A comparison of statistical and machine learning methods for creating national daily maps of ambient PM$_{2.5}$ concentration

Abstract

A typical problem in air pollution epidemiology is exposure assessment for individuals for which health data are available. Due to the sparsity of monitoring sites and the limited temporal frequency with which measurements of air pollutants concentrations are collected (for most pollutants, once every 3 or 6 days), epidemiologists have been moving away from characterizing ambient air pollution exposure solely using measurements. In the last few years, substantial research efforts have been placed in developing statistical methods or machine learning techniques to generate estimates of air pollution at finer spatial and temporal scales (daily, usually) with complete coverage. Some of these methods include: geostatistical techniques, such as kriging; spatial statistical models that use the information contained in air quality model outputs (statistical downscaling models); linear regression modeling approaches that leverage the information in GIS covariates (land use regression); or machine learning methods that mine the information contained in relevant variables (neural network and deep learning approaches). Although some of these exposure modeling approaches have been used in several air pollution epidemiological studies, it is not clear how much the predicted exposures generated by these methods differ, and which method generates more reliable estimates. In this paper, we aim to address this gap by evaluating a variety of exposure modeling approaches, comparing their predictive performance and computational difficulty. Using PM$_{2.5}$ in year 2011 over the continental U.S. as case study, we examine the methods' performances across seasons, rural vs urban settings, and levels of PM$_{2.5}$ concentrations (low, medium, high).