The Gold Standard of Experimental Design | Composite Endpoints, Component Dominance and Interpretability

1. Define the event and the estimand first

Back to contents

A composite endpoint may be the first occurrence of death, non-fatal stroke or hospitalisation within one year. The estimand must state the population, treatment conditions, exact components, follow-up window, summary measure and handling of intercurrent events.

“Any adverse outcome” is not enough. Component definitions, adjudication rules and timing determine which participant counts as having an event and what treatment contrast the analysis estimates.

2. Components need a coherent rationale

Back to contents

Components should be clinically or practically meaningful, related to the problem under study and plausibly affected by the intervention. Combining unrelated outcomes merely to raise event numbers can produce an effect that has no stable interpretation.

Importance need not be identical, but large differences demand care. A composite of death and a mild questionnaire change gives each participant the same event indicator while the consequences are not comparable.

3. Frequency creates component dominance

Back to contents

The most frequent component contributes most composite events and often determines the overall estimate. If that component is less important, more subjective or more responsive to local practice, a statistically persuasive composite may overstate the benefit readers care about.

Estimate expected component frequencies during design. Ask which component would dominate under the smallest clinically plausible effects and whether the composite would still support the intended decision.

4. Ascertainment must be comparable

Back to contents

Death is usually harder to misclassify than discretionary hospitalisation or treatment escalation. Open-label care, site customs and clinician awareness can therefore influence one component more than another.

Use blinded adjudication where appropriate, standard definitions, complete follow-up and site training. Report the number of suspected, confirmed and missing events by component and arm.

5. Time-to-first hides ordering and recurrence

Back to contents

A common analysis uses time to the first component. It gives priority to whichever event occurs first, not whichever is most severe. A hospitalisation can censor a later stroke from the primary analysis, and repeated hospitalisations may be ignored.

Consider supportive analyses for recurrent events, total burden, component-specific cumulative incidence or hierarchical methods when clinically justified. Prespecify them and distinguish confirmatory from descriptive roles.

6. Worked example: the composite improves, but why?

Back to contents

In an illustrative trial with 400 participants per arm and one year of follow-up, 80 control participants and 60 treatment participants experience the composite: 20% versus 15%. The risk difference is −5 percentage points, corresponding to a number needed to treat of 20 over one year if the effect is causal and applicable.

Suppose first events are mutually exclusive for this simplified arithmetic: hospitalisation is 50 versus 30, non-fatal stroke 20 versus 20, and death 10 versus 10. The composite contrast is driven entirely by hospitalisation. It does not show a reduction in stroke or death. In real data, participants can experience multiple component types, so component totals need not sum to first-event composite totals.

7. Show every component with compatible measures

Back to contents

Report counts, risks or rates, effect estimates and uncertainty for every component by treatment arm. Use the same follow-up frame where possible and show whether effects point in similar directions.

Component analyses may be underpowered and, unless protected by a prespecified multiplicity strategy, should not be promoted as separate confirmatory claims. Their role is often to explain the composite and reveal heterogeneity.

8. Changed-context transfer check

Back to contents

Move the design to an education programme where success means passing an exam, maintaining attendance or completing the course. Attendance is common and easy to change, while exam success may be rarer and more consequential.

If the programme improves attendance alone, the composite can rise without evidence of improved learning or completion. The transfer check shows the same dominance problem and adds another: “or” may not represent success coherently when one component is merely a precursor.

9. Limits and common misconceptions

Back to contents

A larger number of composite events does not guarantee greater information when components have different effects or competing risks. Similar direction is reassuring but does not make consequences equal. A non-significant component estimate is not proof of no component effect.

Do not interpret the composite hazard ratio as a common effect on every component. Do not reorder or redefine components after seeing results. Do not claim mortality benefit when the only observed difference is a softer endpoint.

10. Audit design, analysis and reporting together

Back to contents

Compare the protocol, registry, statistical analysis plan and report. Verify that component definitions, hierarchy, follow-up, adjudication, censoring, estimand and analysis method did not change after outcome data were available.

Recalculate the composite from participant-level component dates. Plot component frequencies and first-event order. Run prespecified sensitivity analyses for missing event status, competing events and alternative clinically defensible definitions.

11. Frequently asked questions

Back to contents

Must all components be equally important?

No, but large differences make the combined effect harder to interpret and increase the need for transparent component results.

Can the composite be significant when no component is?

Yes. Pooling events can increase power, while individual components have wider uncertainty. The combined result still needs clinical coherence.

Should components be tested separately?

They should be reported separately. Confirmatory testing of several components requires a prespecified multiplicity strategy; otherwise interpretation is usually descriptive.

12. The gold standard makes the driver visible

Back to contents

A useful composite captures one decision-relevant failure without hiding the outcomes that create it. Design should anticipate dominance; analysis should preserve timing and competing-event logic; reporting should show every component.

Before acting on a composite result, ask which component supplied the events, whether it matters to participants and whether its measurement was equally reliable across arms. The headline is credible only when the driver is clear.