The present study aimed to emulate a pragmatic randomized controlled trial that evaluates the effectiveness of SGLT2 inhibitors for CVD outcomes. To align the timing of eligibility assessment, intervention, and the start of follow-up [25], we implemented an active-comparator prevalent new-user design [26] as implemented in previous target trial emulations on SGLT2 inhibitors [27,28,29]. Dipeptidyl peptidase 4 (DPP4) inhibitors were used as active comparators, in line with previous target trial emulations on SGLT2 inhibitors with prevalent new-user design [27,28,29,30], because of their neutral effect on cardiovascular outcomes [31, 32] and widespread use in Japan [33]. The comparison of the target trial and the present study is listed in Table S1. The study was reported following the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [34] (Table S2).

We used claims data, death records, and annual health screening records of the Japan Health Insurance Association (JHIA) from the beginning of fiscal year 2015 to the end of fiscal year 2021 in Japan (April 1, 2015–March 31, 2022). The JHIA, the largest health insurer in Japan, provides healthcare coverage to over 30 million working-age individuals, mainly employees of small-to-middle-sized companies, representing approximately 40% of the working-age population in Japan [35]. The details of Japan’s government-led annual health screening program have been explained elsewhere [36, 37]. Every government-certified public health insurer in Japan, including the JHIA, is obliged to provide health screening to their members at ages 40–74 every year. The attendance rate of the JHIA members for the screenings was 52.3% in 2020 [38]. The screenings included self-reported lifestyle information, physical examination, and laboratory testing. Claims data consisted of outpatient, inpatient, and pharmacy claims, linked with the International Classification of Diseases, 10th revision (ICD-10) disease codes.

We first identified a base cohort of all individuals who were at least once new users of SGLT2 inhibitors or users of DPP4 inhibitors from April 2015 to March 2022 and have at least once attended health screenings during the same period. Being a new user was defined as having an initial prescription record of the medication with a minimum of six months of preceding insurance coverage to confirm the absence of prior use. Those who became new users of both SGLT2 inhibitors and DPP4 inhibitors during the same month were not included in the base cohort.

From the base cohort, we constructed monthly exposure datasets for every month from October 2015 to March 2022 and analyzed the pooled data of these datasets. For each exposure dataset, the inclusion criterion was defined as those who had not initiated SGLT2 inhibitors up to the respective index month. We then applied exclusion criteria to each exposure dataset. Individuals were excluded if they met at least one of the following conditions: 1) no evidence of diabetes (i.e., no use of glucose-lowering agents, fasting blood glucose [FBG] ≤ 126 mg/dL, and hemoglobin A1c [HbA1c] ≤ 6.4%); 2) annual health screening records were not available from the previous fiscal year up until the index month; 3) had missing data in covariates (see Pre-baseline covariates section); 4) estimated glomerular filtration rate (eGFR) < 30 mL/min/1.73m2; 5) had a hospitalization event due to CVD within two months before the month; and 6) had received a diagnosis of type 1 diabetes (ICD-10: E100–109) or cancer (C000-970, D000–099, or D370–489). As the health screening program requires measuring only either FBG or HbA1c, eligibility criteria regarding these values were assessed based on available measurements.

Within each monthly exposure dataset, participants were classified into two groups: SGLT2 inhibitor new users and DPP4 inhibitor users. The SGLT2 inhibitor new users included incident new users (i.e., those without prior prescription of DPP4 inhibitors) and prevalent new users (i.e., those with prior prescription of DPP4 inhibitors), while DPP4 inhibitor users included incident new users (i.e., those who initiated DPP4 inhibitors in the index month) and prevalent users (i.e., those who initiated DPP4 inhibitors up until the index month).

The primary outcome was the composite of all-cause death and hospitalization due to stroke (I60–63), HF ( I50, I110, I130, or I132), and myocardial infarction (MI; I21 or I22), as validated in Japanese administrative data [39]. The secondary outcomes were the components of the primary outcome. Outcomes were ascertained monthly. All participants were followed until an occurrence of the outcome, disenrollment from the health insurance plan, or the end of study period (March 2022).

For each monthly exposure dataset, the following variables were extracted from the claims data up until the month and the most recent health screening data: age, sex, prescription histories within last six months (insulin, statin, and glucose-lowering agents other than SGLT2 or DPP4 inhibitors), past CVD histories (defined as same as Outcomes), self-reported lifestyle information and objectively measured items in the annual health screening (smoking status [yes if: more than 100 cigarettes lifetime, smoking duration longer than six months, and the last smoking within a month], BMI, HbA1c, FBG, systolic blood pressure, and eGFR). The eGFR was calculated with equations modified for Japanese adults [40]. We categorized BMI into six groups (< 20.0, 20.0–22.4, 22.5–24.9, 25.0–29.9, 30.0–34.9, and ≥ 35.0 kg/m2), systolic blood pressure into three groups (< 110, 110–139, and ≥ 140 mmHg), and eGFR into four groups (30–44, 45–59, 60–89, and ≥ 90 ml/min/1.73m2).

Within each exposure dataset, SGLT2 inhibitor new users were propensity-score matched to DPP4 inhibitor users in a 1:1 ratio. The matching was stratified by BMI category and used a nearest neighbor approach with a caliper of 0.10. Specifically, incident new users of SGLT2 inhibitors were matched with incident new users of DPP4 inhibitors, whereas prevalent new users of SGLT2 inhibitors were matched with prevalent users of DPP4 inhibitors with the same duration of the prevalent use of DPP4 inhibitors [26]. We conducted the conducted the matching chronologically from the dataset of October 2015 to the dataset of March 2022 essentially without replacement, but prevalent new users of SGLT2 inhibitors were included in the match even if they had been once matched as a DPP4 inhibitor users in prior [27].

Propensity scores for initiating SGLT2 inhibitors were also separately constructed for each exposure dataset with pre-baseline covariates. To comprehensively include eligible participants in the matching, those with available HbA1c data were matched based on models with HbA1c, whereas others were on models with FBG. The following three interaction terms were incorporated into the model: 1) HbA1c (or FBG) and insulin use, 2) HbA1c (or FBG) and the use of other glucose-lowering agents, and 3) insulin use and the use of other glucose-lowering agents.

We analyzed the pooled data of all matched datasets. First, we compared the risks of composite outcome events between SGLT2 inhibitor users and DPP4 inhibitor users during the follow-up period after matching by plotting a Kaplan–Meier curve. Second, we fitted multivariable Cox proportional hazard models to estimate hazard ratios (HRs) and 95% confidence intervals (CIs) for outcome events. We used robust standard error to account for correlations in the data for individuals appeared in both arms. In this model, we adjusted for age, sex, previous CVD diagnosis, BMI, systolic blood pressure, eGFR, prescription histories of insulin, statin, and glucose-lowering agents other than SGLT2 and DPP4 inhibitors within the last six months before the matched month, and the duration between the month of entering the base cohort and the month of being matched. Third, we constructed Cox proportional hazard models separately for each BMI category to assess the effect heterogeneity of SGLT2 inhibitors across BMI categories.

We also conducted subgroup analyses according to age (< 60 years, ≥ 60 years), sex (male, female), and user types (with or without prevalent use of DPP4 inhibitors). In the analysis on secondary outcomes and the subgroup analyses, we compared the effectiveness of SGLT2 inhibitors between those with BMI < 22.5 kg/m2 and BMI ≥ 22.5 kg/m2 to secure a sufficient number of events in every comparison arm.

When summarizing patient characteristics before and after matching, pooled data of randomly matched DPP4 inhibitor users in each exposure dataset were used as representative of DPP4 inhibitor users before matching [27]. This approach was necessary because the characteristics of the same patient may differ depending on the index month in which they were matched. The absolute value of standardized difference < 0.10 was considered as an indicator of sufficient balance between groups. All analyses were performed using R version 4.3.1. The study was approved by the institutional review board at Kyoto University (IRB#R3452).