Assessing seasonal confounding and model selection bias in air pollution epidemiology using positive and negative control analyses

Much of the evidence for health effects of particulate air pollution has come from ecologic time series studies that regress mortality or morbidity event counts on pollutant data routinely collected for other purposes. The modelling approach typically involves selecting both a lag at which the effect of particulates should be evaluated and a level of filtering to remove long-term associations confounded with seasonal variations and secular trend. In this paper we investigate the bias introduced by model selection and residual confounding using simulations based on a previous analysis of data from King County, Washington. In comparison with the original analyses we find that the bias is small in absolute terms but of the same order as the estimated health impacts. Copyright © 2000 John Wiley & Sons, Ltd.