Assessing the Health Effects of Air Pollution Using the Case-Crossover Design

The case-crossover design has been widely used to study the association between short term air pollution exposure and the risk of an acute adverse health event. The design uses cases only, and compares exposure at the event time with exposure at other control or "referent" times. By making within-person comparisons, time invariant confounders are controlled. Time varying confounders can also be controlled by design if referents are matched on these factors. Yet, proper referent selection is important for other reasons. The design involves an implicit assumption that the distribution of exposure is constant across referent times. In addition, the statistical method usually used to analyze a case-crossover study, conditional logistic regression, is only unbiased with certain referent strategies. This paper reviews the case-crossover literature in the air pollution context, and clarifies key referent selection issues. It concludes with a set of recommendations for choosing a referent strategy with air pollution exposure data. The time stratified approach to referent selection is advocated, as it ensures unbiased conditional logistic regression estimates and avoids bias due to time trend in the exposure series. The particular stratification can be tailored to control for time varying confounders by design.