Statistical interpretation

Reporting survival analyses: Kaplan-Meier figures, numbers at risk and hazard ratios

Q: Should I report the log-rank P value, the Cox P value, or both?

One is enough if the methods say which. With a single binary covariate the two are close (here 0.0002 from both, reported as P < 0.001). If the primary analysis is specified as a stratified log-rank test, report that P value and take the hazard ratio from a Cox model stratified in the same way.

Q: How many time points should the numbers-at-risk table have?

Enough to align with the axis tick marks, usually five to eight. Each column must sit directly under its time point. Stop the axis where the numbers become too small to support the curve.

Q: Do I need confidence intervals for the median as well as for the hazard ratio?

Yes. SAMPL lists median survival times with confidence intervals, survival probabilities at stated times with confidence intervals, and a hazard ratio with a confidence interval as separate items.

Q: Can I give the hazard ratio for control versus intervention instead?

You can, and it is the reciprocal (here 1.66), but pick the direction that matches your hypothesis, state the reference group in the methods, table heads and figure legend, and keep it the same throughout the paper.

A figure checklist, a text template for median survival and the hazard ratio with its confidence interval, and a proportional hazards statement, shown on one worked example.

Editverse Medical editorial deskPublished 10 November 2026Guideline versions checked November 2026

Key points

Report, for each group: the number of patients and events, the median survival time with its 95% confidence interval, and the survival probability at one or two stated time points with confidence intervals.
Report the comparison as a hazard ratio with a 95% confidence interval, say which group is the reference, name the model and the test, and state how the proportional hazards assumption was checked.
A Kaplan-Meier figure needs a stepped curve, a table of numbers at risk under the time axis, censoring shown, labelled axes with units, and a time axis that stops before the numbers at risk become very small.
Define time zero, the event and every reason for censoring in the methods. Without them the curve cannot be interpreted.

Survival output arrives as a life table several hundred rows long, a log-rank statistic and a Cox model coefficient. The paper needs one figure, one or two sentences and sometimes a small table. This article shows the conversion on a single example and lists what the guidelines ask for at each step. The example data are illustrative: a simulated two-arm trial analysed for real, so that the medians, numbers at risk, hazard ratio and curves below all come from the same dataset. They are not results from an actual study.

What the guidelines ask for

The SAMPL guideline has a dedicated list for survival (time-to-event) analyses.¹ The CONSORT 2025 explanation and elaboration adds that for survival time data the effect size can be the hazard ratio or the difference in median survival time, and that every effect estimate needs a confidence interval.²

SAMPL items for survival analyses and where each is reported
SAMPL item (paraphrased)	Where it goes
Purpose of the analysis	Methods
Dates or events that mark the start and end of the period analysed	Methods: time zero, event definition, data cut-off
Circumstances under which data were censored	Methods
Method used to estimate survival	Methods: Kaplan-Meier
Method used to compare survival curves, with the P value of the comparison	Methods and Results: log-rank test
Median survival times with confidence intervals	Results text or table
Survival probability at appropriate follow-up times with confidence intervals, and the number at risk at each time	Results text; table under the figure
Full results in a graph or table	Kaplan-Meier figure
Regression model used to assess associations with survival	Methods: Cox proportional hazards model
A measure of risk, such as a hazard ratio, for each explanatory variable, with a confidence interval	Results text or table
Confirmation that the assumptions of the analysis were met	Methods and Results: proportional hazards

The illustrative trial

Three hundred patients were randomised 1:1 to intervention or control. The outcome is progression-free survival: time from randomisation to documented progression or death from any cause, whichever came first. Patients without an event were censored at their last disease assessment or at the data cut-off. Recruitment was staggered, so potential follow-up ranged from 12 to 30 months.

The raw output

Raw software output condensed (illustrative data)
Quantity	arm = 1	arm = 0
n	150	150
events observed	97	117
censored	53	33
median_survival_time_	14.4	9.3
median 95% lower, upper	11.0, 16.3	6.0, 11.5
KM_estimate at t = 12	0.545 (0.460, 0.623)	0.381 (0.303, 0.459)
KM_estimate at t = 24	0.239 (0.157, 0.331)	0.120 (0.059, 0.205)
logrank test_statistic; p	13.694; 0.000215
Cox coef (arm); se(coef)	−0.508134; 0.138628
exp(coef); lower 95%; upper 95%	0.601617; 0.458480; 0.789442
z; p	−3.665; 0.000247
proportional_hazard_test (rank); p	1.109; 0.292

Nothing here is wrong, and none of it is publishable as printed. Arms are coded 1 and 0, estimates carry six decimals, the P values are in a form no journal uses, and the reader cannot tell which group is the reference for the hazard ratio.

The figure

Figure (illustrative data). Kaplan-Meier estimates of progression-free survival by randomised group. Tick marks show censored observations. The table gives the number of patients at risk at each time point. Hazard ratio for intervention versus control 0.60 (95% CI 0.46 to 0.79).

Figure checklist

What a Kaplan-Meier figure should show
Element	What to do	Basis
Curve form	A step function. Never smooth the line or join points diagonally.	JAMA instructions for authors⁶
Numbers at risk	A row for each group under the time axis, aligned with the tick marks.	SAMPL; JAMA instructions
Censoring	Mark censored observations on the curve, or add cumulative numbers censored and with events to the at-risk table.	The extended table was the most favoured option in the KMunicate survey
Time axis	State the unit and the origin (months since randomisation). Decide how far to extend it: the right-hand tail rests on few patients and is estimated with much more uncertainty.	Pocock and colleagues; KMunicate
Vertical axis	Label it as the proportion or percentage free of the event, or with the event. Show the full range from zero; if you do not, say so in the legend.	JAMA instructions (axis content); full range is our editorial practice
Direction	Consider plotting cumulative incidence going up when events are uncommon. The STROBE explanation suggests this when the event rate is below about 30%.	SAMPL; Pocock and colleagues; STROBE explanation⁷
Uncertainty	Give confidence intervals for survival at the key time points in the text, or show confidence bands if the figure stays readable.	SAMPL; Pocock and colleagues
Group identification	Label curves directly and distinguish them by line style as well as colour.	Legibility in greyscale
Effect estimate	Put the hazard ratio and its confidence interval in the legend or on the plot.	Reader convenience

The points on direction, extent of the time axis, showing the extent of follow-up and displaying statistical uncertainty are the specific issues raised in the Lancet paper on survival plots by Pocock and colleagues, which also urges caution when reading meaning into the shape of the curves.³ In the KMunicate survey of 1,174 respondents, the two best-received changes to the standard plot were an extended table under the plot showing numbers at risk, censored and with events at regular intervals, and confidence intervals around each curve.⁴ In the example above only 11 and 5 patients remain at risk at 24 months, so a difference between the curves beyond that point should not be discussed.

The text template

Template

After a median follow-up of [x] months, [events] of [n] patients in the [intervention] group and [events] of [n] in the [control] group had [event]. Median [outcome] was [x] months (95% CI [a] to [b]) with [intervention] and [y] months (95% CI [c] to [d]) with [control] (hazard ratio [HR], 95% CI [e] to [f]; P = [p]). At [t] months, [outcome] was [s1]% (95% CI ...) and [s2]% (95% CI ...).

Journal-ready results text (illustrative)

At the data cut-off, median follow-up was 19.3 months. Progression or death had occurred in 97 of 150 patients (64.7%) in the intervention group and 117 of 150 (78.0%) in the control group. Median progression-free survival was 14.4 months (95% CI 11.0 to 16.3) with the intervention and 9.3 months (95% CI 6.0 to 11.5) with control (hazard ratio 0.60, 95% CI 0.46 to 0.79; P < 0.001). Progression-free survival at 12 months was 54.5% (95% CI 46.0 to 62.3) and 38.1% (95% CI 30.3 to 45.9), respectively (Figure).

What changed between output and text:

The hazard ratio 0.601617 became 0.60, and the direction is explicit: intervention is the numerator, control the reference. A hazard ratio without a stated reference group can be read either way.
The interval is taken from the exponentiated limits (0.458 and 0.789). It is not calculated from the standard error on the ratio scale.
P = 0.000247 became P < 0.001, the smallest value SAMPL says needs to be reported.¹
Counts sit beside the percentages, and the medians carry their confidence intervals.
Absolute figures (medians and 12-month probabilities) appear with the relative one. The CONSORT 2025 explanation notes that for survival data it can also help to express results as a number needed to treat.²

The methods paragraph and the proportional hazards statement

Journal-ready methods text (illustrative)

Progression-free survival was defined as the time from randomisation to documented disease progression or death from any cause. Patients alive without progression were censored at the date of their last disease assessment; no other censoring rules were applied. Survival functions were estimated with the Kaplan-Meier method, and medians are reported with 95% confidence intervals. Groups were compared with a two-sided log-rank test. The hazard ratio for intervention versus control and its 95% confidence interval were estimated from a Cox proportional hazards model with randomised group as the only covariate. The proportional hazards assumption was assessed with a test based on scaled Schoenfeld residuals. Median follow-up was estimated by the reverse Kaplan-Meier method. Analyses included all randomised patients in their allocated groups and used lifelines 0.30 in Python.

The corresponding sentence in the results: "There was no evidence against proportional hazards (P = 0.29)." Two cautions on wording. First, a P value of 0.29 does not prove the hazards are proportional; it means the test did not detect a departure, so say "no evidence against", not "the assumption was confirmed". Second, when the curves cross or separate late, a single hazard ratio averages over periods with different effects. In that case report the survival probabilities at fixed time points with their intervals as well, say that the hazard ratio is an average over follow-up, and describe any additional method in the methods section before the reader meets it in the results.

Six problems reviewers raise most often

"Median not reached" reported as a blank. If fewer than half of a group have had the event, the median cannot be estimated. Say so and give the survival probability at a fixed time point instead.
Mean follow-up of all patients offered as "follow-up". State the method. The simple median of observed times is shortened by early events.
Events and censoring not defined. "Survival" may be overall, disease-free or progression-free, each with a different event. SAMPL asks for the start and end of the period and the circumstances of censoring.¹
Hazard ratio described as a risk reduction. A hazard ratio of 0.60 is a 40% lower hazard, that is, rate of events among those still at risk at each moment. It does not mean that 40% fewer patients had the event: here the proportions with an event were 64.7% and 78.0%.
The log-rank P value given without any estimate. The ICMJE Recommendations (updated January 2026) ask authors to quantify findings with measures of uncertainty and not to rely on P values alone.⁵
Figure and text disagree. The numbers at risk at time zero must equal the group sizes, and the medians read off the curve at 50% must match the text.

If you already hold the output and need the figure specification, table and wording, that is what our statistical results interpretation service returns. Where the survival analysis has to be run from the dataset, including the figure as a vector file, see the Advanced and Premium tiers of statistical analysis for clinical studies. Both are part of our statistical interpretation work.

Should I report the log-rank P value, the Cox P value, or both?