A Flexible Spatio-Temporal Model for Air Pollution with Spatial and Spatio-Temporal Covariates

Area of Science:

• Environmental Science
• Epidemiology
• Data Science

Background:

• Accurate spatio-temporal air pollution prediction is vital for public health.
• Existing models often lack precision at small spatial scales.
• The Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air) requires reliable exposure data.

Purpose of the Study:

• To develop and validate a spatio-temporal framework for predicting ambient air pollution.
• To assess the model's performance using nitrogen oxides (NOx) in Los Angeles.
• To compare the predictive accuracy of geographic covariates versus dispersion model outputs.

Main Methods:

• Integrated data from monitoring networks and a deterministic air pollution model (Caline3QHCR).
• Utilized geographic information system (GIS) covariates.
• Implemented the framework in an R package (SpatioTemporal) and employed cross-validation for accuracy assessment.
• Evaluated NOx concentrations over a ten-year period in Los Angeles.

Main Results:

• Achieved good predictive ability with cross-validated R-squared of approximately 0.7.
• Replacing geographic traffic indicators with Caline3QHCR output yielded similar accuracy.
• A more parsimonious and interpretable model was achieved without sacrificing predictive power.
• Adding traffic-related geographic covariates did not further improve prediction accuracy.

Conclusions:

• The developed spatio-temporal framework provides accurate air pollution predictions.
• The R package SpatioTemporal is a valuable tool for environmental health research.
• Dispersion model outputs can effectively substitute for multiple geographic covariates in air pollution modeling.