CASE-CROSSOVER ANALYSES FOR ACUTE EVENTS: APPLICATION TO 14 CITIES

Abstract

Air pollution has been associated with deaths and morbid events in hundreds of studies throughout the world. Most of these studies have used regression analysis to control for weather and season, which might otherwise confound the pollution effects. Recently, problems in the estimation of standard errors from the most commonly used software have been reported. This has lead to many reanalyses, but also more extensive discussion of how best to control for season and weather. Thc case crossover approach constructs a case control study by matching each person on their event day to the same person on a day or days when the event did not occur. Such studies control, by design, for all time invariant risk factors. Bateson and Schwartz have shown that choosing control days close to the event day controls for season (and slowly varying risk factors) by design, avoiding the need for sophisticated models. In addition, by choosing control days with the same mean temperature as the event day, one can control for much of the weather effect by design as well. Further, matching controls for interactions between season and temperature as well. I examined the association between PM10 and daily deaths in 14 US cities (Birmingham, Boulder, Canton, Chicago, Cincinnati, Colorado Springs, Columbus, Detroit, Minneapolis, New Haven Seattle, Spokane, and Provo/Orem) using this approach. Using the Bateson and Schwartz symmetric bidirectional control sampling strategy, I found a significant association with a 10 ug/m3 increase in PM10 on the previous day associated with a 0.36% increase in daily deaths (95% CI 0.22%. 0.50%). Using the time-stratified approach to selecting controls, and also matching on current day temperature to the nearest degree, I found an effect of 0.39%, 95% CI 0.19%, 0.58%). All models controlled for day of week, and three degree of freedom functions of relative humidity and previous day's temperature. The association between PM10 and deaths persists when matching on season and temperature, and the estimated effect size is almost twice the effect estimates presented in NMMAPS.