How combining two diabetes drugs compares to using one drug alone in people with type 2 diabetes

The targeted causal estimand was the effectiveness of combined incident new‐use (simultaneous initiation of SGLT‐2 inhibitors and GLP‐1 RAs) and prevalent new‐use (initiating a GLP‐1 RA while continuously exposed to SGLT‐2 inhibitors). Thus, we addressed the causal question: Among individuals in the matched study cohort, what would have been if initiators of combination therapy who remained on combination therapy had initiated and remained on SGLT‐2 inhibitor therapy or had continued and remained on SGLT‐2 inhibitor therapy?10, 11

We used nationwide health claims data from one of Germany's largest statutory health insurance funds (BARMER, Wuppertal, Germany). The BARMER scientific data warehouse contains longitudinal individual‐level data on approximately 9 million insured individuals and has been used in prior pharmacoepidemiological research.12 The database includes a broad range of routinely collected data such as information on sociodemographic characteristics, health insurance enrolment status, inpatient and outpatient medical encounters coded according to ICD‐10 in its German modification, surgical and procedural codes (Operationen‐ und Prozedurenschlüssel [OPS]), as well as dispensed medications coded according to the German version of the international Anatomical Therapeutic Chemical (ATC) classification system.

We implemented a prevalent new‐user design.13 First, we established a base cohort that comprised individuals who were incident new‐users of SGLT‐2 inhibitors (dapagliflozin, empagliflozin, ertugliflozin, or canagliflozin) between 1 January 2013 and 31 December 2023. The base cohort index date corresponded to the date of an individual's first SGLT‐2 inhibitor dispensing. Therefore, we excluded individuals who were only exposed to GLP‐1 RAs, had an unclear health insurance status, had less than 1 year of insurance coverage before the base cohort index date, had no ambulatory encounter during that year, did not meet the criteria for verifiable type 2 diabetes (definitions provided in Table S1, Supporting Information), had missing information on sex, were younger than 18 years, or received multiple active ingredients or multiple dispensings of SGLT‐2 inhibitors on the base cohort index date.

Next, we formed the study cohort by following these individuals and collecting subsequent dispensings of SGLT‐2 inhibitors within continuous therapy. We defined continuous therapy using an on‐treatment approach with a 180‐day refill interval considering therapy as discontinued if no refill occurred within this interval. Rather than one observation per individual, each dispensing was treated as a separate row, meaning that one individual could contribute multiple observations over time. The unit of analysis therefore shifted from the individual to each of that individual's dispensings.

We then identified individuals who initiated a GLP‐1 RA. This GLP‐1 RA dispensing was classified as either incident new‐use of combination therapy if it occurred on the same date as the first SGLT‐2 inhibitor dispensing or prevalent new‐use if it occurred during continuous SGLT‐2 inhibitor therapy.

To ensure comparability, dispensings—whether GLP‐1 RA or SGLT‐2 inhibitor—were excluded if no ambulatory encounter was recorded in the preceding year or if multiple dispensings or active ingredients occurred on the same date.

For each combination therapy user, a hybrid exposure set was created to identify eligible reference dispensings from individuals who continued SGLT‐2 inhibitor therapy that matched the combination therapy user's treatment history at that specific timepoint. This ensures that individuals are compared at an equivalent stage of their disease progression and treatment history. A detailed definition of hybrid exposure sets and possible matching timepoints is provided in Figure. S1

We adjusted for a total of 64 covariates to account for potential confounding. All covariates were measured prior to or on the date of the drug dispensing that defined (for GLP‐1 RA initiators) or corresponded to (for reference dispensings) membership in the respective hybrid exposure set. These covariates included sociodemographic characteristics, diabetes duration and diabetes‐related complications, comorbidities, procedures, co‐medications, glucose‐lowering therapies, markers of health‐seeking behaviour, healthcare utilization, and cohort entry year. Detailed definitions and assessment windows are provided in Table. S2

We calculated a time‐conditional propensity score (TCPS) for initiating GLP‐1 RA (becoming a combination therapy user) versus continuing SGLT‐2 inhibitors using multivariable conditional logistic regression with 64 pre‐specified covariates, conditional on the hybrid exposure sets. Unlike conventional propensity scores, which are estimated once per individual, the TCPS was recalculated for each dispensing using the updated covariates from the timepoint of inclusion in the hybrid exposure set. By updating the propensity score for each hybrid exposure set, this approach provides a time‐conditional measure of the propensity of initiating GLP‐1 RA versus continuing SGLT‐2 inhibitors, accounting for temporal changes in covariates.

We ensured positivity by trimming hybrid exposure sets where the combination therapy users' TCPS fell outside the TCPS range of the eligible reference dispensings. One‐to‐one matching without replacement was conducted in chronological order, starting with the earliest combination therapy user by calendar time,13 based on both the hybrid exposure set (to ensure identical treatment history) and the nearest TCPS (to balance all measured covariates). Additionally, we used a calliper width of 0.2 times the standard deviation of the linear predictor. Once a reference dispensing was matched, it was no longer eligible for subsequent matching.

Combination therapy was defined as the period of concurrent use of both SGLT‐2 inhibitors and GLP‐1 RAs. Consistent with our definition for SGLT‐2 inhibitors, a 180‐day refill interval was also applied to GLP‐1 RAs to determine continuous use. This period ended when coverage for either drug lapsed.

The primary outcome was all‐cause mortality. Secondary outcomes included a modified cardiovascular composite (all‐cause mortality, myocardial infarction, or stroke), its individual components, heart failure, diabetic nephropathy, and renal failure. All outcomes except all‐cause mortality were defined by ICD‐10‐GM hospital diagnosis codes in primary position capturing both incident and recurrent hospitalizations. For detailed outcome definitions, see Table. S3

Accumulation of follow‐up time started on the study cohort index date (date of first GLP‐1 RA dispensing or date of matched SGLT‐2 inhibitor dispensing for reference individuals) and continued until the outcome, death (unless death was the primary outcome in analyses of all‐cause mortality or the modified cardiovascular composite outcome), end of health plan enrolment, treatment discontinuation, initiation of GLP‐1 RA (for reference individuals), or the end of the study (31 December 2024).

References censored due to GLP‐1 RA initiation could re‐enter as combination therapy users if a match was available, with follow‐up starting 1 day after GLP‐1 RA initiation. No other re‐entries were permitted, and only the first drug exposure episode was used.

We assessed potential effect measure modification by including interaction terms between treatment group (combination therapy versus continuing SGLT‐2 inhibitors) and sex, as well as cardiovascular disease (CVD).

In addition, we conducted 10 sensitivity analyses to evaluate the robustness of the result for our primary outcome. First, we assessed whether the effect varied with different matching procedures: one‐to‐one nearest neighbour matching without a calliper, matching the latest GLP‐1 RA initiator first, and one‐to‐two nearest neighbour matching. In addition, we investigated whether including individuals who were initially matched as SGLT‐2 inhibitor continuers and later re‐entered the cohort as combination therapy users influenced the results. For this analysis, re‐entry into the cohort was not allowed. To address potential bias due to informative censoring from treatment discontinuation, initiation of GLP‐1 RA among reference individuals, or administrative censoring (end of health plan enrolment or end of study), we applied stabilized time‐varying inverse probability of censoring weighting (IPCW).14 To evaluate the impact of potential exposure misclassification, we repeated the analysis using a 120‐day and 90‐day refill interval and an intention‐to‐treat (ITT) exposure definition. Finally, we constructed prescription‐based and time‐based exposure sets as originally described.13 In these analyses, exposure sets were defined solely by the number of prior SGLT‐2 inhibitor dispensings or time since base cohort index date, relaxing the stricter hybrid exposure set criteria.

This observational study was conducted and reported in accordance with the STROBE guidelines for observational research.15 Characteristics of individuals were summarized using counts and proportions for categorical variables and means with standard deviations for continuous variables. Covariate balance was evaluated using absolute standardized differences (ASDs), with values ≥10% considered indicative of meaningful imbalance.16 Event rates per 1000 person‐years were calculated based on the Poisson distribution with corresponding 95% confidence intervals. For the comparative analyses, Cox proportional hazards models were used to estimate hazard ratios and 95% confidence intervals. Individuals initially matched as a reference could potentially contribute two observation periods, the first period as a reference continuing SGLT‐2 inhibitors, and the second period as a combination therapy user upon GLP‐1 RA initiation. To account for this within‐individual correlation of person‐time we calculated robust (sandwich) standard errors.

Absolute risks were derived using Kaplan–Meier methods. Risk differences (RDs) were estimated for up to 5 years of follow‐up, with 95% confidence intervals obtained via bootstrap resampling with 1000 iterations. All statistical analyses were performed using R (R Foundation for Statistical Computing, Vienna, Austria, http://www.R-project.org/↗).

Combination therapy users had a 29% lower hazard of all‐cause mortality compared with SGLT‐2 inhibitor continuers (410 vs. 868 events; 13.4 vs. 18.6 per 1000 person‐years; HR 0.71, 95% CI 0.63–0.80). The RD was −0.4% (95% CI −0.7% to −0.1%) at 1 year, increasing to −2.3% (95% CI −3.7% to −1.0%) at 5 years. Figure 2 shows the corresponding cumulative incidence curves for both treatment groups, and annual RDs up to 5 years are provided in Table S5.

Combination therapy users had a 19% lower hazard for the modified cardiovascular composite outcome (740 vs. 1372 events; HR 0.81, 95% CI 0.74–0.88) and a 22% lower hazard for heart failure (304 vs. 562 events; HR 0.78, 95% CI 0.68–0.89) compared with SGLT‐2 inhibitor continuers. For myocardial infarction and stroke, combination therapy was associated with a 5% (HR 0.95, 95% CI 0.79–1.14) and 14% (HR 0.86, 95% CI 0.72–1.04) hazard reduction, respectively, though both confidence intervals included the null. Regarding renal outcomes, hazards were increased for nephropathy (HR 1.23, 95% CI 0.76–1.99) but 9% lower for renal failure (HR 0.91, 95% CI 0.70–1.17), again with wide confidence intervals. All outcomes are summarized in Table 2, with cumulative incidence curves provided in Figures S3–S8.

No effect measure modification was observed across the examined subgroups (details provided in Tables S6‐S7). Stratified analyses showed comparable protective effects by sex (women: HR 0.64, 95% CI 0.52–0.78; men: HR 0.75, 95% CI 0.65–0.87) and baseline CVD status (history of CVD: HR 0.68, 95% CI 0.59–0.78; without history of CVD: HR 0.77, 95% CI 0.61–0.96). The hazard ratio remained robust across alternative matching procedures, including no caliper (HR 0.80, 95% CI 0.73–0.87), matching the latest GLP‐1 RA initiator first (HR 0.74, 95% CI 0.68–0.81), a one‐to‐two matching ratio (HR 0.75, 95% CI 0.69–0.81), and when excluding individuals who were initially matched as SGLT‐2 inhibitor continuers and later re‐entered the cohort as combination therapy users(HR 0.69, 95% CI 0.61–0.78). To address informative censoring, stabilized time‐varying IPCW revealed similar effects (HR 0.78, 95% CI 0.67–0.89). Altering the fixed refill interval to 120 days (HR 0.80, 95% CI 0.68–0.93) or 90 days (HR 0.71, 95% CI 0.43–1.17) yielded consistent findings, while an ITT exposure definition resulted in no differential effect (HR 1.01, 95% CI 0.95–1.06). Finally, prescription‐based (HR 0.65, 95% CI 0.58–0.72) and time‐based exposure sets (HR 0.64, 95% CI 0.57–0.71) were nearly identical to the original hybrid exposure set approach. A forest plot summarizing the hazard ratios for the primary outcome and all sensitivity analyses is presented in Figure 3.

The observed benefits likely stem from the complementary mechanisms of both drug classes. SGLT‐2 inhibitors block glucose reabsorption in the proximal tubule, promoting urinary glucose excretion.24 Potential cardiorenal protection is driven by natriuresis and osmotic diuresis,25 reduced systolic blood pressure and cardiac pre‐ and afterload, improved cardiac metabolism,26 and modulation of the estimated glomerular filtration rate.27

In contrast, GLP‐1 RAs enhance glucose‐dependent insulin secretion and inhibit glucagon release. They reduce postprandial glucose and insulin concentrations by delaying gastric emptying.28 Furthermore, while SGLT‐2 inhibitors induce weight loss via caloric excretion and fluid reduction,29 GLP‐1 RAs promote weight loss through central nervous system pathways that increase satiety and reduce caloric intake.5 GLP‐1 RAs also improve LDL profiles30 and may inhibit atherosclerotic plaque progression.31 Targeting these distinct pathways of cardiorenal pathophysiology may explain the observed benefits.

While a meta‐analysis shows that SGLT‐2 inhibitors maintain cardiovascular benefits regardless of background GLP‐1 RA use,27 it was limited by a small sample size and a focus on baseline use of GLP‐1 RA, rather than sequential initiation.

Our study has several notable strengths. First, we selected continuation of SGLT‐2 inhibitor therapy as the comparator to reflect contemporary guideline‐supported treatment pathways, in which both SGLT‐2 inhibitors and GLP‐1 RAs are recommended and may be used in combination, whereas DPP‐4 inhibitors are not recommended for use alongside GLP‐1 RAs.2 Second, the prevalent new‐user design captured SGLT‐2 inhibitor users who later initiated GLP‐1 RA, reflecting real‐world prescribing patterns more comprehensively than traditional designs.13

Third, the use of hybrid exposure sets enabled precise matching on treatment history, reducing bias introduced by differences in prior treatment history. Fourth, the large, unselected real‐world population allowed us to examine effectiveness in a real‐world cohort, avoiding restrictive eligibility criteria common in clinical trials.32

Nevertheless, some limitations must be acknowledged. Outcome definitions based on health claims data used in this study may be subject to bias, as no external validation studies exist. However, we restricted outcomes to hospital diagnoses in primary position to enhance validity. Residual confounding remains a concern, as individuals initiating combination therapy may systematically differ from those continuing SGLT‐2 inhibitors. We lacked factors such as HbA1c, BMI, albuminuria, lifestyle, and socioeconomic status. While we used markers of health‐seeking behaviour as proxies for socioeconomic status, these differences may still contribute to the observed associations, despite the prevalent new‐user design and extensive covariate adjustment.

However, the consistent protective effect across subgroups and sensitivity analyses alongside comparable results from the UK study9 that accounted for HbA1c and BMI supports the robustness of our findings, although caution is still warranted.

Further, on‐treatment estimates are susceptible to informative censoring, as individuals who discontinue therapy may differ in their risk of the outcome compared with those who continue therapy.23 Since combination therapy required two refills versus one for SGLT‐2 inhibitor continuers, the resulting differential follow‐up times, as well as differences in treatment persistence and healthy adherer behaviour, could have influenced our results. To address informative censoring, we applied stabilized time‐varying IPCW, which yielded nearly identical results (HR 0.78, 95% CI 0.67–0.89).

Moreover, the impact of therapy sequence is unclear. In future, assessment of the reversed sequence—compared with continued GLP‐1 RA therapy—may clarify whether benefits depend on sequence or on combination therapy per se.

The datasets analysed in this study are not publicly available due to data protection regulations and contractual agreements with the data provider, BARMER (Wuppertal, Germany).