Spatial Statistics Vs Machine Learning: Evaluating Air Pollution Exposure Prediction Models

• Activity Number: 349 - Longitudinal, Spatial, and Bayesian Methods
• Type: Contributed
• Date/Time: Tuesday, July 31, 2018 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Gregory Watson* and Donatello Telesca
• Companies: UCLA and UCLA
• Keywords: Spatial Statistics; Machine Learning; Air Pollution; Prediction; Cross Validation

Abstract

Epidemiologists use prediction models to quantify air pollution exposure where monitoring data are insufficient, traditionally relying upon models that emphasize spatial or spatiotemporal covariance. Recently the computational burdens and inflexible mean structures of these models have prompted practitioners to explore machine learning algorithms as scaleable, more flexible alternatives. These algorithms usually ignore dependence between observations, but perform well on a variety of prediction problems. This study compares popular examples of both model classes on ozone and particulate matter (PM) data collected during a 2008 wildfire event in California, introducing an intuitive framework for model evaluation that accounts for spatial dependence.