Risks of estrogen plus progestin therapy: a sensitivity analysis of findings in the Women's Health Initiative randomized controlled trial.

CONTEXT: The Women's Health Initiative (WHI) randomized controlled trial that compared estrogen plus progestin versus placebo was stopped because of a significantly increased risk of breast cancer in the active treatment group (reason 1), and because a global index supported a finding of overall harm (reason 2). The possibility that the findings could have been accounted for by bias was not considered.

OBJECTIVE: The present analysis was undertaken to determine whether detection bias is a plausible alternative to causality as an explanation for those findings that contributed to the reasons for discontinuing the study.

DESIGN: setting and participants: This work took the form of a sensitivity analysis of the published WHI data to determine the magnitude of the detection bias required to account for those findings that contributed to the two reasons for stopping the study.

MAIN OUTCOME MEASURES: These were differences in incidence rates of breast cancer (reason 1), and breast cancer, coronary heart disease, stroke and pulmonary embolism (reason 2), among estrogen plus progestin and placebo recipients.

RESULTS: In the WHI study, 44.4% of the women on active treatment, as against 6.8% of the placebo recipients, had their treatments unblinded (mainly because of vaginal bleeding). Among them, detection bias could not be excluded. For the three cardiovascular outcomes, bias became a strong likelihood after the women were twice cautioned about possible increased risks observed in the interim data among estrogen plus progestin recipients. On the null hypothesis, bias could have accounted for the observed associations if, on average, it augmented the detection of disease that would otherwise have gone undiagnosed by 0.7-0.8/1000 per year.

CONCLUSIONS: For differences in incidence of the order of 0.7-0.8/1000 per year, it is not possible to discriminate between causation and detection bias as alternative explanations for the findings.