Comparative Effectiveness of Sulfonylurea and Metformin Monotherapy on Cardiovascular Events in Type 2 Diabetes Mellitus: A Cohort Study

Christianne L. Roumie; Adriana M. Hung; Robert A. Greevy; Carlos G. Grijalva; Xulei Liu; Harvey J. Murff; Tom A. Elasy; Marie R. Griffin

doi:10.7326/0003-4819-157-9-201211060-00003

Original Research |6 November 2012

Comparative Effectiveness of Sulfonylurea and Metformin Monotherapy on Cardiovascular Events in Type 2 Diabetes Mellitus: A Cohort Study Free

Christianne L. Roumie, MD, MPH; Adriana M. Hung, MD, MPH; Robert A. Greevy, PhD; Carlos G. Grijalva, MD, MPH; Xulei Liu, MD, MS; Harvey J. Murff, MD, MPH; Tom A. Elasy, MD, MPH; Marie R. Griffin, MD, MPH

Christianne L. Roumie, MD, MPH

From Veterans Health Administration, Tennessee Valley Healthcare System Geriatric Research Education Clinical Center, Health Services Research and Development Center, and Vanderbilt University, Nashville, Tennessee.

Adriana M. Hung, MD, MPH

Robert A. Greevy, PhD

Carlos G. Grijalva, MD, MPH

Xulei Liu, MD, MS

Harvey J. Murff, MD, MPH

Tom A. Elasy, MD, MPH

Marie R. Griffin, MD, MPH

Article, Author, and Disclosure Information

From Veterans Health Administration, Tennessee Valley Healthcare System Geriatric Research Education Clinical Center, Health Services Research and Development Center, and Vanderbilt University, Nashville, Tennessee.
Disclaimer: The authors of this report are responsible for its content. Statements in the report should not be construed as endorsement by the Agency for Healthcare Research and Quality, the U.S. Department of Health and Human Services, or the Department of Veterans Affairs.
Grant Support: This project was funded under contract 290-05-0042 from the Agency for Healthcare Research and Quality, U.S. Department of Health and Human Services, as part of the Developing Evidence to Inform Decisions about Effectiveness program. Drs. Roumie (04-342-2) and Hung (2-031-09S) were supported by Veterans Affairs Career Development Awards. Dr. Roumie was also supported in part by the Vanderbilt Clinical Translational Scientist Award UL1 RR024975-01 from the National Center for Research Resources/National Institutes of Health. Support for Veterans Affairs/Centers for Medicare & Medicaid Services data provided by the Department of Veterans Affairs, Veterans Affairs Health Services Research and Development Service, Veterans Affairs Information Resource Center (project numbers SDR 02-237 and 98-004).
Potential Conflicts of Interest: Disclosures can be viewed at www.acponline.org/authors/icmje/ConflictOfInterestForms.do?msNum=M12-0009.
Reproducible Research Statement: Study protocol and statistical code: Available from Dr. Roumie (e-mail, christianne.roumie@vanderbilt.edu). Data set: Not available.
Requests for Single Reprints: Christianne L. Roumie, MD, MPH, Nashville Veterans Affairs Medical Center, 1310 24th Avenue South, Geriatric Research Education Clinical Center 4A120, Nashville, TN; e-mail, christianne.roumie@vanderbilt.edu.
Current Author Addresses: Drs. Roumie and Liu: Veterans Affairs Tennessee Valley Healthcare System, 1310 24th Avenue South, Geriatric Research Education Clinical Center, Nashville, TN 37212.
Dr. Hung: Vanderbilt University Medical Center, 1161 21st Avenue South, Garland Division of Nephrology, S-3223 Medical Center North, Nashville, TN 37232.
Dr. Greevy: Vanderbilt University School of Medicine, Department of Biostatistics, 1161 21st Avenue South, S-2323 Medical Center North, Nashville, TN 37232.
Dr. Grijalva: Department of Preventive Medicine, 1500 21st Avenue South, The Village at Vanderbilt, Suite 2650, Nashville, TN 37212.
Dr. Murff: Vanderbilt Institute for Medicine and Public Health, 2525 West End Avenue, Sixth Floor, Nashville, TN 37203-1738.
Dr. Elasy: Vanderbilt University Medical Center, Medical Center East, North Tower, Suite 6000, 1215 21st Avenue South, Nashville, TN 37232-8300.
Dr. Griffin: Department of Preventive Medicine, 1500 21st Avenue South, The Village at Vanderbilt, Suite 2600, Nashville, TN 37212.
Author Contributions: Conception and design: C.L. Roumie, A.M. Hung, R.A. Greevy, C.G. Grijalva, T.A. Elasy, M.R. Griffin.
Analysis and interpretation of the data: C.L. Roumie, A.M. Hung, R.A. Greevy, C.G. Grijalva, X. Liu, H.J. Murff, T.A. Elasy, M.R. Griffin.
Drafting of the article: C.L. Roumie, R.A. Greevy, T.A. Elasy.
Critical revision of the article for important intellectual content: C.L. Roumie, A.M. Hung, R.A. Greevy, C.G. Grijalva, X. Liu, H.J. Murff, T.A. Elasy, M.R. Griffin.
Final approval of the article: C.L. Roumie, A.M. Hung, R.A. Greevy, C.G. Grijalva, H.J. Murff, T.A. Elasy, M.R. Griffin.
Statistical expertise: R.A. Greevy.
Obtaining of funding: C.L. Roumie, R.A. Greevy, C.G. Grijalva, M.R. Griffin.
Administrative, technical, or logistic support: C.L. Roumie, T.A. Elasy.
Collection and assembly of data: C.L. Roumie, R.A. Greevy.

Abstract

Background:

The effects of sulfonylureas and metformin on outcomes of cardiovascular disease (CVD) in type 2 diabetes are not well-characterized.

Objective:

To compare the effects of sulfonylureas and metformin on CVD outcomes (acute myocardial infarction and stroke) or death.

Design:

Retrospective cohort study.

Setting:

National Veterans Health Administration databases linked to Medicare files.

Patients:

Veterans who initiated metformin or sulfonylurea therapy for diabetes. Patients with chronic kidney disease or serious medical illness were excluded.

Measurements:

Composite outcome of hospitalization for acute myocardial infarction or stroke, or death, adjusted for baseline demographic characteristics; medications; cholesterol, hemoglobin A_1c, and serum creatinine levels; blood pressure; body mass index; health care utilization; and comorbid conditions.

Results:

Among 253 690 patients initiating treatment (98 665 with sulfonylurea therapy and 155 025 with metformin therapy), crude rates of the composite outcome were 18.2 per 1000 person-years in sulfonylurea users and 10.4 per 1000 person-years in metformin users (adjusted incidence rate difference, 2.2 [95% CI, 1.4 to 3.0] more CVD events with sulfonylureas per 1000 person-years; adjusted hazard ratio [aHR], 1.21 [CI, 1.13 to 1.30]). Results were consistent for both glyburide (aHR, 1.26 [CI, 1.16 to 1.37]) and glipizide (aHR, 1.15 [CI, 1.06 to 1.26]) in subgroups by CVD history, age, body mass index, and albuminuria; in a propensity score–matched cohort analysis; and in sensitivity analyses.

Limitation:

Most of the veterans in the study population were white men; data on women and minority groups were limited but reflective of the Veterans Health Administration population.

Conclusion:

Use of sulfonylureas compared with metformin for initial treatment of diabetes was associated with an increased hazard of CVD events or death.

Primary Funding Source:

Agency for Healthcare Research and Quality and the U.S. Department of Health and Human Services.

Editors’ Notes

Context

Diabetes increases risk for cardiovascular disease, but how metformin and sulfonylureas affect that risk is less clear.

Contribution

In this analysis of a national population of veterans, new use of sulfonylureas seemed to increase incidence of and risk for cardiovascular events and death compared with metformin.

Caution

The findings apply primarily to white men.

Implication

Sulfonylureas seem to increase cardiovascular events and death compared with metformin. Whether sulfonylureas are harmful, metformin is protective, or both is unclear.

—The Editors

Cardiovascular disease (CVD) accounts for most deaths in patients with diabetes mellitus (1–3). Randomized trials have evaluated CVD risk associated with selected thresholds of glycemic control (4, 5), but how specific antidiabetic drugs contribute to CVD risk is less clear. Some studies found that thiazolidinediones increased CVD risk compared with placebo or active comparators (6–8), but the comparative CVD risk associated with the 2 most commonly used drugs, metformin and sulfonylureas, is not well-characterized.

We sought to compare the hazard of CVD outcomes and all-cause mortality in patients who initiated metformin and sulfonylurea therapy by using data from a national cohort that allow for control of important patient characteristics associated with both diabetes treatment and CVD or death (hemoglobin A_1c [HbA_1c] level, body mass index [BMI], serum creatinine level, and blood pressure).

Methods

Study Design and Data Sources

We defined a cohort of patients initiating oral monotherapy for diabetes between 1 October 2001 and 30 September 2008 using data sets from national Veterans Health Administration (VHA) Decision-Support Services: pharmacy data sets for prescription data dispensed by the VHA or a consolidated mail outpatient pharmacy, including medication name, date filled, days supplied, pill number, and dosage (9); medical data sets for patient demographic characteristics and International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM)–coded diagnostic and procedure information from inpatient and outpatient encounters (10); and laboratory data sets derived from Veterans Health Information Systems and Technology Architecture clinical sources. Data on vital signs included all outpatient measurements of height, weight, and blood pressure. We obtained dates of death from VHA Vital Status File. For Medicare- or Medicaid-eligible veterans, we obtained data on supplemental encounters and race from the Centers for Medicare & Medicaid Services (11).

The institutional review boards of Vanderbilt University and the VHA Tennessee Valley Healthcare System (Nashville, Tennessee) approved this study.

Study Population

The study population comprised veterans aged 18 years or older who received regular VHA care (a VHA encounter or prescription fill at least once every 180 days) for at least the past 365 days. Incident users with known birth date and sex and with more than 365 days of baseline data preceding their first eligible prescription fill were identified. Patients were eligible if they filled a first prescription for an oral antidiabetic drug after at least 365 days without any oral or injectable diabetic drug fill (new users) (12). We excluded patients with serious medical conditions identified at baseline (heart failure, HIV, cancer except for nonmelanoma skin cancer, organ transplantation, end-stage kidney or liver disease, or respiratory failure), cocaine use, or a baseline serum creatinine level of 133 µmol/L (1.5 mg/dL) or greater, because these may influence the prescription of specific antidiabetic drugs and risk for outcomes.

Exposures

Incident exposures were to metformin and sulfonylureas (glyburide and glipizide). We excluded thiazolidinediones and combination metformin–sulfonylurea prescriptions because they are uncommon incident regimens in the VHA. Using pharmacy information, we calculated “days' supply in hand,” accounting for early refills. Follow-up began on the incident prescription date and continued until a switch to or addition of another antidiabetic drug, the 90th day with no drugs in hand, an outcome, or a censoring event—whichever came first. Censoring events comprised reaching a serum creatinine level of 133 µmol/L (1.5 mg/dL) or greater (because metformin use is not recommended in this setting), the 181st day of no contact with any VHA facility (inpatient, outpatient, or pharmacy use) or the end of the study (30 September 2008).

Outcomes: CVD and Death

The primary composite outcome was hospitalization for acute myocardial infarction (AMI) or stroke, or death. We defined “AMI” as an ICD-9-CM primary discharge diagnosis for fatal and nonfatal AMI (ICD-9-CM code 410.x) (positive predictive value, 67% to 97% compared with chart review) (13–15). We defined “stroke” as ischemic stroke (ICD-9-CM code 433.x1, 434 [excluding 434.x0], or 436), intracerebral hemorrhage (ICD-9-CM code 431), and subarachnoid hemorrhage (ICD-9-CM code 430), excluding traumatic brain injury (ICD-9-CM codes 800 to 804 and 850 to 854) (positive predictive value, 97%) (16). We determined mortality using the VHA Vital Status File, which combines information from multiple sources (Medicare, the VHA, the U.S. Social Security Administration, and VHA compensation and pension benefits) to determine date of death (sensitivity, 98.3%; specificity, 99.8%; relative to the National Death Index) (17).

Covariates

Covariates were selected a priori on the basis of clinical significance and included age, sex, race, fiscal year of cohort entry, physiologic variables closest to cohort entry (blood pressure; serum creatinine, HbA_1c, and low-density lipoprotein [LDL] cholesterol levels; and BMI), indicators of health care utilization (number of outpatient visits and active medications, hospitalization during baseline [yes or no]), smoking status, selected medications indicative of CVD, and comorbid conditions (MI, obstructive coronary disease or prescription for a long-acting nitrate, stroke or transient ischemic attack, atrial fibrillation or flutter, mitral or aortic or rheumatic heart disease, asthma or chronic obstructive pulmonary disease, or procedures for carotid or peripheral artery revascularization or bypass or lower-extremity amputation [Appendix Table 1]).

Appendix Table 1. Definitions of Comorbid Conditions and Medications, on the Basis of Codes and Prescriptions in 365 Days Before Exposure

We initially stratified the population by previous CVD history, defined as diagnoses or procedures for MI, coronary artery disease, transient ischemic attack, stroke, or surgical procedures for repair of peripheral or carotid artery disease during baseline. A formal test of interaction between CVD history and treatment was not statistically significant (P = 0.98), so we present overall findings. For patients missing covariates, we conducted multiple imputations using the Markov-chain Monte Carlo method and a noninformative Jeffreys prior (SAS software, version 9.2, SAS Institute, Cary, North Carolina) (18). All covariates, survival time, and a censoring indicator were included in 20 imputation models and used to compute final estimates.

Statistical Analysis

The primary analysis was time to the composite outcome of hospitalization for AMI or stroke, or all-cause death. A secondary analysis included a composite of AMI and stroke events only, with death as a censoring event rather than an outcome. We used Cox proportional hazards regression models to compare time to composite outcomes for sulfonylureas versus metformin, adjusting for the covariates previously stated.

Except for the first 90 to 180 days, when censoring was high, the proportional hazard assumptions were met through examination of log (log survival) plots (Appendix Figure 1). We adjusted for clustering of observations within the VHA facility of care and calculated robust SEs (19). Continuous covariates were modeled with third-degree polynomials to account for nonlinearity (age; BMI; HbA_1c, LDL cholesterol, and serum creatinine levels; blood pressure; and number of medications and visits).

Appendix Figure 1.

Examination of the proportional hazards assumption using log(log survival) plots.

We also performed propensity score–matched analyses. The propensity score modeled the probability of metformin use given all other study covariates and the VHA facility of care ( Appendix and Appendix Table 2, shows additional information and logistic regression model). The visual inspection of the distributions of propensity scores among exposure groups showed good overlap (Appendix Figure 2). Sulfonylurea and metformin observations were matched using a 1-to-1 greedy matching algorithm, yielding 80 648 propensity score–matched observations (20, 21).

Appendix Table 2. Odds of Receiving Metformin Compared With Sulfonylureas

Appendix Figure 2.

Distribution of propensity scores, by drug.

* Probability of using metformin.

Sensitivity and Subgroup Analyses

We performed multiple sensitivity and subgroup analyses. In an approach similar to intention-to-treat analyses in clinical trials, we used the incident prescription to define drug exposure and ignored subsequent changes in regimens (persistent exposure not required). We restricted analyses to patients with complete covariates (multiple imputations not used) (22–24). We conducted stratified analyses by CVD history, age (<65 and ≥65 years), and BMI (<30 and ≥30 kg/m²) in the full cohort and proteinuria in a subset of patients with information on baseline urinary protein–creatinine ratio (36 425 of the 253 690 patients [14.3%]), where “proteinuria” was defined as a urinary protein–creatinine ratio of 30 mg/g or more.

Finally, we quantified the strength of the association of a hypothetical unmeasured binary confounder that would be required to eliminate a statistically significant association (25). We assumed a confounder–outcome association similar to that which we observed among measured covariates (hazard ratio, 1.25) and considered a range of confounder prevalence in sulfonylurea and metformin users; we also considered a stronger confounder–outcome association (hazard ratio, 2.0). Analyses were conducted using R, version X64 2.12.1 (R Foundation for Statistical Computing, Vienna, Austria) and SAS software, version 9.2 (SAS Institute).

Role of the Funding Source

The U. S. Department of Health and Human Services and the Agency for Healthcare Research and Quality's Developing Evidence to Inform Decisions about Effectiveness program sponsored this study. The principal investigators and co-investigators had full access to the data and were responsible for the study protocol, statistical analysis plan, progress of the study, analysis, reporting of the study, and the decision to publish. The Agency for Healthcare Research and Quality reviewed the manuscript and had the opportunity to comment before submission.

Results

Study Cohort and Patient Characteristics

Of 364 865 incident prescriptions for oral antidiabetic drugs, 667 (<0.2%) were excluded for missing date of birth, sex, age younger than 18 years, or data errors; 64 175 (17.6%) were excluded for serious medical illness or cocaine use during baseline; and 14 676 (4.0%) were excluded for a serum creatinine level of 133 µmol/L (1.5 mg/dL) or greater. The remaining 285 347 prescriptions were filled by 269 921 patients, approximately 5% of whom met criteria for cohort entry more than once. Our analysis focused on incident prescriptions for metformin (50%) and sulfonylureas (40% [55% glyburide and 45% glipizide]) and excluded combination metformin–sulfonylurea (8%), rosiglitazone (3%), and pioglitazone (<1%) (Figure 1). Ninety percent of patients had an ICD-9-CM–coded encounter for diabetes, and 73% had no history of CVD at the time of their incident prescription.

Figure 1.

Study flow diagram.

CVD = cardiovascular disease; DM = diabetes mellitus.

There were a median 1768 prescriptions (interquartile range [IQR], 1131 to 2306; range, 410 to 6544) per facility among 128 VHA facilities (median, 1030 [IQR, 696 to 1554] in the propensity score–matched cohort). Median follow-up was 0.78 years (IQR, 0.25 to 1.71 years; range, 1 day to 5.5 years) for patients taking metformin and 0.61 years (IQR, 0.25 to 1.50 years; range, 1 day to 5.5 years) for sulfonylurea users. Reasons for censoring were discontinuing therapy (73% metformin and 66% sulfonylureas), changing therapy (18% metformin and 21% sulfonylureas), leaving the VHA or ending the study (5% metformin and 7% sulfonylureas), and reaching a serum creatinine level of 133 µmol/L (1.5 mg/dL) (2% metformin and 4% sulfonylureas); proportions for each reason within drug groups were similar in the propensity score–matched cohort. Censoring was the highest in the first year; however, characteristics of patients who remained at risk after 1, 2, and 3 years were similar to baseline characteristics (Supplement).

Among the patients, 97% were men and 75% were white (Table 1). Median age was 62 years (IQR, 56 to 71 years) among metformin users versus 67 years (IQR, 57 to 76 years) among sulfonylurea users. The HbA_1c level was 7.0% (IQR, 6.4% to 7.8%) among those who began metformin therapy and 7.3% (IQR, 6.6% to 8.2%) among those who began sulfonylurea therapy; metformin users were slightly heavier (BMI, 31.9 kg/m² vs. 30.2 kg/m²) and used statins more often (61% vs. 55%) than sulfonylurea users.

Table 1. Patient Characteristics in Full and Propensity Score–Matched Cohorts, by New Exposure to Metformin or Sulfonylureas

Characteristics of the 2 groups were more similar after propensity score matching. Standardized differences, a more meaningful measure of between-group differences in large samples, were small before matching and became negligible after matching. Baseline characteristics of the subset with complete covariates were similar, with no important between-group differences (Appendix Table 3).

Appendix Table 3. Baseline Characteristics of Patients With Complete Covariates, by Antidiabetic Drug

Cardiovascular Events and Deaths

Unadjusted rates of the composite outcome were 18.2 per 1000 person-years among 98 665 patients starting sulfonylurea therapy and 10.4 per 1000 person-years among 155 025 patients starting metformin therapy (adjusted hazard ratio [aHR], 1.21 [95% CI, 1.13 to 1.30]) (Table 2). Results were consistent for glyburide (aHR, 1.26 [CI, 1.16 to 1.37]) and glipizide (aHR, 1.15 [CI, 1.06 to 1.26]). Unadjusted rates of CVD events (AMI and stroke) excluding deaths were 13.5 per 1000 person-years for sulfonylurea users and 8.2 per 1000 person-years for metformin users (aHR, 1.16 [CI, 1.06 to 1.25]). Using adjusted rate differences, we estimated 2.2 (CI, 1.4 to 3.0) more CVD events or deaths and 1.2 (CI, 0.5 to 2.1) more CVD events per 1000 person-years of sulfonylurea compared with metformin use.

Table 2. Unadjusted Incidence Rates, Adjusted Incidence Rate Difference, and Adjusted Hazard Ratios for Hazard of the Primary Composite Outcome and Secondary Outcome Among Full and Propensity Score–Matched Cohorts of New Users of Sulfonylureas Compared With Metformin

Table 2. Unadjusted Incidence Rates, Adjusted Incidence Rate Difference, and Adjusted Hazard Ratios for Hazard of the Primary Composite Outcome and Secondary Outcome Among Full and Propensity Score–Matched Cohorts of New Users of Sulfonylureas Compared With Metformin

Results from the propensity score–matched analysis were consistent with those of the full cohort. Among 80 648 patients receiving sulfonylureas and 80 648 patients receiving metformin, the unadjusted rate of the composite outcomes was 15.2 per 1000 person-years for sulfonylurea users and 13.0 for metformin users (aHR, 1.15 [CI, 1.07 to 1.25]) (Table 2 and Figure 2 [top]). Cardiovascular event rates were 11.6 for sulfonylurea users and 10.1 per 1000 person-years for metformin users (aHR, 1.13 [CI, 1.03 to 1.23]). Appendix Table 4 shows unadjusted rates and adjusted incidence rate differences by time in follow-up.

Figure 2.

Cumulative incidence (95% CIs) of cardiovascular disease or death.

Top. Propensity score–matched cohort with persistent exposure to oral hypoglycemic medication required. Bottom. Propensity score–matched cohort with persistent exposure to oral hypoglycemic medication not required, in which patients remain in their exposure group regardless of persistence with drug therapy.

Appendix Table 4. Yearly Unadjusted Incidence Rates and Unadjusted and Adjusted Incidence Rate Differences for the Primary Composite Outcome Among a Propensity Score–Matched Cohort of New Users of Sulfonylureas Compared With Metformin

Appendix Table 4. Yearly Unadjusted Incidence Rates and Unadjusted and Adjusted Incidence Rate Differences for the Primary Composite Outcome Among a Propensity Score–Matched Cohort of New Users of Sulfonylureas Compared With Metformin

Sensitivity and Subgroup Analyses

Results were similar in analyses where patients remained in their original exposure group even if they changed their regimen (persistent exposure not required) (Table 2 and Figure 2 [bottom]). Results stratified by CVD history, age, BMI, and proteinuria (in the subset tested for urinary protein levels) were similar to the main findings (P > 0.60 for each interaction term) (Appendix Figure 3 and Appendix Table 5), as were results restricted to patients with complete covariates (Appendix Table 6).

Appendix Figure 3.

Adjusted hazard ratios for the primary composite outcome (CVD or death) and secondary outcome (CVD alone), stratified by CVD history, age, and BMI.

AMI = acute myocardial infarction; BMI = body mass index; CVD = cardiovascular disease.

* CVD defined by diagnoses or procedure codes for MI, coronary artery disease, transient ischemic attack, stroke, or surgical procedures for repair of peripheral or carotid artery disease in the baseline period.

† Results are also presented for a sample of patients (14.3%) tested for proteinuria and found positive or negative.

Appendix Table 5. Incidence Rates and Adjusted Hazard Ratios for Risk for the Primary Composite Outcome and Secondary Outcome Among the Full Cohort of New Users of Sulfonylureas Compared With Metformin, Stratified by CVD History, Age, and BMI

Appendix Table 5. Incidence Rates and Adjusted Hazard Ratios for Risk for the Primary Composite Outcome and Secondary Outcome Among the Full Cohort of New Users of Sulfonylureas Compared With Metformin, Stratified by CVD History, Age, and BMI

Appendix Table 6. Rates and Adjusted Hazard Ratios for Risk for the Primary Composite Outcome and Secondary Outcome Among Those With Complete Covariates Who Were New Users of Sulfonylureas Compared With Metformin

Appendix Table 6. Rates and Adjusted Hazard Ratios for Risk for the Primary Composite Outcome and Secondary Outcome Among Those With Complete Covariates Who Were New Users of Sulfonylureas Compared With Metformin

Our finding of increased hazard for the composite outcome among sulfonylurea users could have resulted from an unmeasured confounder that increased the hazard for this outcome and had a greater prevalence among sulfonylurea users compared with metformin users. Assuming a degree of association similar to that observed among measured covariates, we calculated that an unmeasured binary confounder would need to be at least 53% more prevalent among sulfonylurea users than metformin users to explain our main findings (Appendix Table 7). A stronger confounder with a hazard ratio for the composite outcome of 2.0 would need to be 14% more prevalent in sulfonylurea users than metformin users (Appendix Table 8).

Appendix Table 7. Risk for CVD in the Presence of an Unmeasured Confounder With a Hazard Ratio of 1.25 for CVD and Various Prevalence Levels of the Confounder, by Exposure Group

Appendix Table 8. Risk for CVD in the Presence of an Unmeasured Confounder With a Hazard Ratio of 2.0 for CVD and Various Prevalence Levels of the Confounder, by Exposure Group

Discussion

This national cohort study of veterans initiating oral treatments for diabetes mellitus found that sulfonylurea use was associated with an increased hazard of AMI, stroke, or death compared with metformin use. The findings do not clarify whether the difference in CVD risk is due to harm from sulfonylureas, benefit from metformin (26), or both. Recent comparative effectiveness reviews and meta-analyses (4, 5, 27) concluded that metformin was associated with a slightly lower risk for all-cause mortality compared with sulfonylureas, but results were inconsistent and imprecise. This study provides further evidence of a risk difference in CVD outcomes for sulfonylurea and metformin users and quantifies the difference.

Questions about the cardiovascular safety of sulfonylureas date back to 1970. The University Group Diabetes Program reported an increased risk for cardiovascular death with tolbutamide compared with placebo and insulin (28–30), leading to a controversial U.S. Food and Drug Administration–mandated black box warning for all sulfonylureas (30–33). Between 1977 and 1991, the UKPDS (United Kingdom Prospective Diabetes Study) randomly assigned patients newly diagnosed with diabetes to intensive sulfonylurea or insulin treatment or diet. In 1998, this study reported similar between-group diabetes-related and all-cause mortality at 10 years, allaying concerns about an increase in sulfonylurea-associated cardiovascular risk. In a UKPDS subpopulation of overweight patients randomly assigned to metformin (n = 342) or diet (n = 411), those receiving metformin experienced relative risk reductions of 42% for diabetes-related deaths and 36% for all-cause deaths compared with the diet-alone group, suggesting an advantage of metformin on mortality (26, 34). In the early 2000s, ADOPT (A Diabetes Outcome Prevention Trial) randomly assigned 4360 patients to metformin, rosiglitazone, or glyburide (35) and reported similarly low numbers of cardiovascular events (fatal or nonfatal AMI and stroke) across treatment groups after a median 4 years of treatment.

Compared with metformin, sulfonylureas are associated with increases in weight and lipid levels and greater risk for hypoglycemia but similar glycemic control (4, 36–38). Thus, metformin is recommended as first-line therapy for patients without contraindications (39–41). Nonetheless, sulfonylureas are sometimes preferred because they require little titration and have fewer gastrointestinal adverse effects than metformin. In 2007, more than 10.1 million Americans (approximately 34% of patients with treated diabetes) used a sulfonylurea as part of their diabetes treatment (42).

Our results are consistent with those of several observational studies in diabetic patients. In a smaller propensity score–matched cohort (n = 8977), McAfee and colleagues (43) showed a 23% decrease in AMI or revascularization with metformin compared with sulfonylurea (aHR, 0.77 [CI, 0.62 to 0.96]). Using the United Kingdom general practice research database (n = 91 000), Tzoulaki and associates (44) found that, compared with metformin, sulfonylureas were associated with an increase in all-cause mortality (aHR, 1.24 [CI, 1.14 to 1.35]) but not first AMI (aHR, 1.09 [CI, 0.94 to 1.27]). A study by Corrao and coworkers (45) found that patients initiating sulfonylurea therapy had a higher risk for hospitalization (aHR, 1.15 [CI, 1.08 to 1.21]) and death (aHR, 1.37 [CI, 1.26 to 1.49]) than did those initiating metformin therapy. Finally, the VHA Diabetes Epidemiology Cohort reported all-cause mortality of 2.7% in 2988 metformin users compared with 5.3% among 19 053 sulfonylurea users (adjusted odds ratio, 0.87 [CI, 0.68 to 1.10]) (46). Of note in our study, we were able to measure and adjust for clinical variables, such as HbA_1c, cholesterol, and serum creatinine levels; blood pressure; and BMI; both McAfee and colleagues' (43) and Corrao and associates' (45) studies relied on administrative data alone.

The reason for the difference in risk between metformin and sulfonylurea users remains unknown. Our previous studies evaluating the association of oral antidiabetic medications and intermediate outcomes in a regional VHA cohort reported results similar to those of a comparative effectiveness review of “high-quality evidence.” In that review, metformin compared with sulfonylureas resulted in decreases of 2.7 kg in weight, 0.259 mmol/L (10 mg/dL) in LDL cholesterol levels, and 0.1 mmol/L (8.6 mg/dL) in triglyceride levels and no difference in HbA_1c levels (4). We estimated that after 1 year, those who began metformin therapy compared with sulfonylurea would have decreases of 3.2 kg in weight, 0.130 mmol/L (5 mg/dL) in LDL cholesterol levels (not statistically significant), and 0.1 mmol/L (8.7 mg/dL) in triglyceride levels and no difference in HbA_1c levels (36, 37). Our previous studies also found that metformin users compared with sulfonylurea users had a decrease of 1.2 mm Hg in systolic blood pressure and less likelihood of a decline in kidney function (47, 48). Whether the minor advantages in cholesterol level, weight, and blood pressure among metformin users could account for the differences in CVD and death or whether another mechanism accounts for the risk difference observed, such as ischemic preconditioning (49), is currently unknown.

Our study has limitations. Confounding by indication could occur if patients with certain characteristics that increase CVD risk were also more likely to use metformin or sulfonylureas. There were some differences in the 2 groups at baseline; however, our large sample size allowed us to directly control for many baseline variables in our primary analysis, and a propensity score–matched analysis yielded similar results. We included only baseline clinical variables and did not account for time-varying covariates. Furthermore, the laboratory results came from individual VHA facilities, not a central laboratory, which could lead to imprecision in measurement.

We accounted for the decrease in sulfonylurea prescribing over time (42, 50) by controlling for year of study entry in all analyses. Although we could not exclude residual confounding, we estimated that an unmeasured confounder or an underreported confounder, such as smoking, with a risk for CVD or death of 1.25 would need to have a very large prevalence imbalance among exposure groups to explain our findings. A much stronger confounder with a risk for CVD equal to 2.0 would need to be less imbalanced (approximately 14% more common among sulfonylurea users) to explain our results.

Refill data were used as a proxy for medication taking and may result in exposure misclassification. Nevertheless, prescription fills seem to be a good proxy for medication use (51). Our definitions required patients to refill their prescribed medications (persistence) because they were censored for gaps in medication use greater than 90 days or for a change in therapy. Censoring because of stopping or changing medications was high, especially in the first year; however, censoring was similar between groups, and the results of analyses that did not require persistent exposure were consistent with the main findings. In addition, analyses of results for each year of follow-up were similar (Appendix Table 4 and Supplement).

If persons were admitted to non-VHA facilities for study outcomes, those events could be missed and outcome misclassification could occur. We supplemented our VHA data with national Medicaid or Medicare data to minimize this concern. Furthermore, use of non-VHA facilities is unlikely to be differential by exposure group. Finally, our patients reflect a typical veteran population, with most patients being white and male.

In conclusion, our study suggests a modest but clinically important 21% increased hazard of hospitalization for AMI or stroke or of death associated with initiation of sulfonylurea compared with metformin therapy. This translates into an excess of approximately 2.2 (CI, 1.4 to 3.0) cardiovascular events or deaths per 1000 person-years of sulfonylurea use. These observations support the use of metformin for first-line diabetes therapy and strengthen the evidence about the cardiovascular advantages of metformin compared with sulfonylureas.

Appendix: Propensity Score

We analyzed 2 cohorts. The first cohort comprised all eligible persons who initiated either metformin or sulfonylurea monotherapy after 365 days with no exposure to medications for diabetes. The second cohort is a subset of the first and used propensity scores to match eligible metformin users to sulfonylurea users. The propensity score is defined as the probability of metformin use, given a particular pattern of baseline covariates. We estimated the propensity score by using a logistic regression model in which the dependent variable was 1 for patients who used metformin at baseline and 0 for sulfonylurea users. The model was simple logistic regression, with a third-degree polynomial term for continuous covariates and facility of care in the model.

Appendix Table 1 and Table 1 list baseline covariates included. Appendix Table 2 shows the model for the probability of being a metformin user. Two variables were strongly related to metformin initiation. Metformin use increased relative to sulfonylurea use over time as reflected by odds ratios for fiscal years 2004 to 2007. Initiation of metformin therapy decreased with increasing baseline serum creatinine levels as reflected by odds ratios for 0.54. Table 1 shows the P values for patients who initiated metformin and sulfonylurea therapy before and after propensity score matching; after matching, few standardized differences are statistically significant, indicating good balance.

Another important assumption for propensity score methods is that every cohort member has a nonzero probability of being either a sulfonylurea user or a metformin user. Any cohort members who must always receive a sulfonylurea or who could never receive a sulfonylurea would be excluded, because the relevant comparison is between persons who are eligible for either drug but may or may not actually receive one of them. We tested this assumption by reviewing the overlap in the distribution of the propensity scores in patients who initiated sulfonylurea and metformin therapy. As Appendix Figure 1 shows, this distribution differed slightly for users of metformin and sulfonylureas, but the overlap was nearly complete. The model yielded a c-statistic of 0.71.