Glucagon-like peptide 1 receptor agonist discontinuation and risks of major adverse cardiovascular events in adults with type 2 diabetes: target trial emulation

Abstract

Objectives

To characterise patterns of use of glucagon-like peptide 1 receptor agonists (GLP-1RAs) and assess the associations between various GLP-1RA treatment scenarios and the risk of major adverse cardiovascular events.

Design

Target trial emulation.

Setting

Electronic healthcare databases of US Department of Veterans Affairs, 1 January 2017 to 31 December 2023. Data obtained from domains in the Veterans Affairs Corporate Data Warehouse.

Participants

Veterans Affairs users with type 2 diabetes who started treatment with GLP-1RAs (n=132 551) or sulfonylureas (n=201 136), followed up for three years. Veterans Affairs users were defined as having at least two visits to Veterans Affairs and having used the Veterans Affairs outpatient pharmacy within a year before receiving treatment with GLP-1RAs or sulfonylureas.

Interventions

GLP-1RAs or sulfonylureas. In the GLP-1RA arm, treatment status was reassigned every six months, generating 16 prespecified treatment strategies that varied in length of continued use, discontinuation, or interruption.

Main outcome measures

Three year cumulative incidence of major adverse cardiovascular events (myocardial infarction, stroke, or all cause death).

Results

The cohort included 132 551 incident users of GLP-1RAs and 201 136 incident users of sulfonylureas (defined as no prescription for GLP-1RAs or sulfonylureas within a year of the first prescription). A duration dependent association was found between the use of GLP-1RAs and the cumulative three year risk of major adverse cardiovascular events. Compared with the sulfonylurea reference group, participants who used GLP‐1RAs for 0.5, 1, or 1.5 years, before discontinuing for the remainder of the three years, showed incidence risk ratios close to 1.0, with no significant reduction in the risk of major adverse cardiovascular events at three years. Compared with the sulfonylurea group, the reduction in the risk of major adverse cardiovascular events was significant in people who continued to use GLP-1RAs for 2 and 2.5 years, before discontinuing for the remainder of the three years (incidence risk ratio 0.93, 95% confidence interval (CI) 0.88 to 0.98 and 0.85, 0.81 to 0.90, respectively). Participants who continued to use GLP-1RAs for the whole three year follow-up period had the most pronounced risk reduction (incidence risk ratio 0.82, 95% CI 0.78 to 0.85) compared with the sulfonylurea group. Compared with continued use of GLP-1RA, discontinuing treatment for 0.5 years was associated with an increased risk of major adverse cardiovascular events (incidence risk ratio 1.04, 95% CI 1.01 to 1.08); the risk increased progressively with a longer duration of discontinuation, with an incidence risk ratio of 1.14 (1.09 to 1.18) and 1.22 (1.16 to 1.27) for one and two years of discontinuation, respectively. Compared with continued use of GLP-1RA, 0.5 years of interruption was associated with an increased risk of major adverse cardiovascular events; longer durations of interruption were progressively associated with a higher risk of major adverse cardiovascular events, with an incidence risk ratio of 1.12 (95% CI 1.06 to 1.19) and 1.16 (1.11 to 1.22) for one and two years of interrupted use, respectively.

Conclusions

The cardiovascular benefit of GLP-1RAs accumulated with continuous use, but even brief periods of discontinuations or interruptions might progressively erode and could ultimately reverse this protection, increasing the risk of cardiovascular events.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• Use of glucagon-like peptide 1 receptor agonists (GLP-1RAs) is associated with weight loss and a reduced risk of cardiovascular disease

• Discontinuing GLP-1RA treatment is common, and is often followed by weight regain and a rebound in inflammation, both of which are major drivers of the risk of cardiovascular disease

WHAT THIS STUDY ADDS

• A duration dependent association between the use of GLP-1RAs and risk of major adverse cardiovascular events was found, with longer continuous use linked to greater reductions in the three year risk

• Discontinuations diminished the cardiovascular effectiveness of GLP-1RAs, with longer durations of discontinuation associated with greater loss in effectiveness

• With interrupted use of GLP‐1RAs, longer interruptions were associated with a graded reduction in the protective association with the risk of major adverse cardiovascular events

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE, OR POLICY

• Discontinuing and interrupting GLP-1RA treatment may erode and could reverse the cardiovascular benefits in a duration dependent manner, increasing the risk of cardiovascular disease

• Strategies to reduce treatment discontinuity should be evaluated to maximally realise the cardioprotective benefits of GLP-1RA

Introduction

Use of glucagon-like peptide 1 receptor agonists (GLP-1RAs) is associated with several health benefits, including weight loss and reduced risk of major adverse cardiovascular events.¹,²⁰ This pleiotropic effectiveness profile has caused a major increase in use,^{21 22} with recent estimates suggesting that one in eight adults (12%) in the US has reported ever receiving a GLP-1RA.²³ GLP-1RAs are associated with substantial out-of-pocket costs and adverse events, factors that could contribute to high rates of discontinuing the drug.²⁴ Studies have shown discontinuation rates of 36-81% in the first year after starting treatment.²⁵,²⁹ Discontinuation of GLP-1RAs is associated with weight regain and partial or complete reversal of some of the improvements in cardiometabolic risks (eg, blood glucose levels, blood pressure, lipid profile, and inflammatory markers).³⁰ Little is known, however, about the consequences of discontinuing GLP-1RAs on the risk of major adverse cardiovascular events.³¹

In this study, we used the electronic healthcare databases of the US Department of Veterans Affairs to emulate a target trial,^{32 33} enrolling incident users of GLP-1RAs and sulfonylureas. We followed participants for up to three years. For both groups, and each prespecified GLP-1RA treatment strategy (continued use, discontinuation, or interruption), we estimated the risk of major adverse cardiovascular events at one, two, and three years. The primary estimand was the cumulative incidence risk ratios of major adverse cardiovascular events, comparing each GLP-1RA strategy with sulfonylureas in eligible adults with type 2 diabetes from the Veterans Affairs population. The secondary estimand was the incidence risk ratios of major adverse cardiovascular events comparing discontinuation or interruption strategies of GLP-1RAs with continuous use of GLP-1RAs.

Methods

Emulation of target trial

The study was designed and conducted according to the framework of emulation of a target trial.^{32 33} We first specified a protocol of a randomised pragmatic trial that would answer the research question (online supplemental table S1) and then emulated the trial with data from the electronic healthcare databases of the US Department of Veterans Affairs.

The trial protocol specified enrolment of Veterans Affairs users with type 2 diabetes who had not previously used GLP-1RAs or sulfonylureas within the year before enrolment and had no contraindications to these drugs. After enrolment, the protocol stated that participants would be randomly assigned to receive GLP-1RAs or sulfonylureas and have their prescription record and health status recorded during the follow-up period. Participants in the GLP-1RA arm would be further randomised every six months during follow-up, conditional on their current use status. Those who continued to use GLP-1RAs would be re-randomised to continue or discontinue; those who discontinued use of GLP-1RAs would be re-randomised to remain off or resume use of GLP-1RAs. We then emulated the target trial following the methodological approach detailed below.

Setting

The study was conducted with the electronic healthcare databases of Veterans Affairs. Veterans Affairs operates the largest nationally integrated healthcare system in the US, providing care to more than nine million veterans at 1380 healthcare facilities, including 170 medical centres and 1193 outpatient sites. Veterans Affairs provides a comprehensive package of medical benefits to veterans enrolled in the system: inpatient and outpatient services, preventive and primary care, specialty care, geriatric care, mental healthcare, home care, extended care and rehabilitation services, prescription drug treatment coverage, and medical equipment and prosthetics coverage. The Veterans Affairs system provides its users with a comprehensive medical benefits package that includes prescription coverage for GLP-1RAs.

Data sources

We used data from domains in the Veterans Affairs Corporate Data Warehouse: outpatient encounters domain, inpatient encounters domain, outpatient pharmacy domain, laboratory results domain, vital signs domain, patient domain, Veterans Affairs vital status, and health factors domain.³⁴ Medicare data were obtained from the Veterans Affairs Information Resource Center and were used to collect information on care that occurred outside of Veterans Affairs.³⁵ The area deprivation index, a composite measure of education, employment, income and other poverty measures, and housing quality, with a score from 1 to 100, was collected based on participants' residential location.^{36 37}

Cohort construction

Online supplemental figure S1 shows a flowchart of the study design and cohort selection. We identified 620 168 Veterans Affairs users with type 2 diabetes who used GLP-1RAs or sulfonylureas between 1 January 2017 and 31 December 2023. Veterans Affairs users were defined as having at least two Veterans Affairs visits and having used the Veterans Affairs outpatient pharmacy within one year before receiving GLP-1RAs or sulfonylureas. Within this group, we then selected 211 953 participants with incident use of GLP-1RAs (defined as no prescription for GLP-1RAs within one year before the first prescription during the study period). After removing participants who received sulfonylureas within a year before the GLP-1RA prescription, 141 849 participants remained in the incident GLP-1RA group. We separately selected 215 657 participants with incident use of sulfonylureas (defined as no prescription for sulfonylureas within a year before the first prescription). After removing participants who received GLP-1RAs within a year before the prescription for sulfonylureas, 209 614 participants remained in the sulfonylurea group.

We then removed participants with contraindications to GLP-1RAs or sulfonylureas, including those with a history of medullary thyroid carcinoma, multiple endocrine neoplasia type II, gastroparesis, estimated glomerular filtration rate <30 mL/min/1.73 m², dialysis or kidney transplant, hypoglycaemia with coma, hypoglycaemia requiring hospital admission, pancreatitis, and pancreatic cancer, resulting in 132 551 participants in the GLP-1RA group and 201 136 participants in the sulfonylurea group. Time zero (T₀) was defined as the date of the start of treatment. Participants were followed up until the first occurrence of the outcome, three years after T₀, or 31 December 2024.

Interventions

Use of GLP-1RAs and sulfonylureas was defined based on outpatient pharmacy records. Individual GLP-1RAs were albiglutide (which has since been withdrawn from the market) (0.96%), dulaglutide (13.41%), exenatide (0.98%), liraglutide (18.48%), lixisenatide (0.01%), semaglutide (66.13%), and tirzepatide (0.03%). Individual sulfonylureas were glimepiride (13.94%), glipizide (85.72%), and glibenclamide (glyburide) (0.33%) (online supplemental table S2). After the start of treatment with GLP-1RAs, participants were considered to have stopped using GLP-1RAs if they did not receive a refill for at least 90 days after the end of their prescription (defined based on date of release together with the number of days' supply of that prescription). Participants who stopped GLP-1RA treatment were further categorised into two groups: those who stopped use and never resumed use before the end of the follow-up period were classified as the discontinuation group and those who stopped use and subsequently resumed use (received another prescription after the stop date) were classified as the interruption group.

We defined several treatment strategies for participants who started GLP-1RAs by assigning their treatment status every six months, conditional on their current use status (continue or discontinue); individuals who discontinued treatment could be assigned to remain off or resume GLP-1RA treatment (conditional on their use status). The definition yielded 16 prespecified treatment strategies (online supplemental figure S1B). We also conducted primary comparisons between each of the 16 GLP-1RA treatment strategies and starting treatment with sulfonylureas. We conducted analyses on secondary comparisons within the GLP-1RA group comparing continued use of GLP-1RAs and the rest of the 15 GLP-1RA treatment strategies.

Outcomes

The primary outcome was major adverse cardiovascular events, defined as a composite outcome of the three endpoints stroke, myocardial infarction, and all cause mortality.³⁸,⁴⁴ Stroke and myocardial infarction were defined based on ICD-10 (international classification of diseases, 10th revision, clinical modification) codes during any inpatient or outpatient visit, and all cause mortality was defined based on vital status data.

Covariates

We prespecified a set of covariates based on previous knowledge and guidance from a directed acyclic graph (online supplemental figure S2).⁴¹,⁴⁵ Baseline covariates were measured within a year before T₀ unless otherwise specified. For variables with repeated measurements, the measurement closest to and preceding T₀ was selected.

To balance the differences in characteristics between participants starting GLP-1RAs and sulfonylureas, we adjusted for sociodemographic variables, including age, race (white, black, and other), sex, rurality (urban, rural, highly rural, and isolated island), and area deprivation index. Patient data for sex were self-reported. Covariates also included vital measurements (systolic and diastolic blood pressure and body mass index) and laboratory measurements (estimated glomerular filtration rate, albuminuria, and low density lipoprotein).

We adjusted for diabetes mellitus status including levels of haemoglobin A_1c (HbA_1c) at baseline, average HbA_1c level within a year before T₀, and separately within five years before T₀, duration of diabetes from October 1999 until T₀, use of metformin, sulfonylureas, thiazolidinediones, dipeptidyl peptidase 4 inhibitors, sodium-glucose co-transporter 2 inhibitors, and insulin, as well as length of use of metformin, sulfonylureas, thiazolidinediones, dipeptidyl peptidase 4 inhibitors, sodium-glucose co-transporter 2 inhibitors, and insulin within five years before T₀.

We adjusted for conditions that could influence the probability of receiving GLP-1RAs, including a comprehensive list of circulatory conditions, such as stroke, transient ischaemic attack, atrial fibrillation, tachycardia, bradycardia, ventricular arrhythmias, atrial flutter, pericarditis, myocarditis, acute coronary disease, myocardial infarction, ischaemic cardiomyopathy, angina, heart failure, non-ischaemic cardiomyopathy, cardiac arrest, cardiogenic shock, pulmonary embolism, deep vein thrombosis, and venous thrombotic embolism. We also adjusted for the gastrointestinal conditions nausea, diarrhoea, abdominal pain, dyspepsia, constipation, gastro-oesophageal reflux disease, gastroparesis, gastritis, pancreatitis, non-alcoholic fatty liver disease, ulcerative colitis, intestinal obstruction and ileus, diverticulosis and diverticulitis, haemorrhoids, anal and rectal conditions, peritonitis and intra-abdominal abscess, biliary tract disease, hepatic failure, gastrointestinal haemorrhage, non-infectious gastroenteritis, non-infectious hepatitis, postprocedural digestive system complications, and gastrointestinal cancer.

We adjusted for nutritional deficiencies, bulimia, bariatric surgery, acute kidney injury, urinary tract infections, fatigue, suicidal ideation, thyroid disorders (including thyroid cancer), fluid and electrolyte disorders, pituitary disorders, hypoglycaemia, diabetic ketoacidosis, postprocedural complications, other cancers, HIV, and alcohol use disorder.

Health behaviours and use of healthcare services were also adjusted for: smoking status, influenza vaccination status, use of long term care, number of outpatient visits, number of hospital admissions, number of prescriptions, number of blood panel tests, number of HbA_1c measurements, number of outpatient visits and hospital admissions from Medicare, and the calendar week of treatment assignment. We adjusted for the patient’s history of drug treatment use, including the use of statins, angiotensin converting enzyme inhibitors, angiotensin receptor blockers, β blockers, diuretics, calcium channel blockers, bupropion, naltrexone, orlistat, phentermine, and topiramate. Missing values for HbA_1c, estimated glomerular filtration rate, low density lipoprotein, blood pressure, and body mass index were recorded for 2.86%, 4.93%, 7.85%, 2.36%, and 6.18% of participants, respectively. We applied multivariate imputation by chained equations to assign values based on the predictive mean matching method.⁴⁶ Continuous variables were adjusted in the form of restricted cubic splines with knots at 5th, 35th, 65th, and 95th centiles.⁴⁷

To account for changes in patient status during follow-up that could influence the probability of stopping treatment and the probability of treatment resumption after stopping treatment, we assessed covariates every 90 days during follow-up. In estimating the probability of stopping treatment, we included baseline and time updated covariates in the GLP-1RA group that could be associated with the probability of stopping treatment, such as the specialist providing the prescription (primary care, endocrinology, cardiology, weight management, and other), and the types and doses of GLP-1RAs. We also adjusted for length of time treatment was stopped, and latest type and dose of GLP-1RAs used before stopping treatment, in estimating the probability of resumption. Time updated covariates were used along with baseline covariates to balance differences in characteristics between participants who stopped treatment (discontinued and interrupted use) and those who continued GLP-1RA treatment, and separately between those who resumed use and those who did not, among those who stopped GLP-1RA treatment.

Statistical analyses

Baseline characteristics of the sulfonylurea group, overall GLP-1RA group, and those with continued, discontinued, and interrupted use of GLP-1RAs, are presented. We used absolute standardised differences to evaluate differences in baseline characteristics between the sulfonylurea and GLP-1RA groups. Within the overall GLP-1RA group, absolute standardised differences were used to evaluate differences in baseline and time updated characteristics between those who continued use versus those who discontinued or interrupted use at each time point. For those who discontinued or interrupted GLP-1RA treatment, we evaluated absolute standardised differences in baseline and time updated characteristics between those who resumed use versus those who did not resume use at each time point. An absolute standardised difference value <0.1 was considered evidence of good balance.⁴⁸

We then constructed a series of baseline and time varying weights designed to estimate effects in a weighted population where, after weighting, the likelihood of starting sulfonylureas versus GLP-1RAs was not dependent on baseline characteristics, and the likelihood of discontinuation or interruption among GLP-1RA users was not dependent on baseline or time updated characteristics. Online supplemental figure S3 presents the analytical flowchart.

To balance differences in baseline characteristics between the sulfonylurea and GLP-1RA groups, we constructed baseline inverse probability weighting based on the propensity score. The probability of a participant being assigned to the GLP-1RA group (the propensity score) was estimated based on logistic regression. The baseline inverse probability weights were then constructed as propensity score/(1−propensity score) for the sulfonylurea group and truncated at the 1st and 99th centiles, and as 1 for the GLP-1RA group. During follow-up, participants in the sulfonylurea group were censored at the time they received GLP-1RAs and participants in the GLP-1RA group were censored at the time they received sulfonylureas. Stabilised inverse probability of censoring weights after truncation at the 99.5th centile was applied to account for the informative censors.⁴⁹ The summary weights from baseline weights and inverse probability of censoring weights were applied to the cohort for further estimation of inverse probability of treatment weights and inverse probability of resumed use weights, detailed below.

For the GLP-1RA group, during the continued use period, we further balanced differences in baseline and time updated characteristics between those who continued treatment and those who stopped treatment (discontinued or interrupted use) with marginal structural models. In contrast with direct adjustment of time updated characteristics, the marginal structural model approach retains the causal pathways from previous interventions.^{49 50} Pooled logistic regression was used to estimate the probability of treatment (continued use of GLP-1RAs) at intervals of 90 days. The time dependent probability was estimated from logistic regressions $logit\mathrm{Pr}\left(Z_k=1|\mathrm{V},{\stackrel{-}{Q}}_{k-1}\right)={\beta }_{0k}+{\beta }_{1k}V+{\beta }_{2k}\mathrm{}{Q}_{k-1}$

where Z is an indicator of treatment, V is a vector of covariates at baseline, and $Q_{k-1}$ is a vector of covariates at time point k−1.

The time varying inverse probability of treatment weighting was then computed as 1 /P(treatment) for those who continued use, and as 1/(1−P(treatment)) for those who discontinued or interrupted treatment at the time of discontinuation or interruption. These weights were stabilised with the marginal probability of treatment at each time point. We then constructed the cumulative inverse probability of treatment weights as the product of weights from the start of treatment until each time point to reflect the continued treatment history since the start of treatment. The stabilised inverse probability weights for treatment were estimated as:

$S{W}_k^Z=\prod _{k=1}^t\frac{\mathrm{P}\mathrm{r}(Z_k=1|{\stackrel{-}{\mathrm{Z}}}_{k-1}=1)}{\mathrm{P}\mathrm{r}(Z_k=1|{\stackrel{-}{\mathrm{Z}}}_{k-1}=1,V{\stackrel{-}{,Q}}_{k-1})},$

where Z is an indicator of treatment, V is a vector of covariates at baseline, and ${\stackrel{-}{Q}}_{k-1}\mathrm{}$ is a vector of covariates at time k−1. For the duration of non-use, the cumulative inverse probability of treatment weights was equal to the weights at the date of discontinuation or interruption.

For GLP-1RA participants, during their non-use period, we estimated the probability of resumed use in pooled logistic regression as $logit\mathrm{Pr}\left(R_k=1|\mathrm{V},{\stackrel{-}{Q}}_{k-1}\right)={\beta }_{0k}+{\beta }_{1k}V+{\beta }_{2k}\mathrm{}{Q}_{k-1}$

where R is an indicator of resumed use, V is a vector of covariates at baseline, and $Q_{k-1}$ is a vector of covariates at time point k−1, and constructed the cumulative inverse probability of resumed use weighting following similar methodology. The stabilised inverse probability weights for resumed use were estimated as:

$S{W}_k^R=\prod _{k=1}^t\frac{\mathrm{P}\mathrm{r}(R_k=1|{\stackrel{-}{\mathrm{R}}}_{k-1}=0)}{\mathrm{P}\mathrm{r}(R_k=1|{\stackrel{-}{\mathrm{R}}}_{k-1}=0,V{\stackrel{-}{,Q}}_{k-1})},$

where R is an indicator of resumed use, V is a vector of covariates at baseline, and ${\stackrel{-}{Q}}_{k-1}\mathrm{}$ is a vector of covariates at time k−1.

For these participants, the final weights were constructed as the product of the cumulative inverse probability of treatment and resumed use weights, to reflect the treatment history, including periods of use and non-use of GLP-1RAs since the start of treatment. Weights for all participants at different time points were truncated at the 99.5th centile. Online supplemental table S3 presents the distribution of the time varying inverse probability weights.

We then evaluated the relative risk and event rate of several treatment scenarios among those who discontinued or interrupted treatment with GLP-1RAs by varying the durations of use and non-use. The scenarios of discontinued use were use for 0.5, 1, 1.5, 2, and 2.5 years followed by permanent discontinuation. The scenarios of interrupted use were initial use for 0.5 years, followed by interrupted use for 0.5, 1, 1.5, and 2 years, then resumption of use; initial use for one year, followed by interrupted use for 0.5, 1, and 1.5 years, then resumption of use; initial use for 1.5 years, followed by interrupted use for 0.5 and 1 years, then resumption of use; and initial use for two years, followed by interrupted use for 0.5 years before resumption of use.

To estimate the relative risk and event rate of major adverse cardiovascular events in different treatment scenarios, we first modelled the instantaneous risk in the sulfonylurea group, and in the GLP-1RA group during periods of use, non-use, and resumed use with weighted pooled logistic regression for discrete time survival analyses. For GLP-1RA users, the time varying risk during the initial use period (full period for those who continued use and the period before non-use for those who discontinued or interrupted treatment) was modelled conditional on the number of days since the start of treatment. Time varying risk during the non-use period was modelled conditional on both the duration of previous use before stopping treatment and the number of days during the non-use period (number of days since stopping treatment). Time varying risk during the resumed use period was modelled conditional on both the duration of the preceding non-use before resumption and the number of days during the resumed use period. All time variables were incorporated with restricted cubic spline functions. Interaction terms between treatment status and time variables were included in the model.

The cumulative event rates at one, two, and three years for the sulfonylurea and GLP-1RA groups after different treatment scenarios were then estimated based on the baseline risk and time varying instant risk. We also estimated incidence risk ratios and absolute rate differences per 100 persons at one, two, and three years between different GLP-1RA treatment scenarios and the sulfonylurea group, and also between the different GLP-1RA treatment scenarios and continued use of GLP-1RA.

We evaluated the association between the proportion of days covered, calculated as the cumulative days of use of GLP-1RAs over the follow-up period, and the risk of major adverse cardiovascular events. For each participant, the adjusted cumulative event rates at three years were computed. We then estimated the association between different proportions of days covered and the event rates based on locally estimated scatterplot smoothing. Those with proportion of days covered of ≤2.5% (use of GLP-1RAs for <28 days in three years of follow-up) were the reference group. To evaluate the association between GLP-1RA treatment scenarios and major adverse cardiovascular events in populations with different body mass index values, we conducted subgroup analyses for participants with a body mass index ≤30, >30, and >35.

We conducted the following sensitivity analyses to examine the robustness of the study findings. Because indications and use of GLP-1RAs have changed over time, we conducted analyses in participants enrolled after 4 June 2021 (the date that the US Food and Drug Administration approved semaglutide for weight management). We removed participants with events occurring within the first 180 days of the start of treatment. We only selected participants who continued to use GLP-1RAs or sulfonylureas for at least 180 days after the start of treatment. We compared GLP-1RA treatment strategies with continued use of sulfonylureas by censoring participants in the sulfonylurea group when they stopped using sulfonylureas and applied inverse probability of censoring weight to correct for informative censoring. We removed participants with a history of myocardial infarction or stroke before the start of treatment. We included only participants who started semaglutide treatment (the most commonly used GLP-1RA) from the GLP-1RA group. We censored GLP-1RA users who had further interruption or discontinuation during the resumed use period. We applied the clone-censor-weight approach to estimate the per protocol effect of following each prespecified treatment strategy. Specifically, the GLP-1RA group at T₀ was cloned into 16 copies, corresponding to the 16 treatment strategies. For each strategy, participants were artificially censored at the 90 day interval in which their observed treatment pattern deviated from the assigned strategy. For example, under the strategy of continued GLP-1RAs for one year, then discontinued, a participant's clone was censored at the time of discontinuation or interruption if they had discontinuation or interruption before one year, at one year if they remained on treatment beyond one year, or at the time of resumption if they resumed treatment after one year. At each time interval, we estimated the inverse probability of censoring weight by modelling the probability of remaining uncensored, conditional on baseline and time varying covariates. The cumulative inverse probability of censoring weight for each individual at time T was computed as the product of the probabilities of remaining uncensored from baseline to time T. These weights were then applied in the outcome models to adjust for the informative censoring induced by enforcing adherence to the prespecified strategies.^{32 33} We also adjusted for frailty score composed from 6422 codes representing 31 health deficits.^{51 52} We applied a range of upper (and corresponding lower) truncation cut-off values to the time varying inverse probability of treatment weights and, separately, the inverse probability of resumption weights.⁵³ We applied trimming and varied their cut-off thresholds (instead of truncation which was used in the primary analyses) to the time varying inverse probability of treatment weights and, separately, the inverse probability of resumed use weights.⁵³

Confidence intervals (CIs) of the results were estimated based on parametric bootstrapping. Point estimates and 95% CIs are reported, where a 95% CI of ratio not including one or rate not including 0 was considered significant. Analyses were conducted with Statistical Analysis System Enterprise Guide 8.3, and data visualisations were produced using R version 4.3.3.

Patient and public involvement

We did not directly involve patients or the public in developing the research question, or in the study design or implementation owing to a limited timeline. We have no specific plans to disseminate the findings to study participants. The results of the study, however, will be disseminated by press release with mainstream media and will also be promoted on X (formerly known as Twitter, from the corresponding senior author’s account @zalaly). The results will be shared with advocacy groups and will be presented at scientific meetings and at academic institutions.

Results

The cohort included 132 551 incident users of GLP-1RAs and 201 136 incident users of sulfonylureas. Participants were followed up for a median of 3.00 (interquartile range (IQR) 2.03-3.00) years in the GLP-1RA group and 3.00 (3.00-3.00) years in the sulfonylurea group, corresponding to an overall median follow-up time of 3.00 (2.95-3.00) years and total follow-up of 914 898 person years. Table 1 and online supplemental tables S4 and S5 show the baseline personal characteristics before and after weighting. After weighting, the GLP-1RA group had a similar age, HbA_1c level, and body mass index distribution to the sulfonylurea group (mean age 65.87 (standard deviation 10.58) years v 65.97 (10.91) years; HbA_1c level 8.57% (1.63%) v 8.58% (1.59%); and body mass index 35.43 (6.95) v 35.31 (8.31)). Evaluation of baseline and time updated standardised mean differences suggested that all of the covariates were balanced after weighting (online supplemental figures S4–S6).

Table 1. Selected baseline characteristics of the overall glucagon-like peptide 1 receptor agonist (GLP-1RA) group, and in the subgroups of continued use, discontinued use, interrupted use of GLP-1RAs, compared with the sulfonylurea group, after weighting.

Baseline characteristics	GLP-1RA group				Sulfonylurea group (n=201 136)	Absolute standardised mean difference (GLP-1RAs v sulfonylureas)
	Overall (n=132 551)	Continued use (n=67 553)	Discontinued use (n=34 935)	Interrupted use (n=30 063)
Mean (SD) age (years)	65.87 (10.58)	66.10 (10.29)	66.75 (10.81)	64.31 (10.77)	65.97 (10.91)	0.009
Sex
Men	121 483 (91.65)	62 005 (91.79)	32 099 (91.88)	27 379 (91.07)	181 666 (90.32)	0.05
Women	11 068 (8.35)	5548 (8.21)	2836 (8.12)	2684 (8.93)	19 470 (9.68)	0.05
Race:*
White	100 832 (76.07)	51 963 (76.92)	27 425 (78.50)	21 444 (71.33)	152 440 (75.79)	0.007
Black	26 475 (19.97)	12 998 (19.24)	6210 (17.78)	7267 (24.17)	40 431 (20.10)	0.003
Other	5244 (3.96)	2592 (3.84)	1300 (3.72)	1352 (4.50)	8265 (4.11)	0.008
Mean (SD) area deprivation index†	56.47 (18.44)	56.54 (18.47)	56.49 (18.19)	56.29 (18.65)	56.83 (18.68)	0.02
Mean (SD) body mass index	35.43 (6.95)	35.86 (6.74)	34.57 (7.21)	35.48 (7.02)	35.31 (8.31)	0.02
Mean (SD) estimated glomerular filtration rate (mL/min/1.73m2)	73.44 (21.46)	73.33 (21.27)	72.57 (21.55)	74.68 (21.72)	73.83 (21.34)	0.02
Mean (SD) low density lipoprotein (mg/dL)	81.99 (34.19)	79.36 (32.61)	83.45 (35.03)	86.22 (36.09)	82.16 (33.78)	0.005
Mean (SD) HbA1c (%)	8.57 (1.63)	8.38 (1.52)	8.64 (1.68)	8.91 (1.76)	8.58 (1.59)	0.005
Mean (SD) No of HbA1c measurements	2.59 (1.26)	2.58 (1.24)	2.59 (1.26)	2.60 (1.29)	2.55 (1.22)	0.03
Metformin	88 354 (66.66)	45 933 (68.00)	22 195 (63.53)	20 226 (67.28)	130 148 (64.71)	0.04
Insulin	92 345 (69.67)	44 896 (66.46)	25 155 (72.01)	22 294 (74.16)	139 352 (69.28)	0.008
DPP4i	27 409 (20.68)	14 992 (22.19)	6792 (19.44)	5625 (18.71)	49 407 (24.56)	0.09
Thiazolidinediones	7277 (5.49)	3778 (5.59)	1829 (5.24)	1670 (5.56)	13 414 (6.67)	0.05
SGLT2i	53 662 (40.48)	30 095 (44.55)	12 533 (35.88)	11 034 (36.70)	79 309 (39.43)	0.02
Angiotensin converting enzyme inhibitors or angiotensin receptor blockers	85 296 (64.35)	43 468 (64.35)	22 331 (63.92)	19 497 (64.85)	125 848 (62.57)	0.04
Calcium channel blockers	41 802 (31.54)	21 374 (31.64)	11 013 (31.52)	9415 (31.32)	61 715 (30.68)	0.02
β blockers	62 833 (47.40)	31 773 (47.03)	16 920 (48.43)	14 140 (47.03)	93 685 (46.58)	0.02
Diuretics	59 450 (44.85)	30 908 (45.75)	15 376 (44.01)	13 166 (43.79)	86 459 (42.99)	0.04
Statins	108 677 (81.99)	55 765 (82.55)	28 244 (80.85)	24 668 (82.05)	160 743 (79.92)	0.05
Stroke	4616 (3.48)	1987 (2.94)	1475 (4.22)	1154 (3.84)	7167 (3.56)	0.004
Transient ischaemic attack	1798 (1.36)	785 (1.16)	574 (1.64)	439 (1.46)	2824 (1.40)	0.004
Acute coronary disease	4686 (3.54)	2189 (3.24)	1413 (4.04)	1084 (3.61)	7065 (3.51)	0.001
Myocardial infarction	2113 (1.59)	941 (1.39)	644 (1.84)	528 (1.76)	3450 (1.72)	0.01
Ischaemic cardiomyopathy	2673 (2.02)	1291 (1.91)	781 (2.24)	601 (2.00)	3867 (1.92)	0.007
Angina	2954 (2.23)	1393 (2.06)	823 (2.36)	738 (2.45)	4413 (2.19)	0.002
Heart failure	15 843 (11.95)	7813 (11.57)	4446 (12.73)	3584 (11.92)	23 319 (11.59)	0.01
Non-ischaemic cardiomyopathy	4262 (3.22)	2093 (3.10)	1173 (3.36)	996 (3.31)	6211 (3.09)	0.007

Data are number (%) unless indicated otherwise.

Absolute standardised mean difference <0.1 was considered evidence of good balance.

^*Self-reported race information was collected from electronic health records and used in the study in accordance with the requirement by the funding agency (US Department of Veterans Affairs) and the Office of Management and Budget, which defines minimum standards for maintaining, collecting, and presenting data on race and ethnicity for all federal reporting agencies. Other category includes American Indian and Alaska Native, Asian or Native Hawaiian, and other Pacific Islander.

^†A measure of socioeconomic disadvantage, range 0-100, with higher scores indicating greater disadvantage.

DPP4i, dipeptidyl peptidase 4 inhibitor; HbA_1c, haemoglobin A_1c; SD, standard deviation; SGLT2i, sodium-glucose co-transporter 2 inhibitor.

Rates of discontinuation and interruption

Among incident users of GLP-1RAs, 34 935 (26.36%) discontinued GLP-1RA treatment during the follow-up period; 22 425 (64.19%), 7493 (21.45%), and 5017 (14.36%) participants discontinued GLP-1RA treatment during the first, second, and third year of follow-up, respectively. Median duration of use of GLP-1RAs before discontinuation was 0.61 (IQR 0.20-1.40) years, with 30 063 (22.68%) incident users of GLP-1RAs having interruptions; 20 268 (67.42%), 7381 (24.55%), and 2414 (8.03%) of the interruptions occurred during the first, second, and third years of follow-up, respectively. Median duration of use of GLP-1RAs before interruption was 0.60 (IQR 0.23-1.24) years and median duration of interruption before resumption was 0.40 (IQR 0.30-0.65) years (figure 1 and figure 2).

Figure 1. Patterns of use of glucagon-like peptide 1 receptor agonists (GLP-1RAs). Proportion of participants with continued use, discontinued use, and interrupted use of GLP-1RAs during the three year follow-up period. Proportion during the three year follow-up was estimated based on the Kaplan-Meier estimator. Discontinuation was defined as stopping use without resumption; interruption was defined as stopping use for at least 90 days followed by resumption during follow-up.

Figure 2. Distribution of duration of use of glucagon-like peptide 1 receptor agonists (GLP-1RAs) before discontinuation in participants who discontinued use (top panel), before interruption in participants who interrupted use (middle panel), and interruption among participants who interrupted use (bottom panel). Discontinuation was defined as stopping use without resumption; interruption was defined as stopping use for at least 90 days followed by resumption during follow-up. Mean and median of the distribution are reported. SD=standard deviation; IQR=interquartile range.

Risk of major adverse cardiovascular events after discontinuation or interruption of GLP-1RAs

Based on the framework of emulation of a target trial, we used marginal structural models to examine the association between GLP-1RAs and risk of major adverse cardiovascular events. This model was adjusted for baseline characteristics and time varying covariates by inverse probability weighting. Compared with incident use of sulfonylureas, incident use of GLP-1RAs was associated with a reduced risk of major adverse cardiovascular events (incidence risk ratio 0.87, 95% CI 0.84 to 0.90). Compared with incident use of sulfonylureas, incident GLP-1RA users who continued use throughout the follow-up period had an incidence risk ratio of 0.82 (95% CI 0.78 to 0.85), those whose treatment was interrupted and who subsequently resumed treatment had an incidence risk ratio of 0.88 (0.84 to 0.92), and those who discontinued use and did not resume treatment had an incidence risk ratio of 0.96 (0.92 to 1.01) (table 2 and online supplemental figure S7).

Table 2. Rate of major adverse cardiovascular events at three years in participants with continued use, discontinued use, or interrupted use of glucagon-like peptide 1 receptor agonists (GLP-1RAs), compared with sulfonylurea group.

GLP-1RA use	Rate per 100 persons (95% CI)		Incidence risk ratio (95% CI)	Absolute rate difference (95% CI)
	Sulfonylurea group	GLP-1RA group
Continued use	20.65 (19.93 to 21.40)	16.89 (16.57 to 17.26)	0.82 (0.78 to 0.85)	3.76 (3.00 to 4.63)
Discontinued use	20.65 (19.93 to 21.40)	19.92 (19.46 to 20.40)	0.96 (0.92 to 1.01)	0.75 (−0.12 to 1.60)
Interrupted use	20.65 (19.93 to 21.40)	18.25 (17.72 to 18.75)	0.88 (0.84 to 0.92)	2.41 (1.57 to 3.34)

Model adjusted for baseline characteristics and time updated characteristics by baseline inverse probability weights, inverse probability of treatment adherence weights, and inverse probability of resumed use weights, and estimated by marginal structural models. Event rate for the discontinued use and interrupted use groups were estimated based on the distribution of durations of use and non-use in these groups. Discontinuation was defined as stopping use without resumption; interruption was defined as stopping use for at least 90 days followed by resumption during follow-up. Risk in discontinued use and interrupted use categories estimated based on the marginal distribution of duration of use and non-use within the categories. Absolute rate difference presented as the rate reduction per 100 persons in GLP-1RA group compared with sulfonylurea group.

We then evaluated the cumulative three year rate of major adverse cardiovascular events among individuals who started treatment with sulfonylureas (reference group) compared with those who started GLP-1RA treatment, according to the different treatment scenarios with various durations of continued, interrupted, and discontinued use over three years (figure 3). The results showed a duration dependent association between GLP-1RA use and the risk of major adverse cardiovascular events. A longer duration of GLP-1RA use was associated with a greater reduction in the risk of major adverse cardiovascular events relative to the sulfonylurea reference group. Participants who used GLP‐1RAs for 0.5, 1, or 1.5 years (before discontinuing for the remainder of the three years) had incidence risk ratios close to 1.0, with no significant reduction in the risk of major adverse cardiovascular events at three years. The reduction in the risk of major adverse cardiovascular events was significant in people who continued to use GLP-1RAs for 2 and 2.5 years (incidence risk ratio 0.93, 95% CI 0.88 to 0.98 and 0.85, 0.81 to 0.90; absolute rate difference 1.48, 95% CI 0.42 to 2.53 and 3.07, 2.05 to 4.02, respectively). The reduction in risk of major adverse cardiovascular events was highest among people who had continued use of GLP-1RAs for three years (ie, had no period of non-use; incidence risk ratio 0.82, 95% CI 0.78 to 0.85 and absolute rate difference 3.76, 95% CI 3.00 to 4.63) (figure 3 and online supplemental table S6). Among individuals with interrupted use of GLP‐1RAs, longer interruptions were associated with a reduced protective association on risk of major adverse cardiovascular events (figure 3 and online supplemental table S6).

Figure 3. Incidence risk ratios and absolute rate differences of major adverse cardiovascular events at three years for different glucagon-like peptide 1 receptor agonist (GLP-1RA) treatment scenarios compared with sulfonylureas. Scenarios of continued and discontinued use of GLP-1RAs (A) and interrupted use of GLP-1RAs (B), indicating periods (in years) of GLP-1RA use and non-use. C=Incidence risk ratios, comparing cumulative incidence rates at three years in each treatment scenario with that of sulfonylureas. D=Absolute rate differences per 100 persons at three years between each treatment scenario and sulfonylureas, where a positive value indicates a rate reduction associated with use of GLP-1RAs. Discontinuation was defined as stopping use without resumption; interruption was defined as stopping use for at least 90 days followed by resumption during follow-up. CI=confidence interval.

We then conducted analyses to examine the three year cumulative risk of major adverse cardiovascular events among people with various scenarios of GLP-1RA treatment discontinuation or interruption versus those who continued to use GLP-1RAs throughout the three years of follow-up (figure 4). The results suggested that compared with continued use, even discontinuing treatment for 0.5 years was associated with an increased risk of major adverse cardiovascular events (incidence risk ratio 1.04, 95% CI 1.01 to 1.08); longer durations of discontinuation were progressively associated with a higher risk of major adverse cardiovascular events with an incidence risk ratio of 1.14 (95% CI 1.09 to 1.18) and 1.22 (1.16 to 1.27) for one and two years of discontinuation, respectively. Similarly, compared with continued use, interruption of GLP-1RA treatment for 0.5 years was associated with an increased risk of major adverse cardiovascular events (incidence risk ratio 1.07, 95% CI 1.03 to 1.12; 1.08, 1.04 to 1.12; 1.08, 1.05 to 1.11; and 1.06, 1.04 to 1.09 for 0.5 years of interruption after initial use for 0.5, 1, 1.5, and 2 years, respectively). The greater the duration of interruption, the higher the risk of major adverse cardiovascular events, with an incidence risk ratio of 1.12 (95% CI 1.06 to 1.19) and 1.16 (1.11 to 1.22) for one and two years of interrupted use, respectively (figure 4 and online supplemental table S7).

Figure 4. Incidence risk ratios and absolute rate differences of major adverse cardiovascular events at three years for different glucagon-like peptide 1 receptor agonist (GLP-1RA) treatment scenarios compared with continued use of GLP-1RAs. Scenarios of discontinued use of GLP-1RAs (A) and interrupted use of GLP-1RAs (B), indicating periods (in years) of GLP-1RA use and non-use. C=Incidence risk ratios, comparing cumulative incidence rates at three years in each treatment scenario with continued use of GLP-1RA. D=Absolute rate differences per 100 persons at three years between each treatment scenario and continued use of GLP-1RA, where a negative value indicates a rate increase associated with discontinued or interrupted GLP-1RA treatment scenarios. Discontinuation was defined as stopping use without resumption; interruption was defined as stopping use for at least 90 days followed by resumption during follow-up. CI=confidence interval.

Analyses of the cumulative rate of major adverse cardiovascular events at one year and two years after the start of treatment also showed that longer duration of GLP-1RA use before discontinuation was associated with a greater reduction in the risk of major adverse cardiovascular events. Longer interruption was associated with a reduced protective association with major adverse cardiovascular events outcome (online supplemental figures S8 and S9 and online supplemental tables S8 and S9).

Proportion of days covered and risk of major adverse cardiovascular events

We evaluated the association between proportion of days covered (estimated as the number of days of prescription of GLP-1RAs divided by the number of days of follow-up) and the risk of major adverse cardiovascular events in GLP-1RA users. During the follow-up period, GLP-1RA users had a mean of 62.49% (standard deviation 28.92%; median 72.77%, IQR 40.94-86.03%) for proportion of days covered. As proportion of days covered increased, the risk of major adverse cardiovascular events declined (figure 5).

Figure 5. Proportion of days covered and risk of major adverse cardiovascular events. Association between proportion of days covered and risk of major adverse cardiovascular events based on cumulative incidence rates at three years. Proportion of days covered was defined as cumulative days of use of glucagon-like peptide 1 receptor agonists (GLP-1RAs) over the follow-up period, with a proportion of days covered of ≤2.5% used as the reference group. Top panel=continuous association depicted with locally estimated scatterplot smoothing. Bottom panel=risk of major adverse cardiovascular events for intervals of proportion of days covered between 10% and 100% in increments of 10%. CI=confidence interval.

Discontinuation and interruption by body mass index

Analyses of cohort participants with a body mass index >30, and separately with a body mass index >35 at the start of treatment, were consistent with the main analyses. Analyses in populations with diabetes and a body mass index ≤30 at the start of treatment also showed consistent results (online supplemental figures S10–S12 and online supplemental tables S10–S12).

Sensitivity analyses

We conducted multiple sensitivity analyses to examine the robustness of the study findings. Analyses conducted were: participants enrolled after 4 June 2021 (the date the US FDA approved semaglutide for weight management); removed participants with events occurring within the first 180 days of the start of treatment; selected participants who continued to use GLP-1RAs or sulfonylureas for at least 180 days after the start of treatment; censored participants in the sulfonylurea group at the time they stopped using sulfonylureas (inverse probability of censoring weight was applied to correct for informative censoring); removed participants with a history of myocardial infarction or stroke before the start of treatment; selected participants who started semaglutide treatment (the most commonly used GLP-1RA) from the GLP-1RA group; and censored GLP-1RA users who had further interruption or discontinuation during the resumed use period. We applied the clone method to estimate the effect of following each treatment strategy; participants were cloned across all strategies; and clones were then administratively censored at the first deviation from their assigned regimen and inverse probability of censoring weights were applied to adjust for the informative censoring introduced by these violations. We also adjusted for frailty score obtained from 6422 codes representing 31 health deficits. We applied a range of upper (and corresponding lower) truncation cut-off values to the time varying inverse probability of treatment weights and, separately, the inverse probability of resumed use weights. We applied trimming and varied their cut-off thresholds (instead of truncation which was used in the primary analyses) to the time varying inverse probability of treatment weights and, separately, the inverse probability of resumed use weights. The results of the sensitivity analyses were consistent with the main results (online supplemental table S13).

Discussion

Principal findings

In this study of 132 551 incident GLP-1RA users and 201 136 sulfonylurea users, followed for a median of three years, we showed that GLP-1RA discontinuations (26.36%) and interruptions (22.68%) were common and occurred most often in the first year of starting treatment. Our study extends the previous literature by quantifying how various durations of discontinuations and interruptions were associated with the outcome of major adverse cardiovascular events. Our analyses against two comparators highlighted graded, duration dependent consequences of discontinuation and interruption of treatment with GLP-1RAs. Compared with sulfonylureas, with continuous use of GLP-1RAs over three years, we found the greatest reduction in the risk of major adverse cardiovascular events, whereas (dependent on their duration) discontinuations or interruptions diminished or rendered null the cardiovascular effectiveness of GLP-1RAs. Compared with continued use of GLP-1RAs, discontinuation and interruption of use of GLP-1RAs increased the risk of cardiovascular disease. Our study provides insights into the real world patterns of use of GLP-1RAs and the impact of discontinuing and interrupting treatment on cardiovascular outcomes. Specifically, the study highlighted that the cardioprotective properties of GLP-1RAs were dependent on the duration and continuity of treatment and that even brief discontinuations or interruptions may progressively erode, and could ultimately reverse, this protection, increasing the risk of cardiovascular events.

Our results emphasise the importance of maintaining continuity of treatment, especially in people at high risk of cardiovascular events. Given that discontinuation may be driven at least partially by patients' perceived lack of therapeutic effectiveness and emergence of adverse events, approaches to improve awareness of the substantial long term cardiovascular benefits of remaining on GLP-1RAs and proactive management of side effects (eg, gradual dose titration to reduce gastrointestinal symptoms) may reduce interruptions and discontinuations.^{24 28 54} Because out-of-pocket drug costs are major reasons for stopping GLP-1RA treatment in the US,⁵⁵ reducing these financial barriers may help reduce the risk of interruptions and discontinuations.⁵⁶

The mechanisms underlying the association between discontinuation of GLP-1RAs and risk of major adverse cardiovascular events are not clear. Preclinical studies have shown that GLP-1RAs may improve endothelial function and reduce oxidative stress, platelet aggregation, and thrombus formation.⁵⁷,⁶³ Clinical studies have shown that GLP-1RAs cause weight loss, improved glycaemic control, a reduction in blood pressure, favourable lipid profile changes, and anti-inflammatory effects,^{43 44 64} all of which may lead to improved cardiovascular health.^{65 66} On discontinuation, these improvements may partially or fully reverse, as evidenced by weight regain and worsening cardiometabolic risk factors after stopping treatment,³⁰ likely explaining the increased risk of major adverse cardiovascular events after interruptions and discontinuations. Whether the observed association is mediated by weight regain, deterioration of blood sugar control, inflammatory response, or other unknown factors needs to be confirmed in mechanistic and clinical studies to better understand the consequences of discontinuing treatment with GLP-1RAs.

Strengths and limitations of this study

The study had several limitations. The study was conducted in US veterans who are older and mostly men, which may not represent the general population and could limit the generalisability of our results. Although we adjusted for a comprehensive set of covariates from multiple data domains by inverse probability weighting, we cannot completely rule out the possibility of residual confounding. We relied on the electronic health records of Veterans Affairs to assign interventions, covariates, and outcomes. Although Veterans Affairs is known for its high quality data, the possibility of misclassification bias cannot be excluded. We examined the associations by drug class but we did not examine differences within classes (ie, whether different GLP-1RAs had different effects). More than 99.22% of the GLP-1RAs in this study were injectables and we did not examine differences between routes of administration (eg, injectable or oral). We also did not examine the risks by indications for GLP-1RAs. We relied on prescription records and pharmacy fills to define drug treatment use status. This approach is a widely accepted measure of drug treatment use but is a proxy for actual drug taking behaviour and discontinuation rates. The study protocol was not pre-registered.

The study had several strengths. The Veterans Affairs comprehensive medical benefits package that includes prescription coverage for GLP-1RAs reduced the possibility that starting, interrupting, or discontinuing use was mainly driven by financial considerations. To estimate the association between discontinuation and risk of major adverse cardiovascular events and reduce the influence of factors that led to discontinuation, we accounted for a comprehensive number of baseline and time updated variables from multiple data dimensions to adjust for factors that may be associated with discontinuation or interruption. Unlike direct time varying adjustment, which may block causal pathways from previous interventions, we used marginal structural models to properly account for time updated confounding. This approach produces less biased effect estimates. We not only evaluated risk in participants who discontinued or interrupted use of GLP-1RAs, but also comprehensively assessed the risk according to different treatment scenarios with various durations of discontinuations and interruptions. We conducted comparisons with cumulative risk estimates derived from time varying instant risk under different conditions. This approach avoids reliance on the proportional hazard assumption and gives less biased effect estimates.^{67 68} We used multiple approaches to better understand the association between various patterns of use, including scenarios with different durations and timings of discontinuations and interruptions. We also evaluated the association between proportion of days covered and risk of major adverse cardiovascular events. We tested the robustness of the findings in multiple sensitivity analyses.

Conclusions

This study showed that discontinuing and interrupting GLP-1RA treatment could erode and might reverse the cardiovascular benefits of the drug in a duration dependent manner, increasing the risk of cardiovascular events. Strategies to reduce treatment discontinuity should be evaluated to maximally realise the cardioprotective benefits of GLP-1RAs.

Data availability statement

Data may be obtained from a third party and are not publicly available.