OBJECTIVE

To assess the real-world cardiovascular (CV) safety for sulfonylureas (SU), in comparison with dipeptidyl peptidase 4 inhibitors (DPP4i) and thiazolidinediones (TZD), through development of robust methodology for causal inference in a whole nation study.

RESEARCH DESIGN AND METHODS

A cohort study was performed including people with type 2 diabetes diagnosed in Scotland before 31 December 2017, who failed to reach HbA1c 48 mmol/mol despite metformin monotherapy and initiated second-line pharmacotherapy (SU/DPP4i/TZD) on or after 1 January 2010. The primary outcome was composite major adverse cardiovascular events (MACE), including hospitalization for myocardial infarction, ischemic stroke, heart failure, and CV death. Secondary outcomes were each individual end point and all-cause death. Multivariable Cox proportional hazards regression and an instrumental variable (IV) approach were used to control confounding in a similar way to the randomization process in a randomized control trial.

RESULTS

Comparing SU to non-SU (DPP4i/TZD), the hazard ratio (HR) for MACE was 1.00 (95% CI: 0.91–1.09) from the multivariable Cox regression and 1.02 (0.91–1.13) and 1.03 (0.91–1.16) using two different IVs. For all-cause death, the HR from Cox regression and the two IV analyses was 1.03 (0.94–1.13), 1.04 (0.93–1.17), and 1.03 (0.90–1.17).

CONCLUSIONS

Our findings contribute to the understanding that second-line SU for glucose lowering are unlikely to increase CV risk or all-cause mortality. Given their potent efficacy, microvascular benefits, cost effectiveness, and widespread use, this study supports that SU should remain a part of the global diabetes treatment portfolio.

Type 2 diabetes is associated with increased risk of microvascular and macrovascular disease, with the risk of cardiovascular (CV) mortality more than double in people with type 2 diabetes compared with those without (1). In the last decade, large dedicated CV outcome trials in people with type 2 diabetes and at high risk or with established CV disease have shown that dipeptidyl peptidase 4 inhibitors (DPP4i) do not increase CV risk (2), while sodium–glucose cotransporter 2 inhibitors (SGLT2i) and glucagon-like peptide 1 receptor analogists (GLP-1RA) reduce CV risk (35). As a result, national and international guidelines have favored newer, more expensive glucose-lowering medicines over older, cheaper options such as sulfonylureas (SU) and thiazolidinediones (TZD).

There are huge regional disparities around the world in terms of cost and access to medications, as, notably, 80% of people living with diabetes reside in developing countries, yet these regions account for only 1% of the global diabetes expenditure (6). Therefore, generic medications remain relevant as part of the global diabetes treatment strategy to control cost and effective, accessible care.

SU are inexpensive, potent glucose-lowering agents that have been widely used in the management of type 2 diabetes for over 60 years. There has been a long-standing controversy over the CV safety of SU, originating from the early clinical trials evaluating SU for diabetes management, which were either underpowered or were inadequately designed by current standards to evaluate risks of CV outcomes or mortality (7,8). Although these early trials were designed before current standards were in place, the controversy surrounding SU has been backed up by repeated observational studies, which tend to report increased CV risk for SU versus comparators (often metformin) (911). Given the putative CV risk and other side effects of SU such as hypoglycemia, body weight gain, and limited durability (8,12), there has been a debate whether SU should remain as routine second-line pharmacotherapy for type 2 diabetes (7,13,14). However, the compound annual growth rate of the global SU market continues to rise at a rate of 2.69%, with the fastest growth occurring in developing countries. The U.S. accounts for 44% of the market share of SU, with patients receiving SU having significantly lower total health care costs than those receiving other diabetes medications (15); therefore, the market for cost-effective SU clearly remains strong worldwide.

Few randomized control trials (RCTs) have been conducted that make head-to-head comparisons between SU and other active comparators, in particular with SGLT2i or GLP1-RA. The Thiazolidinediones or Sulfonylureas Cardiovascular Accidents Intervention Trial (TOSCA.IT) study (16), a randomized multicenter trial, compared the long-term CV outcomes of pioglitazone, a TZD, versus SU (2% glibenclamide, 48% glimepiride, 50% gliclazide). The trial was stopped early based on a futility analysis but provided some evidence that SU (mostly glimepiride and gliclazide) and pioglitazone as add-on drugs to metformin were similar in terms of CV safety. More recently, the Cardiovascular Outcome Study of Linagliptin vs. Glimepiride in Type 2 Diabetes (CAROLINA) CV outcome trial (n = 6,042) has demonstrated noninferiority of linagliptin, a DPP4i, versus glimepiride, an SU, in time to first occurrence of CV death, nonfatal myocardial infarction (MI), or nonfatal stroke (adjusted HR: 1.02; 95% CI: 0.88 to 1.19) (17).

Observational studies have attempted to investigate the CV safety of SU in a real-world setting, but many lacked robust designs or appropriate methodologies for data analysis and were therefore criticized for suffering major biases. A meta-analysis of 44 observational studies assessing the CV safety of SU reported several likely sources of bias such as using selected populations with CV complications, performing within-class comparisons, or utilizing a normal glucose tolerance cohort as comparator (10). In general, the biases in observational studies could be classified into one of the two main categories: selection bias and confounding bias. They are different in principle, but both induce incomparability of the exposure groups, which may subsequently lead to biased results for comparison. Therefore, considerable effort is required for the study design and the subsequent analysis to eliminate or at least minimize the potential biases.

In this study, we analyzed a large cohort derived from the entire Scottish population with type 2 diabetes to provide real-world evidence about the CV safety of SU, in comparison with other active comparators, namely, DPP4i and TZD, each being used in combination with metformin for treatment intensification. A robust new-user design of second-line therapies used for treatment intensification was adopted to minimize potential selection bias. Confounding control was achieved by 1) multivariable analysis adjusted for an appropriate set of covariates/confounders, and 2) applying an instrumental variable (IV) approach based upon prescribing preference to emulate the randomization process in RCTs and infer causal treatment effects. The IV approach was originally developed for analyses in economics; however, it has been increasingly applied in medical research, as it explores how a variable influences treatment and has no confounding with the outcome; that is, it accounts for natural randomization (18). Treatment effect is evaluated on the valid instrument (which determines the exposure) rather than the allocated treatment, akin to an intention-to-treat analysis, which is advantageous, as it does not assume the absence of unmeasured confounders to the treatment-outcome relationship. This allows an unconfounded treatment effect to be estimated as in RCT. In this way, we can provide reliable results from analyzing large, routinely collected, real-world health care records, and provide guidance for comparative effectiveness and studying drug safety using observational data.

Data Sources

We conducted a retrospective population-based cohort study using data from a 2018 extract of the Scottish Care Information–Diabetes national register, a clinical database which contains data on all health care encounters in relation to diabetes. Scottish Care Information–Diabetes was rolled out across Scotland in 2000 and captures key diabetes-related data items from all hospitals and around 1,100 general practices in Scotland. The data were also linked by the Information Services Division of National Health Services Scotland to national mortality, cancer registry, and hospital admission records.

Study Cohort

People with an incident diagnosis of type 2 diabetes in Scotland were included in the study cohort if they 1) were aged 18 years or over at diagnosis of type 2 diabetes mellitus (T2DM) and 2) failed to reach target HbA1c level (48 mmol/mol) through first-line metformin monotherapy, and subsequently initiated second-line treatment on or after 1 January 2010 with one of the following classes of drugs: SU, DPP4i, or TZD. Cohort entry (i.e., index date) was defined by the date of the first prescription of the above second-line drugs. To make sure these drugs were prescribed as add-ons to metformin, we required that either metformin was coprescribed on the index date or at least one prescription for metformin was issued within 60 days after the index date and prior to adding other third-line drugs. This was to exclude people who switched from metformin to one of the study drugs (potentially due to intolerance or contraindication) but remained on monotherapy. The study cohort was then further restricted by excluding 1) people aged under 40 years, or above 85 years of age at index date, and 2) people prescribed more than one class of second-line drugs at index date.

Study Outcomes

The primary outcome was the composite major adverse CV events (MACE), including hospitalization for MI, hospitalization for ischemic stroke, hospitalization for heart failure (HF), and CV death. Each individual component of the composite end point as well as all-cause death were analyzed as secondary outcomes. Hospital admission for MI (ICD-9 codes 410.x, ICD-10 codes I21.x), stroke (ICD-9 codes 433.x, 434.x, or 436.x; ICD-10 codes I63.x or I64.x), and HF (ICD-9 codes 428.x; ICD-10 codes I50x, I11.0, I13.0, or I13.2) were identified using The General/Acute and Inpatient Day Case data set (SMR01). CV death (ICD-9 codes 390.x to 398.x, 401.x to 405.x, 410.x to 417.x, 420.x to 429.x [except 427.5], 430.x to 438.x, or 440.x to 447.×; ICD-10 codes I00.x to I77.x [except I46.9]) was identified from all causes recorded in the death certificates from the General Register Office, National Records of Scotland, and Scottish Cancer Registry (SMR06). All-cause death was identified from all three databases, with the date of death defined by the earliest recording of death in any data sets.

Exposures

For the primary analysis, we assembled DPP4i and TZD to be one “non-SU” group and considered a binary exposure, that is, SU versus non-SU. Further subgroup analyses included 1) head-to-head comparisons including SU versus DPP4i and SU versus TZD; 2) study cohort stratified respectively by prior history of MACE, age at index date (< or ≥ 70 years old), and BMI (< or ≥ 30 kg/m2); and 3) SU exposure stratified by individual SU (gliclazide, glipizide, glimepiride). The treatment effects were evaluated in an intention-to-treat framework (19), that is, based on the initiation of second-line treatment irrespective of their discontinuation or subsequent switches to, or additions of, third-line drug classes. This was to 1) avoid the informative (i.e., nonrandom) censoring and the potential time-varying confounding due to the differences in drug response, and 2) make the estimates of treatment effects consistent with analyzing RCT data. Participants included in the study cohort were followed until the occurrence of one of the study outcomes or were censored at the end of the study period, that is, 31 December 2017 (Supplementary Fig. 1).

Covariates

We adjusted our analyses for the following covariates, selected based on the “disjunctive cause criterion” to achieve better confounding control (20):

  1. Demographics: age at cohort entry, sex, ethnicity, quintiles of Scottish Index of Multiple Deprivation, duration of diabetes, smoking status, year of cohort entry

  2. Most recent clinical measurements (on or prior to cohort entry): BMI, estimated glomerular filtration rate by the Chronic Kidney Disease Epidemiology Collaboration creatinine (CKD-EPI) equation, HbA1c, systolic blood pressure, and total cholesterol/HDL cholesterol ratio

  3. Existing (ICD-coded) comorbidities: atrial fibrillation of flutter, coronary artery disease, cancer, chronic obstructive pulmonary disease, diabetic retinopathy, and history of MI, stroke, or HF

  4. Currently used drugs: ACE inhibitors or angiotensin II receptor blockers, β-blockers, calcium channel blockers, diuretics, cardiac glycosides, nitrates, oral anticoagulants, antiplatelets, and lipid-lowering drugs.

Statistical Analyses

Descriptive Statistics

Characteristics of the study participants were summarized using descriptive statistics. Incidence of the study outcomes was reported as number of events per 1,000 person-years. The temporal prescribing pattern of the three classes of drugs was described graphically. SGLT2i were also incorporated when deriving the annual prescribing proportions, as they started being prescribed in Scotland from 2013 and were officially recommended as one of the second-line options by The National Institute for Health and Care Excellence in December 2015.

Cox Proportional Hazards Regression

Cox proportional hazard regression models were used firstly to evaluate the associations between exposures and the study outcomes. Unadjusted and adjusted hazard ratios (HRs) were reported. The 95% CIs were established using robust SEs to address the potential “clustering effect” between practices. In the absence of unmeasured confounding, conventional multivariable analysis adjusting for a reasonable selection of covariates can provide unbiased estimates for treatment effects. Residual confounding, however, may still exist when there are key unmeasured confounding factor(s). Validity of the proportional hazard assumption was assessed by checking the Schoenfeld residuals.

IV Analyses

To account for potential residual confounding, we conducted IV analyses (18,21), with practice-level prescribing preference as an instrument to act as a proxy for exposure. The rationale for the IV analysis of observational data was attempting to re-establish the balance or exchangeability brought by the randomization process in an RCT. Prescribing preference cannot be directly measured; therefore, we used the prescriptions issued to previous patients in the practice as a proxy for the preference. Two different IVs were defined: 1) proportion of SU prescriptions among the 10 most recent prescriptions and 2) proportion of SU prescriptions among all the prescriptions during the last 365 days. Both IVs were evaluated at each patient’s index date to allow the practice-level preference to be time varying. This was important because the utilization of the three classes of drugs varied substantially over the study period.

IV estimates for the exposure effects were then derived using two techniques, namely, two-stage estimation and G-estimation. For the two-stage estimation, the exposure was regressed on the IV and year of cohort entry in the first-stage model. In the second stage, a Cox model including the exposure, the adjusted covariates, and the “control function” was used to estimate the exposure effect. For the G-estimation, a structural model was formed of one linear model for the IV, regressed on year of cohort entry, and one Cox model for the outcome, regressed on the IV, the exposure, and the adjusted covariates. As noted previously, the two-stage estimation for binary or time-to-event outcome is asymptotically biased (22), but the bias can sometimes be reduced by using the control function approach (23,24). G-estimation is an alternative approach in causal inference that can give an unbiased estimate. Here we used a special case of G-estimator and its analytic SEs, which were recently proposed to allow the G-estimation technique to be implemented in IV analysis (22,24). Assessment of essential IV conditions are described in Supplementary Method 3.

Sensitivity Analyses

For sensitivity analyses, we added additional censoring criteria, that is, adding or switching to another class of glucose-lowering drug (different from metformin and the current second-line drug), to evaluate the treatment effects in an on-treatment framework. Again, the outcome event rates were compared between 1) SU and non-SU agents (DPP4i or TZD), 2) SU and DPP4i, and 3) SU and TZD.

All analyses were conducted in R version 3.6 (25).

Data and Resource Availability

These individual patient–level real-world data are only available for analysis on a Trusted Research Environment and therefore cannot be made available.

Patient Characteristics

A total of 31,460 people in Scotland with type 2 diabetes met the study inclusion criteria, where 19,854 initiated second-line treatment by adding SU, 9,591 were prescribed DPP4i, and another 2,015 were prescribed TZD. Baseline characteristics are summarized in Table 1; a flowchart describing the study cohort is shown in Supplementary Fig. 2. People who received SU prescriptions were slightly older, with higher baseline HbA1c but lower BMI, and had more comorbidities compared with those who received non-SU agents. The percentages of missing data were extremely low; therefore, the individuals with incomplete information of baseline covariates were excluded from further analyses. This also guaranteed that our study outcomes would be analyzed on the same cohort of people. The final cohort for analysis included 29,518 people, where 18,531 were SU initiators (gliclazide [n = 16,152, 87.2%], glimepiride [n = 1,540, 8.3%], glipizide [n = 818, 4.4%], and glibenclamide [n = 21, 0.1%]) and 10,987 were non-SU initiators (9,114 DPP4i and 1,873 TZD).

Incidence Rate of Outcomes

Supplementary Table 1 summarizes the number of outcome events, median follow-up time, and the incidence rate, stratified by exposure groups. The median follow-up of the SU group was 3.9 years for composite MACE and was 4.1 years for all-cause death, longer than those of the non-SU group (3.0 years for MACE and 3.1 years for all-cause death). Higher incidence rates per 1,000 person-years were observed for all study outcomes in the SU versus the non-SU group (MACE: 23.4 vs. 18.7; hospitalization for MI: 7.1 vs. 5.5; hospitalization for stroke: 5.1 vs. 4.8; hospitalization for HF: 3.4 vs. 2.1; CV death: 12.2 vs. 9.2; and all-cause death: 21.2 vs. 16.1).

Relative Effect of SU Versus Non-SU Agents

Fig. 1 shows the results of the comparison between second-line SU and non-SU (DPP4i/TZD) agents. For MACE, the multivariable Cox regression and the IV analyses provided consistent estimates showing that prescribing SU as the second-line addition to metformin was not associated with increased overall CV risk. The estimated HR was 1.00 (95% CI: 0.91 to 1.09) from the multivariable Cox regression, was 1.02 (0.91 to 1.13) from the G-estimation using IV-10, was 1.03 (0.91 to 1.16) from the G-estimation using IV-365, was 0.95 (0.77 to 1.16) from the two-stage estimation using IV-10, and was 0.96 (0.77 to 1.20) from the two-stage estimation using IV-365. The upper limits of the 95% CIs were all below 1.3, the noninferiority upper limit suggested by the Food and Drug Administration (FDA) for CV safety trials.

For all-cause death, the estimated HR was 1.03 (95% CI: 0.94 to 1.13) from the multivariable Cox regression, was 1.04 (0.93 to 1.17) from the G-estimation using IV-10, was 1.03 (0.90 to 1.17) from the G-estimation using IV-365, was 1.02 (0.83 to 1.25) from the two-stage estimation using IV-10, and was 1.01 (0.81 to 1.25) from the two-stage estimation using IV-365. All these indicated that prescribing SU for initiation of treatment intensification was unlikely to increase the risk of all-cause death.

Similar results were obtained for the individual MACE end points. For hospitalizations for MI, stroke, and HF, the variation of the estimates from the IV analyses was slightly larger, which could be due to the small numbers of observed events, that is, high censoring percentages. In general, the 95% CIs of the IV estimates were slightly wider compared with those of the conventional multivariable Cox regression. This is a typical characteristic of the IV approach (26).

Subgroup Analyses

The results of the head-to-head comparison between second line SU and DPP4i are shown in Fig. 2. In our analyses, the estimated HR for 4-point (4P)-MACE was 0.98 (0.88 to 1.08) from the multivariable Cox regression, was 0.91 (0.72 to 1.17) and 0.97 (0.86 to 1.10) from the two-stage estimation and G-estimation, respectively, using IV-10, and was 0.97 (0.75 to 1.27) and 1.00 (0.88 to 1.14) from the two-stage estimation and G-estimation, respectively, using IV-365. For all-cause death, our estimate was 1.01 (0.92 to 1.12) from the multivariable Cox regression, was 1.02 (0.80 to 1.29) and 0.99 (0.87 to 1.13) from the estimations using IV-10, and was 1.03 (0.80 to 1.33) and 0.98 (0.85 to 1.12) from the estimations using IV-365.

Fig. 3 shows the results of the comparison between SU and TZD. No significantly higher risks were observed in the SU group. The TZD group was of a small size (n = 1,873), with fewer outcome events observed (Supplementary Table 1). Therefore, relatively wider 95% CIs were obtained for the point estimates.

The results of other subgroup analyses are shown in Supplementary Tables 2–4. The CV safety of SU was consistently supported across all predefined subgroups. In the subgroup analysis of individual SU, HRs for glibenclamide were not evaluated because of the small sample size (n = 21). However, this is reflective of the decline in prescribing of less tissue-specific SU within the Scottish population. Our results showed little difference in outcome rates among different types of SU.

Instrument Variable Assessment

The "relevance" condition (Supplementary Method 3) was satisfied for the two proposed IVs, indicated by the large difference in the deviance (analogous to the F statistic) and the significance results from the likelihood ratio tests (Supplementary Table 6). The point-biserial correlation was 0.497 for IV-10 and 0.516 for IV-365. The crude and adjusted odds ratios shown in Supplementary Table 7 were similarly large with and without year of cohort entry. All these assured strong association between the exposure and the proposed IVs. As shown in Supplementary Table 8, most covariates were balanced across the binary exposure groups (SU versus non-SU), except for age (standardized difference [SDif] = 0.105 > 0.1) and baseline HbA1c level (SDif = 0.220 > 0.1). For the two proposed IVs, all the covariates were well balanced across the quartiles, indicating that the “exchangeability” condition (Supplementary Method 3) was unlikely to be violated.

Sensitivity Analyses

The design of sensitivity analyses is shown in Supplementary Fig. 5. Censoring additionally at adding or switching to another class of glucose-lowering drug reduced the follow-up time. For MACE, the median follow-up time in the SU group was reduced from 3.9 years to 1.8 years, while, in the non-SU group, this was reduced from 3.0 years to 1.4 years (Supplementary Table 5). The outcome rates in the SU group were similar to those obtained in the primary analyses, while lower outcome rates were observed in people who received non-SU agents. These were also reflected in the slightly higher HRs shown in Supplementary Figs. 6–8. However, none of the estimates indicated significantly higher CV risk of SU comparing to DPP4i or TZD.

This study analyzed data for the entire Scottish population with T2DM to systematically assess the CV safety of SU, in comparison with DPP4i and TZD, all being prescribed as second-line add-ons to the first-line metformin. Our findings demonstrate that prescribing SU, compared with the other two non-SU agents, is not significantly associated with higher risks of MACE or all-cause death. Furthermore, the HRs presented in Figs. 2 and 3 show that our approach has produced nearly identical results when compared with those of major RCT involving second-line SU as comparator to DPP4 or TZD: CAROLINA (HR for 3-point [3P]-MACE: 1.02 [0.88 to 1.19]; HR for all-cause death: 1.10 [0.94 to 1.28]) and TOSCA.IT (HR for 3P-MACE: 1.04 [0.79 to 1.35]; HR for all-cause death: 0.91 [0.62 to 1.33]). Given that DPP4i have been established to be neutral for MACE risk (2,17,27,28) and pioglitazone has been found to have cardioprotective effects (29,30), our findings provide real-world evidence to support the conclusion that SUs prescribed as second-line pharmacotherapy are unlikely to increase CV risk or all-cause death.

Substantial changes in the prescribing pattern were found from our drug utilization analysis (Supplementary Fig. 3). SU used to be the most prescribed second-line add-on to metformin. DPP4i was approved in 2007 and was recommended as a second-line option in May 2009 (The National Institute for Health and Care Excellence guideline CG87), together with SU and pioglitazone. Since then, prescribing of DPP4i has increased rapidly, and, in 2017, it had become the most prescribed second-line drug class in Scotland (38% DPP4i, 37% SU, 22% SGLT2i, and 3% TZD). Rosiglitazone was indicated to increase risk of MI, in a systematic review in 2009, and was subsequently suspended from use in the European Union from 2010. Based on the facts described above, we therefore restricted our study cohort to include only eligible individuals on or after 2010 to improve the comparability between exposure groups and minimize potential selection bias.

Over the study period from 2010 to 2017, higher incidence rates of CV outcomes and all-cause death were observed in the SU cohort, compared to those prescribed non-SU agents. The incidence rate ratio was 1.25 for MACE and was 1.32 for all-cause death, consistent with the unadjusted HRs reported in Fig. 1 (1.25 [1.14 to 1.36] for MACE, and 1.30 [1.18 to 1.42] for all-cause death). Higher crude incidence rates would be expected from the systematic differences in baseline characteristics. As demonstrated in Table 1, the SU cohort was slightly older, and had a higher proportion of current smokers, poorer glycemic control, and more existing comorbidities, in comparison with those prescribed DPP4i or TZD. To address these systematic baseline differences, multivariable Cox regression with adjustment and IV approach were applied in further analyses.

The primary analyses of this study addressed the CV outcomes of SU as second-line agents versus DPP4i/TZD; however, within-class differences in SU KATP channel tissue specificities have suggested that second-generation SU are preferable to first generation, particularly in terms of safety (3133). Novel findings of a recent study also suggest SU with high-affinity binding with cardiac mitochondrial KATP channels are associated with increased MACE risk compared with those with low affinity (34). Our subgroup analyses showed little difference between second-generation SU. The difference in CV outcome observed in this study compared with older observational studies could be explained by gliclazide being the SU of choice within Scotland (87.2% of second-line SU users), while other studies included high use of SU with high cardiac KATP and mitochondrialKATP affinity such as glibenclamide (35,36). Furthermore, some existing observational studies reporting higher CV risk of SU included a high proportion of people who switch from metformin to SU; including SU users who switched from metformin but remained on first-line monotherapy may contaminate the treatment effect estimates. In our study, we excluded patients who switched treatment, ensuring the second-line drug was used for treatment intensification as add-on to metformin.

Our analyses demonstrate that observational studies can generate reliable and robust evidence, consistent with RCT findings. When unmeasured confounding is not a major concern, conventional multivariable regression together with a careful study design can minimize or at least reduce the potential biases. If residual confounding is suspected, IV approaches provide a potential way to address this so that covariate balance can be achieved. In particular, preference-based IVs defined at the level of the geographic region, hospital, or individual physician have been used in comparative effectiveness and safety studies in the past two decades (37). However, IV estimates are usually characterized by larger variance(18,26). As a result, null effect of an exposure is often concluded when the IV-exposure association is weak; however, this was not the case for our study where the IV was strong (see Supplementary Tables 6 and 7 for the evaluation of IV strength). Therefore, for comparative effectiveness or drug safety studies aiming for causal treatment effect, we recommend performing both conventional multivariable regression and IV analysis.

To date, this is the first and the only large-scale population study applying IV approach with G-estimation in a survival context to estimate causal treatment effects. Unlike G-estimation, two-stage methods generally give a biased estimate when a Cox model is used at the second stage. Existing studies usually ignore this problem or circumvent it by considering the outcome as binary or even continuous and evaluate the causal treatment effect through structural mean models. In our analyses, the two-stage estimates were obtained by using the “control function” approach, instead of substituting the exposure in the second-stage model with its predicted value from the first-stage model. This reduced the bias and provided the point estimates close to those obtained from the G-estimation (24). Aalen’s additive hazard model is another option under the two-stage setting. However, it may be less attractive for clinical or epidemiological studies, as the interpretation of results is not as intuitive as those from a Cox proportional hazard model. The performances of the two proposed IVs were similar, although the instrument defined using the prescriptions in the previous year (IV-365) is slightly stronger than the one defined by a fixed number of historical prescriptions (IV-10), often over a longer period than a year. We did not consider longer prescribing history, as older prescriptions may be less relevant to the current prescribing preference, especially when the prescribing pattern varies significantly over time.

A limitation of this study is that the potential impact of competing risk was not considered for nonfatal study outcomes. However, the results obtained for these outcomes were in keeping with the findings for all-cause death, which may suggest a negligible impact of competing end points. The power of subgroup analyses by SU type was limited by sample size, which reflects the shift in prescribing preference toward more tissue-selective SU which were associated with lower risk of all-cause CV-related death in a large meta-analysis (31). In this study, given the data governance for large anonymized electronic health record data, outcomes were not adjudicated; however, this is a limitation of all observational studies. This work used ICD codes to establish MACE events, which is widely accepted in epidemiological research. Finally, the focus of this work was to assess real-world CV safety of SU through development of robust methodology for casual inference; while it is acknowledged that these models do not address other clinical risk associated with SU such as durability and the risk and associated costs of severe hypoglycemia, this work does provide support that CV risk is not increased when considering SU against the other second-line agents studied.

Conclusion

In conclusion, our study has provided robust real-world evidence for the CV safety of SU, being prescribed in addition to metformin, in an unselected population with T2DM and with or without high CV risk or established major CV events. Furthermore, we have developed robust methodology for estimating causal treatment effects. We acknowledge that newer noninsulin agents such as SGLT2i and GLP-1RA may carry long-term benefits from reducing risks of CV and renal events. In particular, SGLT2i were suggested to be cost-effective even at the current price (38), and would be prescribed more in the foreseeable future. However, when these newer agents are not accessible or are contraindicated, the concern for CV safety should not be a barrier to prescribing SU. Although other clinical factors such as hypoglycemia risk and durability regarding SU need to be considered (39), our findings from this study support the most recent international guidelines (40), which recommend SU as one of the second-line options after metformin if resources are limited. Therefore, SU should remain as part of the global diabetes treatment portfolio, given its strong efficacy in glycemic control, established microvascular benefits, and the real-world evidence added to trial evidence for CV safety.

This article contains supplementary material online at https://doi.org/10.2337/figshare.22106990.

This article is featured in podcasts available at diabetesjournals.org/journals/pages/diabetes-core-update-podcasts.

H.W. and R.L.M.C. contributed equally to this work.

*

A list of collaborators within the Scottish Diabetes Research Network epidemiology group can be found in the supplementary material online.

Funding. This work was supported by Health Data Research UK which receives its funding from HDR UK Ltd (HDR-5012) funded by the UK Medical Research Council, Engineering and Physical Sciences Research Council, Economic and Social Research Council, Department of Health and Social Care (England), Chief Scientist Office of the Scottish Government Health and Social Care Directorates, Health and Social Care Research and Development Division (Welsh Government), Public Health Agency (Northern Ireland), British Heart Foundation (BHF), and the Wellcome Trust.

Duality of Interest. R.L.M.C. has received honoraria from Sanofi, J.M. has received speaker fees from Napp Pharmaceuticals and has been involved in CV outcome trials funded by Novo Nordisk, Eli Lilly, Boehringer, GlaxoSmithKline, and Medimmune Ltd. H.C. has received grants or institutional fees from Eli Lilly and Company, AstraZeneca LP, Pfizer, and Novo Nordisk. N.S. has consulted for Afimmune, Amgen, AstraZeneca, Boehringer Ingelheim, Eli Lilly, Hanmi Pharmaceutical, Merck Sharp & Dohme, Novartis, Novo Nordisk, Pfizer, and Sanofi, and received grant support paid to his university from AstraZeneca, Boehringer Ingelheim, Novartis, and Roche Diagnostics outside the submitted work. R.J.M. has received royalties or licenses from Elsevier and honoraria from Sanofi and Novo Nordisk, and institutional fees from National Health Services Tayside and Medical Reserve Corps. E.R.P. has received honoraria from Sanofi and Eli Lilly. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. H.W., R.L.M.C., Y.H., E.R.P., D.M., and L.D. were involved in the design of the study. H.W. and Y.H. led the statistical analysis. H.W. and R.L.M.C. wrote the first draft. All authors contributed to further drafts and approved the manuscript. E.R.P. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

1.
Hippisley-Cox
J
,
Coupland
C
,
Brindle
P
.
Development and validation of QRISK3 risk prediction algorithms to estimate future risk of cardiovascular disease: prospective cohort study
.
BMJ
2017
;
357
:
j2099
2.
Green
JB
,
Bethel
MA
,
Armstrong
PW
, et al.;
TECOS Study Group
.
Effect of sitagliptin on cardiovascular outcomes in type 2 diabetes
.
N Engl J Med
2015
;
373
:
232
242
3.
Zinman
B
,
Wanner
C
,
Lachin
JM
, et al.;
EMPA-REG OUTCOME Investigators
.
Empagliflozin, cardiovascular outcomes, and mortality in type 2 diabetes
.
N Engl J Med
2015
;
373
:
2117
2128
4.
Marso
SP
,
Bain
SC
,
Consoli
A
, et al.;
SUSTAIN-6 Investigators
.
Semaglutide and cardiovascular outcomes in patients with type 2 diabetes
.
N Engl J Med
2016
;
375
:
1834
1844
5.
Marso
SP
,
Daniels
GH
,
Brown-Frandsen
K
, et al.;
LEADER Steering Committee
;
LEADER Trial Investigators
.
Liraglutide and cardiovascular outcomes in type 2 diabetes
.
N Engl J Med
2016
;
375
:
311
322
6.
Fralick
M
,
Jenkins
AJ
,
Khunti
K
,
Mbanya
JC
,
Mohan
V
,
Schmidt
MI
.
Global accessibility of therapeutics for diabetes mellitus
.
Nat Rev Endocrinol
2022
;
18
:
199
204
7.
UK Prospective Diabetes Study (UKPDS) Group
.
Effect of intensive blood-glucose control with metformin on complications in overweight patients with type 2 diabetes (UKPDS 34)
.
Lancet
1998
;
352
:
854
865
8.
The University Group Diabetes Program
.
A study of the effects of hypoglycemic agents on vascular complications in patients with adult-onset diabetes. V. Evaluation of pheniformin therapy
.
Diabetes
1975
;
24
(
Suppl. 1
):
65
184
9.
Monami
M
,
Genovese
S
,
Mannucci
E
.
Cardiovascular safety of sulfonylureas: a meta-analysis of randomized clinical trials
.
Diabetes Obes Metab
2013
;
15
:
938
953
10.
Azoulay
L
,
Suissa
S
.
Sulfonylureas and the risks of cardiovascular events and death: a methodological meta-regression analysis of the observational studies
.
Diabetes Care
2017
;
40
:
706
714
11.
Douros
A
,
Dell’Aniello
S
,
Yu
OHY
,
Filion
KB
,
Azoulay
L
,
Suissa
S
.
Sulfonylureas as second line drugs in type 2 diabetes and the risk of cardiovascular and hypoglycaemic events: population based cohort study
.
BMJ
2018
;
362
:
k2693
12.
Viberti
G
,
Kahn
SE
,
Greene
DA
, et al
.
A diabetes outcome progression trial (ADOPT): an international multicenter study of the comparative efficacy of rosiglitazone, glyburide, and metformin in recently diagnosed type 2 diabetes
.
Diabetes Care
2002
;
25
:
1737
1743
13.
Khunti
K
,
Chatterjee
S
,
Gerstein
HC
,
Zoungas
S
,
Davies
MJ
.
Do sulphonylureas still have a place in clinical practice?
Lancet Diabetes Endocrinol
2018
;
6
:
821
832
14.
Cordiner
RLM
,
Pearson
ER
.
Reflections on the sulphonylurea story: a drug class at risk of extinction or a drug class worth reviving?
Diabetes Obes Metab
2019
;
21
:
761
771
15.
Liss
DT
,
Kang
RH
,
Lancki
N
, et al
.
Costs for commercially insured adults prescribed second-line diabetes medications
.
Am J Manag Care
2021
;
27
:
e72
e79
16.
Vaccaro
O
,
Masulli
M
,
Nicolucci
A
, et al.;
Thiazolidinediones Or Sulfonylureas Cardiovascular Accidents Intervention Trial (TOSCA.IT) study group
;
Italian Diabetes Society
.
Effects on the incidence of cardiovascular events of the addition of pioglitazone versus sulfonylureas in patients with type 2 diabetes inadequately controlled with metformin (TOSCA.IT): a randomised, multicentre trial
.
Lancet Diabetes Endocrinol
2017
;
5
:
887
897
17.
Rosenstock
J
,
Kahn
SE
,
Johansen
OE
, et al.;
CAROLINA Investigators
.
Effect of linagliptin vs glimepiride on major adverse cardiovascular outcomes in patients with type 2 diabetes: The CAROLINA Randomized Clinical Trial
.
JAMA
2019
;
322
:
1155
1166
18.
Baiocchi
M
,
Cheng
J
,
Small
DS
.
Instrumental variable methods for causal inference
.
Stat Med
2014
;
33
:
2297
2340
19.
Swanson
SA
.
Instrumental variable analyses in pharmacoepidemiology: what target trials do we emulate?
Curr Epidemiol Rep
2017
;
4
:
281
287
20.
VanderWeele
TJ
.
Principles of confounder selection
.
Eur J Epidemiol
2019
;
34
:
211
219
21.
Tchetgen Tchetgen
EJ
,
Walter
S
,
Vansteelandt
S
,
Martinussen
T
,
Glymour
M
.
Instrumental variable estimation in a survival context
.
Epidemiology
2015
;
26
:
402
410
22.
Sjolander
A
,
Martinussen
T
.
Instrumental variable estimation with the R package ivtools
.
Epidemiol Methods
2019
;
8
:
20180024
23.
Martinussen
T
,
Vansteelandt
S
,
Tchetgen Tchetgen
EJ
,
Zucker
DM
.
Instrumental variables estimation of exposure effects on a time-to-event endpoint using structural cumulative survival models
.
Biometrics
2017
;
73
:
1140
1149
24.
Martinussen
T
,
Nørbo Sørensen
D
,
Vansteelandt
S
.
Instrumental variables estimation under a structural Cox model
.
Biostatistics
2019
;
20
:
65
79
25.
R Foundation for Statistical Computing
.
R, Version 3.6
R Foundation for Statistical Computing
26.
Burgess
S
,
Small
DS
,
Thompson
SG
.
A review of instrumental variable estimators for Mendelian randomization
.
Stat Methods Med Res
2017
;
26
:
2333
2355
27.
Scirica
BM
,
Bhatt
DL
,
Braunwald
E
, et al.;
SAVOR-TIMI 53 Steering Committee and Investigators
.
Saxagliptin and cardiovascular outcomes in patients with type 2 diabetes mellitus
.
N Engl J Med
2013
;
369
:
1317
1326
28.
White
WB
,
Cannon
CP
,
Heller
SR
, et al.;
EXAMINE Investigators
.
Alogliptin after acute coronary syndrome in patients with type 2 diabetes
.
N Engl J Med
2013
;
369
:
1327
1335
29.
Dormandy
JA
,
Charbonnel
B
,
Eckland
DJA
, et al.;
PROactive Investigators
.
Secondary prevention of macrovascular events in patients with type 2 diabetes in the PROactive Study (PROspective pioglitAzone Clinical Trial In macroVascular Events): a randomised controlled trial
.
Lancet
2005
;
366
:
1279
1289
30.
Holman
RR
,
Retnakaran
R
,
Farmer
A
,
Stevens
R
.
PROactive study
.
Lancet
2006
;
367
:
25
26
[author reply 367:26–27]
31.
Simpson
SH
,
Lee
J
,
Choi
S
,
Vandermeer
B
,
Abdelmoneim
AS
,
Featherstone
TR
.
Mortality risk among sulfonylureas: a systematic review and network meta-analysis
.
Lancet Diabetes Endocrinol
2015
;
3
:
43
51
32.
Hong
J
,
Zhang
Y
,
Lai
S
, et al.;
SPREAD-DIMCAD Investigators
.
Effects of metformin versus glipizide on cardiovascular outcomes in patients with type 2 diabetes and coronary artery disease
.
Diabetes Care
2013
;
36
:
1304
1311
33.
Douros
A
,
Yin
H
,
Yu
OHY
,
Filion
KB
,
Azoulay
L
,
Suissa
S
.
Pharmacologic differences of sulfonylureas and the risk of adverse cardiovascular and hypoglycemic events
.
Diabetes Care
2017
;
40
:
1506
1513
34.
Wang
MT
,
Huang
YL
,
Lai
JH
, et al
.
Association between specificity of sulfonylureas to cardiac mitochondrial KATP channels and the risk of major adverse cardiovascular events in type 2 diabetes
.
Diabetes Care
2022
;
45
:
1276
1287
35.
Powell
WR
,
Christiansen
CL
,
Miller
DR
.
Long-term comparative safety analysis of the risks associated with adding or switching to a sulfonylurea as second-line type 2 diabetes mellitus treatment in a US veteran population
.
Diabet Med
2019
;
36
:
1384
1390
36.
Eriksson
JW
,
Bodegard
J
,
Nathanson
D
,
Thuresson
M
,
Nyström
T
,
Norhammar
A
.
Sulphonylurea compared to DPP-4 inhibitors in combination with metformin carries increased risk of severe hypoglycemia, cardiovascular events, and all-cause mortality
.
Diabetes Res Clin Pract
2016
;
117
:
39
47
37.
Brookhart
MA
,
Rassen
JA
,
Schneeweiss
S
.
Instrumental variable methods in comparative safety and effectiveness research
.
Pharmacoepidemiol Drug Saf
2010
;
19
:
537
554
38.
Morton
JI
,
Marquina
C
,
Shaw
JE
, et al
.
Projecting the incidence and costs of major cardiovascular and kidney complications of type 2 diabetes with widespread SGLT2i and GLP-1 RA use: a cost-effectiveness analysis
.
Diabetologia
2023
;
66
:
642
656
39.
Alemayehu
B
,
Liu
J
,
Rajpathak
S
,
Engel
SS
.
Healthcare resource use and associated costs of hypoglycemia in patients with type 2 diabetes prescribed sulfonylureas
.
J Diabetes Complications
2017
;
31
:
1620
1623
40.
Davies
MJ
,
Aroda
VR
,
Collins
BS
, et al
.
Management of hyperglycemia in type 2 diabetes, 2022. A consensus report by the American Diabetes Association (ADA) and the European Association for the Study of Diabetes (EASD)
.
Diabetes Care
2022
;
45
:
2753
2786
Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. More information is available at https://www.diabetesjournals.org/journals/pages/license.