Optimal Referent Selection Strategies in Case-Crossover Studies

Abstract

I n this issue of EPIDEMIOLOGY, Janes et all present a comprehensive review of referent selection strategies in case-crossover analyses of air pollution data. This review is timely, given the increasing popularity of the case-crossover approach, and makes an important contribution to the field. As for other substantive areas of epidemiology, concerns over the validity of estimates of the short-term health effects of air pollution are paramount. Thus, study design decisions and analytic approaches should be engineered to maximize efficiency to the extent possible without introducing inordinate The authors did an excellent job of summarizing insights regarding the importance of careful selection of referent periods to minimize potential biases, including those that arise from time trends in exposure, truncation of the air pollution data at the start and end of the exposure time series, and violations of the assumptions inherent in the conditional logistic likelihood that result in overlap bias. The arguments put forth by Janes and colleagues convincingly show that, although other approaches may be valid under certain circumstances, the time-stratified selection strategy is more generally valid than any of the alternatives thus far proposed. Thus, this strategy should be considered the de facto standard approach to the analysis of data arising in studies of the short-term effect of air pollution and weather. The caution against model shopping is an extremely important one, but it must be distinguished from the concept of sensitivity analysis. Conducting several analyses and cherry picking results (ie, presenting only the subset of results that look the most interesting to the investigator) clearly is problematic and is likely to lead to the presentation of spuriously extreme effect estimates. However, assuming one considers only approaches that have been shown to be generally valid, it is reasonable for investigators to conduct sensitivity analyses. For example, when conducting a time-stratified case-crossover analysis, the length of the time strata should be chosen subject to the best judgment of the analyst. In studies of short-term health effects of air pollution, investigators have most commonly chosen strata defined by calendar month.>6 This is clearly a convenient choice that is easy to implement. However, an implicit assumption is that time-dependent confounders, such as season, or respiratory illness outbreaks are relatively constant within a calendar month. It is possible that strata of 2 or 3 weeks duration may better control for such confounding or that longer periods of up to 6 weeks might control confounding as well, but be more efficient.