National PM_2.5 spatiotemporal model integrating intensive monitoring data and land use regression in a likelihood-based universal kriging framework in the United States: 2000–2019

Nationwide PM_2.5 exposure models typically rely on regulatory monitoring data as the only ground-level measurements. In this study, we develop a high-resolution spatiotemporal PM_2.5 model for the contiguous United States from 2000 to 2019 with dense monitoring data at both regulatory and residential sites. Specifically, we combine publicly-available data from 1843 regulatory monitors with our own set of multiple 2-week measurements at 939 residential locations. As we show, these additional data enhance the spatiotemporal prediction capabilities of the model. The model can handle varying data densities and regional variations; it predicts two-week average PM_2.5 concentrations at fine spatial scale for the contiguous United States. Cross-validation performance indicates a spatial R² of 0.93 and a root mean square error (RMSE) of 1.19 (μg/m³), and a temporal R² of 0.85 and RMSE of 2.05 (μg/m³). Regional spatial R² ranged from 0.80 (northwest) to 0.93 (northeast and central). Over time, the average PM_2.5 across the United Stats decreased from 7.6 μg/m³ in 2000 to 4.7 μg/m³ in 2019. Our model effectively captured local PM_2.5 gradients, highlighting its ability to address fine-scale variations related to local sources and roadways.