AIR-POLLUTION EPIDEMIOLOGY - CONSIDERATIONS IN TIME-SERIES MODELING

Recent epidemiological studies have indicated that ambient air polluti on, including PM-10, is associated with excess mortality and morbidity . These studies have included both cross-sectional comparisons across communities and rime-series analyses over time in single communities. Time-series analysis offers certain advantages, primarily in that the study population is the same over time, so that it acts as its own ''c ontrol.'' However, modeling such data is complicated by the fact that other environmental factors and other causes of illness can confound t he results unless they are adequately addressed. For example, winterti me influenza epidemics cause long-wave peaks in respiratory mortality, and variations in emissions, dispersion, and atmospheric chemistry ca n cause seasonal cycles in pollution. Such superimposed long-wave vari ations in both health outcomes and pollutant concentrations can underm ine the statistical validity of time-series models by inducing autocor relation, and can create long-wave ''noise'' signals that can overwhel m a short-term ''signal'' of interest. Also, model specification can s trongly affect the results of a time-series model. For example, analys es focusing on only one routinely collected pollution metric, to the e xclusion of other possibly more influential pollution components, can cause the effects of the overlooked pollutants to be ascribed to the s tudied pollutant. In addition, the potential effects of nonnormal (e.g ., Poisson) data distributions on time-series results need to be consi dered. It is concluded that how these various time-series modeling fac tors are, or are not, addressed can have a large influence on the stud y conclusions, or the ''message'' resulting from such analyses. Sensit ivity analyses incorporating multiple modeling methods and model speci fications are therefore recommended as part of such an analysis. Moreo ver, in this article exploratory and diagnostic procedures are recomme nded that may aid the modeler in assessing and avoiding the noted prob lems and that will allow the validity of such studies to be more easil y documented and intercompared.