Delayed Entry in Survival Analysis: Why Nobody Is at Risk Before They Enter
Anas H. Alzahrani, MD PhD MPH
Department of Preventive Medicine and Public Health
Faculty of Medicine, King Abdulaziz University
Suppose a cancer registry defines survival from diagnosis, but patients enter the linked genomic dataset only when they receive sequencing months or years later. Anyone who died before sequencing can never appear. The observed cohort has already passed a survival test before follow-up begins.
This is delayed entry, often called left truncation. It is not fixed by placing “time since diagnosis” on the x-axis. The analysis must keep patients out of every risk set before the moment they could actually enter the study. Otherwise it gives them event-free person-time during a period in which the dataset could not have observed their event.
Four Clocks Before One Curve
| Clock | Question | Example |
|---|---|---|
| Origin | When does the scientific timescale begin? | Date of diagnosis |
| Entry | When can this person first join the observed risk set? | Date of genomic testing |
| Event | When does the outcome occur on that timescale? | Date of death |
| Censoring | When does observation end without the event? | Last known follow-up |
A person diagnosed in January 2022, sequenced in January 2024, and alive through January 2026 contributes information from year two onward on the diagnosis timescale. They did not contribute two observed years at risk before sequencing. Including them from diagnosis creates guaranteed survival time by construction: they had to survive long enough to enter.
The Risk-Set Test
Concrete takeaway
At each event time, ask two questions about every patient: had they entered the cohort yet, and were they still event-free? A patient belongs in the risk set only when both answers are yes. “Alive in retrospect” is not the same as “under observation and at risk.”
Delayed-entry Kaplan–Meier and Cox analyses implement this logic by using an entry time and an exit time for each person. The event-time denominator then contains only people whose entry precedes that time and whose event or censoring time has not yet arrived. This is a risk-set correction, not a cosmetic shift of the plotted axis.
The ordinary Kaplan–Meier estimator from the origin treats everyone eventually observed as if they were already present at time zero. In a prevalent cohort, that can make early survival appear too favorable because early deaths are structurally missing while later survivors remain available for enrollment.
Left Truncation Is Not Left Censoring
Left truncation
People whose event occurred before entry are absent from the observed cohort. The study may not know they existed.
Left censoring
The person is observed, and the event is known to have occurred before a recorded time, but its exact time is unknown.
The distinction changes the likelihood, the assumptions, and the software specification. Calling both situations “incomplete follow-up” hides the selection mechanism. Right censoring is different again: the person is known to be event-free until a recorded time, after which the event status is unknown.
Risk-Set Adjustment Solves Only Part of the Problem
Standard delayed-entry methods rely on an independence condition: after accounting for the modeled information, the process determining entry must be sufficiently independent of the future event process. That assumption can fail when entry depends on evolving prognosis.
In the genomic example, sequencing may happen sooner for patients with aggressive disease, later for long-term survivors, or only after referral to a specialist center. Even a correctly constructed delayed-entry Cox model cannot automatically recover the target survival distribution when entry time and survival remain dependent through unmeasured severity, access, or clinical decisions.
Two separate questions
- Was the risk set built correctly? Nobody contributes before entry.
- Is delayed entry conditionally independent? Among measured-alike people, entry timing does not still select on future survival.
Passing the first test is necessary. It does not prove the second. Investigators should describe why entry occurs, compare entry delays across clinically relevant groups, identify variables that jointly predict entry and survival, and state the population to which the estimate applies.
Design Choices Can Clarify the Estimand
An incident cohort that enrolls patients at or near the origin can reduce left truncation, although it may not be feasible in retrospective real-world data. Another option is a landmark question: among patients alive and eligible at a prespecified time after diagnosis, what is survival from that landmark? That does not estimate survival from diagnosis for all diagnosed patients, but it makes the selected population explicit.
Changing time zero to database entry is not a universal fix. It answers survival since observation began and may compare people at very different disease durations. The right approach follows the scientific question, the entry mechanism, and the data that could have captured events—not whichever coding choice produces the longest curve.
A Reviewer Checklist
- What event defines time zero: diagnosis, treatment initiation, eligibility, or database entry?
- When could each person first be observed and become eligible to enter the analytic cohort?
- Could the event occur between time zero and cohort entry—and would that make the person absent from the data?
- Does each risk set include only people who have entered and remain event-free at that event time?
- What assumption makes entry time sufficiently independent of the event process, marginally or after conditioning on measured variables?
- Would an incident-cohort design or a landmark estimand answer a clearer, more defensible question?
What Not to Conclude
- “We measured survival from diagnosis, so follow-up began at diagnosis.” A timescale origin does not establish observability.
- “Everyone was alive at cohort entry, so the pre-entry period is valid survival time.” Survival to entry is the selection condition creating the problem.
- “A Cox model handles censoring.” Ordinary right-censoring syntax does not automatically handle delayed entry.
- “We used start–stop data, so selection bias is gone.” Correct risk sets do not guarantee independent truncation.
Sources and Evidence Maturity
- Cain et al., “Bias Due to Left Truncation and Left Censoring in Longitudinal Studies” (2011) — peer-reviewed epidemiologic methods article distinguishing truncation, censoring, and their biases.
- Kehl et al., “Implications of Selection Bias Due to Delayed Study Entry in Clinical Genomic Studies” (2022) — peer-reviewed clinical methods article with a real-world genomics application.
- Sondhi, “Estimating Survival Parameters Under Conditionally Independent Left Truncation” (2022) — peer-reviewed methodological work on risk-set adjustment and conditional independence.
- Applebaum, Malloy, and Eisen, “Left Truncation, Susceptibility, and Bias in Occupational Cohort Studies” (2011) — peer-reviewed simulation and applied epidemiologic analysis.
Evidence note: this guide summarizes established survival-analysis methodology. It does not estimate a clinical effect or recommend a treatment.
The Practical Bottom Line
A survival clock can begin before observation does. When entry is delayed, keep each person out of the risk set until they actually enter, then examine whether the entry mechanism still selects people on prognosis.
The cleanest review question is simple: who had to survive long enough to become visible? If the paper cannot answer that, its earliest survival time may be the least credible part of the curve.
Where Aqrab Fits
Aqrab helps reviewers turn a survival-analysis label into an auditable timeline: origin, eligibility, cohort entry, risk-set membership, event, censoring, and the assumptions connecting them. Try Aqrab on a real-world survival manuscript, or explore the developer documentation for structured review workflows.
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.
Target Trial Emulation Cannot Randomize Clinical Judgment: A Pertussis Study Audit
A practical target-trial emulation audit using an infant pertussis study. Check clinical-judgment confounding, propensity-score overlap, endpoint timing, sparse outcomes, and claim strength.
Vaccine Effectiveness Without Matching: Why Calendar Time Comes Before Pairing
A practical guide to calendar time in vaccine-effectiveness studies. Learn how changing uptake and infection hazards alter risk sets, estimands, and target-trial conclusions.
Specification Curve Analysis: When One Model Hides a Vibration of Effects
A practical guide to specification curve analysis in clinical research. Learn how to define reasonable analyses, separate robustness checks from estimand changes, and audit results that move across defensible choices.