Application and evaluation of four regression techniques for a chemical mass balance receptor model

Abstract

Chemical mass balance (CMB) receptor models have evolved over the past 15 years as a potential alternative to dispersion models for assessing the source contributions of pollutants and aerosols in the atmosphere. Unlike dispersion models, which require a detailed inventory of emission rates and stack parameters from major sources in addition to meteorological data and empirical dispersion factors, receptor models need only information about the characteristics of samples collected at a site and the chemical composition of source categories. There are many advantages and potential applications of receptor models to contemporary air pollution problems. Their relatively simple mathematics compared to source-oriented dispersion models results in a less time and cost-intensive method of source apportionment. Fugitive and area source contributions to ambient aerosol samples can be predicted without the need to develop emission factors. A major application for receptor models is in the area of criteria pollutant standards attainment, where they can be used to determine the major contributing sources to regional air pollutant levels. State implementation plans can then be created to regulate those sources. Source impacts are estimated with CMB receptor models through application of different regression techniques to solve simple mass balance equations. Variations of ordinary least squares regression, both unbiased and biased techniques, have been used in sourced apportionment studies in major urban airsheds across the country. Unbiased techniques include weight least square (WLS) and effective variance weight least square (EVWLS) regression. Biased techniques that have been considered include ridge regression (RR), principal components and latent root regression. Studies have also been published to directly intercompare the different methods1,2. For the purposes of this study, four different solutions to the CMB receptor model have been developed and evaluated for an environmental data set: two unbiased techniques (WLS and EVWLS) and two biased techniques (RR weighted by the measurement variance of the receptor data and RR weighted by the effective variance). These four solutions were then evaluated and intercompared through statistical analysis and physical validation techniques.