When Death Changes the Question: Competing Risks, Intercurrent Events, and Truncation by Death
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Clinical papers love a tidy sentence like “deaths were censored” or “analysis was limited to survivors.” The trouble is that death is not one problem wearing one name badge. Sometimes death prevents a future event from occurring. Sometimes it is a post-randomization event that changes the treatment question. Sometimes it makes the later outcome literally undefined.
If you blur those situations together, you can run perfectly respectable software on the wrong scientific question. That is exactly the kind of mistake Aqrab should help teams catch: not whether the model converged, but whether the manuscript still knows what it is estimating.
The Fast Distinction
Competing risk
Death prevents the event of interest from happening later, as with stroke when some patients die first.
Intercurrent event
Death occurs after treatment assignment and changes what treatment effect the study can still claim.
Truncation by death
The later outcome does not exist after death, as with quality-of-life or functional status among the dead.
Decision rule:
Before choosing a survival curve, censoring rule, or imputation plan, ask whether the target is a real-world event probability, a treatment-policy effect after randomization, or a survivor-restricted outcome. Those are different questions, not different software defaults.
Why This Gets Messy in Practice
The same study can contain more than one of these frames. In a heart-failure trial, death may be a competing risk for hospitalization, an intercurrent event for treatment-policy interpretation, and a truncation problem for six-month quality-of-life scores. The paper does not get to pick one sentence and reuse it for all three.
This is where method sections start sounding polished while the estimand quietly drifts. “Deaths were censored” can mean a cause-specific hazard analysis. It does not automatically justify a claim about cumulative patient risk. “Survivors were analyzed” can mean the estimand shrank to a post-randomization subgroup. It does not automatically describe treatment benefit in the full randomized population.
A Concrete Clinical Example
Case
A frailty-focused pragmatic trial of post-discharge care
Imagine a pragmatic trial comparing enhanced transitional care with usual discharge planning after heart-failure hospitalization. Investigators report 30-day readmission, six-month quality-of-life, and all-cause mortality.
For readmission, death is a competing risk because a dead patient cannot be readmitted later. For quality-of-life at six months, death truncates the outcome because there is no post-death score to observe. For the treatment-policy question, death is also an intercurrent event because it is part of how assignment played out in real care.
If the paper censors deaths for every endpoint and then writes one triumphant conclusion, the analysis is cleaner than the scientific question. That is not rigor. It is category confusion.
Death-framework explorer
Pick the question first and the death problem stops pretending to be just one thing
The same death can be a competing risk, an intercurrent event, or truncation by death depending on the outcome and the scientific question. This explorer shows the frame that usually fits best and the reviewer trap that comes with it.
A time-to-event clinical endpoint that death can preclude.
Think of an older cardiovascular cohort where treatment might affect both stroke and all-cause mortality during follow-up.
This is the classic competing-risks setup. Death prevents future occurrence of the event, so cumulative incidence thinking is usually the right starting point.
Bad reflex
Treat death as ordinary censoring and then read the Kaplan-Meier curve as if dead patients were still at risk of stroke later.
Better move
Estimate cumulative incidence and choose cause-specific or subdistribution modeling only after naming the exact risk question.
Reviewer check
Ask whether the paper interpreted ordinary Kaplan-Meier curves as real-world risk despite a nontrivial death rate.
Chosen frame
Competing risk
If the paper claims something broader than this frame supports, the interpretation is probably outrunning the estimand.
Fit check
This framing can work if the endpoint definition, estimand, and interpretation all stay aligned.
| If the paper says... | Check whether it really means... | Because the main failure mode is... |
|---|---|---|
| “Death was censored.” | A cause-specific hazard estimand, not direct real-world risk. | Readers may mistake a rate model for a cumulative-risk claim. |
| “Only survivors were analyzed.” | A survivor-restricted question with post-randomization selection. | The paper may present a narrow survivor estimand as if it applied to everyone randomized. |
| “Deaths were handled in sensitivity analyses.” | Potentially several different estimands, not one robustness exercise. | Method labels start replacing scientific-question labels. |
What Usually Belongs to Each Frame
| Frame | Typical endpoint | What the paper should say | Common failure mode |
|---|---|---|---|
| Competing risk | Time to stroke, relapse, readmission, or device failure | Name whether the target is cumulative incidence or an etiologic rate comparison. | Treating ordinary Kaplan-Meier as if it were the patient's real-world event probability. |
| Intercurrent event | Fixed-visit trial outcome after rescue, switching, discontinuation, or death | State whether death is handled with treatment-policy, composite, or hypothetical logic. | Calling a convenience censoring rule an estimand strategy. |
| Truncation by death | Quality-of-life, cognition, functional recovery, biomarker change at a later visit | Admit that the outcome is undefined after death and that survivor-only summaries narrow the claim. | Pretending missing-data machinery can rescue an outcome that no longer exists. |
Three Reviewer Questions That Catch the Problem Fast
1. What outcome exists after death?
If the answer is “none,” you are no longer in ordinary missing-data territory. The outcome may be truncated by death.
2. Is the claim about risk or rate?
Cause-specific hazard ratios and cumulative incidence curves answer different questions once death can happen first.
3. Did the estimand shrink without warning?
If only survivors remain analyzable, the paper needs to say plainly whether it still speaks for everyone randomized.
Where Teams Usually Go Wrong
They reuse one death-handling rule across endpoints
Readmission, survival, and quality-of-life can require different framing even inside the same trial. Consistency in prose is not the same as coherence in estimands.
They let software vocabulary replace question vocabulary
Fine-Gray, censoring, principal strata, and composite endpoints are tools or strategies. None of them tells you the scientific question by itself.
They mistake survivor analyses for neutral restriction
Restricting to survivors is often a post-randomization selection step, not a harmless cleaning pass.
They hide death inside “sensitivity analyses”
If two analyses handle death differently, they may be targeting different estimands. That distinction belongs in the main paper, not as a footnote.
What Aqrab Should Help a Team Do Here
This is exactly the kind of manuscript weakness that slips through when review stays generic. A good methods review should ask what death means for each endpoint, whether the estimand is still stable, and whether the discussion section is claiming more than the analysis supports.
If your team wants a faster way to pressure-test endpoint logic, post-randomization events, and reviewer red flags before they harden into a manuscript, Aqrab's critique workflow is built for that kind of reading. If you need a more programmable pipeline for manuscript screening and structured methods checks, the developer surface is the better place to start.
The Bottom Line
Death is not a single nuisance parameter. It can block an event, redefine a treatment question, or erase the very existence of a later outcome. The honest move is to say which of those problems you have before you start modeling. Everything after that gets easier. Everything before that is usually just neat statistics wrapped around a blurry question.
Keep reading
Don't stop at one method.
Good methods judgment comes from contrast. Read the neighboring guides, see where the assumptions diverge, and avoid treating every observational problem like it needs the same hammer.
Differential Misclassification: When One Study Arm Gets More Chances to Be Wrong
A practical guide to differential misclassification for clinical researchers. Covers arm-specific outcome detection, adjudication asymmetry, false positives, missed events, and what reviewers should demand before trusting an effect estimate.
Adaptive Enrichment Trials: When Precision for One Subgroup Pretends to Be Evidence for Everyone
A practical guide to adaptive enrichment trials for clinical researchers. Covers predictive versus prognostic enrichment, assay timing, multiplicity, external validity, and what reviewers should demand before trusting a biomarker-selected win.
Treatment-Induced Mediator-Outcome Confounding: When Mediation Analysis Starts Chasing the Consequences of Treatment
A practical guide to treatment-induced mediator-outcome confounding for clinical researchers. Covers why natural direct and indirect effects fail when treatment changes later severity, toxicity, adherence, or surveillance that affect both the mediator and outcome.