TITLE:
National Prediction of Ambient Fine Particulates: 2000-2009
AUTHORS:
David J. Shavlik, Sam Soret, W. Lawrence Beeson, Mark G. Ghamsary, Synnove F. Knutsen
KEYWORDS:
Long-Term Air Pollution, GAM, Prediction, Fine Particulates
JOURNAL NAME:
Open Journal of Air Pollution,
Vol.5 No.3,
September
1,
2016
ABSTRACT: A large
body of evidence links ambient fine particulates (PM2.5) to chronic
disease. Efforts continue to be made to improve large scale estimation of this
pollutant for within-urban environments and sparsely monitored areas. Still
questions remain about modeling choices. The purpose of this study was to
evaluate the performance of spatial only models in predicting national monthly
exposure estimates of fine particulate matter at different time aggregations
during the time period 2000-2009 for the contiguous United States. Additional
goals were to evaluate the difference in prediction between federal reference
monitors and non-reference monitors, assess regional differences, and compare
with traditional methods. Using spatial generalized additive models (GAM),
national models for fine particulate matter were developed, incorporating
geographical information systems (GIS)-derived covariates and meteorological
variables. Results were compared to nearest monitor and inverse distance
weighting at different time aggregations and a comparison was made between the
Federal Reference Method and all monitors. Cross-validation was used for model
evaluation. Using all monitors, the cross-validated R2 was 0.76,
0.81, and 0.82 for monthly, 1 year, and 5-year aggregations, respectively. A
small decrease in performance was observed when selecting Federal Reference
monitors only (R2 = 0.73, 0.78, and 0.80 respectively). For Inverse
distance weighting (IDW), there was a significantly larger decrease in R2 (0.68, 0.71, and 0.73, respectively). The spatial GAM showed the weakest
performance for the northwest region. In conclusion, National exposure
estimates of fine particulates at different time aggregations can be
significantly improved over traditional methods by using spatial GAMs that are
relatively easy to produce. Furthermore, these models are comparable in
performance to other national prediction models.