Some methods to address collinearity among pollutants in epidemiological time series.

The aim of this paper is to provide accurate estimation methods for regression models used in epidemiological time series to deduce quantitative morbidity relationships. Such models often include highly correlated variables (pollutant levels and climatic conditions) as well as lagged and unlagged values of the same variables (which also show a high collinearity due to the stochastic dependency of consecutive measurements). We first describe some methods to detect and assess multicollinearity. We recall the drawbacks of usual methods of estimation, and then after briefly mentioning traditional solutions, we explore three alternative methods accounting for multicollinearity: Sclove's estimation; Almon's method; and a combination of Almon's method and principal components procedure. We compare these methods in obtaining efficient estimators on environmental epidemiological data (children's hospital admissions as dependent variable and unlagged and lagged values of outdoor temperature, SO2, NO and CO as explanatory variables.