Composite endpoints in clinical trials: the “win ratio”

I recently came across a paper by Redfors et al. (2020) on using the win ratio in clinical trials - https://doi.org/10.1093/eurheartj/ehaa665. This led me to the question: what is the “win ratio” in medical statistics, and is it useful for clinical trials with composite endpoints?

Context

Clinical trials aim to determine if a treatment or intervention can improve patient outcomes compared to a control (or other treatments). For instance, many cardiovascular (CV) trials investigate the impact of a treatment vs a control on the time until a clinical event (e.g. CV death, myocardial infraction hospitalization, or stroke hospitalization; whichever happened first).

Traditional survival analyses (e.g. Cox regression) then estimate a hazard ratio, reflecting the risk of a CV event over time in the treatment group vs the control group. This approach to analyzing a composite endpoint only looks at the first event, therefore all events are assumed to have equal importance. In reality, CV death is more serious than CV hospitalization, and there may be differences in the severity of sub-types of CV hospitalizations.

Pocock et al. (2012) proposed a new approach to analyzing composite endpoints in clinical trials where the outcomes within a composite endpoint are ranked based on priority. This approach yields the “win ratio”, representing the odds of the treatment resulting in a “win” vs the control. A detailed description of the win ratio approach is given in Pocock et al. here https://doi.org/10.1093/eurheartj/ehr352.

What is the win ratio?

Briefly, in the example above on CV events, all patients in each group are compared, and for each pair, the treatment patient is labelled a “winner” or “loser” depending on who had a CV death first. If the patients are tied on CV death (e.g. neither died, or both died on the same day), then the patient is labelled a “winner” or “loser” depending on who had a CV hospitalization first. The win ratio is the total number of “winners” divided by the total number of “losers”.

The win ratio is a relative estimate of the treatment effect. To obtain an absolute estimate of the treatment effect, the “win difference” can be calculated as the proportion of “wins” in the treatment group vs the proportion of “wins” in the control group (described in more detail here https://doi.org/10.1002/ejhf.2853).

How is the win ratio useful?

It priorities the more important components of a composite outcome (e.g. CV death over CV hospitalisation)
It allows flexible specification of a composite outcome (e.g. can include many different CV hospitalizations in order of severity)
Does not rely on assumptions like in traditional survival analysis techniques (such as proportional hazards)
Does not require specialist software to calculate
Can easily be translated into a measure of absolute effect
Confounding can be accounted for by matching on covariates beforehand

A practical example

Consider a trial investigating the effect of a treatment vs control on time to CV mortality and myocardial infarction (MI) hospitalization over one year.

Suppose there are 7220 participants in the treatment group, and 7969 participants in the control group. Looking at Table 1:

The total number of possible pairs is 7220 x 7969.
The win ratio is the number of “wins” in the treatment group (31 million) divided by the number of “wins” in the control group (19 million).
The win difference is the proportion of overall “wins” in the treatment group (31 million divided by 57.5 million) minus the proportion of “wins” in the control group (19 million divided by 57.5 million).
The number of ties represents patient pairings in which a “winner” could not be determined based on time to CV death or MI hospitalization.

In the presence of confounding, one can match based on a set of covariates beforehand (i.e. form pairs between patients in the treatment and control groups with similar covariate characteristics). This requires first obtaining equal-sized groups.

Table 1: Effect of a treatment vs control on the composite outcome “time to CV clinical events” over 1 year (fictional data for illustrative purposes).

	Sex-matched pairs	All possible unmatched pairs
Total no. of pairs	7220	57,536,180

Wins on death in treatment group	701	5,561,790
Wins on death in control group	353	2,917,464

Wins on MI in treatment group	3293	26,121,522
Wins on MI in control group	2042	16,397,486

Overall wins in treatment group	3994	31,683,312
Overall wins in control group	2395	19,314,950

Tied	1662	6,537,918

Win ratio (95% CI)	1.67 (1.59, 1.75)	1.64 (1.57, 1.71)
Win difference (%)	22.1	21.5

Questions and considerations

I’d like to get the community’s thoughts on how useful the win ratio is for clinical trials with composite outcomes compared to traditional methods of estimating relative effects of treatments.

I’m especially interested in the views of medical statisticians and those involved in the design of clinical trials, though views from non-specialists are also very helpful in understanding its interpretability.

I’d like to learn if anyone has practical experience deriving the win ratio using observational data where confounding can be a major issue, and the sample size is often larger than in trials.

Thanks RHub community :)

Preprint

Preprint

Composite endpoints in clinical trials: the “win ratio”