Accounting for time-dependent covariates whose levels are influenced by exposure status.

When measuring the association between an exposure and disease, one must decide whether to account for confounding or modifying variables whose levels are altered by the presence of the exposure. For example, to assess the impact of cessation of unopposed estrogen therapy on the occurrence of endometrial cancer, a researcher needs to consider the duration of the estrogen therapy, a strong risk factor for endometrial cancer, as a potential confounder or effect modifier. Duration of estrogen therapy, however, is itself influenced by the decision to stop the therapy (the "exposure" of interest). In such a case, two distinct approaches may be taken, depending upon the question being considered. One may wish to assess the degree to which the exposure predicts disease incidence, over and above the additional variable, at some later point in time. In this case, it is appropriate to consider the value of the other variable (for example, duration) at that later time. On the other hand, one may also wish to measure the rate of disease beginning at the time of cessation of the exposure, relative to the corresponding rate in persons with continuing exposure Here, the most appropriate analysis considers the level of the confounding variable (for example, duration) measured only until the time of exposure of interest occurs (for example, cessation of unopposed estrogen therapy). Examples are given to illustrate that the specific question being addressed dictates the handling of covariates of this type.