STATISTICAL ISSUES OF AIR POLLUTION AND DAILY MORTALITY

RICHARD L SMITH

Department of Statistics, University of North Carolina, Chapel Hill


For many years, the Environmental Protection Agency (EPA) has enforced standards for airborne pollutants which are harmful to human health. New regulations introduced in 1997 tightened the existing standards for ozone and particulate matter, including a new standard for fine particles (particulate matter of diameter 2.5 microns or less, known as PM2.5). These new standards were introduced after extensive epidemiological research claiming that airborne particulate matter is harmful at levels permitted by previous standards. The standards have proved very controversial, with many groups challenging their validity on scientific grounds as well as an industry-sponsored lawsuit which has, for the time being, halted their implementation. Nevertheless the EPA is proceeding with its year 2000 review of the standards in the expectation that a new standard will be introduced and enforced in 2002.

This talk will discuss some of the statistical issues associated with the epidemiological research into health effects of particulate matter. A large class of studies is based on time series analysis in which the essential hypothesis is that mortality tends to be higher on days immediately following high air pollution events. Detailed analysis of such data sets, however, often allows for alternative interpretations which make the assertion of a direct causal relationship unclear. A recent study using new air pollution data from Phoenix, Arizona, has suggested that the direct effect may be stronger for coarse particles than for fine particles, and also that the adverse effects of fine particles only occur above a threshold in the region of 20-25 micrograms per cubic meter, which is above the proposed EPA standard for the long-term mean. Finally, I shall also discuss some of the methodological issues associated with regression analysis with large numbers of covariates, suggesting that methods based on ridge regression, principal components regression and various Bayesian possibilities may perform better than the standard regression methods currently employed in these studies.

Questions? Jiming Jiang