Two Different Questions About Trial Data
After randomization, a participant may stop the assigned treatment, begin rescue therapy, or miss a scheduled outcome assessment. These facts can occur together, but they answer different questions. An intercurrent event is a clinical event after treatment initiation that affects how the treatment condition or outcome is understood for the clinical question. A missing outcome is an outcome value that was intended for the analysis but was not obtained.
The distinction matters because an event does not tell us, by itself, whether an outcome was measured. A participant can discontinue treatment and still complete follow-up. A participant can begin rescue therapy and still provide every scheduled outcome. Conversely, an outcome can be unavailable even though neither event occurred. Treating these categories as interchangeable can obscure what happened clinically and how much outcome information is actually available.
The estimand framework distinguishes the clinical question from the procedure used to estimate its answer. Here, the immediate task is narrower: record clinical events and outcome observation separately. The question of how an estimand should handle a particular event is a separate decision; the event’s occurrence alone does not settle that decision.
Represent Events and Observation Separately
For clarity, consider one endpoint scheduled at a specified follow-up time. Let \(D=1\) indicate that treatment discontinuation occurred before that time, and \(D=0\) indicate that it did not. Let \(R=1\) indicate that rescue therapy was initiated before that time, and \(R=0\) indicate that it was not. Let \(M=1\) indicate that the scheduled outcome is missing, and \(M=0\) indicate that it was observed. These indicators describe different dimensions of a participant’s record.
The same participant may have both \(D=1\) and \(R=1\), or either event without the other. Independently, the participant may have \(M=0\) or \(M=1\). Thus, “discontinued” does not imply “missing,” and “received rescue therapy” does not imply “missing.” The definitions of the event and the assessment time must be clear enough that each indicator can be assigned consistently.
| Record | Classification | What it does not establish |
|---|---|---|
| Participant stops assigned treatment | Treatment discontinuation: an intercurrent event | Whether the scheduled outcome was obtained |
| Participant starts rescue therapy | Rescue therapy: an intercurrent event | Whether the scheduled outcome was obtained |
| Scheduled outcome value is not available | Missing outcome measurement | Whether an intercurrent event occurred |
| Outcome is recorded after discontinuation or rescue therapy | Observed outcome, with the event also recorded if it occurred | That the event should be ignored in defining the clinical question |
A useful data structure therefore keeps event dates and outcome availability in separate fields. The event fields record what happened and when; the observation field records whether the scheduled value exists. A single label such as “withdrawn” is not enough to establish that all later outcomes are missing. A participant may stop treatment but remain in follow-up.
Worked Example: Discontinuation With Continued Follow-Up
In a hypothetical trial, 100 participants are assigned treatment. Before the week-12 assessment, 20 discontinue assigned treatment. Of those 20, 18 provide a week-12 outcome measurement; the other 2 do not. Among the 80 who do not discontinue, 76 provide the outcome and 4 do not. Classify discontinuation and missingness separately.
The discontinuation count is 20, so the discontinuation proportion is \(20/100=0.20\), or 20%. This is an event summary, not a missing-outcome rate.
Among participants who discontinued, \(18\) of \(20\) have an observed outcome. Their observation proportion is \(18/20=0.90\), or 90%; equivalently, \(2/20=0.10\), or 10%, have a missing outcome.
Among participants who did not discontinue, \(76\) of \(80\) have an observed outcome, giving \(76/80=0.95\), or 95%; \(4/80=0.05\), or 5%, have a missing outcome. Overall, \(18+76=94\) outcomes are observed and \(2+4=6\) are missing, so the overall missing-outcome proportion is \(6/100=0.06\), or 6%.
The 20 discontinuations are not 20 missing outcomes: 18 participants discontinued but still provided the assessment. Nor are all six missing outcomes attributable to discontinuation: four participants without discontinuation also lack an assessment.
Rescue Therapy and Outcome Availability
Rescue therapy is a clinical event that may change the treatment received after randomization. Its occurrence should be recorded even when the outcome is fully observed. For example, an endpoint collected after rescue therapy remains an observed endpoint. Whether that value is relevant to a particular clinical question is a separate estimand issue; it is not a reason to relabel the value as missing.
Likewise, the absence of rescue therapy does not guarantee complete outcome collection. A participant may miss a visit for administrative, logistical, or other reasons. The event record and the measurement record must therefore be inspected directly rather than inferred from one another.
Worked Example: Rescue Therapy Does Not Mean Missing
In a hypothetical study of 60 participants, 12 begin rescue therapy before the scheduled week-8 assessment. All 12 provide the week-8 outcome. Of the 48 who do not begin rescue therapy, 3 have no week-8 measurement. Calculate the observation rates and classify the records.
Among participants who receive rescue therapy, the observation rate is \(12/12=1.00\), or 100%, and the missing-outcome rate is \(0/12=0\%\). Rescue therapy occurred, but no scheduled outcomes are missing in this group.
Among those who do not receive rescue therapy, the observation rate is \(45/48=0.9375\), or 93.75%, because \(48-3=45\) outcomes were measured. Their missing-outcome rate is \(3/48=0.0625\), or 6.25%.
Across the study, \(12+45=57\) of 60 outcomes are observed, so the overall observation rate is \(57/60=0.95\), or 95%. The three missing measurements occurred among participants without rescue therapy. These calculations describe the records; they do not show why the measurements were missing or how a particular estimand should treat post-rescue outcomes.
Why an Event Is Not a Missing-Data Mechanism
A missing-data mechanism describes how the probability that an outcome is observed relates to data, including potentially the outcome itself. An intercurrent event is a clinical occurrence. Even if an event and missingness are associated, the event label alone does not specify the mechanism. For instance, participants who discontinue may be more likely to miss an assessment, but that association does not establish which factors drove the missing measurements or whether the unobserved outcome values differ systematically.
A simple probability identity makes the distinction explicit. Let \(E=1\) denote a particular event, such as discontinuation, and \(M=1\) denote a missing outcome. The overall probability of a missing outcome can be partitioned by event status:
This identity describes how missingness is distributed across event groups. It does not say that \(E\) causes \(M\), that the event is the missingness mechanism, or that the missingness assumptions are satisfied. It is a bookkeeping relationship: each participant belongs to an event group, and missingness can be counted within each group.
Worked Example: Event Status and Missingness Are Separate
In a hypothetical trial, 80 participants are assessed for treatment discontinuation and a scheduled outcome. Thirty discontinue; 10 of them have a missing outcome. Of the 50 who do not discontinue, 5 have a missing outcome. Calculate the group-specific and overall missingness rates.
For discontinuers, \(P(M=1\mid E=1)\) is estimated by \(10/30=0.3333\), or 33.33%. For non-discontinuers, the corresponding rate is \(5/50=0.10\), or 10%.
Overall, 15 of 80 outcomes are missing, giving \(15/80=0.1875\), or 18.75%. Substitution into the partition gives
The data show that missingness is more frequent among participants who discontinued in this hypothetical sample. They do not show that discontinuation and missingness are the same thing, or identify a missing-data mechanism. Twenty participants discontinued but had observed outcomes, and five missing outcomes occurred without discontinuation.
Diagnostics and Common Pitfalls
A practical diagnostic is a cross-classification of event status by outcome availability. Make a separate table for discontinuation and for rescue therapy; if both events can occur, record them independently rather than forcing each participant into only one event category. At minimum, inspect the number of observed and missing outcomes within each group, along with the relevant denominators.
Check that the endpoint and assessment time are specified, that event timing is recorded relative to the assessment, and that “not measured” is not being confused with “not scheduled.” Compare planned assessments with collected assessments, and verify whether follow-up continued after treatment discontinuation. These checks can reveal data-processing errors, such as coding all post-discontinuation visits as missing despite recorded measurements.
The cross-tabulation is a descriptive diagnostic, not a complete investigation of missingness. A difference in observed rates may motivate further examination of available participant characteristics, visit histories, or reasons for non-observation. It cannot reveal the values of outcomes that were not observed. The standard terminology for missing-data mechanisms, including missing completely at random, missing at random, and missing not at random, concerns assumptions about the observation process; those assumptions cannot be assigned solely from the fact that an intercurrent event occurred.
- Pitfall: “All participants who discontinued have missing outcomes.” This is wrong unless the records show that no outcome was collected after discontinuation. Check each scheduled assessment.
- Pitfall: “Rescue therapy makes the endpoint missing.” Rescue therapy is an event; a measured endpoint after rescue remains observed. Its treatment in the clinical question is a separate matter.
- Pitfall: “No intercurrent event means complete data.” Missing measurements can occur for participants with no recorded event.
- Pitfall: “Higher missingness after an event proves the mechanism.” A difference in rates describes an association in the observed records; it does not by itself establish the process that produced missing values.
The key comparison is not between one event label and another, but between two axes: what happened clinically and whether the intended outcome was measured. Keeping those axes separate preserves the clinical description and makes the limits of the observed data visible.
Check Your Understanding
For each question, identify the event and observation information separately. Full solutions follow the questions.
- A participant stops assigned treatment at week 4 and provides the week-12 outcome. Is the week-12 outcome missing?
- In a hypothetical group of 25 participants who receive rescue therapy, 20 have an observed scheduled outcome. How many have a missing outcome, and what is the missing-outcome rate?
- Among 40 participants without treatment discontinuation, 6 lack the scheduled outcome. Does the absence of discontinuation imply complete outcome data?
- A trial reports that missingness is more common among participants who discontinued. What does this establish, and what does it not establish?
Solutions. (1) No. Discontinuation is an intercurrent event, but the week-12 outcome is observed because it was provided. (2) \(25-20=5\) participants have missing outcomes; the rate is \(5/25=0.20\), or 20%. Rescue therapy and missingness are recorded separately. (3) No. The missing-outcome rate in that group is \(6/40=0.15\), or 15%, despite no discontinuation. (4) It establishes a descriptive association between discontinuation status and outcome availability in the reported data. It does not establish that discontinuation is itself a missing-data mechanism, explain the missing values, or determine how the clinical estimand should handle the event.