Outcome standardization to blood pressure reduction in meta-analyses

In the current issue of the Journal of Hypertension, Brunström and Carlberg [1] question the role of standardization procedure to report the results of meta-analyses in the field of hypertension. Authors are dealing with an important and at the same time intriguing issue, as the standardization to blood pressure (BP) reduction of relative risks (RRs) with or without their standard errors (SEs) may introduce bias in the interpretation of the results. It is important, as the pooled risk estimates after standardization are different from those based on crude estimates, and it is intriguing, because various standardization methodologies have been used previously by different groups of investigators to report their findings [2–4]. However, other investigators including the authors of the present meta-analysis [1] constantly have resisted standardizing to BP reduction [5], whereas others have selected to standardize or not based on the research question [6]. It may be useful to standardize to determine how different outcomes are reduced for the same extent of SBP and/or DBP reduction. This is the case of BP-lowering treatment meta-analyses including trials of intentional (vs. placebo or more vs. less aggressive treatment) and nonintentional BP reduction, as well as the case in which older trials of untreated patients are tested together with new trials of patients under baseline antihypertensive treatment [6]. BP reductions are higher in intentional BP-lowering trials, mostly including patients without baseline treatment at variance with nonintentional trials mostly including patients already treated at baseline. Unstandardized analyses may suggest that different BP reductions result in significant outcome reductions, but standardizing outcome RRs to a preselected (standard) BP reduction makes the result much more straightforward for the usual clinical practice. As carefully pointed out by Brunström and Carlberg [1], the rationale to standardize for the attained BP difference among arms is that treatment effects on various outcomes is expected to be related to the attained BP difference. It has been previously demonstrated through meta-regression analysis [6,7] that the extent of BP lowering is related to the magnitude of outcome reduction in a semilogarithmically manner for most but not for all relevant outcomes. This rationale represents a sine-qua-non criterion to allow standardization for a predetermined range of SBP and DBP difference, meaning that one may standardize only if the logarithmic outcome reduction is related to the extent of BP reduction [6,7]. Any attempt of standardization out of the range, in which the association between outcome and BP reduction has been previously demonstrated, is at least biased and problematic. BP lowering can be achieved by different antihypertensive agents. Standardizing RRs to BP reduction for different classes of antihypertensive agents vs. placebo has no meaning as indirect comparisons in traditional meta-analyses are not allowed [8]. The effect of different classes of antihypertensive agents for the same outcome can only be tested in head-to-head comparisons for a 0/0-mmHg SBP/DBP difference, respectively. Regarding the initiation of antihypertensive treatment (in patients without baseline treatment), one may question whether BP lowering is effective or not across different baseline BP levels or different hypertension grades (or stages). Once again, standardized outcome measures are not clinically important among different hypertension grades (or stages), because the overall cardiovascular risk is by far different across groups [9]. However, standardization might be clinically meaningful to test the differential impact of BP lowering on outcome reduction at different levels of achieved BP or between trials grouped by the level of overall cardiovascular risk (i.e. low/moderate vs. high vs. very high risk trials) [9,10]. To investigate the effect of standardization for SBP reduction on all-death and major cardiovascular events, Brunström and Carlberg [1] used a meta-analysis of 70 BP-lowering trials including almost 300 000 participants. The specific aim of this investigation was to present how the individual trial effect estimate and variance are changing following standardization to a 10 mmHg of SBP reduction for either the mean risk ratio alone or both the mean risk ratio and the SE of each individual trial. Brunström and Carlberg [1] also examined how the implementation of these two different standardization methodologies affect the pooled results (summary estimate and heterogeneity across trials) compared with the unstandardized (crude) results in the same meta-analysis. According to the results of this exemplary analysis [1], standardizing each individual trial risk ratio to a 10-mmHg SBP reduction was associated with changes in individual trial variance and effect estimate. Also, it was accompanied by increased heterogeneity across trials and a very different pooled effect estimate in the random-effects model analysis compared with the analysis without standardization. When risk ratios and SEs were both standardized to a 10-mmHg SBP reduction in addition to the shortcomings observed previously, the confidence intervals become wider, making the validity of the fixed-model analysis unreliable. To better understand why standardization procedures to BP reduction may be problematic, some points deserve attention. The selection of all-cause death and major cardiovascular events as relevant outcomes of the present investigation seems reasonable, because the first is the most objective and the second is the most comprehensive outcome compared with all other individual cardiovascular outcomes [1]. However, all-cause death logarithmic reduction only marginally has been found associated with the extent of SBP lowering [6]. Thus, standardization to BP reduction for all-cause death may a priori be judged as somehow extrapolated. Beyond major cardiovascular events considered here by the authors [1], other outcomes linearly related to the extent of BP reduction are cardiovascular death, heart failure and stroke [6,7] which were not evaluated in the present investigation [1]. Standardization to a preselected BP reduction for coronary heart disease is discouraged, because the extent of BP lowering was not found linearly associated with the resulting outcome reduction [6,7]. The selection to standardize only for SBP and not for both SBP and DBP reduction also seems reasonable, because SBP compared with DBP reduction is better associated with outcome benefits [6]. However, treatment benefits are associated with both SBP and DBP reduction. Thus, standardizing only for SBP and not for DBP reduction may further complicate the interpretation of the standardized results. Although the selection to standardize to a 10-mmHg reduction for SBP is suggested by previous meta-analyses in the field [2,4,6], in the present analysis by Brunström and Carlberg [1], the overall SBP reduction was of only 5.6 mmHg. Standardizing to a much higher SBP reduction (10 mmHg) compared with the observed (5.6 mmHg) introduces extrapolation bias. The increased heterogeneity following standardization to a higher than the observed BP reduction (i.e. 10 vs. 5.6 mmHg) should be interpreted against the reduced heterogeneity observed with standardizations to a BP reduction well below the observed (e.g. 1 vs. 5.6 mmHg). Thus, selecting to standardize far above or below the mean observed SBP difference has a bidirectional (i.e. increasing or decreasing) trailing effect on the ‘standardized statistical heterogeneity’. Statistical heterogeneity cannot explicitly describe the clinical heterogeneity which is almost always substantial and many times is erroneously interpreted by statistical means only. The increased heterogeneity observed after standardization explains – at least partly – the theoretical gap between statistical and clinical heterogeneity. As pointed out by the authors [1], differences in baseline characteristics are more important than statistical heterogeneity in this setting (e.g. trials of hypertensive patients and trials of normotensive patients considered together, age, comorbidities, different ongoing baseline treatment). At variance with the standardization methodologies examined here by Brunström and Carlberg [1], one can avoid the distortion of study weights and imprecision by standardizing to the summary estimate of each individual outcome and not to the risk ratio of each individual trial. Thus, the first step is to estimate the pooled size effect of BP lowering for the outcome of interest, then to estimate the weighted by the inverse variance (before standardization) BP difference across trials and finally standardize the summary RR to the weighted BP difference. This alternative methodology preserves the weight of each individual trial and does not affect the degree of imprecision of each individual trial. However, even in this case, the same general limitations of standardization procedures still remain alive. Standardization to a preselected BP reduction is a problematic issue per se, irrespectively of the implemented methodology. This is because the ‘real-world’ summary effect estimate is transformed to an imaginary estimate. We agree with the conclusion by Brunström and Carlberg [1] that standardization should generally be avoided in meta-analyses of BP-lowering treatment. However, when selecting to standardize to unmask some clinical insights that crude analysis cannot provide, one should decide whether the projected comparison of the standardized RRs has a solid clinical background and whether the outcome to standardize is linearly associated with the extent of BP reduction. Finally, it is important to overcome problems related to the architecture of the meta-analytic procedure like heterogeneity, imprecision and variance changes. ACKNOWLEDGEMENTS Conflicts of interest C.T. declares consultancy fees from Sanofi and Menarini and lecture honoraria from Servier, Menarini and MSD.