Interval Censoring: When the Visit Date Pretends to Be the Event Date
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
A scan finds recurrence at month six. The spreadsheet records recurrence at month six. The Kaplan–Meier curve drops at month six. Yet the study did not observe recurrence happening then. It observed no recurrence at the previous scan and recurrence at this one. The event occurred somewhere between them.
That is interval censoring: the event time is known to lie inside a window, not at an exact point. Treating detection time as event time replaces an honest interval with invented precision. When visit timing differs across patients or groups, that shortcut can also change the comparison itself.
The Calendar Has Two Truths
Consider a patient whose imaging is negative on day 84 and positive on day 168. The clinical record supports one statement: recurrence occurred in the interval (84, 168]. It does not reveal day 84, day 126, or day 168 as the true event time.
The clean metaphor
A scheduled visit is a lighthouse sweep, not a timestamp. It tells you the event became visible between the last beam and this one. It does not tell you when, inside the darkness, the change occurred.
This arises whenever an event requires active detection: radiographic progression, asymptomatic infection, loss of viral suppression, cognitive decline past a threshold, graft dysfunction, or disease onset found at periodic screening.
Interactive detection-window explorer
Same event. Different visit schedule. Different recorded date.
Give two study groups the same unobserved event day, then change how often each group is assessed. The event does not move. The date in a naive dataset does.
Group A
Every 28 days- Last negative
- Day 28
- First positive
- Day 56
The honest record is (28, 56]. Coding the first positive visit invents day 56; midpoint imputation invents day 42.
Group B
Every 84 days- Last negative
- Day 0
- First positive
- Day 84
The honest record is (0, 84]. Coding the first positive visit invents day 84; midpoint imputation invents day 42.
Precision the study never observed
Both groups have the same true event day. If first detection is treated as the event, Group B appears to experience the event 28 days later. An interval-censored analysis keeps each group's last-negative-to-first-positive window intact.
Teaching model only. Real analyses contain many overlapping intervals, right-censored observations, missed visits, covariates, and uncertainty. The true event day is shown here only so the measurement error is visible.
Four Observation Patterns Reviewers Should Separate
| What was observed? | Data structure | Honest statement |
|---|---|---|
| Death recorded continuously | Exact event | T = day 47 |
| Negative test, then positive test | Interval censored | Day 28 < T ≤ day 56 |
| Positive at the first assessment | Left censored | T ≤ first assessment |
| Event-free at the last assessment | Right censored | T > last assessment |
The important distinction is not whether a date exists in the electronic record. It is whether the event process was observed continuously enough for that date to represent occurrence rather than detection.
Why the Convenient Shortcuts Fail
First positive visit
This systematically moves the recorded event to the end of its observed interval. Longer assessment gaps create more room for delay.
Last negative visit
This moves the event to a time at which the study still had evidence it had not occurred.
Interval midpoint
This can be a useful sensitivity scenario, but as a primary fix it assumes away where events fall within each interval and treats an imputed point as observed.
All three shortcuts let standard right-censored software run. That is computational convenience, not methodological validation. They can distort effect estimates and make uncertainty look smaller because the analysis forgets that event times were only localized to intervals.
Visit Timing Can Become Part of the Bias
Interval censoring is simplest when assessment timing is unrelated to the event process after accounting for the analysis model. Clinical follow-up often violates that ideal. Symptoms can trigger an unscheduled scan. High-risk patients may be monitored more frequently. One treatment may require laboratory visits that create more opportunities to detect an otherwise silent outcome.
Now the interval width is informative. Shorter windows may signal worsening health rather than better measurement logistics. A method that handles interval endpoints cannot, by itself, repair a visit process driven by latent disease activity.
Reviewer red flag
The study compares time to detected disease across groups with different scheduled or symptom-triggered assessment intensity, but discusses neither the observation process nor sensitivity to unequal detection opportunities.
What a Defensible Analysis Looks Like
- Store both boundaries. Preserve the last known event-free time and first event-positive time instead of collapsing them into one date.
- Name the estimand. Decide whether the question concerns latent biological onset, first clinical detection, or a decision-triggering confirmed event. Those are different endpoints.
- Match the method to the data. The Turnbull nonparametric maximum likelihood estimator can describe an interval-censored survival distribution. Regression models for interval-censored data can target covariate or treatment contrasts under proportional hazards, proportional odds, accelerated failure time, or parametric assumptions.
- Audit the visit process. Compare assessment schedules, missed visits, interval widths, and symptom-triggered assessments across study groups.
- Stress-test assumptions. Show whether conclusions survive plausible models and transparent boundary or midpoint sensitivity analyses. Agreement is reassuring; disagreement is the result, not an inconvenience.
A Five-Question Reviewer Checklist
1. Was occurrence observed?
Or was it only first detected at a scheduled assessment?
2. Were both interval bounds retained?
A single event-date column may already have destroyed the needed information.
3. Did visit intensity differ?
Check groups, sites, risk levels, symptoms, and treatment-required monitoring.
4. Did the analysis respect intervals?
If not, ask why its imputation assumptions fit the clinical process.
5. Does the conclusion concern onset or detection?
A result about faster detection should not quietly become a claim about faster disease development.
Why This Matters for Aqrab
Interval censoring is easy to miss because the dataset looks complete: every detected event has a date. The critique begins one level earlier. What did the measurement process truly observe, and which precision was added during data preparation?
Use Aqrab Try to pressure-test whether a time-to-event manuscript distinguishes occurrence from detection, preserves event windows, and explains unequal surveillance. A rigorous review does not merely ask whether the Cox model was adjusted. It asks whether the clock ever saw the event it claims to time.
Methods Anchors
Lindsey and Ryan provide a clinical-trial tutorial showing why beginning, midpoint, or endpoint substitutions can produce invalid inference. Turnbull develops the nonparametric maximum likelihood approach for grouped, censored, and truncated observations. Zhang and Sun review interval-censored survival estimation, group comparisons, regression models, and assumptions about the observation mechanism.
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.
The Denominator Illusion: Why 20,000 Measurements May Still Mean 200 Patients
A practical unit-of-analysis guide for clinical researchers. Separate rows, patients, clusters, and target populations before repeated observations create false precision.
Ecological Fallacy: When Hospital-Level Data Become Patient-Level Advice
A practical ecological fallacy guide for clinical researchers. Match the unit of analysis, exposure, outcome, and claim before turning group-level associations into patient-level advice.
Delayed Entry in Survival Analysis: Why Nobody Is at Risk Before They Enter
A practical guide to left truncation in clinical survival analysis. Align the origin, entry time, event time, and risk set before interpreting Kaplan–Meier curves or Cox models from prevalent cohorts.