What is a composite endpoint?

A composite endpoint is a single trial outcome that counts a participant as having had an event when any one of several specified component outcomes occurs — for example death, myocardial infarction or stroke, whichever comes first. Composite endpoints raise the event rate, so a trial reaches adequate power with fewer participants, and they avoid testing several outcomes separately.

That definition is uncontroversial. Everything difficult about composite endpoints follows from one consequence of it: the analysis treats a death and an unplanned revascularisation as the same event, and the reader of the abstract usually cannot tell which one moved the result.

Why trials use composite endpoints

Composite endpoints solve a power problem and a multiplicity problem in the same move. The FDA’s Multiple Endpoints in Clinical Trials guidance for industry, finalised in October 2022, defines a composite endpoint as one “defined as the occurrence or realization in a subject of any one of the specified components”, and gives the reason plainly: “the incidence of each of the events may be too low to allow a study of reasonable size to have adequate power; the composite endpoint can provide a substantially higher overall event rate.”

The multiplicity argument is the second half. If cardiovascular death, myocardial infarction and stroke were three separate primary endpoints, the trial would need to spend its alpha across three tests. Combined into one variable tested once, the FDA notes, “no multiplicity problem will occur for this endpoint.” That is a genuine statistical gain, not a rhetorical one — a trial can be smaller, shorter, or both.

The third reason is scientific rather than statistical. A treatment such as an antiplatelet agent in coronary artery disease is expected to prevent several related events, and the trialist frequently does not know in advance which one will move. The FDA describes exactly this case: prevention of “myocardial infarction, stroke, or death”, “possibly without knowledge of which event(s) may be affected.”

The counting rule that does the work

A composite endpoint is almost always analysed as time to the first occurrence of any component, and that single convention generates most of the interpretive difficulty. The FDA states it directly: “an event is usually defined as the first occurrence of any of the designated component events.”

Three things follow, and none of them is obvious from a Kaplan–Meier curve. First, every component carries equal weight: a participant who dies contributes one event, and a participant who has an unplanned hospitalisation contributes one event, and the analysis cannot distinguish them. Second, a participant who is hospitalised in month three and dies in month seven contributes only the hospitalisation — the death falls outside the primary analysis and does not appear in the component breakdown. Third, recurrent events are discarded entirely; four heart failure hospitalisations count once.

The FDA is explicit about the second point when discussing decomposition: for participants with more than one event type, “events occurring after the first composite event (e.g., end-stage renal disease or death occurring after a doubling of serum creatinine) would not be counted in the decomposition.” A published component table is therefore a table of first events, not of all events, and the two can differ materially. Time-to-event definitions also carry their own hazards around when follow-up starts; see immortal time bias.

Discarding recurrent events is a design choice, not a necessity. When repeat events are themselves the clinical question — a heart failure trial where re-hospitalisation matters independently of whether it is the first or the fourth — some trials instead model all events with a recurrent-event method such as the Lin-Wei-Yang-Ying (LWYY) model, or a joint frailty model that treats death as a terminating event competing with further hospitalisations. These preserve information the first-event convention throws away, at the cost of an effect measure — a mean-events ratio rather than a hazard ratio — that reviewers and readers find less familiar.

Common questions about composite endpoints

What’s a composite endpoint, in plain terms?

A composite endpoint is a single outcome that counts a trial participant as having had an event the moment any one of several specified things happens to them — death, a heart attack, a stroke, whichever comes first. Trials use it to reach an adequate event rate with fewer participants, not because the components are considered clinically interchangeable.

If a patient is hospitalised and later dies, does that count as two events?

No. A composite endpoint counts only the first qualifying event per participant, so a hospitalisation followed later by death contributes exactly one event — the hospitalisation — and the death does not appear in the component breakdown at all. This is why a component table lists first events, not every event that occurred.

Why do trials lump death, heart attack and stroke into one number instead of reporting them separately?

Combining components raises the overall event rate enough that a trial can reach adequate statistical power with a realistic sample size, and it avoids splitting the trial’s significance level across three separate tests. The FDA’s October 2022 multiple-endpoints guidance states that a composite tested once, rather than three endpoints tested separately, avoids a multiplicity problem entirely.

I have three components in my composite endpoint and none of them would be individually powered on their own — is that a problem?

Not by itself — it is the ordinary reason composite endpoints exist. But it does mean any single component’s result in your paper is likely underpowered, so a non-significant component is not evidence the treatment has no effect on it; report it as a signal to interpret in the discussion, not a null result to claim in the abstract.

The gradient problem

The central criticism of composite endpoints is that their components frequently differ in importance to patients and in how strongly treatment affects them, and that these two gradients run in opposite directions. Ferreira-González and colleagues examined this in BMJ in 2007 in a systematic review of randomised cardiovascular trials reporting composite primary endpoints.

The pattern they described has two halves that compound each other. The less severe components — hospitalisation, revascularisation, symptomatic worsening — occur far more often than death, so they supply most of the events in the composite. And the treatment effect was typically larger on those less important components than on mortality. The composite therefore moves largely on the strength of the events patients care about least, while the headline result is read as though it were a statement about the most severe component.

Neither half is a defect in the arithmetic. Both are consequences of pooling outcomes of unequal severity into one count and reporting the pool as a single effect estimate.

The FDA guidance names the same failure mode as an interpretive limit: the effect on the composite “will not be a reasonable indicator of the effect on all of the components or an accurate description of the drug’s benefit if the clinical importance of different components is substantially different and the treatment effect is chiefly on the least important event.”

The gradient does not always run in that direction. When death itself is the most frequent component — as in composites used in critical-care or advanced-oncology trials with short expected survival — the pattern inverts, and the composite ends up dominated by the outcome patients and clinicians weight most heavily, not least. The underlying complaint is unweighted pooling of unequal severities; either direction of mismatch is the same defect, and a reviewer should check which way it runs before assuming the worse case applies.

Decomposition, and why it under-reports

Decomposition means reporting how the first composite events divided among the components, and it is the first thing an informed reader looks for. The canonical shape of the problem is a renal trial whose primary endpoint is the first occurrence of doubling of serum creatinine, end-stage renal disease or death — as in RENAAL (Brenner et al., New England Journal of Medicine, 2001). Doubling of creatinine is the most frequent of the three: in the FDA’s decomposition of RENAAL, 52% of first composite events were doublings of serum creatinine, 29% were deaths and 19% were end-stage renal disease events. A reader who sees only the composite hazard ratio has no way to recover that split.

Two cautions attach to any such table. It counts first events only, so deaths following a non-fatal component are absent; the FDA recommends additionally analysing “all events for the event type of interest (even those that occur after events of other event types)”. And component-level tests are structurally weak: “testing for individual component endpoints is likely to be underpowered as the sample size or total number of events is usually planned for testing the composite endpoint.”

This creates a trap for authors. A non-significant component result is not evidence of no effect on that component, and writing it up as though it were invites the post hoc power objection. The FDA’s own rule on adjustment turns on intent: examining components to understand the composite requires no multiplicity adjustment, while claiming a distinct effect on a component does.

What the reporting evidence shows

The published record on how composite outcomes are actually reported is poor, and it has been examined systematically. Cordoba, Schwartz, Woloshin, Bae and Gøtzsche reviewed randomised trials with composite outcomes in BMJ in 2010 and reported three recurring defects: authors almost never justified why those particular components were combined; the components combined were frequently of very unequal clinical importance; and the composite was described inconsistently across the abstract, the methods and the results of the same paper.

That last finding is the most practically useful one for an author. The defect is usually not a wrong analysis; it is that the composite is defined one way in the protocol, described more loosely in the methods, and reported in the abstract as if it were the most severe of its components. Pre-specification discipline is the remedy, and it is checkable against the registry record — see trial registration and endpoint discipline.

CONSORT 2010 item 6a already required completely defined, pre-specified primary and secondary outcome measures, including how and when they were assessed, and its successor, the CONSORT 2025 statement (Hopewell et al., published simultaneously in JAMA, BMJ and The Lancet in 2025), keeps outcome definition as a core reporting requirement. A composite is completely defined only when each component’s diagnostic criteria and ascertainment window are stated.

MACE is a label, not a definition

“Major adverse cardiac events” and “MACE” name no fixed set of components, and treating the acronym as self-explanatory is a recognised defect rather than a stylistic quibble. Kip, Hollabaugh, Marroquin and Williams set out the consequences in the Journal of the American College of Cardiology in 2008, in a paper whose subject was precisely that the coronary-intervention literature used the same three letters for materially different endpoints.

Their demonstration was to apply several different published MACE definitions to one cohort of patients treated with drug-eluting stents. The definitions did not agree: the same patients, the same follow-up, and a different answer depending on which components the acronym was taken to include. Their recommendation was to evaluate safety and effectiveness outcomes separately rather than to rely on a non-standardised label.

The practical rule for a manuscript is that the acronym must never appear without its expansion at first use, including whether it contains all-cause or cardiovascular death, whether it includes revascularisation, and whether periprocedural myocardial infarction counts and under which universal-definition type. Two papers reporting “MACE” are not comparable until those questions are answered, and a meta-analysis that pools them without answering them is pooling different quantities.

In practice, most papers use one of two conventions without naming it: “3-point MACE” (cardiovascular death, non-fatal myocardial infarction, non-fatal stroke) or “5-point MACE,” which adds hospitalisation for unstable angina and coronary revascularisation. Stating which of the two — or spelling out the actual list if it is neither — resolves most of the ambiguity Kip and colleagues describe, and costs one clause in the methods section.

Composite endpoints and competing risks

A composite endpoint of non-fatal events analysed alongside death raises a competing-risks problem that a standard Cox model does not handle. When death is a component, the competing-risks problem does not arise by construction, because death cannot preclude the endpoint — it is the endpoint. When death is excluded from the composite, participants who die can no longer experience the component events, and censoring them treats death as though it were uninformative loss to follow-up, which it is not.

This is the most common technical error attached to composites of hospitalisation, progression or morbidity that leave mortality out. Cause-specific hazards and subdistribution hazards answer different questions, and the choice must be stated rather than defaulted into by the software. Fine and Gray’s subdistribution hazard model (Journal of the American Statistical Association, 1999) is the standard method when the question is the cumulative incidence of the non-fatal component in the presence of death as a competing risk; a cause-specific Cox model instead estimates the event rate among those who remain alive and event-free, which is a different quantity with a different clinical reading, and neither is simply “more correct” than the other. The consequences are set out in more detail under competing risks reporting, and unexpected crossing curves are frequently the visible symptom — see my survival curves cross.

Hierarchical alternatives to counting first events

Hierarchical composite endpoints rank the components by clinical priority instead of counting whichever came first, and they are the main methodological answer to the gradient problem. The win ratio, introduced by Pocock, Ariti, Collier and Wang in the European Heart Journal (2012;33:176–182), forms matched pairs of treated and control participants, compares each pair on the most important outcome first (cardiovascular death), and only if that is tied moves to the next (heart failure hospitalisation). The win ratio is wins divided by losses; the authors illustrated it with EMPHASIS-HF, PARTNER B and CHARM, and the FDA’s 2022 multiple-endpoints guidance cites the paper when endpoints “are ordered based on clinical importance”.

Two related methods share the logic. The Finkelstein–Schoenfeld test (Statistics in Medicine, 1999) combines mortality with a longitudinal measure through pairwise comparisons on the same hierarchical principle, and the win ratio can be understood as a way of reporting that comparison as an interpretable ratio. Desirability of outcome ranking (DOOR) generalises the idea to an ordinal scale of overall patient outcome, in which each participant is placed on a ranked scale that combines efficacy and harm; Evans and colleagues introduced it for antibacterial trials in Clinical Infectious Diseases in 2015, and the approach has since been applied well beyond cardiology.

State the limitations honestly, because reviewers do. A win ratio is not a hazard ratio and has no risk-difference interpretation; its value depends on the follow-up distribution and on how ties are handled; and the unmatched form can be sensitive to censoring patterns that differ between arms. Hierarchical endpoints fix the weighting problem and introduce an estimand that is harder to explain to a clinician.

Questions about MACE and hierarchical endpoints

Is MACE always death, heart attack and stroke?

No — MACE has no standardised definition. Kip and colleagues showed in the Journal of the American College of Cardiology in 2008 that applying different published MACE definitions to the same drug-eluting-stent cohort changed the result, so a manuscript must expand and define MACE at first use rather than treating it as self-explanatory.

Why does every cardiology paper seem to define MACE differently?

Because MACE was never standardised to begin with — it is a label, not a specification. Papers disagree on whether death is all-cause or cardiovascular only, whether revascularisation counts, and which universal-definition type of periprocedural myocardial infarction is used, so two papers reporting “MACE” can describe materially different endpoints under the same three letters.

What’s a win ratio, and why would I use it instead of a hazard ratio?

A win ratio ranks a composite’s components by clinical priority — comparing matched treated-and-control pairs on death first, then on hospitalisation only if death is tied — rather than counting whichever event happens first regardless of severity. Pocock and colleagues introduced it in the European Heart Journal in 2012; it has no risk-difference interpretation.

My reviewer wants me to add a win ratio instead of my composite endpoint’s hazard ratio — is that actually necessary?

Not automatically. A win ratio answers the gradient problem — your least severe component driving the result — better than a hazard ratio does, but it introduces an estimand that is harder to explain and depends on follow-up length and tie-handling. Offer it as a sensitivity analysis alongside the hazard ratio rather than a replacement, unless the reviewer specifically objects to component weighting.

A composite endpoint is not a composite variable strategy

“Composite endpoint” and the “composite variable strategy” of ICH E9(R1) are different concepts that share a word, and conflating them produces confused methods sections. ICH E9(R1), the addendum on estimands and sensitivity analysis adopted on 20 November 2019, uses “composite variable strategy” to describe one of five ways of handling intercurrent events: “An intercurrent event is considered in itself to be informative about the patient’s outcome and is therefore incorporated into the definition of the variable.” The other four are the treatment policy strategy (use the outcome regardless of whether the intercurrent event occurred), the hypothetical strategy (estimate what would have happened had it not occurred), the while-on-treatment strategy (use only data up to the event), and the principal stratum strategy (restrict the estimand to the subpopulation that would not have experienced it).

The worked case in the addendum is treatment discontinuation for toxicity being scored as treatment failure, or death being assigned a value on a physical-functioning scale that reflects the absence of function. ICH E9(R1) also observes that “progression-free survival in oncology trials measures the treatment effect on a combination of the growth of the tumour and survival” — a construct that is simultaneously a composite endpoint and a composite variable strategy, which is precisely why the terms get muddled.

The distinction that matters in writing: a composite endpoint answers “what counts as an event”; a composite variable strategy answers “what do we do when something happens that changes the meaning of the measurement”. A modern methods section should specify both, and specifying one is not specifying the other.

Composite endpoints in non-inferiority trials

A composite endpoint interacts badly with a non-inferiority margin, and this is the least-discussed of its problems. A non-inferiority margin is justified against the effect size of the active control on the endpoint in question; adding low-severity, high-frequency components inflates the overall event rate, which makes a fixed relative margin correspond to a larger absolute number of severe events that may be tolerated as “non-inferior”.

The asymmetry compounds it. In a superiority trial, adding a component the treatment does not affect biases the estimate towards the null and reduces the sponsor’s power. In a non-inferiority trial, the same dilution moves the estimate towards the margin and works in the sponsor’s favour. Where a composite is used as a non-inferiority endpoint, the margin justification must be tied to the composite as constituted, not borrowed from a trial that used a narrower one.

This is also an assay-sensitivity problem in the ICH E10 sense: a non-inferiority trial only has the power to detect a true difference if one would have shown up under the trial’s conditions, and a diluted composite with a large pool of components the treatment does not move weakens exactly that sensitivity. A margin defended by historical trials that used a narrower, more severe endpoint is not evidence that the wider composite retains the same ability to detect a real loss of effect.

Reviewers of trials read these design choices as claims about clinical equivalence, and overstatement here is treated severely; the general pattern is covered under overclaiming and causal language discipline.

How to detect a mishandled composite endpoint in a manuscript

Detection is a small set of concrete checks against the manuscript text, not a matter of judgement. Run them in this order.

Compare the abstract’s endpoint wording to the methods. If the abstract says “reduced cardiovascular events” and the methods define a composite including revascularisation, the abstract has silently promoted the composite to its most severe component.

Look for the component table. If results are reported only for the composite, the paper is not decomposable and the FDA’s recommendation that “results for each component event should therefore be individually examined and should be included in study reports” is unmet.

Read the component table for direction and share. Compute what proportion of first events came from the least severe component. If the composite is significant and the mortality component points the other way, the discussion must address it explicitly.

Check the registry record. Compare the registered primary outcome to the reported one, component by component, and note any addition, removal or change of ascertainment window.

Check the ascertainment. Look for a clinical events committee, blinded adjudication, and pre-specified diagnostic criteria for each component. Soft components — hospitalisation, revascularisation, worsening symptoms — are the ones most sensitive to unblinded ascertainment, and they are also the ones that usually drive the result.

Check the competing-risks handling if death is not a component, and check whether recurrent events were discarded when the disease is one of recurrence.

What a peer reviewer says when it is mishandled

“The composite is driven by the least severe component.” “The choice of components is not justified, and their clinical importance is not comparable.” “Please report each component separately, with event counts and effect estimates.” “The abstract conclusion refers to cardiovascular events, but the composite includes unplanned revascularisation, which is ascertainment-sensitive in an open-label design.” “The composite as reported differs from the registered primary outcome.”

Two more appear frequently and are harder to answer after the fact. “Deaths occurring after a non-fatal component are not accounted for in the analysis” is a comment about the time-to-first-event convention that cannot be resolved without re-analysis. “Non-inferiority was declared on a composite whose margin was derived from a differently constituted endpoint” is a design objection, and it is generally fatal at review.

If a reviewer’s objection is that the composite was driven by the least important component, the honest response is a component table and a narrowed claim, not a defence of the composite. Guidance on framing that reply without conceding more than the data require is under how to write a response to reviewers, and the broader class of statistical objections under reviewer objections on statistics.

How to report a composite endpoint

Define every component with its diagnostic criteria, its ascertainment window, and who adjudicated it. State whether death is all-cause or cause-specific, and whether periprocedural events count. Name the counting rule — time to first event, recurrent events, or hierarchical — in the methods rather than leaving it to be inferred from the software.

Report the composite and every component in the same table, with event counts and effect estimates, and say in the text which component contributed the largest share of first events. Where mortality is a component, report it separately as a pre-specified secondary endpoint as well; the FDA guidance describes this arrangement for clinically critical endpoints too infrequent to serve as a primary.

Match the claim to the endpoint. If the composite includes revascularisation, the conclusion is about the composite, not about “cardiovascular events” in general. A single sentence stating what the composite does and does not license — written before the abstract is drafted — prevents most of the review comments above.

Frequently asked questions

Reviewer 2 says my composite endpoint is driven by the least important component — what do I write back?

Don’t defend the composite — concede the point with a component table showing event counts and effect estimates for each part, and narrow your abstract’s claim to match it. If revascularisation drove the result, the conclusion is about the composite, not about “cardiovascular events” broadly. A reviewer reads a defended composite as an unaddressed objection, not a rebutted one.

Can a composite endpoint be the primary outcome in a non-inferiority trial?

Yes, but the margin must be justified against the composite as constituted, not borrowed from a trial that used a narrower endpoint. Adding low-severity, high-frequency components inflates the event rate, which lets a fixed relative margin correspond to more absolute severe events being tolerated as “non-inferior” — the opposite direction from a superiority trial, where the same dilution hurts the sponsor instead.

My composite endpoint is significant but none of the individual components are significant on their own — is that normal?

Usually, yes. Component-level tests are structurally underpowered because the trial’s sample size was planned around the composite’s event rate, not each component’s separately. A significant composite with null individual components is not a red flag by itself; it becomes one only if the direction of effect on mortality or the most severe component points the wrong way.

How many components can a composite endpoint have?

There is no fixed limit, but each component should plausibly be affected by the treatment on its own mechanistic grounds and be defined with its own diagnostic criteria and ascertainment window. A component added mainly to raise the overall event rate, without a reason to expect an effect on it, dilutes the estimate toward the null and invites the “why is this one here” review comment.

Related

Competing risks · Trial registration and endpoints · Causal language · When an effect size is not meaningful · Clinical trial manuscripts

Checked before submission by trial registration and endpoint discipline, which compares the reported primary outcome to the registered one component by component and flags composites whose abstract claim is broader than the endpoint supports.

Review my manuscript