Composite Strategies for Intercurrent Events

Why Build the Event Into the Endpoint?

A clinical outcome measured after an intercurrent event may not answer the intended treatment question. For example, a participant who takes rescue therapy may later report improved symptoms, but that improvement occurred after an additional treatment. If the clinical question is whether the assigned treatment achieves improvement without rescue, the endpoint must say what rescue means for success. A composite strategy makes that choice directly: it incorporates the event into the endpoint, often by defining the event as treatment failure.

The resulting estimand is not simply an ordinary outcome with an event ignored. It answers a different, explicit question about the probability of meeting a clinical criterion under the assigned treatment while avoiding the specified event. The event rule must be chosen before analysis and stated precisely enough that two analysts would classify each participant the same way.

This strategy does not require claiming that a participant who experiences the event would have failed to respond in the absence of that event. Instead, the endpoint rule declares that the event itself precludes success for the purpose of the clinical question. The observed post-event clinical outcome may still exist, but it does not reverse the failure classification.

Formal Definition and Identification

Let \(a\) denote a treatment condition, and let \(R^a\) indicate whether a participant would meet the specified clinical response criterion under that condition. Let \(E^a=1\) indicate that the participant would experience the specified intercurrent event within the prespecified window under condition \(a\). Define the potential composite outcome \(S^a\) by

Here \(I(\cdot)\) is an indicator: it equals 1 when its condition is true and 0 otherwise. The multiplication makes \(S^a=1\) only if both conditions hold. A natural estimand for a binary success outcome is the population risk difference

where \(1\) and \(0\) label the two treatment conditions. A positive value favors condition 1 when success is desirable. The estimand’s target population, treatment conditions, response criterion, event definition, and event window must all be specified. These are not details that can safely be left to the data analyst.

The identity \(S^a=I(R^a=1,E^a=0)\) follows because the product of two indicators is 1 exactly when both indicators are 1. Taking expectations gives \(E(S^a)=P(R^a=1,E^a=0)\), since the expectation of a binary indicator equals the probability of its event. Thus the success risk is the probability of response and no event, not the response probability among participants who happened to avoid the event.

In a randomized trial, the treatment-specific composite risks can be identified from the corresponding randomized groups when assignment is randomized, the observed endpoint under assigned treatment agrees with the relevant potential outcome (consistency), participants do not affect one another’s outcomes, and event and endpoint status are ascertained. Randomization does not repair unrecorded events or missing endpoint assessments. If the composite outcome cannot be classified for some participants, additional missing-data assumptions and methods are needed.

A Bayesian analysis may place a probability model on the composite outcomes and use the resulting data to update uncertainty about the treatment-specific risks or their contrast. That analysis does not alter the estimand: the composite endpoint rule determines what success means before a prior or likelihood is chosen.

Constructing the Rule Carefully

A usable composite definition specifies the event, the time window in which it counts, and exactly how it changes the endpoint. It also fixes the clinical response criterion and the assessment time. For example, “rescue counts as failure” is incomplete if the protocol does not say which rescue treatments qualify, when the event window starts and ends, or whether a participant who takes rescue after the scheduled assessment is classified as a failure.

A composite strategy often makes a binary endpoint easy to apply, but the rule is not limited to binary clinical measurements. A continuous measure can be converted into a threshold-based response criterion and then combined with the event rule. That conversion changes the endpoint: it no longer targets the mean continuous outcome. The threshold and its clinical rationale therefore need to be prespecified and communicated.

The event classification can also be the endpoint itself when the scientific question defines the event as failure. For instance, a fixed-horizon estimand could compare the probability of avoiding hospitalization through day 90. This is still a composite-style treatment of the event when the broader endpoint definition declares hospitalization a failure; it is not an analysis of a post-event symptom score. The exact target should be described in words as well as notation.

How This Differs From Other Strategies

The composite strategy is distinct from the treatment-policy and hypothetical strategies. Under a composite strategy, the event changes the endpoint classification according to a prespecified rule. Under a treatment-policy strategy, outcomes after the event remain part of the endpoint as observed. Under a hypothetical strategy, the target concerns an explicitly defined scenario in which the event did not occur. These strategies can lead to different estimands even when they use data from the same trial. The earlier material on treatment-policy and hypothetical estimands provides the context for those alternatives.

In particular, a composite rule does not estimate what the clinical response would have been had rescue therapy or hospitalization been prevented. It assigns an outcome under an explicit endpoint convention. Nor is it generally valid to analyze only participants who avoided the event: that subset is defined after randomization and may differ across treatment groups. Such a comparison need not preserve the protection of randomization.

Diagnostics, Interpretation, and Limitations

Before interpreting a composite result, verify that the event rule was applied consistently in both groups. Check the event definition, timing, adjudication process, and whether treatment assignment could influence how often or how carefully the event was recorded. Confirm that the clinical response measure and assessment time are the ones specified in the estimand. A discrepancy between the protocol rule and the analysis coding changes the estimand actually reported.

A composite result can also conceal important patterns. Two groups may have similar overall success risks while differing in the proportions who failed because of the intercurrent event versus because they did not meet the clinical response criterion. Reporting the total composite contrast alone may therefore be insufficient for clinical interpretation. The component counts can help explain the result, but they do not replace the prespecified composite estimand.

• Ambiguous classification: If it is unclear which treatments, events, or time windows count, different analysts may construct different endpoints. Specify them before examining outcomes.
• Different clinical meanings: A nonresponse and a hospitalization may both be coded as failure but have different severity and consequences. Explain why the common failure classification answers the clinical question.
• Differential ascertainment: If event recording differs by treatment group, the observed composite risks can reflect measurement differences as well as outcome differences. Use consistent surveillance and adjudication.
• Missing status: An unobserved event or response is not automatically a failure. The composite rule handles defined event occurrences; it does not, by itself, resolve missing data.
• Overinterpretation: A lower composite failure risk does not establish that every component improved, or that the treatment caused a particular mechanism of improvement. Report the target and the observed components accurately.

The main limitation is interpretive: the clinical meaning depends heavily on the chosen event rule. A composite can make a question operational and clinically relevant, but it can also obscure whether the contrast reflects response, event frequency, or both. It should not be described as a neutral data-processing choice. It is part of the estimand itself.