Intercurrent Events Versus Missing Outcome Data

Two Different Questions After Randomization

After randomization, participants may discontinue assigned treatment, receive rescue medication, start another therapy, experience a clinical event, or miss an outcome assessment. These developments matter, but they do not all create the same statistical problem. One question is what treatment effect is being asked about when a post-randomization event occurs? A separate question is what can be learned when a defined outcome was not observed?

The distinction matters because an event can be fully observed while the endpoint is also collected, and an endpoint can be missing even though no intercurrent event occurred. Treating these situations as interchangeable can change the clinical question, conceal missing information, or lead to an analysis that does not address the stated estimand.

To classify a situation, first specify the endpoint and its assessment time. Then ask whether the event changes how that endpoint should contribute to the treatment-effect question, and separately whether the endpoint value is available. These are not competing labels: a participant can have an intercurrent event, an observed endpoint, both an event and a missing endpoint, or missing data without an intercurrent event.

Classify the Event and the Observation Separately

The event classification depends on the clinical question, not simply on whether something unusual happened after randomization. For example, treatment discontinuation can be an intercurrent event when the question concerns outcomes regardless of discontinuation. Under a treatment-policy strategy, the participant’s outcome after discontinuation remains part of the treatment-condition comparison. Under a hypothetical strategy, the question instead concerns an outcome under a specified scenario in which the event did not occur. These strategies are defined in the earlier material on intercurrent-event strategies.

Neither strategy answers whether the outcome was collected. If the outcome after discontinuation is observed, it is not missing merely because treatment stopped. If it was not collected, it is missing even if the estimand’s strategy for discontinuation is already specified. Likewise, observing a rescue-medication event does not reveal a subsequent symptom score that was never recorded.

Events Can Change the Question; Missingness Can Obscure the Answer

An intercurrent-event strategy specifies how outcomes associated with the event relate to the target question. A treatment-policy strategy uses outcomes regardless of the event; a hypothetical strategy asks about outcomes under a stated scenario in which the event did not occur. A composite strategy incorporates the event into the endpoint, while a while-on-treatment strategy focuses on outcomes before the event. The appropriate choice depends on the clinical question and endpoint definition; it cannot be selected just by looking at which values are available.

Missing outcome data require a different kind of justification. The outcome value is still the intended endpoint at the specified time, but it has not been observed. Analyses must therefore rely on assumptions or models about the unobserved values, supported where possible by the reasons for missingness and the observed data. A missingness assumption cannot determine what the estimand should mean, and an intercurrent-event strategy cannot recover a value that was never collected.

A Short Justification: Why Missing Values Are Not Events

A simple calculation shows why the observed outcomes alone need not represent all outcomes when some values are missing. Within a treatment group, suppose \(Y\) is a defined outcome and \(R\) indicates whether it is observed. The mean among observed participants is \(E(Y\mid R=1)\), whereas the mean across the group is \(E(Y)\). If observation is independent of the outcome, then

But if participants with worse outcomes are less likely to be observed, then the distribution of \(Y\) among those with \(R=1\) can differ from the full group’s distribution, and the equality need not hold. This issue concerns the relation between the outcome and its observation indicator. It does not tell us whether discontinuation, rescue medication, or another event changes the clinical question. The event and observation status must therefore be recorded and reasoned about separately.

Diagnostics and Common Failure Modes

A useful review starts with the endpoint definition, assessment time, event record, and outcome-collection record. For each participant, compare event timing with the scheduled assessment: did the event occur, was the outcome still defined under the endpoint, and was a value collected? Tabulating missingness by treatment group, event status, event timing, site, and recorded reasons can reveal data gaps or patterns that deserve investigation. These checks can identify inconsistencies, but observed patterns alone cannot prove how unobserved outcomes would have looked.

• Calling every post-randomization event “missing data.” An event may occur with the endpoint fully observed. Record the event and the value’s availability separately.
• Calling every unavailable value an intercurrent event. A failed assay or missed visit may create missing data without changing the treatment-effect question.
• Using an event strategy as if it were a missing-data method. A treatment-policy strategy defines how event-associated outcomes enter the question; it does not supply outcomes that were not measured.
• Changing the clinical question to avoid missingness. Restricting attention to observed or on-treatment outcomes may define a different target, rather than resolve unavailable endpoint values.
• Assuming an event makes the endpoint undefined. This depends on the endpoint. As discussed in the earlier material on death as an intercurrent event or endpoint component, death may be handled as an endpoint component or as an intercurrent event, depending on the clinical question and endpoint definition. It should not automatically be treated as an ordinary missing score.

A practical comparison is useful: an intercurrent-event strategy answers how should the event relate to the treatment-effect question? Missing-data assumptions and methods address what can be inferred about defined outcomes that were not observed? Both may be needed in one trial. Neither category is a substitute for clearly defining the endpoint and assessment time.