Confounding and indication for treatment in evaluation of drug treatment for hypertensionBMJ 1997; 315 doi: https://doi.org/10.1136/bmj.315.7116.1151 (Published 01 November 1997) Cite this as: BMJ 1997;315:1151
- Diederick E Grobbee, professor of clinical epidemiologya (, )
- Arno W Hoes, associate professor of clinical epidemiologya
- a Julius Centre for Patient Oriented Research, Utrecht University Medical School, Academic Hospital, PO Box 80035, 3508 TA Utrecht, Netherlands
- Correspondence to: Professor Grobbee
- Accepted 28 May 1997
In research on the effectiveness of treatments, the randomised controlled trial is considered the best study design because it enables several sources of bias to be removed from the observation. The most important advantage of such a trial is that the baseline prognoses of patient groups are comparable so that unbiased estimates of the effects of a particular intervention can be obtained. In non-experimental (observational) studies allocation to drug treatment is, by definition, not random. This usually means that the prognoses of the patient groups are not comparable and any inferences drawn about the relative effects of treatment are therefore invalid.1 2 In studies of patients who are not randomly allocated to a treatment arm but receive treatment when their doctor decides to prescribe it, the onus is on the investigator to achieve comparability. We discuss the pitfalls of non-randomised comparisons of treatment effects. The results of a recently published study which suggested that hypertension should not be lowered too far by treatment and the subsequent debate provide an example of problems that may arise in validating the conclusions of such studies.
Confounding by indication
A sensible doctor prescribes drugs only to patients who need them—those who have indications that this treatment is required. Moreover, he or she will give priority to treating the most needy patients. When two patients seem to have indications but only one is being treated, the treated patient probably has more compelling indications than the second. The prognosis in any given group of treated patients will be different from that in untreated subjects because the latter will not generally have any indications for treatment. Furthermore, although many drugs can affect the course of a disease positively, the outcome in people with that disease compared with those who do not have it or have a less severe form may be worse or, at best, similar. At first glance, therefore, the evidence seems to show that the drugs induce the disease rather than curing it. This biased observation is considered to arise from confounding by indication.3
Confounding by indications for treatment is a serious threat to the validity of non-randomised comparisons of treatment effects
Conclusions about the efficacy of treatment should not be drawn from comparisons of treated and untreated patients
Under some circumstances, confounding by indication in a non-randomised study can be limited or even removed in the design phase or data analysis
Reported increased risks of heart disease in hypertensive patients with low blood pressure during treatment may be attributable to pre-existing severe atherosclerosis rather than “overtreatment”
Unequivocal rejection of all non-randomised studies assessing treatment effects is unjustified
Confounding by indication commonly creates unsurmountable problems for non-randomised research on treatment effects. Valid inferences can be drawn only under those rare circumstances in which (a) groups of patients with the same indications but different treatment can be compared and (b) residual dissimilarities in characteristics in patients receiving different drugs for the same indications can be adjusted for. For example, Psaty et al compared the effects of several antihypertensive drugs on the risk of angina and myocardial infarction. They used a case-control study design, and took ample measures to exclude confounding by indication.4 An alternative approach would be to include only those subjects who are similar for all prognostic factors (such as a history of disease or presence of other risk factors) except treatment.
“Do not accept chemotherapy because you will die of cancer”
“Refuse to take non-steroidal anti-inflammatory drugs because people who don't use them have less arthritis”
“Avoid antihypertensive treatment as those who are treated have higher stroke rates”
The indications for drug use, because of their natural association with prognosis, may confound comparison so that it looks as if the treatment causes the disease
Research on the risks associated with drugs also gives examples of study designs in which a shared indication, combined with adjustment for any remaining prognostic discrepancies, limits confounding by indication. The putative association between non-potassium sparing diuretic drugs and sudden cardiac death in treated patients with hypertension was recently examined in two case-control studies. These compared the risks of sudden death in patients taking different classes of antihypertensive drugs, especially diuretic drugs.5 These examples are exceptions to the rule that non-randomised comparisons of treatment effects should not be trusted. That rule applies particularly when treated patients are compared with untreated ones rather than with patients treated with different drugs for similar indications. However, even when several confounding factors are taken into account, the validity of findings on drug benefits and risks in non-randomised studies may be questionable. This is illustrated by the debate that followed the publication of a recent case-control study on the putative risks of calcium antagonist drugs in the treatment of hypertension.6 7 8
In theory, the effect of confounding by indication could be completely removed by measuring all the patient characteristics which the doctor considered formed the basis of the indications for treatment and adjusting for these subsequently in the statistical analysis. Table 1 gives an example of adjustment for confounding by indication in a non-randomised study which compared the effectiveness of drug treatments for hypertension over 10 years of follow up in the Netherlands. Here, as a first approach to reducing confounding by indication, a comparison was made between treated and untreated hypertensive subjects rather than between treated hypertensive subjects and the rest of the population.9 Adjustment for differences in patient characteristics that were relevant to the prognosis caused a striking change in the risk estimates (table 1).
For statistical adjustment to be useful, several requirements must be met. Firstly, all relevant determinants of the indications should be known. Secondly, the information on confounding factors should be of sufficient quality (precision) to allow complete adjustment. Finally, the number of patients in the study should be sufficient to allow for statistical modelling of many confounding variables. If only some variables that determine the indication for treatment are available, statistical adjustment will be incomplete. In the example of hypertension shown in table 1, information on confounding by indication is clearly incomplete. A family history of hypertension, for example, could affect the doctor's decision to start treating a patient. Consequently, the rate ratio of disease in treated versus untreated hypertensive patients may still be biased upwards.
Blood pressure should not be lowered too far
Recently, Merlo et al presented an analysis from the study of men born in 1914.10 This showed that those patients treated for hypertension who had a diastolic blood pressure below 90 mm Hg during treatment were at a greater risk of myocardial infarction than the remainder of the population.10 An almost fourfold increase in risk was estimated, even after adjustment for several confounding variables. The authors concluded that blood pressure during treatment should not be reduced below 90 mm Hg.
This conclusion is surprising for several reasons. Since the authors estimated as 1.1 the risk of myocardial infarction in treated hypertensive patients whose blood pressure was greater than 90 mm Hg compared with untreated hypertensive patients, they ought to have concluded that hypertensive patients should not be treated at all. More importantly, the validity of their conclusion in the patients who had lower blood pressures during treatment is highly questionable.
Firstly, since valid measurement of all factors that influence a doctor in prescribing a drug seems impossible, part of the observed increased risk is attributable to residual confounding in relation to indication. This is well illustrated by the 90% increased risk in treated hypertensive patients compared with the rest of the population after extensive adjustment for confounding variables. These variables included previous myocardial infarction, previous cerebrovascular disease, ischaemic heart disease, intermittent claudication, diabetes mellitus, raised systolic blood pressure, duration of hypertension, hypercholesterolaemia, hypertriglyceridaemia, raised serum creatinine concentration, obesity, use of cardiac glycosides, and a history of smoking.
Secondly, a comparison of the risk of ischaemic heart disease events in treated patients with a low blood pressure and treated patients with a high blood pressure, which is therefore conditional on treatment, would be more relevant and valid. The results of this analysis using the data provided by Merlo et al are shown in table 2. These show that the presumed increased risk in treated patients with lower blood pressure compared with those with higher blood pressure is far from statistically significant.
Whether and why a low treated blood pressure puts a patient at risk of myocardial infarction is important. Several studies have suggested that the association between diastolic blood pressure levels and risk is J shaped.1617 One suggests that a low diastolic blood pressure compromises coronary blood flow and subsequently increases the risk of coronary heart disease.21 Results from several studies of hypertensive middle aged people have shown that the use of medication and the reduction in blood pressure after treatment may contribute to the J shaped association.16 17192223 These findings have caused heated debate on the diastolic blood pressure that should be achieved with antihypertensive treatment.16 17 Strong evidence against an important contribution of antihypertensive treatment comes from the observation that the J shaped association is also found in control groups in trials of drugs for hypertension.5 6 5 In the systolic hypertension in the elderly program trial in older people with isolated systolic hypertension, the mean diastolic blood pressure was lowered from 77 mm Hg to 68 mm Hg and the number of deaths from cardiac causes was not increased but reduced.26 Moreover, in observational studies a J shaped association between diastolic blood pressure and cardiovascular risk was also reported in patients who were not taking antihypertensive drugs.1325
Sleight proposed that stiffening of the large arteries in elderly people might lead to high systolic and low diastolic blood pressures, and that stiffer arteries are associated with an increased risk of coronary heart disease.27 Widespread atherosclerosis may be the link between stiffening of large arteries, low diastolic blood pressure, and an increased risk of cardiovascular disease.28 29 In support of this view, Witteman et al reported a J shaped association between the fall in diastolic blood pressure and the progression of atherosclerosis of the abdominal aorta in postmenopausal women.30 The results indicated that a fall in diastolic blood pressure may indeed be a result of stiffening of the vessel wall, as indicated by the progression of aortic atherosclerosis. Recent findings of thicker carotid artery walls in subjects with low untreated diastolic blood pressure who were participating in the Rotterdam study confirmed this.31
In studies on the effects of treatment, non-randomised comparisons can be affected by confounding by indication, and this may result in wrong conclusions. Appropriate methods of limiting confounding exist, however, and this means that non-experimental studies assessing treatment effects should not be rejected unequivocally. Nevertheless their conclusions should not be accepted uncritically,32 particularly when comparisons between treated and untreated patients are made, as in the report of Merlo et al.10 The conclusions of Merlo et al may compromise the quality of care currently given to many hypertensive patients as they may lead to stopping antihypertensive treatment and consequently to stroke or myocardial infarction.10 Before any change in the current goals of blood pressure treatment is made, unequivocal evidence must be provided that the observed association between low diastolic blood pressure and the risk of myocardial infarction is not an artefact resulting from pre-existing severe atherosclerosis. Randomised trials are under way to clarify this issue.33 In the meantime, there is no reason to deny the fact that treated hypertensive patients still have a high incidence of stroke and heart attacks. That, after all, is the reason for treating them in the first place.