Potential Overfitting in a Spatio-Temporal Exposure Model Developed with Few Monitoring Sites

Overfitting is a concern with air pollution estimation models that rely on a limited number of monitoring sites. Here we develop spatio-temporal air pollution exposure prediction models using land use regression in a universal kriging framework with temporally rich but spatially poor monitoring data. Two-week average concentrations for six criteria pollutants (PM10, PM2.5, SO2, NO2, ozone and CO) from 2014 through 2017 were obtained from 23 administrative monitoring sites in the urban area in Beijing. A large array of geographic covariates (GC) was collected. Exposure prediction models for the air pollutants involved three steps: (1) deriving the smoothed temporal trends from a singular value decomposition; (2) reducing the dimensionality of the collected GC using partial least square regression (PLS); (3) developing spatio-temporal models based on the previous smoothed temporal trends and GC-related PLS scores. Leave-one-out cross validation (LOOCV) was used for prediction accuracy at each of the three s...