Ideas for inclusion of trials
Although there are several systematic reviews and meta-analyses regarding the effects of blood pressure lowering treatment on cardiovascular diseases and mortality, the most recent and comprehensive we are aware of are Ettehad et al. [
13], which was published in 2016, and Brunström and Carlberg [
14], which was published in 2018.
The Ettehad paper included a whole range of randomized controlled trials on blood pressure lowering treatment. The authors searched MEDLINE from January 1, 1966 till July 7, 2015. They also searched the medical literature and identified trials till November 9, 2015. Their search strategy was compatible to that of Law et al. [
12] and used search terms related to hypertension, blood pressure, major classes of blood pressure lowering drugs (for example, diuretics), and the specific names of blood pressure lowering drugs listed in the British National Formulary. The studies Ettehad et al. included had a minimum of 1000 patient-years of follow-up in each trial group. The paper selected trials which compared active drug treatment groups with control groups, which mostly used placebo. The paper also selected trials which compared groups with different blood pressure targets. Although the paper did not exclude trials based on the existence of comorbidities at baseline, they excluded trials in patients with heart failure or left ventricular dysfunction.
The Brunström and Carlberg paper identified previous systematic reviews on the effect of blood pressure lowering from PubMed, the Cochrane Database of Systematic Reviews, and the Database of Abstracts of Reviews of Effect. They identified trials by scrutinizing reference lists from relevant reviews. They additionally searched PubMed and Cochrane Central Register for Controlled Trials in order to find trials after November 1, 2015 up till February 2017. The Brunström and Carlberg paper included mostly the same type of trials as the Ettehad paper, except for the following points. Unlike the Ettehad paper, the Brunström and Carlberg paper excluded trials in the acute phase after myocardial infarction from meta-analysis. They also excluded trials they deemed to be at high risk of bias. Unlike the Ettehad paper, which required a minimum of 1000 patient-years of follow-up for each trial group, the Brunström and Carlberg paper required a minimum of 1000 patient-years of follow-up for the whole study sample. We constructed our dataset for meta-analysis and meta-regression by adopting the trials included in these two papers. We used the names of included trials as used in the Ettehad paper and/or the Brunström and Carlberg paper.
One of the authors (YS) extracted the following data from each trial: trial start year, sample size of trial (number of total participants), number of events (all-cause mortality and stroke), mean age at baseline, mean systolic blood pressure (SBP) at baseline, difference in SBP reduction between the two groups, SBP reduction from the baseline in each group, attained SBP in each group, and the region in which each trial was carried out.
As defined in both the Ettehad paper and the Brunström and Carlberg paper, stroke included fatal and non-fatal stroke and excluded transient ischemic attack (TIA). In addition to data on stroke as a whole, we aimed to extract data on its subtypes (e.g., cerebral infarction, cerebral hemorrhage, and subarachnoid hemorrhage) from each trial and, where possible, performed analyses based on the subtypes. In the process of data extraction, we found that several papers covered only fatal stroke or non-fatal stroke. For these papers, we recorded the cases shown in each paper. We also found that several papers included TIA in the number of cases, and exclusion of TIA was not possible. In these cases, we extracted the number of cases including TIA.
As the Ettehad paper showed the number of events (all-cause mortality and stroke) in its appendix, we compared our extraction results with those of the Ettehad paper. When there were differences, we explored the causes of the differences. This exploration revealed that some data were different from those in the Ettehad paper. Regarding all-cause mortality, there were three differences compared with the Ettehad paper (Supp. Figure 11 [
13]). In PREVEND IT [
15], we calculated the number of non-cardiovascular deaths from the indicated ratio and added them to the cardiovascular death number. In BBB [
16], we could not extract all-cause mortality data because it was not indicated in the paper. In EWPHE [
17], we used the number from the intention-to-treat analysis. Regarding stroke, there were five differences compared with the Ettehad paper (Supp. Figure 8 [
13]). In PHARAO [
18], we added hemorrhage. In Oslo [
19], we added fatal subarachnoidal hemorrhage. In JATOS [
20], we subtracted the number of TIA. In EWPHE [
17], we used the numbers from the intention-to-treat analysis. In the Hunan Province study [
21], we used the data indicated in the original publication.
Regarding the variable to identify when each trial was carried out, we used the data of trial start year. To extract the trial start year, we first checked the papers for each trial. When we could not find the trial start year, we checked
ClinicalTrials.gov. When these did not work, we emailed the corresponding or first author of the original papers.
Differences in the attained SBP reduction between intervention groups and control groups (hereinafter, “SBP difference”) were determined in the following way. We retrieved three patterns of figures for SBP difference: (1) difference in SBP reduction between the two groups, (2) SBP reduction from the baseline in each group, and (3) attained SBP in each group. We prioritized (1) when reported. If (1) was not reported, we calculated the difference from (2). If neither (1) nor (2) were reported, we calculated the difference from (3). We used mean figures of the whole measurements when reported. When plural measurements were reported and mean figures were not reported, we calculated the mean of plural measurements. When only graphs of SBP transition were shown, we calculated the SBP levels using an online tool [
22]. Regarding the Lewis study [
23], SBP data were extracted from Black et al. [
24]. In BBB [
16], the number in each group was not mentioned, so we split the total amount (
N = 2127) and assumed both groups included 1064 participants.
The regions in which trials were carried out were categorized as follows: Asia, Europe, North America, Oceania, South America, and mixed regions. Trials carried out in two or more regions were included in the mixed region category. Turkey and Israel were included in Europe.
Statistical analysis
The statistical analyses in this study were carried out in the following two steps. First, we conducted single meta-regressions, in which the dependent variable was the log RR of all-cause mortality or stroke and the explanatory variable was each of the extracted variables, namely trial start year, sample size, age, baseline SBP, SBP difference, and region (Model 1M for all-cause mortality and Model 1S for stroke). RR was calculated by dividing the risk of outcome events of intervention groups by the risk of outcome events of control groups. Risk of outcome events of each group was calculated by dividing the number of participants experiencing outcome events by the number of participants. With regard to the variable of region, there were several variables for regional dummies in the model, but we included it in the table of single meta-regression because they all belonged to the regional category.
Second, we conducted multiple meta-regressions, in which the dependent variable was the log RR and the explanatory variables were all of the abovementioned variables (Model 2M for all-cause mortality and Model 2S for stroke). We judged whether the effects of blood pressure lowering treatment on all-cause mortality and stroke have diminished over time from Model 2M and Model 2S. Diminishing was confirmed only when the variable of trial start year was significant at the 5% level.
When there was no event in either group in a trial, we added 0.5 to the number of events and the number of non-events in both groups before performing the meta-regressions. All analyses were performed using STATA 15 (Stata Corp., College Station, TX, USA). For meta-regressions, we used the “metareg” command [
25], which estimates the random-effects model via the restricted maximum likelihood method.
For trials with three or more arms, we followed the classification performed in the Ettehad paper or the Brunström and Carlberg paper. The two papers defined the control group in the ONTARGET study [
26] differently; here, we followed the Brunström and Carlberg paper’s classification.