Intercurrent Events in Clinical Trials: When Rescue Therapy and Death Are Not Missing Data
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Trials often spend months polishing eligibility criteria, endpoints, and sample-size assumptions, then start improvising the moment patients need rescue therapy, switch treatment, discontinue because of toxicity, or die before a planned assessment. That improvisation usually gets described as an analysis detail. It is not. It changes the causal question.
These post-randomization events are called intercurrent events. They happen after treatment assignment but before the endpoint is fully observed or interpreted. The main mistake is to act as if they merely create inconvenient missing data. In many trials, they define whether the result is about treatment assignment in practice, biological effect absent switching, failure including rescue, or something narrower still.
The Core Decision Rule
When an intercurrent event occurs, do not ask first how to keep the model running. Ask what treatment effect the paper still wants to estimate after that event enters the story.
Decision rule:
If the analysis rule for rescue, switching, discontinuation, or death is chosen because it is convenient rather than because it matches an explicit clinical question, the trial is probably estimating the wrong thing with great precision.
This is why estimand language matters. Not because regulators demanded a new glossary, but because it forces a trial to say what counts as part of treatment, what counts as failure, and what world the estimate is trying to represent.
Why Intercurrent Events Are Not All the Same Problem
Rescue therapy
Rescue medication often means the assigned strategy failed to control symptoms well enough on its own. Censoring at rescue can hide that failure.
Treatment switching
Switching can blur the effect of initial assignment, but that does not automatically mean the right answer is a heroic adjustment to recover a no-switching world.
Death or discontinuation
Some later outcomes become undefined after death, while discontinuation may be part of the real tolerability profile. Neither should be waved away as a bookkeeping nuisance.
A Concrete Clinical Example
Imagine a 24-week randomized rheumatoid arthritis trial comparing a new biologic with standard therapy. Patients with uncontrolled symptoms may receive rescue steroids by week 8. Patients with serious infections may discontinue the biologic. A subset of control patients may switch to open-label biologic after prespecified failure criteria. The primary endpoint is low disease activity at week 24.
Bad reflex
Censor everyone at rescue or switching, then call the remaining week-24 scores the treatment effect.
What that really asks
A vague hybrid question about patients who stayed on path long enough to remain easy to analyze.
What the trial should ask instead
Is the goal to estimate the effect of assignment in real practice, the effect without rescue, or a composite definition where rescue counts as failure? Those are different estimands, not sensitivity labels.
Once that question is explicit, the analysis path becomes narrower. A treatment-policy estimand keeps rescue and switching inside the strategy. A hypothetical estimand needs defensible assumptions about what would have happened without the event. A composite estimand counts the event itself as part of the clinical outcome.
Intercurrent event strategy explorer
The right analysis depends on the question, not on how annoying the event was
Pick a post-randomization event and the scientific question you actually care about. The explorer maps that pair to the estimand strategy that usually fits best, along with the reviewer warning that should come with it.
Rescue therapy changes the patient path after randomization.
Keep post-randomization events inside the estimand. Rescue, switching, or discontinuation are part of the real treatment strategy rather than reasons to erase follow-up.
Why this event matters
A naive censoring rule can turn clinical deterioration into missing data, which usually answers a different question from the one in the abstract.
Practical default
For a pragmatic question, treatment policy is often the starting point. If rescue counts as failure clinically, a composite strategy can be more honest.
Reviewer caution
If the paper censors at rescue without naming the estimand, assume the authors may be replacing a clinical question with a convenience question.
Recommended estimand strategy
Treatment policy
Prefer analyses that preserve randomized follow-up and describe intercurrent-event frequencies clearly.
What to watch for
Review whether the endpoint and summary still answer a clinically interpretable policy question instead of blending incompatible states.
Best fit if the analysis, endpoint definition, and discussion are all aligned with this exact question.
| Bad reflex | Why it fails | Better move |
|---|---|---|
| Censor at the event because follow-up became messy | Messiness is not a scientific question. Censoring can silently redefine the target effect. | State the estimand first, then choose whether the event stays inside the outcome, inside the strategy, or inside a hypothetical world. |
| Treat death like ordinary missing data | Some outcomes are undefined after death, not merely unmeasured. | Consider a composite strategy or a clearly justified alternative that clinicians would recognize as meaningful. |
| Call every post-event adjustment “sensitivity analysis” | Different methods often answer different causal questions, so the outputs are not interchangeable robustness checks. | Label each analysis by its estimand and explain which one supports the headline claim. |
The Five Strategy Labels and What They Usually Mean
| Strategy | What it asks | Typical use | Main danger |
|---|---|---|---|
| Treatment policy | What happens after assignment as care actually unfolds? | Pragmatic trials, real-use decisions, tolerability-inclusive questions. | Can sound diluted if readers expected a no-switching biological effect. |
| Hypothetical | What would happen if the event never occurred? | No rescue, no switching, no discontinuation counterfactual questions. | Easy to request, hard to justify, often paired with weak assumptions. |
| Composite | Should the event itself count as part of failure or success? | Death before symptom endpoint, rescue-as-failure, treatment failure definitions. | Components may differ in clinical importance and frequency. |
| Principal stratum | What is the effect in patients who would avoid the event? | Specialized causal questions about latent subgroups. | The target subgroup is partly unobservable and the claim becomes narrow fast. |
| While on treatment | What happens before treatment stops? | Occasionally useful as supportive analysis for exposure-period questions. | Often overused as if it were the main causal answer instead of a different, fragile estimand. |
Failure Modes That Should Make a Reviewer Slow Down
Red flags
- The paper censors at rescue, switching, or discontinuation without naming the estimand.
- Death before outcome assessment is handled like ordinary dropout.
- The discussion interprets a while-on-treatment estimate as if it were the primary trial effect.
- Multiple post-event analyses are presented, but none is tied to the headline claim.
- The supplementary appendix lists methods, but the main paper never explains the clinical question each one answers.
Better reviewer questions
- What exact treatment effect is the authors' main claim trying to describe?
- Are rescue, switching, discontinuation, and death part of that estimand, or excluded by design?
- Does the chosen analysis method actually identify that estimand under stated assumptions?
- Would clinicians recognize the resulting question as meaningful for practice or policy?
- Which result should survive if the abstract had to use only one sentence?
What to Do in Practice
Start the protocol with a short table: likely intercurrent events, why they matter clinically, which estimand strategy each one gets, and what estimator will support it. This is a better investment than adding a heroic post hoc adjustment after the data turn unruly.
For reviewers, the key move is even simpler. Refuse to let the paper hide behind generic language like “patients were censored at rescue” or “a sensitivity analysis adjusted for switching.” Ask what question those choices answer. If the authors cannot say it in one clean sentence, the trial probably has not earned an interpretable headline.
If you want a fast critique of whether a manuscript's intercurrent-event handling, estimand language, and analysis plan actually agree, Aqrab is built for that kind of methods stress test. If you want to wire those checks into your own review workflow, the developer tools are the cleaner place to start.
The Bottom Line
Rescue therapy, switching, discontinuation, and death are not distractions from the treatment effect. They are part of the logic that defines it. A trial that handles intercurrent events casually may still produce a precise estimate, but precision is not the same thing as asking the right question.