Modeling Elevated Children’s Blood Lead Levels across Five U.S. States: A Statistical Approach to Improve Predictions and Understand Key Drivers

Gaps remain in the interpretability and real-world applicability of children's elevated blood lead level (EBLL) predictive models, which are important for fostering efficient health intervention. We produced predictive statistical models using publicly available population data and children's EBLL rates (≥5 or ≥3.5 μg/dL) at the United States (U.S.) census tract resolution from Michigan (2014-2016), Ohio (2014-2016), California (2018-2022), Minnesota (2017-2021), and Wisconsin (2018-2021). Spatial and nonspatial random forest and logistic regression models were developed for each state to predict whether a census tract was above or below the 75th percentile EBLL rate. Spatial logistic regression models performed the best based on overall predictive performance, evaluated through a 10-fold cross-validation approach. Predictive accuracy ranged from 82% to 90%. The most influential final predictive variables for each optimal model varied, but there were consistencies across states. For example, the percent of homes built before 1940 was significant (<i>p</i> < 0.05) across all models, increasing the odds of a tract having higher EBLL rates (between 1 and 7%). By accounting for spatial autocorrelation, our modeling approach improves predictions by capturing local clustering effects. This approach can be applied elsewhere across the U.S. to inform prevention and mitigation actions and make predictions where measured EBLLs are sparse.