Composite and Hierarchical Endpoints With Unequal Component Effects

Why component effects matter

A trial endpoint may combine several outcomes to increase the number of informative events or to reflect more than one way a patient can benefit. But a single summary can hide important differences: components may occur at very different rates, overlap in the same patients, differ in clinical importance, and respond differently to treatment. An overall treatment contrast is therefore not interpretable from the endpoint label alone. Its meaning depends on the rule used to combine components and on the joint distribution of those components.

This tutorial focuses on two approaches. A composite endpoint defines one event or score from multiple components, commonly counting a patient as having the event if any component occurs. A hierarchical endpoint compares outcomes in a prespecified order of clinical priority; a comparison on a lower-priority component matters only when the higher-priority comparison is tied. These rules answer different questions and generally have different power characteristics.

Composite endpoints: prevalence, overlap, and unequal effects

Let \(A\) and \(B\) denote binary component events measured on each participant. For an “\(A\) or \(B\)” endpoint, the probability of a composite event in a treatment arm is

The subtraction avoids counting participants with both events twice. The composite probability is not determined by the two marginal probabilities \(p_A\) and \(p_B\) unless their overlap is also known. Across randomized arms, the treatment contrast in composite risk is

where \(T\) and \(C\) denote treatment and control. Thus, even the component risk differences do not by themselves determine the composite risk difference: the change in overlap also matters. A component that is common can contribute many events, while a rare component may contribute little to the total event count. Neither fact establishes its clinical importance.

For a simple planning calculation, suppose participants are independently randomized in equal numbers, each participant contributes one binary composite outcome, and the planned analysis compares the two arm-specific event proportions. Assume a fixed assessment time, complete outcome ascertainment, and no clustering. Under these conditions, a usual large-sample approximation for the number \(n\) per arm for a two-sided test is

Here \(\alpha\) is the type I error rate, \(1-\beta\) is the target power, and the \(z\) values are standard normal quantiles. This approximation plans power for the specified composite contrast, not for each component separately. It relies on the proposed risks being credible and on the analysis matching the stated comparison.

Why dependence changes the planning picture

Dependence affects how many distinct participants experience at least one component. For fixed marginal risks, more overlap means fewer distinct composite events. It can also affect the treatment contrast if overlap differs between arms. This is why planning from component prevalences alone can misstate the composite risk and the expected treatment effect.

A practical justification follows directly from the identity \(P(A\cup B)=p_A+p_B-p_{AB}\): subtracting the two arms’ identities gives the composite contrast formula above. If overlap is unknown, a planner must either obtain defensible joint data, consider plausible overlap scenarios, or use simulation under an explicitly specified joint model. Assuming independence is one possible scenario, not a default truth; under independence \(p_{AB}=p_Ap_B\), an assumption that should be checked against available clinical knowledge or data.

Hierarchical endpoints: priority, wins, and ties

A hierarchical endpoint retains an explicit priority order rather than treating every component event as interchangeable. In a pairwise comparison, one participant from treatment is compared with one from control. If the treatment participant has a better outcome on the highest-priority component, that pair is a treatment win; if worse, it is a loss. Only a tie on that component permits comparison on the next component. A final tie is possible if the pair ties on all ranked components.

The hierarchy’s power depends on how frequently comparisons are decided at each level, as well as on the treatment effects at those levels. A common, highly prioritized component may decide many pairs before lower-priority outcomes are considered. A less common component may influence relatively few pairs even if its treatment effect is large. Conversely, a lower-priority component can contribute useful information when many pairs tie on higher-priority outcomes. Component dependence affects the frequency of those ties and the outcomes observed among them.

Planning, interpretation, and diagnostics

For either endpoint type, begin by stating the clinical question and the exact outcome rule. A composite asks about the probability, rate, or other summary of the defined combined event. A hierarchy asks how often treatment wins or loses under a specified ordering. These are different estimands, so choosing one should not be treated as a purely statistical way to increase power.

For binary composites, useful diagnostics include checking that proposed joint probabilities are feasible, calculating the implied composite rate separately in each arm, and confirming that the assumed risk difference has the intended sign and size. Compare implied rates with relevant historical or pilot data. For hierarchical endpoints, tabulate or simulate the expected proportions of wins, losses, and ties at each priority level. A large fraction of pairs decided at the first level means later components have limited influence on the overall comparison.

When dependence or the hierarchy’s win distribution makes a simple formula unreliable, simulate complete participant outcomes under the proposed joint distributions, apply the prespecified endpoint algorithm to each simulated trial, and estimate the proportion meeting the planned rejection rule. The simulation must represent the analysis and type I error target actually intended. It should include plausible alternatives for component effects and dependence, because a single assumed joint distribution can give a misleading impression of power.

No combined endpoint makes unequal component effects disappear. Report and interpret the overall result alongside component-specific event rates or comparisons, while recognizing that these component analyses may have different precision and may not have been powered for confirmatory conclusions. A statistically favorable composite may be driven mainly by a frequent, less consequential component; a neutral composite may also conceal divergent component effects. The endpoint should be judged against the clinical importance and interpretation of its components, not only against the anticipated power.

Common mistakes and limitations

• Adding component risks without subtracting overlap. For an “A or B” endpoint, use \(p_A+p_B-p_{AB}\). Adding marginal risks double-counts participants with both events.
• Assuming marginal effects determine the composite effect. They do not if overlap changes across arms. Specify or vary the joint probabilities.
• Calling a composite effect a uniform component effect. The contrast is for the combined endpoint; components can have unequal, null, or opposing effects.
• Interpreting a hierarchy as an ordinary average. Priority rules decide which component is considered for each pair; they do not assign a fixed numerical weight to each component.
• Using pair counts as independent observations. In a trial, many pairwise comparisons share participants. Treating all pairs as independent generally gives an incorrect variance and power calculation; use an appropriate variance method or trial-level simulation.

Composite endpoints can be sensitive to component definitions, event ascertainment, and the chosen assessment period. Hierarchical endpoints depend on the priority order and may leave lower-ranked outcomes with little influence when higher-ranked outcomes frequently decide pairs. Both approaches summarize multiple outcomes, so they can trade a single interpretable endpoint for less visibility into the separate clinical effects. Their power estimates are conditional on the assumed component distributions, dependence, analysis method, and trial design; they are not guarantees of realized power.