Time order

If allowance is made for the time from the diagnosis of breast cancer to its recording in a registry, virtually all the cases identified at 0.1 years of follow-up (HRT: RR, 1.37)22 or at 4 months (ET: RR, 1.19; E+P: RR, 1.41),24 were already present when the women were recruited (see: Detection bias) and time order was violated. In a properly designed cohort study breast cancers already present at baseline should have been excluded.

Time order was further violated in respect of the timing and duration of HRT use. In the third report23 follow-up information on HRT use was unavailable for 62% of the women. In the fourth report,24 by December 2002 the follow-up questionnaire had been received by about 66% of the women, among whom the response rate was 65%. Hence follow-up information on HRT use [and on menopausal status (see: Detection bias) and on confounders (see: Confounding)] was missing for about 57% [1 – (±0.66×0.65)×100] of the women. Following publication of the WHI findings7 there was a rapid and marked decline in the use of HRT.27 For that reason, as well as for other reasons (e.g. HRT-induced breakthrough bleeding),7 since 66% of ever-users of HRT at baseline were current users [our calculation: derived from Figure 1 (current use) and Figure 2 (past use) in Reference 21], a substantial proportion could have become past users by the end of 2002.

How unreliable were the data? Recruitment commenced in 1996 and follow-up ended in December 2002.24 For about 50% of the women the time from last contact to diagnosis was >1.2 years,21 and to the end of follow-up >2.7 years.23 For women enrolled in 1996 that interval could have been as much as 6 years. Thus it is likely that much of what was defined in the analysis as current HRT use became past use during follow-up. In addition, the duration data were incorrect (see: Duration-response), as were the data on menopausal status and confounding (see: Detection bias and confounding).

Information bias

Information bias in a cohort study is unusual, but it can occur, and in the MWS it was likely. At recruitment HRT users already aware of as yet undiagnosed breast lumps, or of suspect mammographic changes identified before recruitment (see: Detection bias), could have tended to overestimate the total duration of use. Had women who already had breast cancer at baseline been excluded, that bias could largely have been avoided (see: Time order).

A defect in the study design may also have facilitated the occurrence of information bias. Ethinylestradiol (EE), listed as one of 34 memory-prompts in the questionnaire25 as an HRT preparation, is a synthetic estrogen present exclusively in oral contraceptives. Women who were aware of breast lumps at recruitment, or who had suspect mammographic changes (see: Time order and detection bias), could erroneously have identified EE as HRT. Soon after publication of the MWS report21 the authors stated in an erratum that what was meant by ‘ethinylestradiol’ was ‘estradiol’.28 Yet the error was not corrected in the second questionnaire,26 administered 2–3 years after the first questionnaire.25

Detection bias

The design of a study of the risk of breast cancer in relation to the use of HRT in which the women were recruited from a screening programme guaranteed that it would be biased. By definition, women who decided to have mammograms were alerted to the possibility of breast cancer, as has also been acknowledged in an earlier study based on mammographic screening,29 and concern that HRT may cause the disease has been widespread, and has increased over time. The MWS invitation was explicit in the first questionnaire:25 “We have a unique opportunity … to learn about the way different types of HRT … [affect] a woman's health, particularly her breasts”. That wording ensured that HRT users already aware of breast lumps, or of suspected breast cancer, would selectively participate (see: Time order).

There was quantitative evidence of detection bias. First, HRT users were selectively enrolled: 32% of the women who participated and 19% or those who did not were HRT users.30 Second, the data suggested that women already aware of breast lumps, or of suspected breast cancer, tended selectively to participate (see: Time order): whereas the incidence of breast cancer in the MWS population was 2.8 per 1000 WY,31 in the population at large it was 2.0 per 1000 WY.21 Third, the baseline RRs of 1.3722 or 1.4124 (‘screen-detected’ breast cancer) indicated that women who both used HRT and who were also aware of breast lumps, or of suspect lesions, or of suggestive precancerous changes identified in earlier mammograms, were the most likely to participate. Fourth, the average time from recruitment to breast cancer diagnosis was 1.2 years,21 and 1.7 years thereafter the RR of fatal breast cancer was 1.22. An increased risk of fatal cancer among HRT users within 2.9 (1.2+1.7) years of recruitment was not plausible (see: Biological plausibility), and it could have been due to the selective enrolment of HRT users with pre-existing suspected or diagnosed breast cancer. Fifth, the RRs declined with increasing BMI,24 a known risk factor for breast cancer in postmenopausal women (see: Biological plausibility), and the larger the breasts, the less likely was it that otherwise occult breast cancer would selectively have been detected among HRT users by mammography.

Detection bias could also have occurred during follow-up, as previously described in our critique of the CR.6 Briefly, HRT users are advised to have regular breast examinations and mammograms, and in the MWS users more frequently underwent mammography than did non-users;30 when mammograms are performed HRT use is routinely recorded, and about 30% of breast cancers actually present go undetected;32 about 5% of postmenopausal women have ‘clinically silent’ breast cancer;33 and HRT diminishes the sensitivity of mammography.32 The mammograms of HRT users could have been more intensively scrutinised than those of non-users, especially if they were radiologically dense, and otherwise occult breast cancer could selectively have been detected among the users.

For both ET users and E+P users the RRs were lower during the first 4 months of follow-up than subsequently.24 The investigators stated that “it has been suggested that part of the increased hormone therapy-associated risk … observed in this study may have resulted from the selective recruitment of hormone therapy users who already had symptoms of breast cancer. If that had happened there would have been a greater hormone therapy-associated excess of breast cancer soon after recruitment than subsequently. However, the opposite was found”. They argued that these findings “largely [reflected] the lower hormone therapy-associated risks observed for screen-detected breast cancers than for non-screen-detected breast cancers”. That claim ignored the likelihood that during follow-up HRT users could more commonly have had repeat mammograms than non-users (see: Duration-response), and because of a further defect in the study design that possibility could not be assessed: information on repeat mammograms was not solicited in the second questionnaire26 (see: Confounding).

There was further evidence to suggest detection bias. The RRs were consistently lower for ET users than for E+P users.21,–,24 Unopposed ET causes uterine cancer, and ET is preferentially prescribed to hysterectomised women, among whom vaginal bleeding does not occur. By contrast, E+P is preferentially prescribed to women with a uterus, among whom breakthrough bleeding is common;7 and bleeding makes it mandatory to rule out endometrial cancer. HRT users alerted to the risk of that cancer would have become worried about breast cancer as well, and have sought to rule it out. Hence, it was to be expected that detection bias would be greater for E+P users than for ET users.

The RRs for invasive lobular and tubular tumours were higher than for ductal tumours.23 Lobular and tubular tumours are more highly differentiated, smaller, and more slow-growing than ductal tumours,34 35 and the detection of lobular tumours by mammography is also more difficult.36 More intensive scrutiny of mammograms of HRT users than of non-users could have resulted in the selective detection of lobular tumours, especially in radiologically dense mammograms, that might otherwise have gone undetected.

The RR for in situ breast cancer was 1.55.23 In situ tumours are seldom clinically detectible, usually they are identified by mammography, and the investigators acknowledged that detection bias was likely (see: Detection bias). Yet in the fourth report24 in situ and invasive breast cancers were considered together. In that report the RRs were higher if HRT had commenced within 5 years of the menopause than subsequently. However, since the data for in situ breast cancer were biased, the combination of in situ and invasive breast cancer was also biased. In addition, most of the women who were premenopausal at recruitment would have reached the menopause during follow-up, among the 57% of women not followed that information was missing, and there was substantial misclassification of menopausal status, and of the time since menopause (see: Time order).

The RRs declined with increasing tumour grade, and were higher for ER-positive than for ER-negative tumours, and for node-positive than for node-negative tumours.24 As shown in Table 1 (our calculations: derived from Figures 1 and 3 in Reference 24) unknown values for tumour grade, ER status and nodal status among current users of ET and E+P, and among never-users of HRT, ranged from 49.5% to 74.1%. Such high rates cast doubt on the validity of the evidence. In addition, the declining RRs with increasing tumour grade could have been biased if more common use of mammography by HRT users than by non-users resulted in the selective detection of low-grade tumours; an association with ER-positivity could have occurred if breast cancers in HRT users were more commonly tested, and if ER-positive, more commonly documented in the registries; and the higher RRs for node-positive than for node-negative tumours could readily have been due to detection bias.

How much bias would it have taken to account for the findings? In the first report21 the investigators estimated that among women aged 50–64 years the use of HRT would result in “five to six extra cancers per 1000 women with 5 years' use and 15–19 … per 1000 with 10 years' use”. Thus if detection bias resulted in the identification of 1–1.2 (5–6/5) otherwise occult cases each year among 1000 women exposed for 5 years, or 1.5–1.9 (15–19/10) cases each year among women exposed for 10 years, that bias would have nullified the findings. In the second report,22 among women aged 50–55 years the respective absolute risks for ET or E+P use for 5 years were estimated to be 1.5 and 6.0 per 1000. That is, if detection bias resulted in the identification of 0.3 (1.5/5) additional cases in ET users each year, or 1.2 (6.0/5) additional cases in E+P users, that bias would have nullified the findings. Absolute risks ranging from 0.3 to 1.9 per 1000 women per year could plausibly have been due to detection bias.

Confounding

Confounding was incompletely controlled. In the first,21 second22 and third23 reports the factors allowed for included age, time since menopause, parity, age at first birth, family history, BMI, region and socioeconomic status. In the fourth report24 age at menopause and alcohol consumption were also allowed for. During follow-up factors such as menopausal status, time since menopause, age at menopause and BMI changed, and for about 57–62% of the women the information was missing (see: Time order). In addition, information on the receipt of a mammogram during follow-up was not solicited in the second questionnaire26 (see: Detection bias).

Statistical stability and strength of association

In our critique6 of the CR5 we alluded to the relationship between the statistical stability and strength of any given association: if a RR is ‘large’ (say >5.0), a 95% CI that excludes 1.0 (i.e. a ‘statistically significant’ association) can be documented in a relatively small study. But if a RR is ‘small’ (say <2.0), usually it can only be documented in a massive study. The difficulty however, is that “if a massive study is sufficiently massive, any deviation of the RR from 1.0, no matter how small, becomes ‘significant’”; but it may be impossible to discriminate among bias, confounding and causation as alternative explanations. By contrast, “in a well-conducted study, when a RR is large, it may be reasonable to judge that it might perhaps be reduced, but not be obliterated, even if it were possible to entirely eliminate all sources of bias and confounding. But if an association is small it may be impossible to judge. In the latter circumstance ‘statistical significance’ may not equate with causality: given a massive amount of data, all that may be accomplished is to rule out chance as one possible explanation, but not bias or confounding”.

In the four reports the highest overall RR for HRT users was 1.74,23 and RRs in excess of 2.0 were identified only in subgroups. For ET users the overall RR was 1.30, and <2.00 in all subgroups. Such small RRs could have been due to bias or confounding. For E+P users the overall RR of 2.00 was again small,21 but significantly higher than the estimate of 1.30 for ET (p<0.0001). Or put another way, the RR for E+P versus ET was 1.54 (2.00/1.30). Such a small association could readily have been biased or confounded (see: Detection bias and confounding), illustrating how in a massive study, virtually any deviation of RR from 1.0, no matter how small, can yield a p value of <0.0001.

Dose/duration-response

Under a promotional hypothesis it might reasonably be expected that the use of HRT would confer a greater risk of breast cancer, the higher the dose or the longer the duration of use (see: Biological plausibility).

Dose-response

Dose-response was not analysed.

Duration-response

In the first report,21 for women who were using HRT at baseline (defined in the MWS as current users) the total duration of use of all episodes of use, current plus past, was analysed. That analysis was incorrect. Since the RR approximated unity within 2 years of stopping,24 the duration of past use was irrelevant, and only the duration of the current episode of use should have been analysed. In addition, the analysis of duration of use, as represented at baseline, misrepresented the actual duration of use, since follow-up information was missing for 62% of the women (see: Time order).

A further defect in the study design made it impossible to analyse the duration of current HRT use among women who used more than one product. In the baseline questionnaire25 five relevant questions were asked: “32. Have you ever used [HRT]?”; “35. For about how many years in total have you used HRT?”; “36. Are you now using HRT?”; “37. What is the name of the most recent HRT you have used?”; and “38. For how many years did you use the most recent type of HRT?”.

Based on questions 32 and 35, among current HRT users at baseline who used more than one product the total duration of ever-use could be analysed, but based on questions 36, 37 and 38 the duration of current HRT use could not be. To illustrate, consider a current HRT user who at baseline had used E+P for 9 years, and following hysterectomy, ET for 1 year: the current 1-year duration of ET use would have been recorded, but the current 10-year duration of HRT use (E+P, 9 years + ET, 1 year) would not have been.

In the second report22 duration data were not given. In the third report23 “there was no significant difference in the trends in [RR] with duration of use of either type of hormone therapy [ET or E+P] for ductal, tubular or lobular cancer”. In the fourth report24 the RRs for ≥5 years of use of ET and of E+P at baseline were higher than for <5 years of use, and higher for E+P than ET users. Those differences could have been due to detection bias; trends according to durations of <1, 1–4, 5–9 and ≥10 years were not presented; ‘total duration’ of use again referred to all episodes of use, not to current use; and the duration data were again misclassified, because of follow-up information on HRT use was missing for 57% of the women.

Finally, the RRs for increasing duration of follow-up were inconsistent with a duration-response effect (see: Internal consistency). Among women who used HRT, ET or E+P at baseline the RRs were highest at 0.7 years of follow-up, after which they declined.22 Yet under causal assumptions, the longer the duration of follow-up, the higher should the RRs have been. A plausible explanation of these inconsistent findings is that violation of time order and detection bias could have been greatest during the first year of follow-up.

Internal consistency

As described above, the RRs according to duration of follow-up were inconsistent (see: Duration-response).

External consistency

For ET users the MWS findings were inconsistent with those of the WHI clinical trial15 16 in which the evidence suggested that unopposed ET does not increase the risk of breast cancer. In the MWS there was quantitative evidence of bias, whereas in the WHI trial women were randomly assigned, ‘double-blind’, to ET or placebo, all participants were hysterectomised, vaginal bleeding did not occur, ‘unblinding’ was seldom necessary, the ‘unblinding’ rate was <2.0%, and there was little or no bias.20

For E+P users the MWS findings were inconsistent with those of the CR. In the MWS the RRs approximated unity within 2 years of stopping HRT;24 in the CR the RR only declined to unity 5 years after stopping.5

Biological plausibility

Elsewhere we have considered relevant pathological and experimental evidence for and against the possibility that HRT may cause breast cancer.6 19 20 Briefly, the hypothesis is not that HRT causes genetic mutation (initiation), but that estrogens, and probably progestogens as well, accelerate the proliferation of otherwise slowly growing malignant cells (promotion). Possible mechanisms are the proliferative effects of estrogens on estrogen-sensitive cells,37 or the excessive metabolism of estrogens to highly active compounds38 with strong proliferative as well as possibly genotoxic effects. However, estrogens also have antiproliferative and pro-apoptotic effects,39 which could possibly reduce the risk of breast cancer. In addition, estrogens can be metabolised not only to potentially genotoxic metabolites, but also to carcino-protective metabolites, such as 2-methoxy-estradiol.40

In short, some mechanisms could possibly increase the risk of breast cancer in HRT users, and other mechanisms could decrease it. However, under a promotional hypothesis, for the most aggressively multiplying cells it is generally accepted that on average it takes at least 10 years to attain a tumour diameter of about 1 cm, which is about the smallest lesion that can be diagnosed clinically.38 In the MWS the average total duration of HRT use at baseline was 6.1–6.9 years,22 and the duration of current use would have been appreciably less. Since the RR approximated unity within 2 years of discontinuing HRT use, among current users of HRT the duration of past use cannot have had any effect. It is implausible that the current use of HRT at baseline for less than 6 years could have increased the risk of breast cancer. It is also implausible that cancer cells, once already promoted, and once already invasive, could have ‘unpromoted’ within 2 years of stopping HRT.24

Obesity is a risk factor for breast cancer in postmenopausal women, perhaps because of increased endogenous estrogen secretion,37 and the RRs declined with increasing BMI.24 Under a causal hypothesis, however, although obesity itself increases the risk of breast cancer, the RRs among HRT users should have been higher than among non-users within strata of BMI, and the decline in the RR was explicable by the diminished sensitivity of mammographic screening with increasing BMI (see: Detection bias).