Author + information
- Received May 12, 2016
- Revision received June 25, 2016
- Accepted July 20, 2016
- Published online November 1, 2016.
- Stuart J. Pocock, PhDa,∗ (, )
- George Bakris, MDb,
- Deepak L. Bhatt, MD, MPHc,
- Sandeep Brar, MDd,
- Martin Fahy, MSd and
- Bernard J. Gersh, MB, ChB, DPhile
- aDepartment of Medical Statistics, London School of Hygiene & Tropical Medicine, London, United Kingdom
- bComprehensive Hypertension Center and Department of Medicine, University of Chicago Medicine, Chicago, Illinois
- cBrigham and Women’s Hospital Heart and Vascular Center and Harvard Medical School, Boston, Massachusetts
- dClinical Department, Medtronic, Santa Rosa, California
- eDepartment of Cardiovascular Diseases, Mayo Clinic College of Medicine, Rochester, Minnesota
- ↵∗Reprint requests and correspondence:
Prof. Stuart J. Pocock, Department of Medical Statistics, London School of Hygiene and Tropical Medicine, Keppel Street, London WC1E 7HT, United Kingdom.
Regression to the mean (RTM) describes the tendency for an extreme measurement on 1 occasion to become less extreme when measured again. RTM may affect clinical trial data interpretation when the outcome measure has high variability. We investigated RTM in the SYMPLICITY HTN-3 (Renal Denervation in Patients With Uncontrolled Hypertension) trial of renal denervation versus a sham procedure. Analysis of covariance was performed on the 6-month change in systolic blood pressure, estimating a mean treatment difference of −4.11 mm Hg (95% confidence interval: −8.44 to 0.22 mm Hg; p = 0.064), which was similar to the unadjusted difference but with a smaller confidence interval. RTM occurred in both arms, but it had a negligible effect on the observed treatment difference. A second example concerns changes in hemoglobin A1c in a nonrandomized study. These findings emphasize the importance of incorporating RTM and analysis of covariance into the design and reporting of clinical studies of how treatments affect time changes in quantitative outcomes. (Renal Denervation in Patients With Uncontrolled Hypertension [SYMPLICITY HTN-3]; NCT01418261)
- blood pressure
- hemoglobin A1c
- randomized controlled trials as topic
- renal denervation
- statistical regression
The statistical concept of regression to the mean (RTM), first introduced by Francis Galton in 1886 (1), is the tendency for a quantitative variable that is extreme on 1 occasion to become less extreme when measured again. This statistical phenomenon happens in clinical studies when repeated measurements are made on the same subject with random variation (intrapatient variability) around a true mean (2–4). If a variable is extreme on its first measurement, it will tend to be closer to the average on its second measurement (2), and therefore can create a false impression that an ineffective intervention has an effect. RTM is of great concern when the variable analyzed has substantial within-subject variability over time. This is especially true when there is a high threshold for trial enrollment, making some subjects’ entry measurements above their true means (5).
Blood pressure (BP) is an example of a measurement with large intraindividual variations (6), creating the potential for RTM to be easily mistaken for a treatment effect on BP change. BP measurements, even if taken only minutes apart, will vary due to both biological variability and measurement error (7), and will also vary by time of day, between visits, and seasonally (8). Thus, subjects with a high initial systolic BP measurement will tend to have lower systolic BP at follow-up, regardless of whether the treatment is effective or not. A high BP threshold to enter the trial can lead some subjects to have a baseline BP well above their underlying mean BP (Central Illustration). The greater the variability in a participant's BP readings, the greater the patient’s risk for future cardiovascular events (9). That is, those patients with high BP variability, who are therefore most susceptible to RTM, are those most in need of treatment options. Examples of RTM on BP measurements have been previously discussed (10–12). However, when a trial randomizes subjects between a treatment and a control arm, we expect only chance differences between treatment arms (13); RTM is expected to occur in both treatment arms, and should not bias the primary analysis of trial findings.
The SYMPLICITY HTN-3 (Renal Denervation in Patients With Uncontrolled Hypertension) trial in patients with treatment-resistant hypertension was the first large randomized controlled trial of catheter-based renal denervation (RDN) to have a sham control group in which both the patient and BP assessor were blinded to treatment arm. The trial showed no significant treatment difference in the 6-month change in systolic BP in patients treated with RDN versus sham control (−14.13 ± 23.93 mm Hg in the RDN arm compared with −11.74 ± 25.94 mm Hg in the sham control; p = 0.26 for superiority with a margin of 5 mm Hg) (14). These surprising results contrasted sharply with a previously reported randomized controlled trial (15) that had no such blinding and showed dramatic pressure improvements only in the therapy arm. Previously examined potential explanations for the results of the SYMPLICITY HTN-3 trial include drug changes and adherence, study population, technical and procedural methods (16), as well as the potential for the effect of the placebo and Hawthorne effects on outcomes (17). RTM is another potentially important concept that is examined in the present analysis.
We sought to assess the effect of RTM on the treatment difference observed in the SYMPLICITY HTN-3 trial using analysis of covariance (ANCOVA). ANCOVA increases the statistical power by recognizing that each subject’s change in a measurement over time (in this case, 6-month change in office systolic BP) depends on the subject’s baseline measurement, thereby increasing the precision of the estimated treatment effect (18–20). We also present a second example from an observational study of 6-month change in hemoglobin A1c (HbA1c) in patients with type 2 diabetes mellitus comparing 2 second-line antidiabetic combinations.
SYMPLICITY HTN-3 study design
The design and primary outcomes from the SYMPLICITY HTN-3 trial have been previously published (14,21–23). In brief, patients 18 to 80 years of age who had baseline office systolic BP ≥160 mm Hg and 24-h ambulatory systolic BP ≥135 mm Hg despite ≥3 antihypertensive medications, including a diuretic agent, and who met renal artery anatomy requirements were enrolled. Office BP was measured at both an initial screening visit (screening visit 1) and a subsequent confirmatory screening visit (screening visit 2), which occurred 2 weeks later. However, only the office systolic BP at screening visit 2 served as the baseline BP measurement for comparison. Final eligibility on the basis of renal anatomy was confirmed at the time of randomization, performed in a 2:1 ratio to RDN or a sham procedure. Both the patient and the BP assessor were blinded to treatment arm. Patients were to remain on a stable antihypertensive regimen between the screening visit and the baseline BP measurement, as well as during the 6-month follow-up period to the primary endpoint, unless changes were considered to be clinically necessary. Office BP was determined from the average of 3 measurements, performed while the patient was seated using an automated device. At 6 months, patients were unblinded to the assigned treatment arm.
Analysis of covariance
ANCOVA was used to calculate the treatment difference while adjusting for baseline SBP. In addition, DOUBLE ANCOVA was used to calculate the treatment difference after adjusting for both screening visit 1 and screening visit 2 (the baseline systolic BP measurement). All analyses were performed on the basis of the intention-to-treat principle. Missing data were not imputed. The trial primary endpoint analysis (14) applied a modified intention-to-treat principle: pre-change BP measurements were carried forward to 6 months in patients with qualifying medication changes. The present analyses did not apply this carryover principle, and hence, patients who did not have a systolic BP measurement at 6 months were excluded. Continuous variables are presented as mean ± SD or as mean and 95% confidence interval (CI). Analyses were performed using SAS version 9.3 (SAS Institute, Cary, North Carolina).
HbA1c changes in patients with type 2 diabetes mellitus
Our second example is from an observational study of patients with type 2 diabetes mellitus. From electronic health records in 4 countries (United Kingdom, Germany, Canada, and France), patients were identified at the time of starting their second-line antidiabetic drug treatment regimen. For each patient, HbA1c values were identified at baseline (permitted time window within 3 months prior to or 2 weeks after the start date of new treatment), and at 6 months (permitted time window within 3 to 9 months after new treatment start date). Patients were classified into 9 possible second-line treatment types. Here, we concentrate on the 2 most common choices, metformin plus sulfonylureas (M + S) and metformin plus dipeptidyl peptidase-4 inhibitors (M + D). To illustrate the option of presenting changes over time on a relative scale (% change) as well as on an absolute scale, this example uses ANCOVA on both log-transformed (% change) and original (untransformed) values of HbA1c.
Crude analyses of BP changes
The 6-month changes in office systolic BP in the RDN (n = 350) and sham control (n = 169) arms were −15.3 ± 23.9 mm Hg and −11.2 ± 26.4 mm Hg, respectively, with large variations in BP, as noted in Figure 1. The treatment difference was −4.07 mm Hg (95% CI: −8.62 to 0.48 mm Hg; p = 0.080) (Table 1). A simple alternative to this approach is to compare the mean systolic BP at 6 months (RDN vs. sham control), ignoring the baseline systolic BP (also in Table 1). This is statistically inefficient because the positive association between individual 6-month and baseline values is not utilized; therefore, the end result has less precision (i.e., a wider CI) for the estimated treatment difference.
Application of ANCOVA
Figure 2 shows individual patient-level measurements of the 6-month change in office systolic BP plotted against baseline office systolic BP in both the RDN and sham control arms. ANCOVA was used to assess the effect of baseline office systolic BP on the treatment difference (Figure 2A). The slope is −0.46 (p < 0.001), meaning that the estimated reduction in 6-month systolic BP change from baseline (in both treatment groups) increases by 4.6 mm Hg for every 10-mm Hg increase in baseline systolic BP. The 6-month treatment difference, using ANCOVA to take into account baseline office systolic BP, was −4.11 mm Hg (95% CI: −8.44 to 0.22; p = 0.064) (Table 1), similar to the unadjusted difference, but with an anticipated slight increase in precision (i.e., the CI is smaller and the p value is lower).
The corresponding analyses for diastolic BP (Table 2) reveal the following interesting point. There was a slight baseline imbalance in diastolic BP, with the mean baseline 2.68 mm Hg lower in the RDN group. Because baseline and 6-month BP are positively correlated, the crude mean treatment difference in 6-month diastolic BP of −4.59 mm Hg is biased toward an exaggerated treatment effect. However, the crude mean treatment difference in the 6-month change in diastolic BP of −1.91 mm Hg has a bias in the opposite direction because there is an inverse correlation between BP change and baseline BP. ANCOVA takes account of this correlation in producing an unbiased baseline-adjusted mean treatment difference in the 6-month change in diastolic BP: −2.53 mm Hg (95% CI: −4.72 to −0.34 mm Hg; p = 0.024).
The conventional ANCOVA model forces the slopes of the RDN and the control arms' curves to be parallel. Figure 2B shows the separate regression lines for RDN and sham on the 6-month change in office systolic BP on the basis of baseline systolic BP. The 2 slopes appear to differ (−0.55 for RDN and −0.29 for the control sham: difference −0.26 [95% CI: −0.53 to 0.003; p = 0.053]), suggesting that the effect of baseline systolic BP on the change in systolic BP in RDN patients may be greater than in the sham patients. To explore this further, Figure 3 shows the 6-month mean change in office systolic BP for both the control and RDN arms for patients grouped into tertiles of baseline systolic BP. Even in the highest tertile, the treatment difference was not significant; however, the observed treatment difference was numerically larger (6.0 mm Hg) compared with the lowest tertile (2.1 mm Hg). A post hoc interaction test for trend has p = 0.47.
Regression to the mean in the screening period
The effect of RTM was also evident in the trial screening process. Figure 4 shows both patients who were randomized (in blue) and those who were not (in gray). The change in office systolic BP from screening visits 1 and 2 (2 weeks apart) is plotted on the basis of office systolic BP at screening visit 1. Note that all patients at screening visit 1 had office systolic BP ≥160 mm Hg, and those randomized in the study were also required to have systolic BP ≥160 mm Hg at screening visit 2; otherwise, they were ineligible for randomization. Some patients were not randomized for other reasons. Note that the regression slope for all patients in Figure 4 (−0.35) is less than the slope in Figure 2 (−0.46). This difference indicates that RTM is present between screening visits (2 weeks apart), but not surprisingly, is less pronounced than between on-trial visits (6 months apart).
Value of multiple baseline measurements
Thus, patients required 2 qualifying systolic BP measurements ≥160 mm Hg to enter into the trial. It is useful to simultaneously model how both of these measurements influence the change in systolic BP at 6 months, as shown in the DOUBLE ANCOVA model (Table 1). Both screening visit 1 and 2 systolic BP measurements were highly predictive of the 6-month office systolic BP (p = 0.001 and p < 0.001, respectively). The latter visit was more predictive because it was nearer in time to the 6-month systolic BP. The 6-month treatment difference using a DOUBLE ANCOVA to adjust for both screening visits 1 and 2 was −3.89 mm Hg (95% CI: −8.18 to 0.40 mm Hg; p = 0.077). Note that this model produces the most precise estimate (smallest SE) of treatment difference across all of the models fitted in Table 1.
Regression to the mean in a nonrandomized study
We now turn our attention to treatment comparisons for the change in a quantitative measurement in an observational study. Given the lack of randomization, the baseline distributions of the measurements may well differ between treatments. Ignoring this fact by using a crude comparison of mean changes will lead to bias, so that proper allowance for RTM using ANCOVA is essential to reach sensible conclusions (24). Of course, there may still be other biases due to the nonrandomized allocation of patients to treatment.
From electronic health records in 4 countries, 12,168 patients were identified at the initiation of their second-line treatment for type 2 diabetes mellitus as having recorded their HbA1c values at baseline and 6 months later. Here, we compare the 2 most common choices of second-line treatments: metformin plus sulfonylureas (M + S; n = 4,834) and metformin plus dipeptidyl peptidase-4 inhibitors (M + D; n = 3,403). A crude comparison of mean HbA1c changes yielded mean falls of 1.701% (SE: 0.026) and 1.037% (SE: 0.031), respectively (p < 0.0001). Thus, at face value, we appear to have strong evidence that the former is more effective in lowering HbA1c.
However, the baseline HbA1c had a higher mean and a greater variation in the M + S group: mean 9.269 ± 1.949% compared with the baseline mean 8.395 ± 1.543% in the M + D group. Also, as in our earlier BP example, patients with higher baseline HbA1c values tended to have greater falls in HbA1c at 6 months. The use of ANCOVA should take both of these issues into account, but first we consider the need to transform the data to get a valid regression model.
Use of a log transformation
For any use of ANCOVA, one should consider whether applying a log transformation gives a more valid model fit to the data. This is likely to be the case if, at higher baseline values, there is a greater spread (variance) in the observed changes, and furthermore, if the magnitude of the mean fall at higher values is greater than a linear relation with the baseline value would indicate. This is what happened in this example concerning the 6-month change in HbA1c from the start of second-line antidiabetic treatment. Various diagnostics for lack of model fit were carried out. As a simple summary, we split the data into 2 equal halves, below and above the median baseline HbA1c of 8.4%. For subjects above the baseline median, the slope of the regression of the 6-month change in HbA1c on the baseline HbA1c was more than 50% steeper than for subjects below the baseline median. Also, above the baseline median, the residual SD about the regression line was more than 70% greater than below the baseline median. This was found for both treatment groups.
Accordingly, we perform ANCOVA on a log scale as follows: for M + S and M + D, the baseline-adjusted mean changes in loge(HbA1c) were −0.171 (SE: 0.002) and −0.167 (SE: 0.003), respectively (p = 0.57). Exponentiating log values leads to estimated geometric means and 95% CIs for the percent fall in HbA1c over 6 months in each group, which are baseline-adjusted. These are mean falls of 15.7% (95% CI: 15.4% to 16.0%) and 15.4% (95% CI: 14.9% to 15.9%) for the M + S and M + D groups, respectively. This similarity of effects on HbA1c is compatible with a systematic review of the evidence from randomized trials (25).
Note that the units of HbA1c are also a percentage. Thus, if a patient’s HbA1c is 10.0% at baseline and 7.5% at 6 months, this is an absolute fall of 2.5% and a relative fall of 25%.
Figure 5 shows scatter plots of the relation between the 6-month relative change in HbA1c and baseline HbA1c for patients in each treatment group. Each plot also shows the fitted regression line for these associations, demonstrating the extent to which the expected relative fall in HbA1c increases with the baseline HbA1c. For instance, in the M + S group, for baseline HbA1c of 7%, 10%, and 12%, the expected relative falls in HbA1c are 2.0%, 23.0%, and 32.1%, respectively.
Use of ANCOVA to assess the effect of RTM
BP is an example of a clinical measurement associated with large intraindividual variations, creating the potential for an RTM effect. A BP value extreme in 1 measurement (e.g., at the baseline visit) will likely be less extreme when measured again (e.g., at the follow-up visit). Furthermore, when there is a high baseline threshold to enter a clinical trial, patients at randomization will tend to have an observed BP that is above their true underlying mean BP. Given the marked variability in sequential measurements of BP, large randomized controlled trials are required to measure changes in BP.
The SYMPLICITY HTN-3 trial was the first large trial of RDN that randomized patients to RDN versus a sham control, with treatment blinded to both the BP assessor and the patient. The trial did not find a significant treatment effect, and the role of RTM in resistant hypertension was questioned as a possible contributor to the neutral results. We observed wide between-subject variation in systolic BP change over 6 months. Within that variation, we studied, for both RDN and the control patients, the association between a patient’s 6-month change in office systolic BP and the baseline systolic BP. Our main finding is that the RTM was present in both the RDN and the sham control arms; however, the effect of RTM on the treatment difference (RDN vs. sham control) was small, as evidenced by a similar treatment difference for both the Student t test and the ANCOVA test (Table 1). That is, performance of a randomized controlled trial minimizes the effect of RTM on any treatment effects because this is present in both arms of the trial. Therefore, RTM was likely not responsible for the neutral findings of the SYMPLICITY HTN-3 trial. However, ANCOVA slightly improved the precision of the treatment effect, as evidenced by a narrower CI.
The use of double ANCOVA to adjust for 2 baseline measurements
It is debatable whether repeating the qualifying measurement at a second screening visit helps reduce the effects of RTM on study outcome. The screening process in the SYMPLICITY HTN-3 trial required patients to have an office systolic BP ≥160 mm Hg at consecutive screening visits 2 weeks apart. The second screening visit then served as the actual baseline BP measurement. Enrolling patients with BP consistently above a screening threshold will help to reduce, but will not eliminate, RTM (2). The extent of random variation generally increases with a longer time interval between repeat measurements; thus, as expected, the variability between screening visit 2 and the 6-month measurement was larger than between screening visits 1 and 2. Note that utilization of a DOUBLE ANCOVA, adjusting for both screening systolic BP values, produced the treatment effect estimate with the narrowest CI in SYMPLICITY HTN-3. A general statistical principle for clinical trial design is that any important baseline features that influence the outcomes (6-month change in systolic BP) should be included in the best covariate-adjusted model to optimize the precision of the treatment effect. Future clinical trials should use ANCOVA (and possibly DOUBLE ANCOVA) to increase statistical power.
Suggestions for future trials of RDN
Future trials of RDN versus a sham control would benefit from a primary analysis based on ANCOVA, rather than a crude comparison of mean changes (using a 2-sample Student t test), to take into account the effect of RTM. It is also helpful to fully use the multiple BP measurements at pre-randomization screening visits. Furthermore, ambulatory BP measurements may also reduce the effect of RTM in a clinical study of BP, because within-patient variability is lower than for office BP measures (10), and therefore removes some of the “big day selection bias” (11). Ambulatory BP is also less susceptible to measurement bias and placebo effects, permits discrimination between “white coat” and “true” hypertension, and is more strongly associated with cardiovascular risk (12–14). Many of these recommendations are included in the ongoing SPYRAL HTN Global Clinical Trial Program that is investigating the effect of RDN in uncontrolled hypertensive patients in the absence of and, separately, in the presence of antihypertensive medications (26). Other RDN technologies are also being investigated in prospective, randomized, sham-controlled clinical trials (including REDUCE HTN-REINFORCE [Renal Denervation Using the Vessix Renal Denervation System for the Treatment of Hypertension] [NCT02392351] and RADIANCE-HTN [A Study of the ReCor Medical Paradise System in Clinical Hypertension] [NCT02649426]); however, detailed statistical plans have not yet been published.
RTM in HbA1c for type 2 diabetes mellitus
Our second example studies the effect of RTM on the change in HbA1c over 6 months for patients initiating either M + S or M + D as second-line treatment. This illustrates 2 interesting features. First, in nonrandomized studies of alternative treatments, there may well be baseline differences in the measure of interest (HbA1c). If this is ignored, then a crude comparison of mean changes could lead to misleading results: here, an apparent treatment difference when truly there was none. In such a situation, ANCOVA is essential in correcting for the strong influence of RTM, as well as any other confounding factors (24).
Second, we illustrate the choice one faces of whether to express change on an absolute scale (as in the SYMPLICITY HTN-3 example) or on a relative percent scale, by use of a log transformation. The choice often depends on which ANCOVA model gives the better fit to the data: for HbA1c, this appeared to be the relative scale because for higher values of baseline HbA1c, there was both a greater variability of absolute HbA1c change and also a greater mean absolute change than a linear fit would indicate.
RTM in other disease states
The presently identified statistical issues around RTM also apply to many other diseases and clinical measurements. For instance, Thompson and Pocock (27) reported a 1-year within-person coefficient of variation of hypercholesterolemia measurements that was similar to the between-person coefficient of variation. RTM is also evident in cholesterol screening (28,29), QT prolongation after antiarrhythmic drugs (30), bone densitometry in monitoring osteoporosis therapy (31), and in many other clinical variables with large within-person variability, which, if ignored, may lead to misclassification of subjects.
In addition to RTM and Hawthorne effects, antihypertensive medication changes during the screening process and during patient follow-up may affect the results of clinical studies, including in the presented examples. Our analyses of SYMPLICITY HTN-3 suggest that RDN may have a stronger treatment effect in patients with higher baseline office systolic BP, but these findings are hypothesis-generating, given their post hoc nature and the nonsignificant interaction p value. Our second example, concerning 6-month changes in HbA1c, is from electronic health records, and hence data quality (e.g., timing of HbA1c recordings) is less precise than in a prospectively planned study.
The existence of substantial RTM was similar in both the RDN and the control sham arms of the SYMPLICITY HTN-3 trial. However, RTM did not appear to be responsible for the nonsignificance of the finding for the primary study outcome, 6-month change in systolic BP. Key trial design features, including randomization with a sham control, blinding both patient and BP assessor, and taking multiple BP measurements before randomization, helped to minimize the RTM effect on the primary outcome. Statistical testing using ANCOVA instead of the 2-sample Student t test improves the precision of treatment effect estimation. Furthermore, ANCOVA is essential in correcting for the influence of RTM in nonrandomized studies of treatment options with baseline differences in the measure of interest (in our example, HbA1c). These findings emphasize the importance of incorporating RTM and ANCOVA into the design and reporting of clinical studies that focus on how treatments affect time changes in quantitative outcomes.
The authors thank Nicole Brilakis, MS, MBA, and Douglas A. Hettrick, PhD, for editorial support; Vanessa DeBruin, MS, Denise Jones, RN, BSN, and Melissa Schatz for research support on the SYMPLICITY HTN-3 trial (all of Medtronic); and Thomas Godec, MSc, of the London School of Hygiene & Tropical Medicine, for his statistical analyses of the HbA1c data. The authors are also grateful to AstraZeneca (Medical Evidence and Observational Research team) for permission to use the HbA1c data in our second example.
The SYMPLICITY HTN-3 trial was funded by Medtronic. Prof. Pocock is a consultant/adviser to Medtronic and AstraZeneca. Prof. Bakris is a consultant/advisor to Medtronic, AbbVie, Janssen, Relypsa, Novartis, Boehringer-Ingelheim, NxStage, AstraZeneca, Bayer, and Takeda; and payments were made to University of Chicago Medicine for his work as principal investigator or steering committee member from Bayer, Janssen, and AbbVie. Prof. Bhatt has served on the advisory board of Cardax, Elsevier Practice Update Cardiology, Medscape Cardiology, and Regado Biosciences; has served on the Board of Directors of Boston VA Research Institute and the Society of Cardiovascular Patient Care; has served as Chair of the American Heart Association Quality Oversight Committee; has served on the data monitoring committees of Duke Clinical Research Institute, Harvard Clinical Research Institute, Mayo Clinic, and Population Health Research Institute; has received honoraria from the American College of Cardiology (Senior Associate Editor, Clinical Trials and News, ACC.org), Belvoir Publications (Editor-in-Chief, Harvard Heart Letter), Duke Clinical Research Institute (clinical trial steering committees), Harvard Clinical Research Institute (clinical trial steering committee), HMP Communications (Editor-in-Chief, Journal of Invasive Cardiology), Journal of the American College of Cardiology (guest editor; associate editor), Population Health Research Institute (clinical trial steering committee), Slack Publications (Chief Medical Editor, Cardiology Today’s Intervention), Society of Cardiovascular Patient Care (Secretary/Treasurer), and WebMD (CME steering committees); is a Deputy Editor of Clinical Cardiology; has served as chair of the NCDR-ACTION Registry Steering Committee and VA CART Research and Publications Committee; has received research funding from Amarin, Amgen, AstraZeneca, Bristol-Myers Squibb, Eisai, Ethicon, Forest Laboratories, Ischemix, Medtronic (including for his role as co-PI of SYMPLICITY HTN-3), Pfizer, Roche, Sanofi, and The Medicines Company; has received royalties from Elsevier (Editor, Cardiovascular Intervention: A Companion to Braunwald’s Heart Disease); has served as a site coinvestigator for Biotronik, Boston Scientific, and St. Jude Medical; is a trustee of the American College of Cardiology; and has performed unfunded research for FlowCo, PLx Pharma, and Takeda. Dr. Brar and Mr. Fahy are employees of Medtronic. Prof. Gersh is a consultant for Medtronic, Boston Scientific, Celyad, Johnson & Johnson, and Janssen Pharmaceuticals; has served on the data safety monitoring board of Mount Sinai St. Lukes, Boston Scientific, Teva Pharmaceutical Industries, St. Jude Medical, Jansen Research & Development, Baxter Healthcare, and Cardiovascular Research Foundation; and has served on the advisory board of Medtronic.
- Abbreviations and Acronyms
- analysis of covariance
- blood pressure
- confidence interval
- hemoglobin A1c
- M + D
- metformin plus dipeptidyl peptidase-4 inhibitors
- M + S
- metformin plus sulfonylureas
- renal denervation
- regression to the mean
- Received May 12, 2016.
- Revision received June 25, 2016.
- Accepted July 20, 2016.
- American College of Cardiology Foundation
- Barnett A.G.,
- van der Pols J.C.,
- Dobson A.J.
- Bland J.M.,
- Altman D.G.
- Senn S.
- Warren R.E.,
- Marshall T.,
- Padfield P.L.,
- et al.
- ↵Senn S. Control in clinical trials. In: Reading C, editor. Data and Context in Statistics Education: Towards an Evidence-Based Society. Proceedings of the Eighth International Conference on Teaching Statistics (ICOTS8, July 2010), Ljubljana, Slovenia. Voorburg, the Netherlands: International Statistical Institute, 2010. Available at: http://iase-web.org/documents/papers/icots8/ICOTS8_6D2_SENN.pdf. Accessed August 15, 2016.
- Esler M.D.,
- Böhm M.,
- Sievert H.,
- et al.
- Kandzari D.E.,
- Bhatt D.L.,
- Brar S.,
- et al.
- Pocock S.J.,
- Gersh B.J.
- Twisk J.W.R.
- Vickers A.J.,
- Altman D.G.
- Bakris G.L.,
- Townsend R.R.,
- Liu M.,
- et al.,
- for the SYMPLICITY HTN-3 Investigators
- Bakris G.L.,
- Townsend R.R.,
- Flack J.M.,
- et al.,
- for the SYMPLICITY HTN-3 Investigators
- Pocock S.J.,
- McMurray J.J.,
- Collier T.J.