Why the study?
Do DOACs (dabigatran and rivaroxaban) increase the risk of myocardial infarction compared to VKAs in patients with nonvalvular atrial fibrillation?
Do DOACs (dabigatran and rivaroxaban) increase the risk of myocardial infarction compared to VKAs in patients with nonvalvular atrial fibrillation?
Answers are grounded in this paper and cite their sources.
This commentary highlights methodological flaws in a recent observational study and reaffirms through an updated meta-analysis that DOACs do not increase MI risk compared to warfarin in AF patients.
Patients with nonvalvular atrial fibrillation (AF) require life-long anticoagulation to prevent embolic stroke. Anticoagulation with vitamin K antagonists (VKAs, such as warfarin) can be challenging due to interindividual and intraindividual variability in clearance and effect, which requires frequent coagulation monitoring and dose adjustments due to the drugs' narrow therapeutic index. In contrast to VKAs, the direct-acting oral anticoagulants (DOACs) have fixed-dose regimens that can more easily be followed by the typically older target population. As the prevalence of AF increases rapidly with age 2, age-related comorbidities are frequently present in AF populations and myocardial infarction (MI) is thus an equally relevant outcome (and which is a potentially competing event if death is the consequence). For the DOACs dabigatran and rivaroxaban, randomized trials 3, 4 and several observational pharmacoepidemiological studies 5-12 have provided comparative evidence regarding the incidence of MI in AF patients on oral anticoagulation. Our network meta-analysis of current evidence, shown in Figure 1A, demonstrated no increased risk of MI for these DOACs compared with warfarin. Using the Clinical Practice Research Datalink (CPRD, www.cprd.com), Stolk and colleagues also addressed this research question and suggested a twofold increase in the risk of MI for dabigatran and rivaroxaban compared with VKA 13. We read their paper with interest, as their analytical procedures and the conclusions they reached deviated from those of previous observational studies 5-12. As a result of its retrospective nature, any pharmacoepidemiological approach should entail measures to minimize bias. We believe that the analysis presented by Stolk et al. is prone to various forms of potential bias and could be expanded to refine and strengthen the paper's conclusion. As an example, these conclusions are based on only 10 events under treatment with the DOACs dabigatran or rivaroxaban that were observed in a nonrandomized sample and occurred within the clinically rather short follow-up period of only 1 year. First, selection bias (resulting in confounding) is an ever-present issue in observational studies. It is reasonable to expect highly individualized decisions for initiating such diverse AF treatments (VKAs, DOACs and the former second-line, and now obsolete choice of aspirin 14). The fact that more AF patients were on aspirin than on oral anticoagulants sheds justifiable doubt on an unselected allocation of comparable patients to the different treatment groups. The authors chose a new-user design 15 to compare DOACs, warfarin and aspirin with regard to MI incidence rates. A design based on treatment-naïve users has a number of advantages; for example, there is a clear temporal relationship between the start of treatment and the outcome incidence, thus facilitating treatment comparisons in patients with a similar health need 16. However, the necessary comparability of patient groups is not achieved by a proper design alone; even if AF is newly diagnosed, it still requires accounting for confounding by indication due to nonrandom decisions for the initiation of specific treatments 17. While all but one of the other observational studies on this topic applied the preferred propensity score (PS) adjustment 6-12, here the authors chose covariate adjustment, with covariates being selected by their ability to change the treatment estimate upon (manual) inclusion, a rather uncommon procedure, especially because alternatives for time-dependent variable selection exist that are less prone to bias (e.g. penalized variable selection 18). Numerous imbalances in the patient characteristics of their Table 2 may not be adequately reflected by the small set of covariates used for adjustment. Of note, PS approaches can efficiently include a large number of (time-dependent) covariates without the necessity of any troublesome preselection 17, 19. Second, bias from missing data is a growing concern with increasing proportions of missing values. Therefore, PS alone may not have performed better. Handling missing data in nonrandomized studies by the missing-indicator method will almost always give biased results 20; this issue appears crucial because of nonsignificant associations with otherwise strong cardiovascular risk factors in observational studies 21. Therefore, robustness should be shown in additional analyses, such as complete-case analysis and multiple imputation procedures 22. Third, the time scale for exposure–outcome relationships is a critical issue in pharmacoepidemiological research. The authors applied a time-dependent regression model that has ample justification for studying medication-outcome relationships 23. However, to assess drug exposure on a monthly scale, as done for this time-dependent model, is a rather crude approach because evidence from DOAC discontinuation indicates immediately rising risks after exposure changes 24. Moreover, applying mean prescription durations as a highly generalizing assumption of individual exposure also lacks the necessary granularity for immediate exposure–outcome relationships 25. A daily scale with individually approximated treatment durations would facilitate the precise incorporation of adherence 26, persistence 27 or treatment switches 16 as further critical issues in the model. Fortunately, with prevalent new-user designs and time-conditional PS approaches, promising advanced methods are available to address these aspects 28. Comparing the paper's results with those in the literature, the findings appear surprising, in terms of both direction and magnitude. However, when we update previous evidence (Figure 1A) with the results of Stolk and colleagues 13, this does not change the pooled effects (Figure 1B). New findings should always be placed into the context of prior evidence while taking due account of the limitations of old and new data. As the presented results address an important question and appear to contradict current evidence, it will be crucial to confirm these findings using further approaches that are less susceptible to bias. All authors have completed the Unified Competing Interest form at http://www.icmje.org/coi_disclosure.pdf (available on request from the corresponding author).
No takes yet. Share an insight, caveat, or question.
Meid et al. (2017) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: