A critique of the statistical evidence associating estrogens with endometrial cancer.

The epidemiological problem of detection bias occurs when the evidence needed to diagnose a particular dis ease is sought more intensively in the group of people exposed to a particular agent than in the comparative group, who did not receive the agent. The bias can occur during community surveillance before hospitalization, during the ordering of diagnostic tests for hospitalized patients, or during the interpretation of the tests. In postmenopausal women with endometrial cancer, detection bias is produced because of the dilation and curettage (or other diagnostic tests) ordered when bleed ing occurs as a side effect of estrogen therapy. In estro gen takers, these diagnostic explorations allow the detec tion of endometrial cancers that might otherwise be silent and undetected in women who do not take estrogen. Since this bias occurs before patients are hospitalized, its removal requires special sampling and analytic tech niques that have not been used in most epidemiological case-control studies where high risk ratios were found for the association between estrogens and endometrial can cer. When cases and controls have been selected with the use of a diagnostic procedure as the sampling frame, the odds ratios are much closer to 1 ; with appropriate stratification for the diagnostic stimulus of bleeding, the odds ratios approximate 1. Other sources of bias for the estrogen-endometrial cancer association can be the "protopathic" prescription of estrogen for women with unrecognized endometrial cancer, the absence of "double-blind" interviewing tech niques, and arbitrary definitions of "exposure to estro gens." The difficulties in interpreting the relationship of estro gens and endometrial cancer indicate the need for devel opment of rigorous scientific standards in the epidemio logical methods used for case-control research.