Background and aims: Prior studies have shown mixed results on the cardiovascular safety of insulin analogs compared to human insulins. This observational study uses rigorous causal inference methods of Comparative Effectiveness Research to emulate a trial evaluating the cardiovascular safety of analog versus human insulins. Materials and methods: Study subjects were 127,600 adults with type 2 diabetes receiving care between 1/12005 and 12/31/2013 at 4 U.S. health care delivery systems, classified on date of insulin initiation (index date) as using (a) analog insulins with or without human insulin, versus (b) human insulins only. Subjects were followed from index date to first occurrence of: myocardial infarction (MI), death, health plan disenrollment, or 12/31/2013. In order to estimate the differences between the two groups had patients remained on their insulin regimen (a perprotocol analysis), we used Inverse Probability Weights (IPW) and marginal structural modeling (MSM) analyses in which patients' data were right-censored if they crossed-over to the other exposure group or interrupted insulin therapy during follow-up. Effect estimates were obtained
based on two logistic MSM parameterizations for counterfactual hazards. To address concerns over potential residual confounding, we replicated the same analytic approach using three nested sets of covariates assumed to potentially affect (a) both the outcome and the exposure/ censoring events, (b) the outcome, and (c) either the outcome or exposure/censoring events. Results: The mean follow-up time was 21 months. Hazard ratios and cumulative risk differences comparing the two groups were not statistically significant. The estimates of the hazards ratio and its 95% confidence interval were 1.2214 [0.9877;1.4551]; 1.08 [0.8642;1.2959]; and 1.0779 [0.7692;1.3866] for each covariate adjustment set, respectively. The estimates of the cumulative risk difference at 2 years and its 95% confidence interval were 0.0037 [-8e-04;0.0082]; 0.0011 [-0.0034;0.0057]; and -6e-04 [-0.0055;0.0044] for each covariate adjustment set, respectively.
Conclusion: In this large comparative effectiveness analysis using marginal structural modeling with three covariate adjustment sets, we were not able to identify statistically significant excess occurrence of myocardial infarction in those using analog, compared to human insulins. Supported by: US National Heart & Lung & Blood Institute.