
Does not support: providing new occurrence data on lead concentrations in school drinking water; establishing the magnitude of lead exposure at untested schools; replacing systematic testing with model predictions for regulatory compliance purposes.
Limitations
This is a modeling study; it does not produce new water lead measurements and cannot substitute for systematic testing. The positive class (above 15 µg/L) was chosen for consistency with LCR compliance reporting but excludes sub-threshold lead levels that may still pose health risk (particularly given the AAP and WHO thresholds of 1 to 10 µg/L). State testing protocols differ (sample volume 250 mL to 1 L; stagnation time 6 to 18 hours), which affects comparability of the “positive” classification across states. The model’s utility depends on the completeness and representativeness of the training data; states with fewer than 10 percent of schools tested had lower predictive power.
Related evidence
Update history
No substantive edit history is available in this build. The full commit record is available in git.