Commentary: The HRT story: vindication of old epidemiological theory

In 1995, a long time before any result of any randomized controlled trial would be known, I wrote that about half of the alleged 35–45% reduction of myocardial infarction with hormone replacement therapy (HRT) in observational studies would not be real.1 Still, I should confess that the total reversal of the effect—a slight increase of myocardial infarction in the first years of use—took me also by surprise. Why was I so certain that a large part of the seeming protection from myocardial infarction would not be true—even in the face of an overwhelming number of epidemiological studies?2 I guess, in essence, because during my training in epidemiology at the Harvard School of Public Health in 1978/79, I was much influenced by Miettinen's argument about the difference between studying the ‘intended and unintended’ effects of treatments.3 The view of the believers of the HRT protection was summarized in a lecture by a leading gynaecologist that I attended in the 1990s: he showed beautiful coloured slides about the protective effect of HRT, telling his audience that all possibility of bias and confounding was ruled out by careful statistical adjustments in well-designed studies. He also had one slide in black and white, showing the possibility of an increase of venous thrombosis with HRT, but added that this would most certainly be due to bias. Actually, the theory on the difference between studying intended and unintended effects predicts exactly the opposite. Adverse effects are unintended, therefore generally unexpected and unpredictable—at least in the usual clinical consultation. Thus, there is little likelihood of bias and confounding for adverse effects of HRT, such as venous thrombosis. One can even increase the unpredictability of adverse effects by limiting a study to people without any risk factor for the adverse effect, as was described already in 1978 by Jick and Vessey4—for example, by limiting studies on venous thrombosis and different oral contraceptives to otherwise healthy young women without any risk factor for thrombosis. In contrast, signs were all over the wall that prescription of HRT was highly selective against risk factors of myocardial infarction. Physicians who prescribed HRT in the 1970s and 1980s were risk avoiding about coronary risk factors, as was described by many authors (see ref. 1 for overview of critical opinions from the time before the RCT on HRT). This risk averse prescribing behaviour has persisted until the year 2000.5 Coronary disease has many clinically recognisable risk factors, such as hypertension, hypercholesterolaemia, and diabetes. If prescription is with an eye on such risk factors, and in particular if many complex and interdepedent risk factors come into play, the confounding bias introduced by this prescription process cannot be grasped any more by simply adjusting for a few variables in a logistic regression.3 A similar argument was used by Rubin in defence of randomization: he explained that a clinician's treatment decisions are just too complex to be modelled, and that therefore randomization is always preferable.6 In later works, Rubin and Rosenbaum coined the concept of ‘strong ignorability’ of the allocation mechanism: in specific subgroups, or in specific circumstances, the allocation mechanism might be ‘ignored’: it can be left out of a model when it plays no role in determining the outcome.7 In general, adverse effects will meet that condition because they are most of the time unpredictable, or not taken into account, when prescribing.3,4 Thus, for adverse effects observational studies will be as good as randomized trials. Indeed, for the adverse effects of HRT, such as venous thrombosis and breast cancer, observational and randomized studies concurred almost ‘on the dot’, at least in terms of relative risk.8,9 The main lesson of this episode to me is a vindication of the old theories on ‘intended and unintended’ effects and the investigation of adverse drug reactions.3,4 To study the effect of a treatment that is prescribed on the basis of complex rules of indication and contraindications, observational studies run a high risk of getting the answer wrong. If there is no form of sufficiently ‘haphazard allocation’ of the exposure at baseline, it will not help to throw a few variables into a model for adjustment, and then hope that the causal risk has been quantified.10 In observational research, such haphazard allocations—i.e. allocations that are not intrinsically tied to prognosis—are most likely for unexpected and undesired effects. ‘Haphazard allocation’ is not the same as physical randomization. However, as stated by Rosenbaum: ‘Haphazard is not random … Still, haphazard or ostensibly irrelevant assignments are to be preferred to assignments which are known to be biased in ways that cannot be measured and removed analytically.’11