In a case-control study of estrogens and endometrial cancer, alternative sampling methods were used to eliminate the detection bias that arises from the increased diagnostic attention received by women with uterine bleeding after estrogen exposure. In a set of cases and controls chosen by conventional procedures the odds ratio was 11.98. In an alternative set of cases and controls at the same institution, consisting of patients who had all received dilatation and curettage or hysterectomy because of uterine bleeding, the odds ratio was 1.7. A methodologic analysis demonstrates detection bias arising from the pattern of hospital referral and shows the way in which the bias is neglected or increased by conventional sampling procedures, but reduced by the alternative procedure. The magnitude of the association between estrogens and endometrial cancer has been greatly overestimated because of detection bias; when an appropriate compensation for the bias is introduced, the odds ratio approaches a value much closer to 1.