Regarding baseline characteristics of the patients, continuous variables were reported as mean with standard deviation (SD) while categorical variables were reported as number with percentage. Multiple imputation by chained equation was used to impute missing baseline covariates [14 , 15 (link)]. Each missing datum was imputed using known or imputed variables by five times. Five complete datasets were generated and analyzed independently, and the analyzed results were pooled to an overall estimate by using the Rubin’s rule [16 ]. One-to-one propensity score matching was performed to compare patients who had initiated SGLT2i or GLP-1RA at clinical equipoise. A logistic regression model was fitted to estimate the propensity score of patients by using the baseline covariates. Patients in SGLT2i and GLP-1RA groups were pairwise matched using the propensity score with a caliper of 0.001. An absolute standardized mean difference (ASMD) ≤ 0.1 of baseline covariate indicated the covariate balance between the two groups [17 (link)].
Each included patient had complete follow-up and was observed from the index date until any of the following: (i) the occurrence of events; (ii) death; (iii) addition or switching of treatments to another study exposures (e.g., patients who were initiated on SGLT2i first and switched to GLP-1RA were censored at the start date of GLP-1RA, and vice versa); or (iv) the date of the end of study (i.e., 31st December 2020), whichever was earlier. Cox proportion hazard regression models were constructed to estimate hazard ratio (HR) and 95% confidence interval (CI) of each outcome between the two groups. In addition, incidence rate ratio (IRR) for each outcome was estimated by Poisson regression models.
Subgroup analyses were performed by stratification according to sex (male and female), age group (< 60 and ≥ 60 years), HbA1c (< 8 and ≥ 8%), history of stroke, history of AF, and duration of diabetes (< 10 and ≥ 10 years). Tests for interaction between the novel anti-diabetic medications and subgroups were conducted.
Sensitivity analyses were conducted to check whether consistent results were obtained with different analytic approaches: (i) analysis without censoring on switching treatments; (ii) inclusion of patients with follow-up duration ≥ 1 year; (iii) inclusion of patients with at least two dispensing records within 12 months; (iv) inclusion of only patients who were initiated on SGLT2i or GLP-1RA in or after 2015; (v) ‘as-treated’ analysis, censoring when patients stopped being dispensed the index treatment; (vi) analysis using regression adjustment to control for confounders instead of propensity score matching; and (vii) analysis without adjustment of covariates using either propensity score matching or regression adjustment.
All statistical analyses were performed using Stata version 14.0 (StataCorp, College Station, Texas). p values < 0.05 were considered statistically significant. The current study was reported according to the guideline of STrengthening the Reporting of OBservational studies in Epidemiology (STROBE).
Each included patient had complete follow-up and was observed from the index date until any of the following: (i) the occurrence of events; (ii) death; (iii) addition or switching of treatments to another study exposures (e.g., patients who were initiated on SGLT2i first and switched to GLP-1RA were censored at the start date of GLP-1RA, and vice versa); or (iv) the date of the end of study (i.e., 31st December 2020), whichever was earlier. Cox proportion hazard regression models were constructed to estimate hazard ratio (HR) and 95% confidence interval (CI) of each outcome between the two groups. In addition, incidence rate ratio (IRR) for each outcome was estimated by Poisson regression models.
Subgroup analyses were performed by stratification according to sex (male and female), age group (< 60 and ≥ 60 years), HbA1c (< 8 and ≥ 8%), history of stroke, history of AF, and duration of diabetes (< 10 and ≥ 10 years). Tests for interaction between the novel anti-diabetic medications and subgroups were conducted.
Sensitivity analyses were conducted to check whether consistent results were obtained with different analytic approaches: (i) analysis without censoring on switching treatments; (ii) inclusion of patients with follow-up duration ≥ 1 year; (iii) inclusion of patients with at least two dispensing records within 12 months; (iv) inclusion of only patients who were initiated on SGLT2i or GLP-1RA in or after 2015; (v) ‘as-treated’ analysis, censoring when patients stopped being dispensed the index treatment; (vi) analysis using regression adjustment to control for confounders instead of propensity score matching; and (vii) analysis without adjustment of covariates using either propensity score matching or regression adjustment.
All statistical analyses were performed using Stata version 14.0 (StataCorp, College Station, Texas). p values < 0.05 were considered statistically significant. The current study was reported according to the guideline of STrengthening the Reporting of OBservational studies in Epidemiology (STROBE).
Free full text: Click here
