Statistical considerations in determining the health significance of constituents of airborne particulate matter.

Because of the U.S. Environmental Protection Agency's (EPA) new ambient air quality standard for fine particles, the need is likely to continue for more detailed scientific investigation of various types of particles and their effects on human health. Epidemiology studies have become the method of choice for investigating health responses to such particles and to other air pollutants in community settings. Health effects have been associated with virtually all of the gaseous criteria pollutants and with the major constituents of airborne particulate matter (PM), including all size fractions less than about 20 microns, inorganic ions, carbonaceous particles, metals, crustal material, and biological aerosols. In many of the more recent studies, multiple pollutants or agents (including weather variables) have been significantly associated with health responses, and various methods have been used to suggest which ones might be the most important. In an ideal situation, classical least-squares regression methods are capable of performing this task. However, in the real world, where most of the pollutants are correlated with one another and have varying degrees of measurement precision and accuracy, such regression results can be misleading. This paper presents some guidelines for dealing with such collinearity and model comparison problems in both single- and multiple-pollutant regressions. These techniques rely on mean effect (attributable risk) rather than statistical significance per se as the preferred indicator of importance for the pollution variables.