Statistics from Altmetric.com
Goldacre et al1 use an impressively large database (28 616 cases and 325 456 controls) and an observed/expected ratio significantly below unity (0.83; 95% CI 0.74 to 0.93) in their record linkage study, to conclude that induced abortion “does not increase the risk of breast cancer.”
In their discussion, however, the authors acknowledge a massive deficiency—that is, that their “data on abortions are substantially incomplete because they only include women admitted to hospital (and) only include those in the care of the National Health Service (NHS)”. Considering that the majority of English abortions do not occur in NHS hospitals, most of the women in the study who did indeed have an induced abortion are probably misclassified as not having had any. The even more egregious nature of this flaw is reflected in the fact that a mere 300 cases—just over 1% of the total—are classified as having had an induced abortion. As the overall induced abortion rate in England and Wales averaged more than 1% per year during the study period (1968–1998),2 it is conservatively estimated that approximately 15% of the women in the cohort underwent an induced abortion in their lifetime. Consequently, more than 90% of the women in the study cohort who underwent induced abortion were misclassified as not having an induced abortion. Therefore, the Goldacre et al dataset is wholly inapplicable to the question of an association between induced abortion and breast cancer.
Such inappropriate use of a large dataset is reminiscent of a similar report from 1982 by an Oxford group with authorship overlapping that of the present study (D Yeates).3 In the 1982 study, Vessey et al, using a retrospective dataset consisting of 1176 case-control pairs, reported an odds ratio 0.84 (95% CI 63 to 1.12) for “miscarriage/termination” before first full term pregnancy. In the 1982 paper, the authors concluded: “The results are entirely reassuring, being in fact, more compatible with protective effects than the reverse.” This “reassurance” was claimed despite the fact that the only quantitative description, in that 1982 paper, of the population of women who actually had an induced abortion was “only a handful” (on which basis the authors justified their combination of induced and spontaneous abortion).3
In regard to possible “protective effects', it is also noteworthy that in their 2001 study, Goldacre et al attribute the slightly but significantly lower than expected number of abortions among cases “to confounding with other reproductive or lifestyle variables”. We would suggest that women of higher socioeconomic status—who are known to suffer higher breast cancer incidence—are underrepresented among NHS abortion patients, and thus more likely to be misclassified. This follows from the fact that only non-NHS abortions cost the patients any money.
Yet no less troubling than the use of an inappropriate dataset for discounting the repeatedly observed positive association between induced abortion and breast cancer, is the misrepresentation of the published record by Goldacre et al. In their discussion,1 they claim: “None of the cohort or record linkage studies have shown a significant increase in breast cancer risk after exposure to induced abortion.” In this context, they cite three studies, but omit the 1989 study of Howe et al,4 which was, in fact, the only other case-control study linked to medical records of induced abortion. We discussed this point in some detail in our 1996 review5 in this journal, but Goldacre et al erroneously indicate that we only reviewed studies based on retrospective data. Importantly, Howe et al reported an overall odds ratio of 1.9 (95% CI 1.2 to 3.0) for induced abortion.4
Such misguided attempts, as exemplified by Goldacre et al, to “reassure” the public about the safety of induced abortion in regard to breast cancer risk, do not serve to fulfill the need articulated by Stuart Donnan6 in his December 1996 editorial in this journal, for researchers “to have a view which might be called `pro-information', without excessive paternalistic censorship (or interpretation) of the data.”
Brind and Chinchilli suggest that incompleteness of ascertainment of abortion histories, and misclassification, are reasons for our not finding an increased risk of breast cancer associated with abortion.1 “Misclassification” is all one way: women identified as having had an abortion can all be assumed to be correctly classified. In some of the low ascertainment subgroups—for example, older women with short recorded histories—we readily accept that only (say) 85% are correctly classified as not having had an abortion. Incompleteness of recording is, unfortunately, a design characteristic of the dataset and method—based on NHS hospital cases only, and without a full lifetime history of the women—which is nevertheless the same for cases and controls. To maximise the number of cases, we included a wide range of ages and included periods of short as well as long recorded history. However, older women and those with short recorded histories would have contributed very little to either the observed values of prior abortion (Brind's calculation of 1%) or to the expected values. The important point is that, because the analysis was stratified by age and length of history, the cases and controls were the same in these respects. In subgroup analyses, subdividing by the women's age, birth cohort and year of breast cancer diagnosis, there are very different levels of recording of prior abortion. For example, considering women aged 30–39 years with breast cancer diagnosed between 1989–98, and their corresponding controls, 11.1% (1609 of 14 529) had a record of abortion and 5.9% (857 of 14 529) were specifically recorded as induced abortion. We think that many of the women whose record simply stated “abortion” were in fact cases of induced abortion but we report the data in precisely the way that they were recorded. In women aged 40–49 at the time of breast cancer between 1989–98, the corresponding figures for prior abortion were 8.7% (1199 of 13 734) and 4.3% (589 of 13 734). As shown in table 1, the relative risks in these women were very similar to those reported overall on lower levels of ascertainment. If under-ascertainment itself was important in comparing cases and controls, one would expect to find a divergence of relative risks at different levels of ascertainment. We did not.
The next issue is therefore whether bias—related either to ascertainment or other causes—obscured a true increase of risk. Brind and Chinchilli suggest social class as a biasing factor. The results within social class subgroups show no support for this (table 2). Because of the interest shown in the topic, we will publish results of further subgroup analyses.
If abortion increased the risk of breast cancer, one would expect the increased risk to have an identifiable timescale to reflect a latent period. An increase in risk over time would be expected even if, through bias or confounding, the underlying background risk was underestimated.
As we reported,1 subgroups defined by time interval between breast cancer and abortion show very similar relative risks and, specifically, there was no increase in “risk” with time since abortion. For example, the relative risk of breast cancer associated with induced abortion five years or less before the cancer was 0.80 (95% CI 0.6 to 1.0); and that associated with induced abortion 15 years or more before the cancer was 0.81 (95% CI 0.7 to 1.0).1
Brind and Chinchilli speculate about our interest in this subject. In fact, it stems mainly from a methodological interest in the use of record linkage to study topics that are hard to study using other methods.2 Case-control studies of abortion prior to breast cancer, using personal interview, are problematic because of responder bias. Cohort studies based on personal long term follow up of women who undergo abortion would be formidable and, for most women, very unwelcome. Record linkage is therefore attractive and, for reasons emphasised above, the dataset and methods used are appropriate.
Conflicts of interest: none.
Funding: South East Regional Office of the National Health Service Executive.
Conflicts of interest: none.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.