Recurrent Events in Clinical Trials: When Time to First Event Hides Disease Burden
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
In a heart-failure trial, one patient is hospitalized once and another is hospitalized four times. A time-to-first-event analysis lets the first admission speak, then silences the next three. That may match the trial question. It may also discard the outcome burden treatment was meant to prevent.
Recurrent events—hospitalizations, infections, falls, exacerbations, seizures, or adverse events—need an estimand before they need a model. The central question is not “Which recurrent-event method is best?” It is “Which part of the patient journey is the study trying to compare?”
The Clean Metaphor: The First Alarm Is Not the Fire Record
A building’s first fire alarm matters. It does not tell you how many fires happened that year.
Time to first event asks how long patients remain event-free. Recurrent-event analysis asks about the process after patients re-enter the risk set. Those are related but different clinical summaries.
Interactive outcome explorer
Let the First Event Speak—or Let Every Event Speak
Switch among synthetic 12-month trials. Each dot is a hospitalization; × is death. The examples teach how the outcome definition changes what the same patient histories say. They are not effect estimates from a real trial.
Same first event, different burden
Both arms have the same first-event pattern. Repeat events accumulate only in Arm B.
Arm A
Arm B
What time to first sees
Median first event: Arm A 4.5 months · Arm B 4.5 months
All events after each patient’s first are discarded.
What total burden sees
Observed events: Arm A 4 · Arm B 10
Events after the first: 0 vs 6. Deaths: 0 vs 0.
Methods reading
A time-to-first analysis sees no separation here. A recurrent-event summary sees six additional events in Arm B. Neither summary is automatically primary; the clinical question decides which information matters.
Start With the Estimand, Not the Software Menu
| Clinical question | Possible summary | What it leaves out |
|---|---|---|
| Does treatment delay the first event? | Time to first event | Every later event. |
| Does treatment reduce event frequency? | Event rate or rate ratio | Timing and patient-level concentration unless reported separately. |
| How many events accumulate over time? | Mean cumulative count curve | Event severity unless severity is built into the outcome. |
| What is the path through recurrence and death? | Joint, multistate, or terminal-event-aware summary | Nothing automatically; interpretation depends on the states and assumptions chosen. |
A primary time-to-first analysis can be entirely defensible when avoiding any first event is the clinical objective. A recurrent-event analysis can be more faithful when repeated episodes are meaningful, reliably captured, and preventable. Calling one approach “more complete” does not make its estimand more relevant.
Five Failure Modes That Change the Claim
- The method is selected after the results are visible. First-event and recurrent-event analyses can disagree because they summarize different features. The hierarchy and interpretation should be prespecified.
- Every event is treated as exchangeable. A first hospitalization may differ from a fourth in severity, ascertainment, treatment, and susceptibility. Pooling needs clinical justification.
- Within-patient dependence disappears from the discussion. Events from one patient are correlated. A robust standard error can address one uncertainty problem; it cannot define the right risk set, repair event misclassification, or make the rate ratio answer a different question.
- Death is ordinary censoring. Death stops future recurrences and may be affected by treatment. A lower observed hospitalization count can reflect fewer admissions, earlier death, or both.
- A relative effect replaces the burden. A rate ratio without events per patient, cumulative counts over time, follow-up, deaths, and concentration among frequent-event patients is hard to translate clinically.
What the Common Models Are Actually Asking
A negative-binomial model often compares event rates while allowing more variation than a simple Poisson model. Andersen–Gill models treat recurrences as a counting process and commonly use robust variance for within-person dependence. Prentice–Williams–Peterson models condition later-event risk sets on having experienced earlier events. Marginal rate models target a population-level event-rate contrast.
These names are not interchangeable upgrades. They differ in risk sets, time scale, dependence assumptions, and effect interpretation. If the protocol says only “recurrent events were analyzed” and the paper reports one ratio, the reader cannot reconstruct the clinical question.
Death Is Not Just the End of Follow-Up
Imagine a treatment that increases early mortality. Survivors have little time left to accumulate admissions, so a naive recurrence count could look favorable. The opposite problem also occurs: a life-prolonging treatment can create more time alive in which nonfatal events can occur.
There is no universal correction. The trial must state whether it targets recurrences while alive, a composite involving death, a joint recurrence-and-death process, or another clinically justified summary. It should show deaths by arm and test whether conclusions depend on how the terminal event is handled. The estimand makes the value judgment visible; the model then estimates it under assumptions.
A Reviewer Red-Flag Checklist
- The endpoint is naturally recurrent, but only the first event is reported without justification.
- The recurrent-event method was not prespecified or differs from the registered analysis.
- Event definitions, severity, washout rules, and adjudication are unclear for later events.
- The analysis does not say when a patient becomes at risk again after an event.
- Death, withdrawal, and treatment discontinuation are all labeled non-informative censoring.
- A few patients contribute many events, but their influence is not shown.
- Only a rate ratio appears; absolute cumulative burden and uncertainty are missing.
- “Using all events increases power” is asserted without checking the event process or simulations used for planning.
A Practical Reporting Minimum
Report the number of patients with zero, one, and multiple events; total events and follow-up by arm; timing of events; deaths; event-definition and re-entry rules; the estimand; the model and its risk set; absolute cumulative burden with uncertainty; and a sensitivity analysis that changes the most consequential assumption. If first-event and recurrent-event summaries differ, explain why rather than promoting whichever produces the cleaner headline.
Why This Matters for Aqrab
Recurrent-event papers are a document-linking problem. The endpoint definition lives in the protocol, risk-set rules in the statistical plan, deaths in a separate table, and the claim in the abstract. A rigorous critique reconnects them before judging the model name.
Use Aqrab Try to pressure-test whether a manuscript’s recurrent-event conclusion matches its estimand, event process, and terminal-event handling. The credibility test is simple: could a reviewer explain exactly what patient journey the headline summarizes?
Methods Sources
- Amorim LD, Cai J. Modelling recurrent events: a tutorial for analysis in epidemiology. International Journal of Epidemiology. 2015;44(1):324–333.
- Rogers JK, et al. On reporting results from randomized controlled trials with recurrent events. BMC Medical Research Methodology. 2008;8:33.
- Ghosh D, Lin DY. Nonparametric analysis of recurrent events and death. Biometrics. 2000;56(2):554–562.
- Stegherr R, et al. The challenge of time-to-event analysis for multiple events: a guided tour from time-to-first-event to recurrent time-to-event analysis. Biometrical Journal. 2026.
- International Council for Harmonisation. ICH E9(R1): Addendum on estimands and sensitivity analysis in clinical trials.
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.
Multi-State Models: When One Time-to-Event Endpoint Hides the Clinical Path
A practical multi-state models guide for clinical researchers. Learn how state occupation, transition hazards, and time in state reveal clinical paths hidden by one survival endpoint.
Nonproportional Hazards: When One Hazard Ratio Pretends the Treatment Effect Never Changes
A practical guide to nonproportional hazards for clinical researchers. Covers delayed effects, crossing curves, waning benefit, why one hazard ratio can mislead, and what reviewers should demand instead.
The Hazard-Ratio Autopsy: When Delayed Effects Make One Number Misleading
A practical trial autopsy for non-proportional hazards. Read the curves, risk sets, landmarks, and restricted mean survival time before translating one hazard ratio into a constant treatment effect.