Comparative Effectiveness of SGLT2i vs DPP-4i

Propensity score (PS) matching was used to control for confounding. The PS for initiating SGLT2i vs DPP-4i therapy was calculated within each HbA_1c subcohort separately through a logistic regression model with 128 prespecified covariates. Laboratory data, except for HbA_1c and eGFR_Cr, were not included in the model because of the substantial proportion of missing information. Initiators of SGLT2i therapy were 1:1 matched to initiators of DPP-4i therapy on their estimated PS within each HbA_1c subcohort using the nearest neighbor approach with a caliper width of 0.01 on the PS scale. Covariate balance was assessed with standardized differences, with meaningful imbalances set at values higher than 10%.^{29 (link),30 (link)} We also reviewed the balance in laboratory test results not included in the PS model, to evaluate potential residual confounding after PS matching.
We tabulated numbers of events, incidence rates (IRs), and rate differences (RDs) per 1000 person-years. Hazard ratios (HRs) and 95% CIs were estimated by Cox proportional hazard models. We used Kaplan-Meier methods to plot cumulative incidence of primary outcomes and log-rank tests to compare hazard rates between drug classes. Two-sided P values for homogeneity were obtained by performing Wald tests and values <.05 were considered indicative of treatment heterogeneity.
We inspected the robustness of the main findings through sensitivity analyses (see eMethods in Supplement 1), addressing potential informative censoring, time-lag bias,^{31 (link)} unmeasured confounding for high risk for recurrence, and DPP-4i effects on HHF (since saxagliptin and alogliptin showed an increased HHF rate in CVOTs,^{32 (link),33 (link)} which resulted in an FDA warning,³⁴ we conducted a sensitivity analysis for the HHF outcome redefining the comparator group as sitagliptin only).
All analyses were implemented using Aetion Evidence Platform (Aetion Inc) and Stata statistical software, version 15.1 (StataCorp LLC).