Multivariate Statistical Models for Effects of PM and Copollutants in a Daily Time Series Epidemiology Study
Most analyses of daily time series epidemiology data relate mortality or morbidity counts to PM and other air pollutants by means of single-outcome regression models using multiple predictors, without taking into account the complex statistical structure of the predictor variables. It is particularly important to assess the extent to which adverse health effects of air pollution can be attributed to separate components such as PM, or to an inseparable mixture of pollutants. Several approaches are possible, including: 1) empirically derived combinations of predictors via principal components analyses; 2) theoretical combinations of predictors based on multiple equation models for observed variables; 3) combinations of predictors as manifest outcomes of latent (hidden) variables. We illustrate these for a set of mortality time series data in which the observed PM indicator is the coefficient of haze (CoH). Several methods identify CoH as a highly significant predictor in some models, and is largely independent of CoH in winter. A number of air pollutants are highly correlated with CoH, possibly in a single component derived from motor vehicles. Much of the common variation of these pollutants can be attributed to underlying meteorological variables, such as wind speed and temperature. Separation of PM and CO health effects is difficult in this data set, whereas separation of ozone effects from those of CoH and other pollutants is relatively easy.