We congratulate Lim et al1 for their sophisticated Cox multivariable regression analysis performed in a retrospective cohort study of 2083 diabetic kidney transplant recipients demonstrating a significantly favorable effect of “sodium-glucose cotransporter 2 inhibitor (SGLT2i) use for >90 days” on the primary endpoint, a composite of all-cause mortality, death-censored graft failure, and serum creatinine doubling. However, we believe that their reported statistical analysis was not accurate, with time-dependent (also known as immortal time) bias2-4 existing that was not properly controlled/avoided. Specifically, our Figure 1 shows that in a proper configuration, all patients at baseline (initiation of antidiabetic medication after kidney transplant [KT]) belong to the “Pre-SGLT2i use for >90 days” state (patients who received an SGLT2i for <90 d were excluded from the study). Patients in the “Pre-SGLT2i use for >90 days” state contribute primary outcome events and person-time of follow-up in that state until they have used an SGLT2i for >90 d. Once this “intervening event” has occurred, patients would then belong to the “Post-SGLT2i use for >90 days” state, contributing primary outcome events and person-time of follow-up in the latter state. Unless we misunderstood their analysis, Lim et al1 appeared to have treated “SGLT2i use” as a baseline variable, leading to time-dependent bias of the results. Both Beyersmann et al2 and Suissa3 show that severe underestimation of the true hazard ratio can occur in the presence of time-dependent bias. What is ideally needed is for the authors to rerun their statistical analysis treating “SGLT2i use for >90 days” as a time-dependent covariate (equaling 0 while the patient belongs to the first state, and 1 following entry into the second state), thereby eliminating all time-dependent bias from their study.
Lim et al1 stated in their Results section that the mean follow-up duration was 5.24 y (62.9 mo) posttransplant; however, we see no mention of the distribution of time-to-initiation of antidiabetic medication after KT. We expect that most of the pretransplant diabetic patients (77.2% of the cohort) began antidiabetic medication at or shortly after KT, whereas time-to-initiation of antidiabetic medication would have begun later among the 22.8% who developed posttransplant diabetes mellitus (new onset diabetes after transplant). Although the authors also did not provide the distribution of time-to-initiation of an SGLT2i following antidiabetic medication initiation, they showed in their Supplemental Table 5 that the mean time from KT to SGLT2i initiation was 3.84 y (1403.5 d). Thus, it appears that most of the clinical follow-up of the 226 SGLT2i recipients occurred before their starting an SGLT2i, for is, while in the “Pre-SGLT2i use for >90 days” state. Furthermore, although the distribution of “total time receiving an SGLT2i after KT” was not reported, results of their Supplemental Table 3 suggested that most of the 226 “SGLT2i users” had not used SGLT2i for >1 y in this study. Again, since the range of transplant dates was not reported (only reported was the last transplant date, December 31, 2019), it appears that most of the SGLT2i users started on an SGLT2i at 3–4 y posttransplant. The recent meta-analysis by McGuire et al5 of 6 major randomized trials comparing an SGLT2i versus placebo among (nontransplant) type 2 diabetics, with established cardiovascular disease in over 50% of patients and reduced renal function in approximately 25% of patients, reported an adjusted hazard ratio of 0.62 for a similar composite endpoint. Thus, is it possible that Lim et al1 have correctly demonstrated a noticeably lower adjusted hazard ratio (0.43–0.45) among healthier patients who were treated with an SGLT2i for clearly less time in comparison with the approximately 47 000 patients in the McGuire et al5 report?
1. Lim JH, Kwon S, Jeon Y, et al. The efficacy and safety of SGLT2 inhibitor in diabetic kidney transplant recipients. Transplantation. 2022;106:e404–e412.
2. Beyersmann J, Gastmeier P, Wolkewitz M, et al. An easy mathematical proof showed that time-dependent bias inevitably leads to biased effect estimation. J Clin Epid. 2008;61:1261–1221.
3. Suissa S. Immortal time bias in pharmacoepidemiology. Am J Epid. 2008;167:492–499.
4. Shariff SZ, Cuerden MS, Jain AK, et al. The secret of immortal time bias in epidemiologic studies. J Am Soc Nephrol. 2008;19:841–843.
5. McGuire DK, Shih WJ, Cosentino F, et al. Association of SGLT2 inhibitors with cardiovascular and kidney outcomes in patients with type 2 diabetes – a meta-analysis. JAMA Cardiol. 2021;6:148–158.