Compositional Source Apportionment

• Activity Number: 160
• Type: Invited
• Date/Time: Tuesday, August 13, 2002 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics & the Environment*
• Author(s): Dean Billheimer*+
• Affiliation(s): Vanderbilt University
• Keywords: compositional data ; MCMC ; model identifyability

Abstract

Source apportionment models, or receptor models, seek to apportion an ambient mixture of pollutants to the contributing pollution sources. Often, neither the number of sources nor their chemical profiles are known precisely. The dual goals of modeling are to estimate the chemical ``signature'' of the sources, and to characterize the mixing process. I develop a novel modeling approach for receptor data where all model components are compositions (i.e., vectors of proportions). This approach maintains positivity and summation constraints for source contributions and chemical profiles. Further, it incorporates available prior knowledge regarding the source chemical profiles. Including prior knowledge allows model identification while avoiding restrictive assumptions regarding presence or absence of chemical tracers. I illustrate this approach by modeling air pollution data collected from a receptor near Juneau, AK. The compositional model produces point estimates of source profiles and mixing proportions similar to those obtained in a previous study. However, interval estimates for mixing proportions are roughly 30% shorter than those found previously.