Survival analysis is the heart of most oncology papers, and it is where I look hardest as a reviewer. Not because I enjoy it, but because it is where strong-looking results most often fall apart. A Kaplan-Meier curve can look convincing and still rest on a definition that was never stated, a censoring pattern that quietly biases the result, or a hazard ratio from a model whose central assumption was never checked. When the survival reporting is careless, an experienced reviewer starts to distrust the rest of the paper. When it is clean, it signals that the whole study was done with the same care.

This guide shows how to report survival analysis so that a reviewer has nothing to flag. It is written from the other side of the desk. Each section ends with the specific mistakes that draw a comment, and sometimes a rejection.

1. Define your endpoints before anything else

The most common problem is not a statistical error. It is an endpoint that was never defined precisely. Every survival endpoint needs three things stated in plain words: the start (the index date), the event, and what happens to patients who never have the event.

  • Overall survival (OS): from a defined index date (for example randomisation, diagnosis, or the start of treatment) to death from any cause. The cleanest endpoint, because death is unambiguous.
  • Progression-free survival (PFS): from the index date to disease progression or death from any cause, whichever comes first. You must say how progression was defined, for example by RANO in neuro-oncology or RECIST 1.1 in solid tumours.
  • Time to progression (TTP): like PFS, except that deaths without progression are censored rather than counted as events. PFS and TTP are not interchangeable, and the difference changes the result.
Index date Progression Death OS = event PFS = event TTP death without prior progression → censored (○) 0 6 12 18 24 Months since index date
Figure 1. The same patient, three endpoints. OS stops only at death; PFS stops at progression or death, whichever comes first; TTP stops at progression but censors a death that occurs without prior progression. PFS and TTP coincide here because progression came first, they diverge exactly when a patient dies before progressing.
What I flag as a reviewer

An endpoint with no stated index date. Progression described only as “clinical or radiological” with no named criteria. PFS and TTP used as if they were the same thing. State the start, the event, and the criteria, one short sentence each.

2. Censoring: say what you did, and check that it is honest

Censoring means a patient had not had the event by the time you stopped observing them, whether because the study ended or because they were lost to follow-up. Two things must appear in the paper.

  • How many patients were censored, and why. A high rate of loss to follow-up is itself a finding the reader needs to see.
  • The median follow-up, and how you calculated it. The accepted method is the reverse Kaplan-Meier, where events and censoring are swapped, not simply the median observation time.

The deeper issue is informative censoring. Kaplan-Meier assumes that patients who are censored carry the same future risk as those who remain under observation. If sicker patients drop out, so that censoring is linked to prognosis, the curve is biased toward looking better than reality. You cannot fully prove that censoring was uninformative, but you should show that you considered it.

What I flag as a reviewer

No median follow-up, or one computed as plain observation time. A large, unexplained loss to follow-up. No acknowledgement anywhere that censoring might be informative.

3. The Kaplan-Meier curve: the conventions reviewers expect

A survival figure has a few non-negotiable elements. Leaving them out is the fastest way to signal inexperience.

  • A numbers-at-risk table beneath the x-axis. This is not optional. Without it the reader cannot judge how reliable the later part of the curve is.
  • Censoring marks on each curve, so the reader sees where patients were censored rather than having an event.
  • Median survival for each group, with its 95% confidence interval, reported in the text or on the figure itself.
  • A tail that is not over-interpreted. Once only a handful of patients remain at risk, that part of the curve is the least reliable, and firm conclusions should not rest on it.
100 75 50 25 0 Survival probability (%) 0 6 12 18 24 30 36 Months since randomisation 16-mo median 30-mo median ticks = censored Treatment (n = 120) Control (n = 120) Number at risk Treatment 120 112 96 78 60 44 30 Control 120 92 64 40 26 16 10
Figure 2. A reviewer-ready Kaplan–Meier plot. Four elements are non-negotiable: the numbers-at-risk table beneath the axis, censoring marks on each curve, median survival read against the 50% line, and a tail that is not over-interpreted once few patients remain at risk. Data are illustrative.
What I flag as a reviewer

No numbers at risk under the curve. Authors drawing a confident conclusion from the far tail, where three patients remain. These two together are the most frequent survival-figure problems I see.

4. Comparing groups: log-rank, hazard ratios, and the assumption everyone forgets

To compare two survival curves you will usually report a log-rank p-value and a hazard ratio from a Cox proportional hazards model. Both are standard. One step is skipped far too often.

  • Report the hazard ratio with its 95% confidence interval, not a p-value alone. “HR 0.62 (95% CI 0.44 to 0.88, p = 0.007)” tells the reader the size of the effect and its precision. A bare “p < 0.05” tells them almost nothing.
  • Check the proportional hazards assumption. A Cox hazard ratio is a single number that assumes the relative risk between groups stays constant over time. If the curves cross or converge, that assumption is broken and the hazard ratio is misleading. Test it, for example with Schoenfeld residuals or a log-log plot, and say that you did.
  • Pre-specify the covariates in any adjusted model. A multivariable Cox model with covariates chosen after seeing the data is a classic way to manufacture significance, and reviewers are trained to suspect it.
Hazards stay proportional survival time → One hazard ratio fits the whole follow-up Survival curves cross time → A single hazard ratio is misleading
Figure 3. The proportional-hazards assumption in one picture. When the curves stay proportional (left), a single Cox hazard ratio summarises the difference honestly. When they cross or converge (right), the relative risk changes over time and one hazard ratio hides it, which is why the assumption must be tested, not assumed.
What I flag as a reviewer

A hazard ratio with no confidence interval. No mention of the proportional hazards assumption, especially when the published curves visibly cross. A suspiciously tidy multivariable model with no sign of pre-specification.

5. The bias that sinks observational papers: immortal time

If your study is not a randomised trial, one bias deserves its own section because it is both common and damaging. Immortal time bias arises when patients are grouped by something that can only happen after the index date, for example “patients who received treatment X” versus “those who did not.” The treated group must, by definition, have survived long enough to receive the treatment. That guaranteed stretch of survival is the immortal time, and it makes the treatment look protective when it may do nothing at all.

The usual fix is a landmark analysis, where patients are classified by their status at a fixed time point, or a time-varying covariate in the Cox model. If your design has any of this structure, address it before a reviewer does. For an experienced reader it is an immediate red flag, and an unaddressed one often ends the review.

Cohort entry / index date Treatment starts Event (death) Immortal time must survive to here to be classed “treated” Bias: the treated group is guaranteed to survive the immortal window so the treatment looks protective even if it does nothing. Fix: a landmark analysis, or a time-varying covariate in the Cox model.
Figure 4. Immortal time bias. Grouping patients by something that can only happen after the index date (here, receiving treatment) builds a guaranteed survival window into the treated group, making the treatment look beneficial when it may do nothing.

6. A pre-submission checklist

Before you submit, confirm that the manuscript contains each of these.

  • Each endpoint defined with its index date, its event, and its censoring rule.
  • Progression criteria named (RANO, RECIST 1.1, or the relevant standard).
  • Median follow-up, calculated by the reverse Kaplan-Meier method.
  • Number of events and number censored, per group.
  • Kaplan-Meier curves with numbers at risk and censoring marks.
  • Median survival with 95% confidence intervals.
  • Hazard ratios with 95% confidence intervals, never a p-value alone.
  • A statement that the proportional hazards assumption was checked.
  • Covariates for any adjusted model pre-specified, and the statistical software and version named.

None of this turns a weak result into a strong one. What it does is remove every easy reason for a reviewer to doubt you, so that your finding is judged on its merits. That is the whole game.