Weiss, Noel S.

An inference as to the possible causal role of an exposure in the development of a disease is strengthened when there is a large disparity in the incidence of that disease between exposed and nonexposed persons. If that disparity is particularly great when we would predict it to be so (perhaps on the basis of a suspected susceptibility factor), then a causal inference is strengthened further.¹ For example, the conclusion that exogenous estrogen can predispose to the development of endometrial cancer is based heavily on the very much higher incidence of this disease in hormone users that is present when, on pharmacologic and endocrinologic grounds, it would be predicted, ie, in the absence of exogenous progestogen (which can give rise to differentiation of endometrial cells that have proliferated in response to estrogen stimulation). The inference that a genetic characteristic associated with reduced catabolism of estrogen influences endometrial carcinogenesis would be strengthened if the association were particularly large in the presence of exogenous unopposed estrogen, if it is believed that it requires the relatively high levels of circulating estrogens that result from use of hormonal agents to exceed the metabolic capability of a woman with a slow-catabolizing genotype. (Alternatively, there may be reasons to believe that the relative impact of hormone catabolism on risk of endometrial cancer may be appreciable only when there is not a quantity of exogenous estrogen available to overwhelm a woman's catabolic ability, regardless of her genotype. In this instance, the stronger gene–cancer association would be anticipated in hormone nonusers.)

In the instance of endometrial cancer, in which we are already secure in our knowledge of the deleterious impact of unopposed estrogen, attention has recently focused on the possible etiologic role of various genetic characteristics. The most straightforward means of investigating such a role would entail a comparison of the incidence of the disease in women with and without a given genotype, first in users of unopposed estrogens (of a duration and recency adequate to produce an increased risk) and then again in other women. However, a number of studies of endometrial cancer (and of other diseases with known nongenetic etiologies) assessing the possible etiologic influence of a genetic characteristic, after examining the overall gene–disease association, approach the issue of potential gene–environment interaction by asking whether the size of the association between the environmental factor and disease varies according to genotype. This approach does not provide information as to the size and direction of the association between the genetic factor and disease within strata of the environmental factor, and thus does not help in evaluating whether the genetic factor may play a causal role in the development of the disease. Tables 1 and 2 seek to illustrate this point by depicting results from 2 hypothetical cohort studies of a potential interaction between a genetic and environmental factor (E) with regard to their influence on risk of disease. In both study 1 and study 2 presented in Table 1, E is associated with a 5-fold increased risk in persons with genotype X at a given locus but with no altered risk in persons with genotype Y at the same locus. From this we can infer that the genotype–disease association must vary as a function of E, but we cannot infer the pattern of that variation. Indeed, in the 2 hypothetical studies, the nature of the relation between genotype X and disease, conditional on E, differs considerably (Table 2). In Study 1 persons with genotype X are at a 5-fold increased risk of the disease if they have sustained the environmental exposure, but at the same risk as persons with genotype Y if they have not. Yet in Study 2, the presence of genotype X is associated with no altered risk in those with the environmental exposure and with a 5-fold reduction in risk in other persons. Only this latter analysis, and not the one presented in Table 1 that focuses on the association between E and disease, allows us to see if the association between the genetic characteristic and disease is in the predicted direction and is greatest in the predicted subgroup (based on the presence or absence of exposure E). Thus, only this analysis conveys information that bears directly on the possible inference of a causal or protective role of genotype X on disease risk.

Table 1

Table 2

The above recommendation, which focuses attention on the association between genotype and disease within strata of an environmental exposure known to be a cause of the disease, would appear to be inconsistent with the recommendations of Botto and Khoury² for the analysis of potential gene–environmental interactions. These authors advocate the use of a single referent category—persons with the low risk genotype and no (or the lowest level of) environmental exposure. This approach allows the computation of risk estimates for each factor alone and for the combination of genotype and exposure, which the authors refer to as the “basic, direct estimates of association.” However, what the approach does not allow is the computation of risk associated with the genetic characteristics among persons who also have sustained the environmental exposure (and vice versa) and the confidence interval surrounding that risk estimate. As argued above, these, along with the corresponding parameters among persons without the environmental exposure, are required to inform an inference regarding a possible causal role of the genetic factor. Now, if a judgment has been made that a genetic factor could plausibly be influencing disease occurrence, whether in all persons or in just those with or without a particular environmental exposure, I believe that as a second step it would be appropriate to move to an analysis of gene–environmental interaction that employs a single referent category. Only this means of data presentation allows an examination of the important question of the possible nonadditivity of the influence of 2 factors on disease occurrence.³

ACKNOWLEDGMENTS

I would like to thank Chu Chen, Jennifer Doherty, Carolyn Hutter, Bruce Psaty, and Lori Sakoda for the comments they provided on an earlier draft of this manuscript.

REFERENCES