Testing and optimizing two factor-analysis techniques on aerosol at Narragansett, Rhode Island

Abstract Elemental data for aerosol at Narragansett, RI, USA, were used to compare the source-identification power of positive matrix factorization (PMF), a new variant of factor analysis, with that of conventional factor analysis (CFA) and to investigate how much each technique can be “tuned” for best results. The techniques generally yielded similar results. Although both were degraded by weak elements and gave factors that always differed somewhat from known sources, they nonetheless provided substantial insight into sources of elements. PMF was harder to use than CFA but resolved crustal and marine components up to an order of magnitude better. Best results were generally obtained when the data were log-transformed, when missing data were replaced by means, and when various numbers of factors were tried and their results carefully evaluated for physical reasonableness. But the most important consideration was found to be the choice of elements, which outweighed all differences between techniques. Therefore, to maximize the source-identification power of factor analysis, the two most important steps appear to be selecting the optimum set of elements and selecting the basic technique, in that order.