and the fnal decision process by the primary physicians. Consequently, we referred to the detailed information available in the claims data to refect the patient disease stages and physician practice styles for the propensity score calculation. More precisely, we identifed the use patterns for pharmaceutical prescriptions recommended in the existing clinical guidelines for the treatment of diabetes and related cardiovascular complications [23, 24], and the utilization patterns of outpatient and inpatient services (frequency of physician visits and days of in-hospital service use per year) covered by the universal public health insurance scheme. Te prescription list included oral diabetes medications (biguanides, sulfonylureas, thiazolidinediones, sodiumglucose transporter 2 inhibitors, meglitinides, dipeptidyl peptidase-4 inhibitors, alpha-glucosidase inhibitors), insulin treatment agents of any type, medications for cardiovascular risk control (e.g. anti-platelet, anti-hyperlipidemic, and anti-hypertensive agents of any kind), and cardiorenal protection medications such as angiotensinconverting enzyme and angiotensin II type 1 receptor blockers. Propensity scores were calculated by logistic regression for participation in the disease management program regressed on age, sex, prescription patterns, comorbidities measured in the Charlson comorbidity index [25], and annual numbers of medical care utilization, outpatient visits, and hospitalization days over the past year from the baseline date. We matched the treatment and control groups for each entry year. We performed balanced 1:3 matching using the nearest-neighbor approach with replacement and a caliper width of 0.2 of the pooled standard deviation of the logit of the propensity score. Outcome measures Te major endpoints were occurrence of macroangiopathy (ischemic heart disease, heart failure, stroke, and other cardiovascular disease such as cerebral aneurysm and chronic peripheral arterial disease) and microangiopathy (diabetic retinopathy, neuropathy, and nephropathy) within 5 years of follow-up. Te identifcation of these event diseases relied on diagnosis-related therapeutic medication/device use listed in the claims data to avoid misclassifcation due to upcoding. Other endpoints included all-cause mortality, all-cause hospitalization, and intensive and emergency care use. We also included dependent living conditions in daily activities such as toileting, bathing, clothing, and eating as evaluated in the eligibility criteria for long-term care insurance [26]. We considered that eligibility level of≥2 indicated loss of independence in daily life activities. Finally, the utilization of medical and long-term care services was evaluated by referring to the reference prices set in the standardized item-by-item fee schedule in the Japanese universal public health insurance scheme. Watanabe et al. BMC Endocrine Disorders (2022) 22:135 Page 4 of 9 Fig. 1 Flow chart for selection of the study subjects. Note: The number of samples and patients are described by person-years, as matching was done by multiple years Watanabe et al. BMC Endocrine Disorders (2022) 22:135 Page 5 of 9 All costs were expressed in US dollars (USD) with an exchange rate of 1 USD=108 Japanese yen. Statistical analysis Descriptive statistics were compared between the treatment and control groups before and after propensity score matching. Unmatched patients based on common support of the propensity scores were excluded from the analysis. Standardized diferences were evaluated to confrm the efectiveness of the balancing. Next, a Cox proportional hazard model was used to account for time-to-event with censoring. Multivariate analyses were adjusted for age, sex, comorbidities, and use of oral diabetes medication and/or insulin. We regarded the major endpoints and all-cause death as competing risks, and treated observations as censored at the time when a competing event occurred. Finally, we compared medical and long-term care cost, number of outpatient visits, and number of days of hospitalization using a t-test and log-transformed variables. All analyses were performed using STATA version 15 software (Stata Corporation, College Station, TX, USA). Results Table 1 shows the baseline characteristics of the treatment group and the subpopulations for the two control groups before propensity score matching. Te patients in the treatment group were more likely to be male, receive treatment with insulin, and have medication for cardiovascular risk control and cardiorenal protection, and had a lower Charlson comorbidity index than the patients in the two subpopulations for the control groups. Te treatment group had more frequent physician visits, fewer days of hospitalization, and lower medical cost utilization. Table 2 presents the comparisons between the treatment group and the control groups after propensity score matching. After the propensity score matching, the diferences in demographic characteristics and prescription patterns largely