Some methodological issues in statistical estimation of air pollution and health

This paper introduces a new method for estimating a dose-response relationship from spatially averaged time series of air pollution and health data. Because time is perceived as a nuisance parameter to be eliminated, least-squares regression and traditional time series methodology (e.g., spectral analysis Box-Jenkins methods) are rejected in favor of a nonparametric estimation procedure based on observing health effects in times of nearly equal pollution. The method requires estimating the ratio of two density functions and avoids problems of aggregation, linearity and normality. In spite of the formal tests described, the procedure seems most useful at present as a data analytic and data display device rather than as an inferential tool.