We appreciate the discussion initiated by Badrick et al. (1) and thank them for the questions raised. In an ideal world, we would have a randomized trial with the enormous sample necessary to answer this question as well as decades of follow-up. In an imperfect world, our analysis reveals the experience of a significant fraction of U.S. patients with diabetes using insulin glargine over 3 years of observation.
Regarding our prevalent users cohort and fixed treatment group assignment, we acknowledged the reasonable but unproven assumption that our observed exposure reflects unobserved exposure, preceding 2006 (Part D inception). Under this assumption, assignment to treatment observed in the earliest observation window is the best exposure classification approach. Our intention-to-treat analysis is valid if there is limited treatment crossover, something we observed in our follow-up period.
Badrick et al. suggest propensity score use (for matching or stratification); unfortunately, this approach does not avoid potential shortcomings of adjustment (2,3).
Regarding our comparison group of nonglargine users, our study question, which focused specifically on glargine, was informed by published basic science and population research. Although the potential for short-acting analogs to promote cancer is intriguing, we are aware of no basic science or clinical data suggesting this association. We believe that observational research must always be founded on established science. Nonglargine insulin is used as a comparator by many studies on this topic (including the article by Suissa et al. [4] that Badrick et al. cite); our work thus does permit comparison.
We reported an unexpected association between metformin and breast cancer. Badrick et al. suggest this finding invalidates our study. We disagree and applaud the editors for their consideration of work that may not conform to “conventional wisdom,” thereby avoiding the problem of publication bias. One distinction between our article and others is that we explicitly report the estimate associated with this important coexposure. The omission of such estimates by others denies us a greater context for interpreting our finding.
Our work complements the study by Suissa et al. that Badrick et al. cite. Suissa et al. used “other insulin” as a comparator (as we did); their glargine group included users of glargine and users of glargine plus nonglargine. A portion of this cohort thus is comparable to our combination insulin group. Suissa et al. found a breast cancer association after 5 years of exposure. We also found a higher risk of breast cancer between 2006 and 2008 in a cohort of combination insulin users who had access to glargine beginning April 2000. Allowing for new drug uptake and given limited crossover, it is likely many or most of our patients were exposed for 4–6 years, the time frame in which Suissa et al. found an association.
As this science evolves, diverse weaknesses of available data will challenge all of us. Collectively we must address this by advancing methodology, asking the same questions of various datasets, and openly discussing the limitations of our studies. Although this will require some sophistication from readers, the alternative—to not query imperfect data—is unacceptable given the infeasibility of large randomized trials and the importance of these questions.
Article Information
No potential conflicts of interest relevant to this article were reported.
