Most studies of an association of visit-to-visit variability of blood pressure with increased risk of future adverse cardiovascular events are of short duration and rarely include a placebo group. Using data from the double-blind, placebo-controlled Systolic Hypertension in the Elderly Program, the authors examined mortality from cardiovascular causes up to 17 years of follow-up using the National Death Index. Visit-to-visit blood pressure variability was associated with cardiovascular death after adjustment for sex, age, serum creatinine, diabetes, body mass index, smoking status, left ventricular failure, and high-density lipoprotein cholesterol. The relationship was significantly stronger in the active treatment group compared with the placebo group. Although this could be the result of an effect of the medications used unrelated to visit-to-visit variability, the data are compatible with the hypothesis that inconsistent adherence leading to missing active medication doses may be an additional explanation for the relationship of visit-to-visit variability with cardiovascular death.
Randomized clinical trials, meta-analyses, and clinical guidelines indicate a clinical benefit in controlling hypertension.[1-9] In recent years, emphasis has been placed on the relationship of blood pressure (BP) visit-to-visit variability (VVV) with morbid and mortal events.[10-13] This relationship may be a causative effect of VVV, eg, very low BP precipitating stroke or myocardial infarction (MI), high BP causing intracranial bleed, or raised inflammatory markers associated with VVV impairing endothelial function.[10, 14-19] Another possible explanation of the link between VVV and adverse clinical events is inconsistent adherence to active medication. This question can be addressed by examining the effects of VVV in a randomized trial where inconsistent adherence to active therapy would have a stronger effect than inconsistent adherence to placebo. There are no reports of a significant relationship of VVV to untoward events in clinical trials using placebo or studies with long-term (>10 years) follow-up. In addition, a significant association of self-reported adherence and VVV was not observed in a study of 1391 individuals taking antihypertensive medication. However, these investigators examined average adherence per year rather than consistency of the adherence from visit to visit.
To test the hypothesis that a factor mediating the relationship of high BP VVV with long-term cardiovascular (CV) morality is inconsistent adherence to active antihypertensive therapy, we examined the effects of VVV on CV death in the active treatment and placebo randomized groups at the 17-year follow-up of the Systolic Hypertension in the Elderly Program (SHEP).[22, 23]
SHEP, a study of 4736 persons with isolated systolic hypertension (ISH), defined as systolic BP (SBP) ≥160 mm Hg and diastolic BP (DBP) <90 mm Hg, took place between March 1, 1984, and January 15, 1991. Men and women with baseline SBP between 160 mm Hg and 219 mm Hg and DBP <90 mm Hg were randomized to stepped-care antihypertensive treatment with a target SBP goal of a decrease of SBP by at least 20 mm Hg or to <160 mm Hg. Randomized participants were given chlorthalidone 12.5 mg/d that was doubled if SBP goal was not achieved, followed by the step 2 medication, atenolol 25 mg (or reserpine 0.05 mg) with the dose doubled if target BP was not achieved.[22-24] At the final visit of the randomized phase of SHEP, all participants were advised to take active therapy. Long-term follow-up for mortality and cause of death through December 2006 was obtained by the National Death Index using the birthdate and an acrostic of the initial letters of the last name, first name, and middle initial of the participants. The study was approved by the UMDNJ-Robert Wood Johnson Medical School and the University of Texas Health Science Center institutional review boards.
At each visit, heart rate and BP were measured using random zero manometers by certified trained personnel after 5 minutes of quiet rest. BP was measured twice in the sitting position 1 minute apart. Visits occurred at baseline, at 1, 2, and 3 months, and every 3 months thereafter unless additional visits were needed for BP medication titration. The total duration of SHEP was an average of 4.5 years with a range of 4 to 6 years. For the purposes of this report, the duration of this clinical trial was divided into two periods (Figure 1): Period I, the BP variability determination period lasting 2 years for each patient (averaging 15 visits: minimum 9 and maximum 33) followed by period II, the observation period beginning the first day of year 3 after randomization and ending at the end of year 17 (the maximum separation between the active treatment and placebo survival curves). This ranged from a minimum of 4261 days (11.7 years) to a maximum of 5476 days (15 years). Participants who died during the first 2 years (BP variability determination period) were censored.
Selection of Covariates (Confounders)
Selection of covariates was performed using a 3-step process. The first step evaluated variables hypothesized to be related to the occurrence of CV death in the present dataset considering only the second (observation) period of the investigation. The second step correlated these variables with measures of BP variability to identify and exclude those with correlation >0.25. Confounders with correlations >0.25 were excluded in order to use only variables that were not highly correlated. The cutoff was chosen at 0.25 because it is unlikely that correlations above that value can be produced by chance for sample sizes in the range of SHEP. The third step involved selecting important groups of variables by clinical relevance. Using these procedures, we selected the following variables: treatment group, sex, age, serum creatinine, diabetes, body mass index, smoking status, left ventricular failure, and high-density lipoprotein cholesterol. SBP at baseline or at follow-up was correlated with BP variability and was not included as a covariate in the analysis. The reason for deleting these variables from the model is that if covariates are highly correlated then their corresponding parameter estimates are also highly correlated and the interpretation of statistical significance of an individual parameter alone becomes conservative and noninterpretable. A statistical model analysis with a set of covariates that are not highly correlated is more interpretable. Since SBP is a predictor of morbid and mortal events, we performed additional analyses of the association of VVV with CV death in the entire SHEP cohort and separately in the active treatment and in the placebo groups adjusting for mean SBP during the VVV determination period using separate Cox models. We also examined the relationship of CV death to compliance measures available in SHEP. Average compliance to study medication as assessed by pill count for step 1 and step 2 medications among all SHEP participants and separately for those in the active treatment and the placebo groups was measured. In this analysis we excluded 25% of the values of adherence that were higher than 100%.
The following three models were used to evaluate the relationship of VVV to CV death: First, the relationship of each variability measure to CV death was analyzed in a univariate fashion. Second, and separately for the active treatment and placebo groups, the relationship of each variability measure to CV death was analyzed using Cox proportional hazards regression adjusting for the confounders listed above. Finally, in order to examine whether the relationship between VVV and CV death is different by randomization group, we performed the analysis in the entire population of SHEP adjusting for the above covariates as well as for treatment allocation (active vs placebo) group. The analyses were performed twice: in the whole dataset and after exclusion of SHEP participants who did not crossover from active to placebo or vice versa.
In previous studies, investigators of VVV have used a variety of measures of variability. We employed two concepts in statistics to measure VVV: the residual variance and smoothness penalty. Consequently, we tested the following measures of variability obtained for each individual participant from SBP measurements in the VVV determination period (randomization to the end of 730 days). (1) The sum of squared deviations between each daily average SBP value and the trend-predicted SBP (rSSR). This was selected because it represents the direct estimate of residual (error) variability and, unlike in other VVV measures, does not show a variance-mean relationship. The square root of rSSR was used to make it interpretable in mm Hg units. (2) The variance of the absolute values of the second differences between successive daily average SBP values (VABS2). VABS2 was chosen because it is related to the penalty for smooth function estimation (eg, in splines, spectral functions, and wavelets).[25, 26] (3) Variance independent of mean (VIM) expressed as standard deviation/meanx, where x is derived experimentally such that the absolute value of the correlation between VIM and mean is minimum. VIM was chosen in order to compare our findings to previously published reports of the relationship of VVV to outcomes.[10, 27] The exponent in SHEP was 0.314, indicating a weak (near 0) relationship between the SBP standard deviation and the mean.
Results of the Randomized Phase of SHEP
Of the 4736 participants enrolled in SHEP, 2365 were randomized to active therapy and 2371 to placebo. The average age of patients was 72 years, 57% were women, and 14% were black. Mean SBP at baseline was 170.1±9.2 mm Hg in the placebo group and 170.3±9.4 mm Hg in the active treatment group. Mean SBP was lower throughout the trial for the active treatment group by 11 mm Hg to 14 mm Hg and DBP by 3 mm Hg to 4 mm Hg. The 5-year average BP was 143/68 mm Hg in the active treatment group vs 155/72 mm Hg for placebo. The main outcome measure, fatal and nonfatal stroke, was lower at 4.5 years of follow-up. At the 17-year follow-up, 665 (28.1%) patients in the active treatment group and 730 (30.8%) patients in the placebo group had died from CV causes.
In addition, allocation to active treatment resulted in a 13.8% adjusted reduction in the risk of CV death over 17 years (hazard ratio, 0.874; 95% confidence interval, 0.783–0.977; P=.0237).
VVV and Associations of VVV With CV Death in the Active Treatment and the Placebo Groups of SHEP
Table 1 shows the statistics of variability measures in the active treatment and placebo groups. rSSR and VABS2 were significantly higher and VIM was significantly lower in the placebo group than in the active treatment group (P<.0001 for all).
Table 1. Variability Measures in the Active Treatment and Placebo Groups of SHEP (All Patients)
Active treatment (n=2188)
Placebo – Active Estimate (95% CI)
Abbreviations: CI, confidence interval; IQR, interquartile range; rSSR, sum of squared deviations between each daily average systolic blood pressure (SBP) value and the trend-predicted SBP; SD, standard deviation; SHEP, Systolic Hypertension in the Elderly Program; VABS2, variance of the absolute values of the second differences between successive daily average SBP values; VIM, variance independent of mean.
Wilcoxon rank-sum test of difference
−0.224 (−0.266 to −0.182)
There was a significant association between high VVV and the occurrence of CV death (Table 2). In most cases and for both adjusted and unadjusted comparisons, VVV was a stronger predictor of CV death in the active treatment group than in the placebo group. When all data for both the active treatment and the placebo groups were analyzed together, there was an interaction of the effect with the randomization group indicating a different (more pronounced) effect of VVV in the active group. In adjusted analyses, VVV (rSSR and VABS2) was associated with CV death at 17 years in the active treatment group while no significant association was observed in the placebo group. Also, unadjusted associations were more common and had lower P values in the active treatment than the placebo group. The association of VVV with CV death after adjusting for mean SBP during the SBP VVV determination period was statistically significant in the entire SHEP cohort (P<.001) and in the active treatment group (P=.001) but not in the placebo group (P=.127). Percent adherence based on pill count was very high in both the active treatment and the placebo groups (median, 97.1; interquartile range [IQR] 92.5–99.4 and 97.5 and IQR 93.9–99.5, respectively, for the step 1 drug and median, 97.0; IQR 91.56–100.0 and 97.8 and IQR 93.9–100.0, respectively, for the step 2 drug after exclusion of values >100). A relationship between the adherence measures available in SHEP including average adherence to step 1 and step 2 medications and their respective standard deviations was not observed in either the active treatment or the placebo groups.
Table 2. Relationship of Variability Measures With the 17-Year CV Death in the Active Treatment and Placebo Groups of SHEP
Type of Coefficient
HR (95% CI)
HR (95% CI)
HR (95% CI)
Abbreviations: CI, confidence interval; CV, cardiovascular; HR, hazard ratio; rSSR, sum of squared deviations between each daily average systolic blood pressure (SBP) value and the trend-predicted SBP; SHEP, Systolic Hypertension in the Elderly Program; VABS2, variance of the absolute values of the second differences between successive daily average SBP values; VIM, variance independent of mean. aP<.001 bP<.05. cP<.01.
When patients were stratified into 4 groups according to the magnitude of VVV (<40 mm Hg, 40–<60 mm Hg, 60–<80 mm Hg, and >80 mm Hg), there was an increase in the rate of CV death at 17 years in the active treatment group (from 16.0% to 30.1%, P<.001) while a significant effect was not observed for placebo (P=.143, Table 3). Active treatment participants with the highest VVV (>80 mm Hg) had a higher CV death rate than those in the placebo group, reversing the overall effect.
Table 3. Number of Deaths Between Year 2 and Year 17 by rSSR Category
rSSR Categories, mm Hg
Rate Ratio Active/Placebo
CV Deaths, No.
CV Deaths, No.
CV Deaths, No.
Abbreviations: CV, cardiovascular; rSSR, sum of squared deviations between each daily average systolic blood pressure value and the trend-predicted systolic blood pressure.
40 to <60
60 to <80
Analysis of the 4008 SHEP participants who did not crossover from active to placebo or vice versa showed directionally similar effects.
In this study, we observed a relationship of VVV measure as rSSR and VABS2 with CV death at 17-year follow-up. This relationship was evident among persons randomized to chlorthalidone-based stepped-care therapy while there were very few significant effects among those randomized to placebo. A relationship of CV events to VVV was observed in the majority of previous studies.[10-13, 25, 28-33] Hsieh observed a significant association of VVV to all-cause mortality independent of mean BP and after adjusting for baseline data in patients with type II diabetes. Similar findings have been reported among 144 patients undergoing renal dialysis. The effects of VIM were present only in unadjusted analyses in SHEP. In the SHEP database, the exponent in the denominator is <1 (close to zero) indicating an effect different from that reported by Rothwell and colleagues. This difference may the result of the lack of a mean-variance relationship combined with the strong effect of treatment on SBP.
VVV is related to baroreflex function and may be influenced by carotid endarterectomy. Barorelfex sensitivity diminishes with increasing arterial stiffness caused by aging and risk factors. Sinoaortic denervation causes an increase in BP variability without changing mean BP levels. Studies in rats with sinoaortic denervation have shown an impaired endothelial function, thus providing experimental evidence supporting the hypothesis that high VVV may directly contribute to the pathogenesis of impaired endothelial function. Long-term anti-inflammatory and antioxidant treatment prevents organ damage suggesting that inflammation could be the underlying mechanism facilitating endothelial dysfunction in rats. In humans, Diaz and colleagues reported that increased BP variability impairs endothelial function independently of average BP. In another study, VVV of SBP independently correlated with log urinary albumin excretion and ankle-brachial index even after adjustment for known risk factors, including average SBP and age.
Two mechanisms may underlie the relationship between VVV and outcomes. First, VVV may directly cause morbid and mortal events (eg, very low BP may precipitate a stroke or MI, high BP may result in intracranial bleeding, and raised inflammatory markers caused by VVV may impair endothelial function). Second, an indirect association of VVV with CV death may be explained by risk factors (eg, arterial stiffness, sympathetic overactivity, hostility, diabetes, smoking, low glomerular filtration rate, albuminuria, and left ventricular hypertrophy) causing both VVV and CV death.[10, 14-19] We hypothesize that, in addition to the above, inconsistent adherence to active therapy may be causally related to VVV as well as to the occurrence of CV death (Figure 2). We did not observe a statistically significant effect of average adherence on CV death, an observation similar to that of Muntner and colleagues who suggested that low average adherence to antihypertensive medication explains a small proportion of VVV of SBP. However, adherence, as usually reported, is not necessarily directly related to VVV since it measures the mean (frequently annual) adherence rather than the consistency of taking medication from day to day. For example, a patient may have low adherence and low VVV if they consistently take the same number of pills each day, or low adherence and high VVV if the number of pills taken varies significantly from day to day. The same may occur for persons with high adherence.
As expected, the rate of CV death increased monotonically with increasing VVV stratum in both the active treatment and the placebo groups (Table 3). Also, the ratio of CV death rate (active/placebo) was lowest in the lowest VVV stratum, implying higher importance in receiving active therapy than the VVV effect. This became less important with increasing VVV, and among participants with the highest VVV, the CV death rate was higher in the active treatment group, implying a stronger effect of VVV than randomization to active therapy.
This analysis has significant limitations. The predictor variables included in the models had a rather small impact on CV death. The relationship of VVV to CV death that was observed primarily in the active treatment group may be caused by the effects of the active medications unrelated to VVV, eg, the occurrence of lower lows on diuretic treatment or increased VVV as a result of study medications. However, the VVV in SHEP was significantly higher in the placebo group (Table 1). In addition, the lack of a relationship of adherence to CV death does not disprove our hypothesis that omitting active medication intermittently without a specific pattern may underlie the relationship between VVV and CV death primarily in the active treatment group.
SBP visit-to-visit variability defined as rSSR, VABS2 predicted CV death at 17 years of follow-up of SHEP participants. This was observed primarily in the active treatment group. Although this could be caused by an effect of the medications used unrelated to VVV, the data are compatible with the hypothesis that inconsistent adherence leading to missing active medication doses may be an additional explanation for the relationship of VVV with CV death.
This work was supported in part by a grant from the National Heart, Lung, and Blood Institute (NHLBI); the National Institute on Aging (NIA); and the Robert Wood Johnson Foundation (RWJ).
John B. Kostis received grant support from the NHLBI, NIA, and RWJ Foundation, and Barry Davis and Sara Pressel received grant support from the NHLBI and NIA.