Epidemiological studies increasingly rely on exposure prediction models. Predictive performance of satellite data has not been evaluated in a combined land-use regression/spatial smoothing context. We performed regionalized national land-use regression with and without universal kriging on annual average NO2 measurements (1990-2012, contiguous U.S. EPA sites). Regression covariates were dimension-reduced components of 418 geographic variables including distance to roadway. We estimated model performance with two cross-validation approaches: using randomly selected groups and, in order to assess predictions to unmonitored areas, spatially clustered cross-validation groups. Ground-level NO2 was estimated from satellite-derived NO2 and was assessed as an additional regression covariate. Kriging models performed consistently better than nonkriging models. Among kriging models, conventional cross-validated R2 (R2cv) averaged over all years was 0.85 for the satellite data models and 0.84 for the models without satellite data. Average spatially clustered R2cv was 0.74 for the satellite data models and 0.64 for the models without satellite data. The addition of either kriging or satellite data to a well-specified NO2 land-use regression model each improves prediction. Adding the satellite variable to a kriging model only marginally improves predictions in well-sampled areas (conventional cross-validation) but substantially improves predictions for points far from monitoring locations (clustered cross-validation). (Figure Presented).
Bibliographical noteFunding Information:
Supported by the National Institute of Environmental Health Sciences (NIEHS) F31 Predoctoral Fellowship (1F31ES025096-01), the NIEHS Biostatistics, Epidemiologic, and Bioinformatic Training in Environmental Health Training Grant (T32ES015459), and the Environmental Protection Agency (RD831697, MESA Air).