2812

Accurate prediction models for ambient air pollution concentrations are a key component of modern air pollution epidemiology, and epidemiological studies regarding long-term effects of ozone are urgently needed. Large-scale models are necessary for cohort studies with national recruitment, and models are also needed to accurately characterize fine-scale spatial gradients in a region. We applied a spatiotemporal modeling approach, previously used in individual cities for ozone and nationally for other pollutants, to predict weekly ambient ozone concentration over the contiguous United States for the period from 2005 through 2014. Methods: Using long-term regulatory ozone monitoring data collected daily (n=669 sites) and short-term spatially-rich cohort-specific data collected as two-week time integrated averages (n=1958 sites) divided into nine regions across the country, we fit a regionalized likelihood-based spatiotemporal model using land-use regression on dimension reduced land-use covariates combined with universal kriging. Ten-fold cross-validated predictions, leaving out sites, were used to compute R² (from root mean-squared error) in three different ways: spatial R² to assess accuracy in predicting long-term average exposures at locations without data, temporal R² to assess accuracy in predicting the time series of weekly exposures at locations without data, and spatiotemporal R² to assess overall prediction accuracy. Results: Overall cross-validated performance was high. Spatial R² was 0.69, spatiotemporal R² was 0.78, and median site-specific temporal R² was 0.80. Spatial performance varied by modeling region and season. Conclusion: We report a novel, successful national-scale spatiotemporal model for outdoor ozone concentrations for the contiguous U.S. Strengths of this model include its national scale and its ability to make accurate fine-scale predictions spatially at all arbitrary points and temporally at one-week intervals, suggesting future usefulness to a wide variety of epidemiological studies.