Re: "The 'Case-Chaos Study' l Adjunct or Alternative to Conventional Case-Control Study Methodology"

Abstract

In a recent article, Gillespie et al. (1) proposed a permutation-based method for identification of factors that may be associated with risk of illness in an outbreak setting. They did this by comparing potential risk factors with other potential risk factors in the case population, rather than comparing factors in cases with those in controls. They suggested that in some cases this method may provide an alternative to case-control studies (1). Frequency of exposure among cases can in some situations provide valuable and timely clues about risk, but interpreting these frequencies sensibly depends on information or assumptions about frequency in the general population. For example, a population of newly diagnosed lung cancer cases in the United States would show a much higher frequency of male sex than of asbestos exposure; to evaluate these observations, we need either some sort of control or baseline group or common-sense assumptions about baseline frequencies of these factors. In the absence of data from controls, investigators still need to assess the likelihood that a factor in the cases occurs at an elevated frequency relative to the same factor in a baseline population. The statistical test proposed by Gillespie et al. instead assesses whether a factor is significantly more common in cases than other factors considered as potential risk factors. Thus, their suggested statistical test does not enhance interpretation of the frequency of exposure, and in fact may interfere with it, by adding unnecessary complications. Furthermore, whether the prevalence of a given factor is elevated relative to the prevalences of other factors depends on which other factors are chosen for the study. For example, asbestos exposure is rare even among newly diagnosed lung cancer cases. Not only would a “case-chaos” comparison fail to identify it as a risk factor, but its inclusion in the study would increase the likelihood that other factors would be identified as risk factors. To illustrate the above points for an existing outbreak data set, let us use the case-chaos design to analyze case data from the well-known outbreak of foodborne illness that occurred at a church supper in Oswego County, New York, in 1940 (Figure 1). In addition to the measured risk factors, we introduce 2 hypothetical ones: “church member,” which is a common but randomly distributed factor, and “ate dessert first,” which is rare but an absolute risk factor for illness. As expected, the methodology shows that the case-chaos odds ratio of the common factor is significantly elevated, and that of the rare factor is significantly reduced, in comparison with the other measured factors. Note that the estimated odds ratio of the rare factor is significantly reduced here because most of the other factors considered happen to be relatively common. The same factor, with the same effect on the population, would not show significance in another study, where different sorts of factors had been tested along with it. Figure 1. Illustration of some characteristics of the case-chaos approach proposed by Gillespie et al. (1), using data from an outbreak of foodborne illness that occurred in Oswego County, New York, in 1940 (4, 5). A box-and-whisker plot is shown for each exposure ... Given these limitations, we believe that the proposed case-chaos methodology is not an appropriate alternative to the conventional case-control study design. The analytic formulation suggested by Hohle (2) in a recent letter was intended to clarify the fundamental problem: that the case-chaos approach compares frequencies across different exposures rather than with respective baseline frequencies. McCarthy et al.'s response to Hohle (3) missed this point; in fact, neither a statistical test nor a confidence interval is appropriate, because the hypothesis being addressed is irrelevant to the question at hand. In the event that identifying high-frequency factors among cases is required in the course of an outbreak investigation, we recommend that this be done using the more straightforward approach of calculating and examining the proportions of cases with various exposures.