1 Million Women Study Collaborators. Breast cancer and hormone-replacement therapy in the Million Women Study. Lancet 2003; 362: 419–27. preparations for less than 1 year might really relate to a mean duration of nearly 3 years ([2· 878]/100+ 0· 5= 2· 7). We do not agree with the authors that this bias away from zero is meaningfully counteracted by bias to zero from current users who stopped use and past users who started use during follow-up, since these groups are much smaller. Therefore, with respect to the effect of taking HRT for short durations, the results of the WHI study seem more informative.

Second, an advantage of cohort versus case-control studies is general freedom of selection bias caused by preferential accrual of women who have the disease and the exposure of interest. Was the MWS entirely free of this type of bias though? Invited women were told that “... the study is looking at how hormone replacement therapy affects a woman’s breasts...”(http://www. millionwomenstudy. org). Of those concerned about some breast abnormality when they received the invitation, therefore, HRT users were probably most likely to agree to participate. The same selection mechanism might apply to HRT users who were previously referred for further breast investigations after an abnormal or difficult to interpret mammogram. This bias might concern a sizeable proportion of women; according to results of the WHI trial, women on oestrogen-progestagen combinations more often have abnormal mammograms that need further assessment than do controls (9· 4% vs 5· 4%, respectively). Preferential selection of women who have an increased risk of breast cancer would lead to overestimation of the RR associated with HRT. Third, the increased breast density of HRT users and the resulting delayed diagnosis also has a differential effect on number of women-years included in the analysis. Under the given circumstances, with such short followup, the standard cohort analysis (Cox regression), as used by the investigators, allocates more persontime to the late-diagnosed cancers in HRT users, which would tend to bias risk estimates downwards. A logistical regression analysis as used in casecontrol studies (assuming all cases were present at study entry) might have been more appropriate.