Weiss responds to "Screening as a nuisance variable in cancer epidemiology": for which aspects of a study subject's screening history should we control?

Abstract

Our goal in adjusting for cancer-screening history in nonrandomized studies of cancer etiology is straightforward: to ensure that the groups of persons who do and do not have cancer are comparable with regard to receipt of screening during the period of time the malignancy was detectable during its preclinical phase. Joffe (1) correctly points out that, in the presence of confounding by cancer-screening history, controlling simply for the presence of one or more screens during that period of time may fail to remove all of the confounding. This is because of the following: 1) We may not correctly estimate the mean detectable preclinical duration; and 2) even if we are able to correctly estimate this parameter, there may be substantial variability around it, and so for many cases the true duration could be substantially underor overestimated. However, I believe we have no alternative measure of screening history that is better for this purpose. Specifically, I believe that Joffe’s suggestion of controlling for time since last screen would be problematic. First, it is the fact of screening during the period of time corresponding to the detectable preclinical phase of the tumor that is relevant, not its timing within that period. Second, for a screening test of high sensitivity, the distribution of time since last screen between persons with and without cancer would be completely different. The large majority of cases who had undergone screening would have done so close to the time of diagnosis: Had the test been performed earlier during the detectable preclinical phase of the tumor, the diagnosis most likely would have been made at that time and not later. On the other hand, in persons without the cancer in question, there would be a much more even distribution of receipt of testing during the interval. Effectively, the large majority of persons whose cancer had been diagnosed as a result of screening would contribute little to the analysis because of the paucity of persons in the comparison group who had been screened just as recently. As I point out in my commentary (2), misclassification of screening history—differentially between persons with and without cancer because of the inability to reliably distinguish tests done for screening from tests done in response to symptoms or signs—can impair our ability to control for confounding by screening history. It is my impression (with no data to support it!) that in most instances this problem will be of greater importance than the one arising from the use of the dichotomy—yes or no, screening or no screening, during the presumed detectable preclinical interval—as our variable for which to adjust.