Ranking of the most effective treatments for cardiovascular disease using SUCRA: Is it as sweet as it appears?

The number of clinical trials registered has increased dramatically over the past two decades. Most of these studies are designed to compare a treatment with placebo, treatment as usual, or no treatment. As a result, clinical trials that provide direct comparisons among two or more effective treatments are lacking despite the possibility that such comparisons could readily inform clinical practice. To address this limitation, network meta-analyses are now being routinely conducted to evaluate all treatments simultaneously and generate pooled estimates for all comparisons, even when few or no such comparisons exist in the literature. The analysis of these networks frequently relies on Bayesian principles to estimate measures of association (e.g. the odds ratio (OR)) along with their associated interval estimates. These interval estimates are denoted as credible intervals (CrIs), instead of the more frequently encountered confidence intervals, to reflect the way Bayesian estimates are derived from the distribution of plausible (or credible) values generated by the model. In this issue, Khan et al. report the findings of a network meta-analysis of statins, ezetimibe with or without statins, proprotein convertase subtilisin-kexin type 9 (PCSK9) inhibitors and placebo in adult patients with hypercholesteremia. Thirty-nine randomized controlled trials with 189,116 patients were included in the analyses. The authors found that PCSK9 inhibitors significantly reduced the risk of major adverse cardiac events (MACEs) (myocardial infarction (MI), stroke and all-cause mortality), compared to statins (OR 0.78, 95% CrI 0.62–0.97), ezetimibe plus statins (OR 0.72; 95% CrI 0.55–0.95) and placebo (OR 0.63; 95% CrI 0.49–0.79). Statins were associated with a significantly reduced risk of all-cause mortality (OR 0.88; 95% CrI 0.83–0.94). The surface under the cumulative ranking curve (SUCRA) metric was used to rank the effectiveness of each treatment and identify the best treatment. In this brief editorial, some points to consider are presented to help aid the interpretation of their, along with other, network meta-analysis. Despite the ability of a network meta-analysis to generate paired comparisons, the findings from network meta-analyses are often not sufficient enough for clinicians and patients to choose the best treatment and avoid the worst treatment, especially when there is no obvious winner. Although PCSK9 inhibitors showed advantages over other treatments on MACEs in the study by Khan et al., it is unclear whether PCSK9 inhibitors are still the best for other outcomes, including MI, stroke and mortality, and it is unclear which treatment should be the secondary choice after PCSK9 inhibitors. The authors tried to aid the interpretation by using SUCRA summaries of the data. Methods have been developed to rank the probability of being the best treatment (i.e. treatment A has a 70% probability of being most effective) and order all treatments from the best, second best, and so on. However, treatments being within a range of ranks are, sometimes, equally important, as the best treatment may be too expensive, may not have been approved by regulatory agencies, or may have serious side effects. Instead of ranking the probability of the best treatment, Salanti et al. developed a method to estimate the probability of all possible ranks, from being the best, second best,. . . to the worst, for each treatment. A graphical presentation can be generated to plot the probabilities against all possible ranks (a ‘rankogram’). The SUCRA line shows the percentage of effectiveness of each treatment accounting for all possible rankings and uncertainties in treatment effects. SUCRA values range from 1, being the best without uncertainty, to 0, being the worst without uncertainty. In their study, Khan et al. found that PCSK9 inhibitors were ranked the best for MACEs, MI and stroke,