Cardiovascular trials often include non-fatal events (in addition to mortality) in their primary endpoints maintain the study power i.e., to capture enough events to detect a treatment effect. The evaluation of the treatment under investigation usually involves time-to-first event analysis in which only the first event to occur “counts”. For example, if a patient has a hospitalization for heart failure and 2 weeks later dies from cardiovascular causes, only the hospitalization is counted in a time-to-first event model. This traditional approach has several advantages, including simplicity and reproducibility. However, an important problem is that the components of the endpoint are given equal statistical “importance”, even though they may not have equal clinical importance i.e., fatal and non-fatal events are considered “equal” in a time-to-first event model.
The “win ratio” is an adaptation of the Finkelstein and Schoenfeld method proposed by Pocock and colleagues which takes into account the clinical importance of the components of composite outcomes, as well as the relative timing of the component events by considering the “most important” component first. In other words, the win ratio starts with the last and “most important” outcome, usually cardiovascular death, and determines the “winner” (i.e., the treatment or the placebo in a pairwise manner and determines the group with less events – “winner”), then all the patients that have survived are compared for the next event (e.g., heart failure hospitalization) to determine the “winner”. This process can go on for more events (e.g., myocardial infarction or stroke) and even for “softer” outcomes such as quality of life measures or biomarkers such as NT-pro BNP. Although this approach may seem attractive it also has many caveats. For example, many drugs have a more pronounced effect on non-fatal outcomes (e.g., hospitalizations) and on many occasions death is difficult to delay because the patients are already “too sick” to experience a survival benefit. In such cases, the win ratio may provide a less pronounced effect (because it considers death first) estimate and does not reflect the full effect on non-fatal outcomes (that may also be of clinical importance). It can be impossible to establish the relative clinical “importance” of an event. For example, patients may value staying out of the hospital with a good quality of life more than living longer. The win ratio is also less flexible than the Cox model, regarding some statistical methods such as stratification, interactions, or covariate adjustment. In a time to first event analysis, mortality is also reported individually. In practical terms the win ratio and the time-to-first event Cox model provide very similar results, because in most instances the effect on non-fatal and fatal outcomes is well aligned i.e., patients who do experience a reduction in non-fatal events with a drug, usually do not experience excess deaths with the same drug. Thus, the overall treatment effect will not differ much between the win ratio and the Cox model.
Using the EPHESUS trial as an example, time-to-first event analysis found a hazard ratio of 0.87 (95% CI: 0.79 to 0.96). The win ratio for the same composite outcome tested in the order death from cardiovascular causes, stroke, myocardial infarction, and heart failure hospitalization was 1.15 (95% CI: 1.05 to 1.27; z-score = 2.8; p = 0.0026). Changing the hierarchy to death from cardiovascular causes, heart failure hospitalization, myocardial infarction, and stroke gave a win ratio of 1.18 (95% CI: 1.06 to 1.28; z-score = 3.0; p = 0.0012) the WR had a greater z-score or smaller p value reflecting a higher magnitude of the eplerenone effect on death from cardiovascular causes and heart failure hospitalizations.
One must be very careful in incorporating “soft” endpoints in the composite outcome. For example, examining the effect of a drug on a biomarker such as NT-pro BNP does not mean that that drug will reduce hospitalizations or mortality, it simply means that some drugs may have a more pronounced effect on the biomarker than others (e.g., sacubitril/valsartan reduces NT-pro BNP much more than sodium glucose co-transporter inhibitors, but both drugs reduce morbidity and mortality in heart failure with reduced ejection fraction, and sodium glucose co-transporter inhibitors in many other patient-populations). As in most circumstances most patients are “ties” i.e., fortunately in a trial the majority of patients do not experience any event, the win ratio is not necessarily better than a time to first event model to analyse such “soft” end points, as most comparisons will end up being made on the same patients in either case.
For all these reasons, the Cox model is here to stay, and the win ratio can certainly be used with sense and sensibility under some circumstances in which one believes that is important to consider death as the “most important” event. This should obviously be adequately justified and prespecified in the statistical analysis plan.