Statistical analysis

The data are shown as the number (%) or median (interquartile range). Categorical variables were compared using the χ²‐test, and continuous variables were compared using the Mann‐Whitney U‐test. Glycated albumin levels were converted to hemoglobin A1c using the following formula ¹³ : hemoglobin A1c (%) = 0.216 × glycated albumin (%) + 2.978. The eGFR was calculated by the formula developed for the Japanese population ¹⁴ . For the cross‐sectional analyses, non‐linear or logistic regression models were fitted with restricted cubic splines to investigate the association between hemoglobin levels or the probability of anemia and eGFR among users and non‐users of SGLT2i. The data were adjusted for age, sex, history of smoking, types of diabetes (type 1 or type 2), hospitalization during a study period, diagnosis of malignancy, and the use of angiotensin converting enzyme inhibitors and or angiotensin receptor blockers, dosage of iron supplementation, average monthly dose of darbepoetin or epoetin β pegol, mean corpuscular volume (MCV), mean corpuscular hemoglobin (MCH), mean corpuscular hemoglobin concentration (MCHC), red cell distribution width (RDW), platelet counts and white blood cell counts. The data for ferritin and transferrin saturation were available for only a small portion of participants, so MCV, MCHC, MCH and RDW were used instead as assessments of iron status. Additional adjustment for duration of diabetes, the presence of retinopathy, dipstick proteinuria levels as a categorical variable and C‐reactive protein levels (log‐transformed) were also carried out. C‐reactive protein levels and urine dipstick proteinuria levels were included in a separate model as the number of participants with both data were small. For the case–control study, cases (with anemia as defined above) and controls were matched 1:1 by age within 5 years, sex and eGFR within 1 mL/min/1.73 m². Associations between the use of SGLT2i until the last visit and cases (anemia) were analyzed using conditional logistic regression. The data were further adjusted for the duration of diabetes, body mass index, history of smoking (missing data were treated as a separate category), comorbidities (including active or previous malignancy), medications and hemoglobin A1c level nearest the SGLT2i start time. Active malignancy was defined as treatment for malignancy (including surgery, radiation and chemotherapy) within 3 months or the presence of malignancy (before treatment, under observation or under best supportive care). Previous malignancy was defined as the presence of malignancy at any time before the last visit (those with active malignancy were also included). Missing values for body mass index and CRP were imputed by multiple imputation by a chained equation using predictive mean matching. Five imputed datasets were created. The variables included in the model were age, sex, history of smoking, cardiovascular morbidities, active and previous malignancy, the use of chemotherapy, antihypertensives, antidiabetic medications, statins, aspirin, white blood cell counts, hemoglobin levels, and platelet counts.

The following analyses were also carried out. Changes in hemoglobin levels after initiation of SGLT2i were compared with temporary changes in hemoglobin levels among non‐users who were followed up more than a year. Propensity scores for SGLT2i use were estimated by a logistic regression model. Variables included in the model were age, sex, an interaction term for age and sex, baseline eGFR, baseline hemoglobin levels, MCV, MCHC, MCH, platelet counts, RDW, history of smoking, the use of antidiabetic medications and the average dose of ESA. SGLT2i users and non‐users were matched on the logit of propensity scores (±0.25 SD). Baseline eGFR and hemoglobin levels were defined as those before and closest to the initiation of the SGLT2i for SGLT2i users and the first measurements of these values during the study period for SGLT2i non‐users. Adjusted mean values and adjusted mean differences in hemoglobin levels at 3, 6 and 12 months were estimated by analysis of covariance with baseline hemoglobin levels as a covariate. In addition, among SGLT2i users, variables associated with an increase in hemoglobin levels, defined as an increase in hemoglobin more than the mean change in hemoglobin at 6 months after initiation of SGLT2i, were examined by logistic regression analysis. Association between increase in hemoglobin at 6 months and baseline hemoglobin levels were examined by restricted cubic spline analyses. Statistical analyses were carried out using Stata version 15 (StataCorp, College Station, TX, USA).