Intercurrent Events Versus Missing Data

Why the Distinction Matters

After randomization, a participant may stop assigned treatment, start rescue medication, die, or miss a scheduled assessment. These occurrences are not all the same kind of problem. Some are intercurrent events that affect how an outcome is interpreted in relation to the assigned treatment. Others are simply missing data: the outcome was not recorded when intended. An event can also lead to missing data, so the two categories can occur together.

The distinction matters because an estimand defines the target question, while an analysis uses available data to estimate that target. An intercurrent event may require the target to specify how the outcome is interpreted when that event occurs. Missing data, by contrast, concern the availability of measurements needed to estimate the target. Treating missing values as if they were events can alter the question; treating an event as if it were only a recording problem can leave the question unclear.

Classify the Occurrence, Then the Measurement

A useful first question is, “What happened to the participant or treatment?” A second, separate question is, “Was the outcome value observed?” Keeping those questions distinct prevents one occurrence from being forced into a single category when it has two consequences.

• Treatment discontinuation: stopping the assigned treatment after initiation is generally an intercurrent event. It does not, by itself, mean that later outcomes are missing. A participant can discontinue treatment and still attend visits and provide outcome measurements.
• Rescue medication: starting additional medication to manage symptoms or deterioration is generally an intercurrent event because it can affect the outcome and its interpretation relative to the assigned treatment. Outcomes may still be measured after rescue medication begins.
• Death: its classification depends on the endpoint. For a symptom score intended to be measured at a later visit, death before that visit may make the score undefined and is an intercurrent event. For an endpoint that is death itself, death is an observed outcome, not a missing value.
• Missed assessment: a missed visit or unrecorded measurement is missing data, not automatically an intercurrent event. The reason may be an intercurrent event, such as hospitalization or treatment discontinuation, but the missed measurement and the underlying occurrence are distinct.

The word “generally” is important. Classification is not determined by the name of an occurrence alone. It depends on the outcome variable and the question the estimand is intended to answer. A death is directly observed for a mortality endpoint, while the symptom score that would have been recorded after that death is not merely an uncollected score: the participant is no longer in a state for which that score is defined.

The Target and the Available Data Are Different Things

The five attributes of an estimand include the population, treatment conditions, outcome variable, population-level summary, and a specification for handling intercurrent events. These attributes define the question of interest. Whether a participant’s outcome was recorded is a separate issue that arises when estimating that question from trial data.

For example, suppose the intended outcome is a symptom score at week 12 for participants randomized to each treatment condition. If a participant stops treatment but still supplies a week-12 score, the event has occurred and the outcome is observed. If another participant attends the visit but the score is accidentally not recorded, there is no such treatment event, but the outcome is missing. An analysis that drops both participants for the same reason would conflate event status with data availability.

The following short derivation shows why an observed-case average need not estimate the average for the intended group. Let \(Y\) denote an outcome that is defined for everyone in the target group, and let \(R=1\) indicate that \(Y\) is observed. Write \(q=P(R=1)\), \(\mu_1=E(Y\mid R=1)\), and \(\mu_0=E(Y\mid R=0)\). When both groups are present, the overall mean is a weighted average:

Subtracting the overall mean from the observed-case mean gives:

Thus, the observed-case mean equals the overall mean if all outcomes are observed (\(q=1\)) or if the mean outcome is the same among those with and without observed values (\(\mu_1=\mu_0\)). Otherwise, the two means differ. This algebra does not tell us which mean is the estimand; the estimand determines the target, and assumptions about missing outcomes determine whether and how the available data can estimate it.

Death and Endpoint Definition

Death deserves special attention because it can change whether an outcome exists, not just whether it was recorded. Suppose the endpoint is a symptom score at week 12. If a participant dies at week 8, there is no week-12 symptom score for that participant to collect. Calling this only “missing data” suggests that the same score exists but was not observed, which may not match the endpoint’s definition. In contrast, if the endpoint is death by week 12, the event is an observed outcome.

What the Distinction Means for Adaptive Trials

In an adaptive trial, treatment arms, allocation, population, or follow-up may change according to the trial’s design. The target estimand still needs a clear definition, including the treatment conditions and outcome. If the event classification or measurement process changes across trial stages, simply pooling records can obscure whether the underlying target remains the same.

For example, a protocol change that increases follow-up completion may reduce missing assessments without changing the outcome variable. A change in how rescue medication is used may instead affect the interpretation of outcomes under the treatment conditions. These are different implications. The first concerns how completely the target outcome is observed; the second concerns an intercurrent event and the target’s specification for it. A change in data collection is not automatically a change in the estimand, and a change in event patterns is not automatically a change in missingness.

Missing-data assumptions also matter. In broad terms, data are missing completely at random when the chance an outcome is missing does not depend on observed or unobserved outcomes. They are missing at random when, conditional on specified observed information, missingness does not depend on the unobserved outcome. If missingness still depends on the unobserved outcome after conditioning on observed information, it is missing not at random. These are assumptions about the missingness process, not types of intercurrent events. They cannot be confirmed from the observed outcomes alone and should be considered in light of the trial context.

Diagnostics, Alternatives, and Common Pitfalls

A practical audit should compare event records with outcome availability. Tabulate discontinuation, rescue medication, and other relevant events by assigned treatment and trial stage; separately tabulate the number and timing of missing assessments, with reasons where known. Check whether post-event outcomes were sought and recorded. An event that frequently coincides with a missed assessment may require attention both in the estimand definition and in the missing-data analysis.

There are two different questions to answer. One is, “What target should be defined when the event occurs?” That is part of specifying the estimand. The other is, “How can unavailable measurements be handled while estimating that target?” That requires a statistical analysis and assumptions about the missing data. Complete-case analysis, for instance, uses observed outcomes only; it does not automatically resolve whether those outcomes represent the target group. Alternative estimation methods may use observed participant information to address missing values, but none can recover information without assumptions.

Common mistake: “Participants who stopped treatment have missing outcomes, so discontinuation is just a missing-data problem.” The reasoning is wrong because discontinuation is an event after treatment initiation, while the outcome may still be measured afterward. The correction is to record discontinuation as an intercurrent event and separately determine whether each scheduled outcome was observed. A related mistake is to call every absent score an intercurrent event. A missed appointment is missing data; the reason for that absence may or may not be an intercurrent event.

The principal limitation of this classification exercise is that labels alone do not settle the scientific target. The endpoint definition and intended interpretation must be explicit, especially when death makes an outcome undefined. Nor does classifying an event solve missing-data bias: estimation still depends on the amount, pattern, and plausibility of assumptions about unavailable outcomes.