SEARCH

SEARCH BY CITATION

Abstract

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

Blood pressure (BP) visit-to-visit variability (VVV) influences the risk of vascular events and mortality. Research has suggested that antihypertensive medication classes may differentially impact VVV. This study evaluated whether antihypertensive medication class differentially impacted BP VVV among hypertensive individuals in a clinical, “real-world” setting as well as the association between VVV and patient characteristics. Clinical observational data were extracted for adults (mean age, 63; 56% female, 86% Caucasian) with hypertension from the Mercy EpicCare EHR-Derived Database (MEDD) (n=183,374) who had at least 4 outpatient visits with BP readings. A multilevel mixed model for change over time estimated between- and within-subject effects on the absolute real VVV of systolic BP. Diuretics significantly lowered VVV (β=−0.32[−0.39 to−0.25]) and α-/β-blockers resulted in the highest VVV (β=0.89 [0.77–1.00]). Being older, female, and having a higher systolic BP and certain comorbid conditions significantly raised VVV (P<.001). The findings from the MEDD were consistent in general with other research on BP VVV. However, the magnitude of effect of antihypertensive medication class and patient characteristics was relatively low (<10% of the BP VVV variance for any one variable). More research is needed to evaluate the extent to which the class of antihypertensive medication class impacts BP VVV in the outpatient setting.

Long-term poorly controlled blood pressure (BP) in hypertensive patients is associated with increased cardiovascular, renal, and cerebrovascular morbidity and mortality.[1-3] Although the individual's systolic BP level is considered of primary importance in assessing risk of vascular events, it is possible that other BP parameters might also be influential. One such parameter that has been proposed is BP visit-to-visit variability (VVV).[4-7] Although BP VVV has at times been considered a nuisance variable,[8-11] there is burgeoning literature on the topic.[6, 7, 12] Rothwell and colleagues reported that in two large trials (United Kingdom Transient Ischaemic Attack [UK-TIA aspirin] trial and the Anglo-Scandinavian Cardiac Outcomes Trial-Blood Pressure Lowering Arm [ASCOT-BPLA]) systolic BP (SBP) VVV was found to be predictive of stroke and coronary events, including mortality (and was independent of mean SBP).[6, 13] Munter and colleagues found that higher SBP VVV is associated with an increased risk of negative cardiovascular outcomes and end-organ damage.[7] BP VVV has been previously shown to be associated with adverse outcomes such as mortality and cardiovascular events,[12, 14] yet there remains a gap to demonstrate what factors may explain higher levels of VVV.

Antihypertensive medications may have class-specific effects on BP VVV.[13, 15] A recent meta-analysis considering VVV across multiple randomized trials reported that some antihypertensive classes (calcium channel blockers, diuretics) decrease VVV, while others (angiotensin-converting enzyme inhibitors, β-blockers) increase VVV.[15] Within the ASCOT-BPLA trial it was reported that using a calcium channel blocker resulted in a 24% decrease in cardiovascular mortality and a 23% decrease in stroke compared with β-blocker therapy.[16] In at least one study, VVV was associated with being older and female and a predictor of myocardial infarction.[6] Given the potential impact of BP VVV on vascular events, it is important to evaluate whether certain patient characteristics are associated with BP VVV. Due to the evolving understanding of BP VVV on clinical outcomes, more work needs to be conducted to evaluate whether BP VVV is differentially influenced by the various classes of antihypertensive medications as well as patient characteristics.

In order to appreciate these relationships in a “real-world” setting, it would be beneficial to examine a clinical electronic health record (EHR)–derived database. The current study had two main objectives. The first objective was to investigate whether antihypertensive class differentially affected VVV among individuals in an EHR-derived database, while the second objective was to investigate the effect of patient characteristics on VVV.

Methods

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

Study Population

Patients were selected from the Mercy EpicCare EHR-Derived Database (MEDD), which included data from 1275 Mercy primary care physicians in Missouri, Arkansas, and Oklahoma. The MEDD was queried to identify patients who met the following criteria for inclusion in the VVV analysis: (1) older than 18 years, (2) coded diagnosis of hypertension (the International Statistical Classification of Diseases and Related Health Problems, Ninth Revision [ICD-9] 401.01–404.99), (3) all included observations (BPs taken in the clinic) occurring between January 2007 and June 2011, and (4) having at least 4 observations beginning at least 120 days following the initial coded hypertension diagnosis and continuing for 365 days. This was selected to eliminate the outlying readings that occurred during the initial stabilization phase of hypertension treatment as illustrated in Figure 1. Patients with gestational hypertension or <4 BP readings were excluded from the analysis. A total of 1,536,695 unique visits were captured in the MEDD during the timeframe of the study. After applying the above inclusion criteria, the resulting population included in the analysis totaled 183,374 patients with 922,868 BP readings. The study was conducted in accordance with the Mercy health system guidelines for human subject research and was approved by the local institutional review board.

image

Figure 1. Systolic blood pressure (BP) and absolute real visit-to-visit variability (VVV) during medication titration among diagnosed hypersensitive patients (n=374,105 BP readings).

Download figure to PowerPoint

Study Measures

Demographics

Race, age, and sex were extracted from the MEDD. Age and sex were populated for all patients. Among the 183,374 patients, 8 were missing a race designation; these individuals were included in the “other” category. Calculated body mass index demonstrated higher than expected inter-rater variance due to chart etiquette differences across sites requiring data quality processes to increase the reliability of height and weight.

Comorbid Conditions

Comorbid medical conditions were considered ongoing after the initial diagnosis unless otherwise indicated in the medical record with an end date. Identification of comorbid conditions was accomplished using ICD-9 codes (Figure 2) obtained from the problem list or encounter diagnoses within the EHR. In addition, sensitivity analysis was conducted to determine the proportion of patients who would be coded with a comorbid condition using the least restrictive relative to the most restrictive set of ICD-9 codes as described in the Physician Quality Reporting System measures.[17] Overall, the patient population did not significantly vary between the most to least restrictive definitions.

image

Figure 2. International Statistical Classification of Diseases and Related Health Problems, Ninth Revision (ICD-9) codes for comorbid conditions. EMR indicates electronic medical record.

Download figure to PowerPoint

Antihypertensive Medications

Medication class, as defined by the Systematized Nomenclature of Medicine,[18] was coded for each individual at each visit for either monotherapy or combination therapies. The type of antihypertensive medication class assigned to the individual was based on the medication they were prescribed (or continuing to be prescribed) at the previous visit [yt−1] or anytime between any two consecutive visits [yt−1, yt]. Other measures such as time using an antihypertensive medication class and percent of time between visits were considered, but were not used due to uncertainty in prescription end dates (2.4% of inactive patient medications were missing end dates) and an inability to gather reliable adherence data between visits.

Blood Pressure

One BP reading was captured in the MEDD following standard operating procedure by the clinical staff during outpatient visits. As stated previously, each individual was required to have at least 4 office visits (thus allowing for at least 3 measures of VVV) to be included. To ensure clinical relevance and remove the potential bias related to frequent office visits, only the first 7 office visit BP readings (allowing for 6 measures of VVV) were used for the analysis. In order to control for faulty data representation or erroneous data entries at the point of care, unreliable or nonsensical readings defined as SBP (>240 or <70) or diastolic BP (>130 or <40) were excluded from analysis.

BP VVV

The outcome of interest was the absolute difference in SBP from one visit to the next |ytyt−1| (termed the absolute real VVV), which was recently outlined as a technique to assess VVV by Muntner and colleagues.[14] We used absolute real VVV because[14] we found a relationship between absolute real VVV and mortality based on the Social Security Administration Death Master File.[19] A chi-square test for trend showed that there was a statistically significant difference between all quartiles (P value<.05, for each comparison). Mortality for each of the 4 quartiles (Qs) was as follows: Q1 (lowest VVV)=2.25%, Q2=2.61%, Q3=2.82%, Q4 (highest VVV)=3.77%.

Statistical Methods

Following the methodology of Singer and Willet,[20] we used a multilevel mixed model for change over time to estimate the between- and within-subject effects. All analyses were completed using SAS 9.2 (SAS Institute, Cary, NC).[21] The data in the analysis fit a number of the model's assumptions well. A repeated measures model was considered appropriate and a significant F test of the ordinary least-squares residuals demonstrated the dependency of the data both at the individual and provider level. This model controlled for variability in the number of patient visits as well as multiple drug therapies. No variables demonstrated high levels of multicollinearity (coronary artery disease and myocardial infarction had the highest Spearman correlation [ρ=0.30] among the comorbid conditions). All cells were adequately populated and thus allowed proper estimation of coefficients and standard errors.[22] The observations were repeated measures with varying lengths of time between each measure and these data demonstrated autocorrelation particularly as time between visits moved towards zero. As such, the first-order autoregressive was selected as the covariance structure. The selection was supported by the quasi-information criteria. The purpose of this model was to estimate the associated effects of antihypertensive medication class on BP VVV in order to assess any differences between patients taking different antihypertensive medication classes.

A propensity model was fit using a standard logistic regression in order to reduce selection bias related to medication prescription and improve the ability to draw valid conclusions. Rosenbaum and Rubin first described the propensity score method to reduce selection bias when treatment is not random.[23] In the current study, the propensity score reflected the odds of a patient taking any hypertensive medication conditional upon their characteristics (demographics, comorbid conditions).

Results

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

The population was predominantly female (56.1%) and Caucasian (86.3%), with an age range of 45 to 75 years (mean, 63.4 years). The most common comorbid condition was diabetes (26.9%). Table 1 shows the population differences between patients taking antihypertensive medications (monotherapy or combination therapies) at the first observation and patients not taking antihypertensive medications. Results show that the most notable and significant differences (P<.05) in BP were found when comparing comorbid conditions and age.

Table 1. Comparing Patient Characteristics by Antihypertensive Medication Prescription Status
 Total Population (n=183,374)Prescribed Antihypertensive Medication (n=105,010)Not Prescribed Antihypertensive Medication (n=78,365)
MeanMeanMeanP Value
  1. Abbreviations; BP, blood pressure; BMI, body mass index; CAD, coronary artery disease; CHF, congestive heart failure; MI, myocardial infarction; PVD, peripheral vascular disease; TIA, transient ischemic attack; VVV, visit-to-visit variability. P Value tests differences between medicated and unmedicated populations. Bold P values indicate significance.

Baseline systolic BP131.7132.0131.3.526
Baseline diastolic BP77.377.477.0 <.0001
Systolic absolute real VVV
Visit 1 – baseline13.813.813.7.542
Visit 2 – visit 113.713.813.6 .0002
Visit 3 – visit 213.613.813.4 <.0001
Visit 4 – visit 313.713.813.4 <.0001
Visit 5 – visit 413.713.813.4 <.0001
Visit 6 – visit 513.813.913.4 <.0001
Age, y63.463.763.3 <.0001
Female sex, %56.156.355.9.149
BMI31.031.330.7 <.0001
Race, %
African American4.04.73.0 <.0001
American Indian0.20.20.2
Asian0.50.50.4
Caucasian86.385.886.9
Hawaiian0.00.00.0
Hispanic0.80.70.8
Multiracial0.00.00.1
Other6.36.16.5
Refused2.01.92.1
Conditions, %
Dyslipidemia53.956.051.1 <.0001
Diabetes26.927.526.1 <.0001
CAD without MI20.320.320.4.886
Stroke or TIA10.19.910.4 .001
Atrial fibrillation7.17.17.1.867
CHF6.56.76.4 .015
PVD5.14.85.4 <.0001
MI3.43.43.4.476
Renal disease2.82.73.0 .002

Table 2 shows the breakdown of patient medications, identifying the number of prescribed medications, medication classes, and common combination therapies prescribed. In the study population, 37.0% of patients were prescribed 1 antihypertensive medication, 21.2% were prescribed 2 antihypertensive medications, and <10% were prescribed ≥3 antihypertensive medications. For patients taking >1 medication, the most common combinations were diuretic+angiotensin-converting enzyme inhibitors (3.5%) and diuretic+β-blocker (3.0%).

Table 2. Patient Medication Descriptives
 No. (%)
  1. Abbreviations: ACE, angiotensin-converting enzyme; ARBs, angiotensin receptor blockers.

Prescribed Antihypertensive Medications, No.
0297,767 (32.27)
1341,530 (37.01)
2195,791 (21.22)
370,743 (7.67)
415,287 (1.66)
51,698 (0.18)
66 (0.001)
Medication class
ACE inhibitors229,542 (24.87)
Diuretics218,382 (23.66)
β-Blockers212,250 (23.00)
Calcium channel blockers151,834 (16.45)
ARBs142,444 (15.43)
α-/β-blockers60,839 (6.59)
Common combinations
Diuretic+ACE inhibitor32,366 (3.51)
Diuretic+β-blocker27,685 (3.00)
ACE inhibitor+β-blocker25,566 (2.77)
ARBs+β-blocker16,000 (1.73)
ACE inhibitor+calcium channel blocker15,666 (1.70)
Diuretic+calcium channel blocker15,216 (1.65)
Diuretic+ACE inhibitor+β-blocker14,122 (1.53)

A propensity model was used to reduce selection bias and improve population comparisons in this observational design. The propensity model evaluated patient characteristics to estimate the probability of taking an antihypertensive medication. The actual rate of patients taking any antihypertensive medication was 68% and the propensity model estimated the probability of taking an antihypertensive medication to fall between 50% and 80% for nearly all patients (99.69%). The appropriateness of drawing inferences by comparing the medicated patients relative to the unmedicated patients was well supported given similar probability of taking an antihypertensive medication.

The estimates of the independent association of various antihypertensive medication classes on VVV after the initial treatment period are captured in Table 3 and visually demonstrated in Figure 3. Positive β coefficients indicate higher VVV. The mean effect varied among the various antihypertensive medication classes. As can be seen in Table 3 and Figure 3, both diuretics and calcium channel blockers significantly lowered VVV relative to all other classes (β=−0.32 [−0.39, −0.25], P<.0001, β=−0.11 [−0.19, −0.03], P=.005, respectively) whereas α-/β-blockers (β=0.89 [0.77, 1], P<.0001) showed the highest increase in VVV. Angiotensin-converting enzyme inhibitors, β-blockers, and angiotensin receptor blockers showed a statistically significant VVV increase relative to all other classes.

Table 3. Systolic BP Absolute Visit-to-Visit Variability Model
 β Estimate95% CI t P Value
  1. Abbreviations: ACE, angiotensin-converting enzyme; ARBs, angiotensin receptor blockers; β estimate, mean effect of independent variable on systolic absolute real visit-to-visit variability; CAD, coronary artery disease; CHF, congestive heart failure; 95% CI, represents the lower and upper confidence intervals; MI, myocardial infarction; propensity score, the odds that a patient is taking any hypertensive medication in the time prior to that visit; PVD, peripheral vascular disease; TIA, transient ischemic attack; visit count, control variable that represents which reading in the sequence of readings. Bold P values indicate significance.

Medication class
ACE inhibitors0.34(0.27, 0.41)10.04 <.0001
β-Blockers0.24(0.17, 0.31)6.93 <.0001
ARBs0.34(0.26, 0.42)8.46 <.0001
Calcium channel blockers−0.11(−0.19, −0.03)−2.8 .005
α-/β-Blockers0.89(0.77, 1)14.99 <.0001
Diuretics−0.32(−0.39, −0.25)−9.33 <.0001
Patient characteristics
Systolic BP0.12(0.12, 0.12)156.95 <.0001
Sex (female vs male)0.75(0.68, 0.81)24.01 <.0001
Age (centered at 62)0.04(0.04, 0.04)29.05 <.0001
Atrial fibrillation0.11(0, 0.22)2.03 .042
CAD without MI0.54(0.46, 0.62)13.62 <.0001
Dyslipidemia−0.39(−0.46, −0.31)−10.37 <.0001
MI0.29(0.13, 0.45)3.55 .0004
Renal disease0.72(0.57, 0.88)9.12 <.0001
Stroke or TIA0.64(0.55, 0.73)13.46 <.0001
PVD0.69(0.56, 0.82)10.72 <.0001
CHF0.64(0.52, 0.76)10.52 <.0001
Diabetes−0.09(−0.16, − 0.02)−2.57 .010
Control variables
Visit count – 1−0.07(−0.26, 0.12)−0.73.463
Visit count – 2−0.08(−0.24, 0.07)−1.06.291
Visit count – 3−0.13(−0.26, −0.01)−2.06 .040
Visit count – 4−0.13(−0.23, −0.02)−2.32 .020
Visit count – 5−0.12(−0.20, −0.03)−2.66 .008
Visit count – 60
Propensity score−0.67(−1.63, 0.30)−1.35<.177
Intercept−1.98(−2.72, −1.25)−5.27 <.0001
image

Figure 3. Medication class effect on systolic visit-to-visit variability (VVV). Error bars represent 95% confidence intervals. BP indicates blood pressure; ACE, angiotensin-converting enzyme inhibitors; ARBs, angiotensin receptor blockers..

Download figure to PowerPoint

Table 3 also shows the demographic and comorbid influences on VVV. Sex (β=0.75 [0.68–0.81], P<.0001), age (β=0.04 [0.04–0.04], P<.0001), and SBP (β=0.12 [0.12–0.12], P<.0001) showed statistical significance. All comorbid conditions significantly increased VVV, with the exception of dyslipidemia (β=−0.39 [−0.46 to −0.31], P<.0001) and diabetes (β=−0.09 [−.16 to 0.02], P=.010). Having peripheral vascular disease (β=0.69 [0.56–0.82], P<.0001), renal disease (β=0.72 [0.57–0.88], P<.0001), stroke/transient ischemic attack (β=0.64 [0.55–0.73], P<.0001), or chronic heart failure (β=0.64 [0.52–0.76], P<.0001) resulted in the highest levels of VVV. Atrial fibrillation, coronary artery disease without myocardial infarction, and myocardial infarction also resulted in significantly higher levels of VVV.

Table 3 also includes the β estimates of the control variables (visit count and propensity of taking an antihypertensive medication). Visit sequence was analyzed; visit count 3, 4, and 5 had a significant effect on BP VVV (P<.052) with similar effect sizes to the medication and patient characteristic variables. The propensity score effect was not statistically significant (β=−0.67 [−1.63 to 0.30], P=.177), indicating that taking specific medications, rather than the characteristics of taking hypertensive medication (eg, demographics, comorbid conditions), accounted for the differences in BP VVV. Furthermore, including the propensity score improved the ability to properly estimate the effect of medication class after accounting for the dependency between patient characteristics and medication use. In the initial model we considered additional variables, including time between visits, time taking drug class, percent of time taking drug class (time taking drug class/time between visits), month of visit (seasonal trend indicator), and number of medications. However, these factors did not significantly contribute to VVV.

Discussion

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

Although the differences were small, we demonstrated diuretics and calcium channel blockers to be associated with lower VVV and α-/β-blockers to be associated with higher VVV, consistent with previous research.[24] Additionally, antihypertensive medication class accounted for a small proportion of the variance. Given the strong statistical significance, it is tempting to assign clinical significance. However, in situations where the number of observations is extremely large, attention to statistical significance is necessary, but not sufficient. For example, being prescribed an α-/β-blocker raised VVV nearly 1 mm Hg on average, which represents 3.7% (.89/13.8 mm Hg) of the mean VVV. Similarly, the largest difference between any 2 classes (diuretics and α-/β-blockers) was 1.21 mm Hg, which equated to 8.8% on the average VVV. Furthermore, we did not assess clinical outcomes as they pertain to medication class. There are an array of issues that the clinician must consider when prescribing antihypertensive medication, including drug interactions, side effects, and cost. This analysis confirms that in a large clinical database, higher levels of VVV are associated with increased mortality, with the lowest level of variability associated with 2.25% mortality and the highest level of variability associated with 3.77% mortality. However, more work is needed prior to establishing VVV as another criterion for selecting optimal medical treatment regimens.

Several patient characteristics influenced VVV. Advanced age, female sex, and higher SBP were all associated with greater VVV. In terms of comorbid conditions, all evaluated conditions significantly affected VVV with the exception of diabetes and atrial fibrillation. These findings are consistent with previous research,[7] except that in our study, diabetes was not associated with higher VVV. It has been suggested that increased VVV is caused by increased arterial stiffness,[25] but the lack of effect on VVV when diabetes is present appears inconsistent with this notion. Similar to the impact of medication class on VVV, the magnitude of effect in absolute terms for both patient demographics and comorbid conditions raises questions as to the clinical significance of the findings.

Limitations

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

Although the current study provides insight into VVV in a database of real-world observations, there are limitations inherent to data of this nature. Selection bias exists given that antihypertensive medication prescriptions reflect a decision based on patient characteristics, provider preference, clinical concerns, and financial considerations. The impact of selection bias was mitigated by the use of propensity scoring. The EHR system has limited coverage of pharmacy prescription fill data causing a lack of patient medication compliance. Another limitation is the lack of a standardized method to measure BP due to differing standard operating procedures within each practice. This was partially eliminated by the fact that individual patients would likely have been assessed using similar procedures. Inaccurate or incomplete diagnosis coding practices and problem list completion are potential limitations to the accurate identification of hypertension and/or comorbid conditions. Sensitivity analysis was conducted to determine what portion of patients would be coded as having a comorbid condition under the least restrictive relative to the most restrictive set of ICD-9 codes and found that the patient population identified did not significantly vary between approaches. Lifestyle factors were not included in the analysis due to nonuniform collection in the EHR; however, future studies could address this limitation. Using a large clinical EHR-derived database posed potential limitations; however, the results of this study can be considered hypothesis-generating for future investigations of VVV.

Previous retrospective research[6, 12, 13, 16] has found that as VVV increases, the occurrence for vascular events increase;[7, 26, 27] however, it is unclear the extent to which reducing VVV via switching medication classes reduces the risk for vascular events. Future studies could prospectively evaluate the magnitude of effect of switching medication class. Furthermore, the interplay between VVV and SBP, medication class, nonhypertensive medications, and comorbidities on the risk of vascular events and other chronic-disease related events requires further investigation. Additionally, no clear parameters have been established that articulate what level of variability is considered too high and what amount of reduction in VVV would be of clinical benefit. Once a clinical benefit is demonstrated, translational studies, which determine the maximally efficient methodology for incorporating VVV in the EHR for clinical decision support, are needed, because multiple measures are required to establish a reliable VVV determination.[14]

Conclusions

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

This study of nearly 200,000 clinically observed hypertension patients, extracted from an EHR-derived database, revealed that all medication classes were associated with a statistically significant effect on VVV. However, the magnitude of effect in the current study raises questions as to the clinical significance of specific medication class differences in VVV. The current study is the largest published evaluation of VVV and allows for a clearer picture of BP VVV in the real world.

Conflict of Interest

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References

Novartis provided unrestricted grant support for programmatic costs of this study's conduction. There was no sponsor input on study design or conduct. Robert A. Nicholson, PhD, was receiving a grant from the National Institutes of Health (NS04288-08) unrelated to this project. No other conflicts were declared for the remaining authors.

References

  1. Top of page
  2. Abstract
  3. Methods
  4. Results
  5. Discussion
  6. Limitations
  7. Conclusions
  8. Conflict of Interest
  9. References
  • 1
    Lawes CM, Vander Hoorn S, Rodgers A. Global burden of blood-pressure-related disease, 2001. Lancet. 2008;371:15131518.
  • 2
    Lewington S, Clarke R, Qizilbash N, et al. Age-specific relevance of usual blood pressure to vascular mortality: a meta-analysis of individual data for one million adults in 61 prospective studies. Lancet. 2002;360:19031913.
  • 3
    Willmot M, Leonardi-Bee J, Bath PMW. High blood pressure in acute stroke and subsequent outcome: a systematic review. Hypertension. 2004;43:1824.
  • 4
    Rothwell PM. Limitations of the usual blood-pressure hypothesis and importance of variability, instability, and episodic hypertension. Lancet. 2010;375:938948.
  • 5
    Howard SC, Rothwell PM. Reproducibility of measures of visit-to-visit variability in blood pressure after transient ischaemic attack or minor stroke. Cerebrovasc Dis. 2009;28:331340.
  • 6
    Rothwell PM, Howard SC, Dolan E, et al. Prognostic significance of visit-to-visit variability, maximum systolic blood pressure, and episodic hypertension. Lancet. 2010;375:895905.
  • 7
    Muntner P, Shimbo D, Tonelli M, et al. The relationship between visit-to-visit variability in systolic blood pressure and all-cause mortality in the general population: findings from NHANES III, 1988 to 1994. Hypertension. 2011;57:160166.
  • 8
    Klungel OH, de Boer A, Paes AH, et al. Estimating the prevalence of hypertension corrected for the effect of within-person variability in blood pressure. J Clin Epidemiol. 2000;53:11581163.
  • 9
    Turner MJ, van Schalkwyk JM. Blood pressure variability causes spurious identification of hypertension in clinical studies: a computer simulation study. Am J Hypertens. 2008;21:8591.
  • 10
    Marshall T. When measurements are misleading: modelling the effects of blood pressure misclassification in the English population. BMJ. 2004;328:933933.
  • 11
    MacMahon S, Peto R, Cutler J, et al. Blood pressure, stroke, and coronary heart disease. Part 1, Prolonged differences in blood pressure: prospective observational studies corrected for the regression dilution bias. Lancet. 1990;335:765774.
  • 12
    Howard SC, Rothwell PM. Reproducibility of measures of visit-to-visit variability in blood pressure after transient ischaemic attack or minor stroke. Cerebrovasc Dis. 2009;28:331340.
  • 13
    Rothwell PM, Howard SC, Dolan E, et al. Effects of beta blockers and calcium-channel blockers on within-individual variability in blood pressure and risk of stroke. Lancet Neurol. 2010;9:469480.
  • 14
    Muntner P, Joyce C, Levitan EB, et al. Reproducibility of visit-to-visit variability of blood pressure measured as part of routine clinical care. J Hypertens. 2011;29:23322338.
  • 15
    Webb AJ, Fischer U, Mehta Z, Rothwell PM. Effects of antihypertensive-drug class on interindividual variation in blood pressure and risk of stroke: a systematic review and meta-analysis. Lancet. 2010;375:906915.
  • 16
    Dahlof B, Sever PS, Poulter NR, et al. Prevention of cardiovascular events with an antihypertensive regimen of amlodipine adding perindopril as required versus atenolol adding bendroflumethiazide as required, in the Anglo-Scandinavian Cardiac Outcomes Trial-Blood Pressure Lowering Arm (ASCOT-BPLA): a multicentre randomised controlled trial. Lancet. 2005;366:895906.
  • 17
  • 18
    Unified Medical Lanaguage System (UMLS). National Library of Medicine. http://www.nlm.nih.gov/research/umls/index.html. Accessed December 23, 2011.
  • 19
    Department of Commerce (US) Social Security Administration's Death Master File. https://www.ssdmf.com/FolderID/1/SessionID/%7BE31974AE-C490-43AA-8765-83DB3FDEA1FD%7D/PageVars/Library/InfoManage/Guide.htm. Accessed 12 December 2012.
  • 20
    Singer JD, Willett JB. Applied Longitudinal Data Analysis: Modeling Change and Event Occurrence. New York: Oxford University Press; 2003.
  • 21
    SAS 9.2 [Computer Program]. SAS Institute. Cary, NC: SAS; 2008.
  • 22
    Maas CJM, Hox JJ. Sufficient sample sizes for multilevel modeling. Methodology. 2005;1:8692.
  • 23
    Rosenbaum P, Rubin D. The central role of the propensity score in observational studies for causal effects. Biometrika. 1983;70:4155.
  • 24
    Webb AJS, Fischer U, Mehta Z, Rothwell PM. Effects of antihypertensive-drug class on interindividual variation in blood pressure and risk of stroke: a systematic review and meta-analysis. Lancet. 2010;375:906915.
  • 25
    van den Bogaard B, Westerhof BE, Van den Born BJ. Prognostic significance of blood-pressure variability. Lancet. 2010;376:414415.
  • 26
    Rothwell PM, Howard SC, Dolan E, et al. Prognostic significance of visit-to-visit variability, maximum systolic blood pressure, and episodic hypertension. Lancet. 2010;375:895905.
  • 27
    Pringle E, Phillips C, Thijs L, et al. Systolic blood pressure variability as a risk factor for stroke and cardiovascular mortality in the elderly hypertensive population. J Hypertens. 2003;21:22512257.