Can the “Specificity” of an Association be Rehabilitated as a Basis for Supporting a Causal Hypothesis?

ArticlePlus Click on the links below to access all the ArticlePlus for this article. Please note that ArticlePlus files may launch a viewer application outside of your web browser. https://links.lww.com/EDE/A19 In 1964, the Advisory Committee to the Surgeon General of the U.S. Public Health Service described several guidelines that they used when evaluating the scientific evidence regarding the possible etiologic role of cigarette smoking in the development of lung cancer. 1 One of these pertained to the “specificity” of the association seen in epidemiologic studies. By “specificity,” the Advisory Committee was referring primarily to the idea that the more broadly an exposure is associated with a variety of outcomes, the less likely it is to be causally associated with any one of them. This notion of specificity, also considered by Hill 2 one year later, may have been borrowed from Yerushalmy and Palmer. 3 They had contended that “on an intuitive basis...when a given characteristic is found to be associated with one, or at most a few diseases, then the evidence for a causal relationship is more convincing.” Berkson 4 had raised a similar concern upon seeing that the initial cohort studies of smoking and mortality reported elevated death rates among smokers for lung cancer and a variety of other causes. (The Advisory Committee took issue with Berkson’s interpretation of the data in those studies, contending that “an excess risk for smokers does not occur for every one of the causes of death,” and, thus, “it is reasonable to conclude that the association between cigarette smoking and lung cancer has a high degree of specificity.”) Nowadays, the notion of specificity as a consideration in causal inference doesn’t get much respect. For example, Rothman and Greenland 5 refer to it as being “wholly invalid.” Based on the manner in which this criterion was originally characterized, one must agree with this harsh appraisal, because we have reasons to believe that some agents (such as ethanol) can cause impairment of or damage to more than one tissue or organ. Also, we believe that some exposures can contain numerous harmful substances that collectively could give rise to a variety of effects throughout the body. Even in 1964, before the characterization of the many potentially toxic agents in cigarette smoke, the Advisory Committee to the Surgeon General stated that “it would not be surprising to find that the diverse substances in tobacco smoke could produce more than a single disease.”1 (Then, as now, committees did not necessarily speak with one voice; it was but four lines later in their report that the Advisory Committee pronounced that the criterion of specificity, just shown to be irrelevant, had been fulfilled.) Notwithstanding the above, there are instances in which specificity is invoked—appropriately in my view—as an argument in support of a causal hypothesis. These occur in three general circumstances, none of which was addressed when the criterion was first being put forward. Specificity of Outcome Specificity of outcome refers to a situation in which there is a sound basis for simultaneously predicting that an exposure ought to have an association with one outcome and ought not to have an association with another, and the relevant data correspond to both predictions. For example, screening sigmoidoscopy was observed in case-control studies to be associated with a 50% to 70% reduction in mortality from tumors of the rectum and distal colon that were within the reach of the sigmoidoscope but not with reduced mortality from more proximal colon tumors. 6,7 Despite the potential limitations of the case-control approach in assessing cancer screening efficacy, 8 these results have been widely interpreted as indicative of a genuine protective effect of screening; not only was there a strong negative association found in several studies, but it was specific to that part of the large intestine for which the protection had been anticipated. Similarly, imagine that in a cohort study the rate of head injury was observed to be lower in bicycle helmet users than helmet nonusers. This association could be attributable to a true degree of protection afforded by the helmet but also to confounding factors—perhaps the more careful riders were the same ones who chose to wear helmets. If the association with helmet use was specific to head injury, i.e., it was not present for injuries of other parts of the body, the inference of a protective influence of the helmet would be strengthened. Specificity of Exposure In this circumstance, only the exposures (or exposure subcategories) initially predicted to be associated with an outcome actually are observed to be associated with it. For example, among Swedish women hospitalized with a diagnosis of ovarian endometriosis between 1969 and 1983, 17 went on to develop ovarian cancer, whereas only 5.5 would have been expected based on incidence rates among Swedish women in general. 9 The association of ovarian cancer with ovarian endometriosis appeared to be specific, in that women with endometriosis restricted to non-ovarian sites had no excess risk. As another example, there are reasons to believe that the association between use of long-acting hypnotic/anxiolytic drugs by elderly persons and the incidence of hip fracture is a causal one. 10,11 This causal inference can be made, in part because of the specificity of that association, in that no similar association has been observed for short-acting hypnotic/anxiolytic drugs. 10,11 Because the latter medications are relatively less likely to produce residual daytime sedation and/or orthostatic hypotension, it was predicted that their use should be associated less strongly (if at all) with the incidence of falls and fractures. The lack of specificity of associations between an outcome and several different exposures has been a basis for rejecting a causal interpretation for each of the separate associations. Schumacher et al.12 observed an approximate doubling of the risk of gastric cancer in persons who had been prescribed the H2 blocker cimetidine at least two years before diagnosis. However, they saw virtually the identical elevation in risk in persons who had been prescribed antacids. While it is conceivable that simply reducing hyperacidity predisposes to the development of gastric cancer, the lack of specificity seen in this study argues more strongly in support of the hypothesis that as-yet-undiagnosed gastric tumors or their antecedent lesions prompt patients to seek therapy directed at symptoms of gastric irritation. Specificity with Regard to Susceptibility This type of specificity is present when an association between an exposure and an outcome is observed only in persons who have a particular characteristic that would have been predicted to render them susceptible to an effect of the exposure. The red blood cells of persons deficient in the enzyme glucose-6-phosphate-dehydrogenase (G6PD) cannot respond to an oxidative stress by generating reduced glutathione, allowing the cells’ sulfhydryl groups to become oxidated. Ingestion of sulfa drugs, one source of an oxidative stress, is associated with the incidence of hemolytic anemia only in G6PD-deficient individuals. 13 The specificity of the association, in just the manner predicted based on an understanding of toxicology, supports the hypothesis that it is a causal association. It has been suggested 14 that variation in the size of an exposure-disease association on the basis of the dose of the exposure particularly supports an inference of cause-and-effect when a good basis is present for predicting the observed pattern of variation. The idea being advanced in this commentary is similar. Occasionally, we have a good basis for predicting that only some outcomes ought to be related to a given exposure, that only one or several types of exposure ought to be related to a given outcome, or that only in persons with a certain characteristic should there be a relation between exposure and outcome. When the observations made in epidemiologic studies conform to those expectations—when the associations are specific—the credibility we assign to a causal hypothesis increases. Conversely, when a causal hypothesis predicts specificity, but that is not what is observed, the credibility of that hypothesis is diminished. No, some of our epidemiologic forefathers did not get the notion of “specificity” quite right. But I suggest that a modified version of that notion has a useful place, after all, as we seek to sort out causal from noncausal associations. Acknowledgments The clarity of this manuscript was improved following the incorporation of the suggestions of Thomas Koepsell, Peter Cummings, and Sander Greenland.