Childhood lead exposure through drinking water has long-term effects on cognition and development, and is a significant public health concern. The comprehensive lead testing of public schools entails high expense and time. In prior work, random forest modeling was used successfully to predict the likelihood of lead contamination in the drinking water from schools in the states of California and Massachusetts. In those studies, data from 70% of the schools was used to predict the probability of unsafe water lead levels (WLLs) in the remaining 30%. This study explores how the model predictions degrade, as the training dataset forms a progressively smaller proportion of schools. The size of the training set was varied from 80% to 10% of the total samples in four US states: California, Massachusetts, New York, and New Hampshire. The models were evaluated using the precision-recall area under curve (PR AUC) and area under the receiver operating characteristic curve (ROC AUC). While some states required as few as 10% of the schools to be included in the training set for an acceptable ROC AUC, all four states performed within an acceptable ROC AUC range when at least 50% of the schools were included. The results in New York and New Hampshire were consistent with the prior work that found the most significant predictor in the modeling to be the Euclidean distance to the closest school in the training set demonstrating unsafe WLLs. This study further supports the efficacy of predictive modeling in identifying the schools at a high risk of lead contamination in their drinking water supply, even when the survey data is incomplete on WLLs in all schools.
Keywords: drinking water; environmental justice; lead; machine learning; open-source data mining.