Observational epidemiological studies

Past description of the association between blood pressure and cardiovascular risk In a landmark paper, McMahon et al. [1] reviewed several epidemiological studies and offered very clear answers to two major questions about the association between blood pressure and risk: (i) is it monotonous and continuous? (the answer is yes); and (ii) is there a threshold under which blood pressure differences are not any longer associated with risk modifications? (the answer is no). The direct consequence is the lack of objective, scientific possibility to define two populations, one with hypertension, the other without: all the definitions of hypertension, based on blood pressure measurement, that have been and will be proposed, will stay arbitrary. Is this association different in women compared to that in men? Does it vary across decades? A recently published systematic review of studies including over than one million participants [2] has clearly shown the lack of significant changes between genders, but a decrease in the strength of the association with ageing. In addition, the authors have given convincing evidence that the prognostic importance of systolic is superior to that of diastolic blood pressure, and that diastolic blood pressure conserves its positive relationship across the different decades. What does this paper add to this sum of knowledge? The review presented in this issue by Rodgers et al. [3] deals with the subgroup of the Prospective Studies Collaboration from Asia and Australasia, pooled at the individual level, and concerning over 3 million person-years of follow-up. This has allowed exploration of the relationship between blood pressure level and subsequent risk, in adjusting on sex and age accurately in all cohorts. Other risk factors were sought in data bases, and used when they were available, to adjust the analyses. The results have been based on systolic blood pressure, better correlated to risk than the diastolic one. They have confirmed the lack of blood pressure threshold, and the lack of interaction between sex and blood pressure. The association has been found steeper at younger ages, but the differences in risk between quintiles of blood pressure were increased in the elderly. These results were similar for stroke and ischaemic heart disease, with a steeper association for stroke. The exploration of stroke subtypes has confirmed a slightly steeper association between blood pressure and haemorrhagic stroke, at least at younger ages. It did not find significant association between blood pressure and death caused by pulmonary circulatory disease, whereas the Prospective Studies Collaboration did. The role of the region of origin was assessed in multivariate models by introducing an interaction term between blood pressure and region. Age showed the greater contribution to the heterogeneity between regions, but significant differences in the association strength remained unexplained. Such explorations should be carried out more systematically because they tell us about the general value of the strength of the association. The more universal an association is found, the more valid is its use in every place. By contrast, significant changes of an association strength according to the origin of the study should make us cautious when we want to use this association (e.g. in risk scoring). For example, we found significant differences in the magnitude of gender association with cardiovascular death, with a dipper gap between men and women in UK compared to the other cohorts of hypertensives recruited in controlled trials [4]. Rodgers et al. [3] found a significant interaction between Australasian and Asian cohorts for stroke and ischaemic heart disease, with a significantly steeper association in Asian cohorts. This implies that a risk score established on one population may not be validly used on another. Publishing results based on such general overviews shows: (i) that the meta-analytical approach on individual patient data is feasible, even applied to epidemiological data; (ii) that narrower confidence intervals can be obtained, which suggests a rather general value for the summary results, concerning the association between inter-individual differences of blood pressure and cardiovascular risk; (iii) that exploring heterogeneity is a very important step to ensure definitely this general value: the summary results from the landmark paper of MacMahon et al. [1] cannot be properly used if not accounting for age and region of origin. What can not be inferred directly from these observational data? Burden of disease attributable to elevated blood pressure This question will remain a theoretical one. The answer would require the possibility of manipulating blood pressure, and blood pressure only. Blood pressure-lowering drugs are efficacious, but they do not change blood pressure only, as shown by the differences in risk reduction between drugs, which remain unexplained by blood pressure changes. A more interesting question is the magnitude of benefit to be expected from blood pressure lowering, through drugs or other interventions. Therapeutic benefit The data reported by Rodgers et al. [3] do not reveal which benefit is associated with the decrease of blood pressure. Numerous reasons are available to date, as outlined below. Therapeutic inferences from observational associations may be contradicted by further evidence from unbiased assessment (i.e. from randomized controlled trials). The discrepancy of results between cohorts and trials concerning betacarotene [5] or vitamin E indicated that observational results should not be translated directly into practice. Results obtained concerning pharmacological manipulations of another risk factor, cholesterol, illustrate how this simple reasoning with extrapolation of observational data may be misleading. Indeed, epidemiological studies did not reveal any significant link between cholesterol levels and stroke when considered altogether, and this may be due to the masking of a positive association with ischaemic stroke by a negative association with haemorrhagic ones. Despite the overall lack of association, the prescription of statins is followed by a significant and thus unexpected reduction in stroke rate. Even the link observed between treatment exposure and outcome may be misleading, as illustrated by experience with hormonal treatment after menopause: case–control and cohort studies gave coherent (thus apparently convincing) results, suggesting that hormonal treatment was associated with cardiovascular risk reduction. Unfortunately, to date, randomized controlled trials have shown exactly the reverse outcome [6,7]. If there was a perfect fit between inter-individual and intra-individual relationship, the weaker association observed between blood pressure and risk in the elderly should translate in a weaker treatment effect in this subgroup, for stroke and coronary heart disease (CHD) as well. However, if the results from clinical trials appear to correspond to this pattern for stroke, they do not for CHD, which appears to be better prevented in the elderly than in younger hypertensive individuals [8]. The risk reduction of congestive heart failure (fatal or not) observed in clinical trials of blood pressure-lowering drugs (−42%; 95% confidence interval −29 to −53%) [9] is stronger than that expected from these observations on fatal heart failure (−17%). The relationship between blood pressure reduction and risk reduction within clinical trials appears to be a J-shaped one, as shown by the meta-regression of Staessen et al. [10], and not a log-linear one, as suggested by observational data summarized by Rodgers et al. [3]. As a consequence, the proportionality between blood pressure and risk reductions can not be considered as established, and the suggested conclusion that a lower blood pressure is better may be wrong. Eventually, beyond the data available to date, there are theoretical reasons for a discrepancy among the between- and within-individuals relationships. (i) The between-individuals relationship is estimated from observational data, such as those analysed by Rodgers et al. [3], which, for the best, adjust on available risk factor measures, at the individual level, on different populations recruited in various settings; nevertheless, all these analytical procedures do not definitively warrant that the relationship finally observed reflects only the effect of blood pressure. (ii) The within-individuals relationship is estimated from clinical trials and their syntheses, meta-analyses or systematic overviews; clinical trials protect against confusion for the causality between prescription of a drug and risk reduction (or blood pressure reduction) that is observed. However, it does not demonstrate that the entire risk reduction observed is due to blood pressure reduction, and several other drug effects may explain a part of the risk reduction, such as metabolic effects, other mechanisms of action on the arterial wall, left ventricular muscle, etc., which are not mediated through blood pressure. True intensity of the association between all the risk factors and risk Taking into account the regression dilution bias is a logical technique for making the results closer to the ideal truth, which is expected to be of a direct causal nature, with a universal strength. In the study by Rodgers et al. [3], the resulting estimate has been adjusted on some associations with other risk factors (those which were measured in these studies). However, the regression dilution bias is not considered for these other risk factors, such as cholesterol or glucose level. Similar correction may lead to stronger associations with those factors and risk, but these may explain a greater part of the association between blood pressure and risk, if they contribute to unmask a confounding effect. In addition, this technique, allowing a modelling of ‘true’ or ‘usual’ blood pressure levels, increases the gap between scientific results and what can be inferred from measurements obtained in daily clinical practice. Blood pressure levels needed to benefit, and target blood pressure to be reached with treatment A simple look at the continuous positive association between blood pressure and risk, without any threshold, naturally raises both the issue of the minimal level to obtain a benefit from blood pressure lowering, and of the target to be reached with treatment. If formulating such hypotheses is invaluable per se, we should be aware of the dangers of considering the questions as if they were the answers. To date, the lack of threshold to benefit has received experimental validation, in secondary prevention after stroke, represented essentially by the PATS and PROGRESS trials. On the other hand, attempts to show the proportional benefit from a further decrease in blood pressure did not offer convincing evidence [11]. Conclusion: the values of epidemiology, in its place The role of observational studies in contributing to the current knowledge on cardiovascular disease is a major one. Rigorous approaches to synthesize that knowledge, and to explain its sources of heterogeneity, constitute further progress that must be appreciated: they force us to raise rational hypotheses (e.g. to explain why blood pressure–risk association is steeper in younger individuals, or in some regions), and they allow us to make better use of risk factors to predict the risk of a given individual, within his own background. However, we must always keep in mind the general limitations of observation, which will never replace the unbiased estimates of the effect from interventions, such as those obtained from properly conducted clinical trials.