Performance of Prediction Algorithms for Modeling Outdoor Air Pollution Spatial Surfaces

Abstract

Land use Regression (LUR) models for air pollutants are often developed using multiple linear regression techniques. However, in the past decade linear (stepwise) regression methods have been criticized for their lack of flexibility, their ignorance of potential interaction between predictors and their limited ability to incorporate highly correlated predictors. We used two training sets of ultrafine particles (UFP) data (mobile measurements (8200 segments, 25 seconds monitoring per segment), and short-term stationary measurements (368 sites, 3x30 minutes per site)) to evaluate different modelling approaches to estimate long-term UFP concentrations by estimating precision and bias based on an independent external dataset (42 sites, average of three 24-hour measurements). Higher training data R2 did not equate to higher test R2 for the external long-term average exposure estimates, arguing that external validation data are critical to compare model performance. Machine learning algorithms trained on mobile measurements explained only 38-47% of external UFP concentrations whereas multivariable methods like stepwise regression and elastic net explained 56-62%. Some machine learning algorithms (bagging, random forest) trained on short term measurements explained modestly more variability of external UFP concentrations compared to multiple linear regression and regularized regression techniques. In conclusion, differences in predictive ability of algorithms depend on the type of training data and are generally modest.