Seasonal confounding and residual correlation in analyses of health effects of air pollution



To investigate the health effects of air pollution via a partially linear model, one must choose an appropriate amount of smoothing for accurate estimation of the linear pollution effects. This choice is complicated by the dependencies between the various covariates and by the potential residual correlation. Most existing approaches to making this choice are inadequate, as they neither target accurate estimation of the linear pollutant effects nor handle residual correlation. In this paper, we illustrate two new adaptive and objective methods for determining an appropriate amount of smoothing. We construct valid confidence intervals for the linear pollutant effects, intervals that account for residual correlation. We use our inferential methods to investigate the same-day effects of PM10 on daily mortality in two data sets for the period 1994 to 1996: one collected in Mexico City, an urban area with high levels of air pollution, and the other collected in Vancouver, British Columbia, an urban area with low levels of air pollution. For Mexico City, our methodology does not detect a PM10 effect. In contrast, for Vancouver, a PM10 effect corresponding to an expected 2.4% increase (95% confidence interval ranging from 0.0% to 4.7%) in daily mortality for every 10 µg/m3 increase in PM10 is identified. Copyright © 2006 John Wiley & Sons, Ltd.