Reduced Mortality With Sodium-Glucose Cotransporter–2 Inhibitors in Observational Studies
Avoiding Immortal Time Bias
Sodium-glucose cotransporter–2 inhibitors (SGLT2i), the most recent class of medications to treat type 2 diabetes mellitus, have been the object of large randomized controlled trials and observational studies to assess their effectiveness on major disease outcomes. The EMPA-REG OUTCOME randomized trial (Empagliflozin, Cardiovascular Outcomes, and Mortality in Type 2 Diabetes) of >7000 patients with type 2 diabetes mellitus who had established cardiovascular disease reported a significant 32% reduction in all-cause mortality (hazard ratio, 0.68; 95% confidence interval, 0.57–0.82) with the SGLT2i empagliflozin. The CANVAS randomized trials program (Canagliflozin Cardiovascular Assessment Study) with >10 000 patients with type 2 diabetes mellitus, 66% with cardiovascular disease, reported a lesser 13% nonsignificant reduction in all-cause death (hazard ratio, 0.87; 95% confidence interval, 0.74–1.01) with the SGLT2i canagliflozin.
Two observational studies of these effects in real-world clinical practice diverged from the randomized controlled trials by reporting astonishing reductions in mortality with these drugs.1,2 The CVD-REAL study (Comparative Effectiveness of Cardiovascular Outcomes in New Users of SGLT-2 Inhibitors), using data from 6 countries on >150 000 new users of SGLT2i matched to the same number of new users of other antihyperglycemic agents (AHAs), reported a significant 51% lower rate of all-cause death with SGLT-2i in comparison with other AHAs (on-treatment hazard ratio, 0.49; 95% confidence interval, 0.41–0.57).1 In the EASEL study (Evidence for Cardiovascular Outcomes With Sodium Glucose Cotransporter 2 Inhibitors in the Real World), using data from the US Department of Defense on >12 000 new users of SGLT2i matched to the same number of other AHA new users, SGLT-2i was associated with a significant 56% lower rate of all-cause death than other AHAs (on-treatment hazard ratio, 0.44; 95% confidence interval, 0.35–0.55).2
The discrepancy in results between the randomized controlled trials and observational studies is particularly enigmatic for several reasons. First, canagliflozin was the predominant SGLT2i in the observational studies, with empagliflozin having the lowest usage. Thus, the nonstatistically significant 13% reduction in all-cause mortality with canagliflozin from the randomized controlled trial is of a very different magnitude than the >50% reductions reported in the observational studies. Second, the all-cause mortality curves in the EMPA-REG trial only start to separate meaningfully some time after treatment initiation. For the EASEL observational study (not provided in CVD-REAL), the mortality curves separate appreciably immediately after treatment initiation.
What could cause such discrepancies? This question is puzzling because, in both observational studies, the authors used state-of-the-art propensity score techniques to make the SGLT2i and other AHA groups comparable, and thus control confounding. However, time-related biases that are frequent in such observational studies of drug effects could have exaggerated the association between SGLT2i use and mortality. Immortal time bias is particularly important to examine in these studies.3
The 2 observational studies used a similar design in forming the study cohort. In the EASEL study, for example, new users of SGLT2i or of non-SGLT2i AHAs were defined as patients whose first exposure during 2013 to 2016 was not preceded by a prescription within the same AHA medication class in the prior 365 days.2 However, if a patient was “a new user of both SGLT2i and non-SGLT2i AHA, the patient would be classified as a SGLT2i new user, and the non-SGLT2i AHA would be considered a baseline or concomitant therapy.”2 Therefore, the study cohort creation skipped over the initial treatment period treated with a non-SGLT2i AHA for patients subsequently switched to a SGLT2i. Excluding or misclassifying the time between initiation of the comparator non-SGLT2i AHA and initiation of the study drug SGLT2i will result in immortal time bias.3
Evidence for the extent of this problem is found in the article’s table, showing that the new users of SGLT2i were more frequent users of many of the other AHAs, such as sulfonylureas (47% versus 24%), dipeptidyl peptidase–4 inhibitors (59% versus 15%), glucagon-like peptide–1 receptor agonists (22% versus 3%), and insulin (25% versus 7%) in the year before initiation.2 This implies that most new users of SGLT2i had been prior initiators of other AHAs. For these patients, the time between the first other AHA prescription (comparator drugs) and the first SGLT2i prescription (study drugs) is called immortal because the patient first using another AHA must remain alive to subsequently receive their first SGLT2i prescription; had they died during that time, they would have been classified in the other AHA group.3 Immortal time bias is introduced by omitting this other AHA-exposed time from the design and the analysis.3
The Figure depicts 3 patients who received a first prescription at the same time point during the study period. Whereas the second (SGLT2i) and third (other AHA) patients are directly comparable, the first is not because the other AHA was followed by SGLT2i. By the study’s design, the time they were on the other AHA and survived to receive the SGLT2i (immortal time) was not considered in the other AHA mortality rate calculations. This rate is overrepresented by the patients who died during this time; excluding the AHA-exposed immortal time results in an overestimate of the mortality rate in the comparator group, and thus an apparent protective effect.3
To avoid this bias, one can use Poisson-type regression techniques or Cox-type models with time-varying exposure that allow the proper classification of this immortal person-time. Alternatively, if the study requires matching on propensity scores such as the CVD-REAL and EASEL studies, a prevalent new-user design with time-conditional propensity scores can be used to avoid this bias.4 This design would also result in more accurate matching on prior AHA use, something not fully achieved in the EASEL study between the SGLT2i and other AHA groups (prior dipeptidyl peptidase–4 inhibitors: 58% versus 30%; prior glucagon-like peptide–1 receptor agonists: 20% versus 8%).2
In all, the remarkable >50% lower risk of all-cause death reported by the recent observational CVD-REAL and EASEL studies of SGLT2i effectiveness, inconsistent with the effects observed in the recent large randomized trials, is likely largely the result of immortal time bias.6 This and other time-related biases are known to exaggerate the potential effectiveness of many drugs on mortality and other major outcomes reported in observational studies.3,5 Consequently, the question of whether the potential mortality benefits of these drugs translate to real-world clinical practice remains uncertain. More accurate answers will come from well-designed observational studies that not only address confounding issues but, foremost, are also free of immortal time and other time-related biases.
Dr Suissa is the recipient of the James McGill Chair award. Dr Suissa has received research grants or participated in advisory board meetings or as a speaker at conferences for AstraZeneca, Bayer Pharma, Boehringer-Ingelheim, Bristol-Myers Squibb, Merck, and Novartis.
The opinions expressed in this article are not necessarily those of the editors or of the American Heart Association.
- © 2018 American Heart Association, Inc.
- Kosiborod M,
- Cavender MA,
- Fu AZ,
- Wilding JP,
- Khunti K,
- Holl RW,
- Norhammar A,
- Birkeland KI,
- Jørgensen ME,
- Thuresson M,
- Arya N,
- Bodegård J,
- Hammar N,
- Fenici P
- Udell JA,
- Yuan Z,
- Rush T,
- Sicignano NM,
- Galitz M,
- Rosenthal N
- Suissa S,
- Azoulay L
- Suissa S