By HTA Course · Published

Evidence synthesis

Kaplan-Meier Curve: How to Read and Use It in HTA

The Kaplan-Meier curve is the standard picture of survival in clinical trials. This guide explains the product-limit estimator with a worked example, how to read steps, censoring ticks, numbers at risk and median survival, how curves are compared, and why HTA models need more than the observed curve.

12-minute read Intermediate level Survival analysis for HTA and HEOR

A Kaplan-Meier curve is a step-shaped graph of survival over time, and oncology submissions to HTA bodies typically include several. It shows the estimated proportion of patients who have not yet had an event, such as death or disease progression, at each point in time, and it does so correctly even though many patients are still event-free when the trial data are cut. Reading these curves well is a core skill for anyone critiquing clinical or economic evidence.

A Kaplan-Meier curve is a step-shaped plot of the Kaplan-Meier estimator, also called the product-limit estimator: a non-parametric estimate of the survival function S(t), the probability of remaining event-free beyond time t, from data in which some observation times are censored.

The method was published by Edward Kaplan and Paul Meier in the Journal of the American Statistical Association in 1958. They named it the product-limit estimate because it is the limiting case of the older actuarial (life-table) estimates. Non-parametric means it assumes no particular shape for the survival curve: it simply follows the observed data.

“Survival” is a convention. The same method estimates time to progression, relapse, conception or any other single event, so trial reports show Kaplan-Meier curves for overall survival (OS), progression-free survival (PFS) and many other time-to-event endpoints.

Censoring: why survival data need their own method

Censoring means that a patient’s event time is not known exactly, usually because the patient was still event-free when last observed. Simple percentages handle this badly; Kaplan-Meier analysis uses each patient’s information for as long as they are followed.

The common form is right censoring. A patient is right-censored when the study ends before their event, when they are lost to follow-up, or when another event stops further observation. Clark and colleagues also describe left censoring (the event happened at an unknown time before it was observed) and interval censoring (patients come in and out of observation); standard Kaplan-Meier analysis deals with right censoring.

Censored patients are not failures and not survivors. Counting them as survivors to the end of the study overstates survival; dropping them throws away the time they were known to be event-free. The worked example below shows both errors.

The key assumption

Kaplan-Meier estimates are valid only if censoring is non-informative: patients who are censored must have the same prospects as those who remain under observation. If patients leave a trial because of toxicity or a worsening condition, the assumption fails and the curve is biased. Bland and Altman note that it cannot easily be tested.

The Kaplan-Meier estimator: the product-limit formula

The Kaplan-Meier estimator splits follow-up at each time an event occurs, estimates the conditional probability of surviving that moment among those still at risk, and multiplies these probabilities together. Survival to any time is the product of all the conditional survival probabilities up to it.

Survival just after event time tj S(tj)
=
Survival at the previous event time × share of those at risk who did not have the event S(tj−1) × (1 − dj ÷ nj)

Here dj is the number of events at time tj, and nj is the number at risk just before it: patients who have neither had the event nor been censored. The estimate starts at S(0) = 1 and changes only at event times. A censored patient never causes a drop; they simply leave the risk set, so later drops are calculated over fewer people. With no censoring at all, the estimate reduces to the plain proportion of patients still event-free.

Confidence intervals for each survival probability are usually calculated with Greenwood’s formula for the standard error, which statistical software applies automatically.

Worked example: a Kaplan-Meier survival analysis by hand

The arithmetic is short enough to do on paper. Take a hypothetical cohort of 10 patients followed for up to 12 months. Five die (at 2, 4, 4, 7 and 10 months) and five are censored (at 3, 6, 9, 12 and 12 months). These are invented round numbers for teaching, not data from any trial.

Month At risk (n) Deaths (d) Censored 1 − d ÷ n S(t)
2 10 1 0 9/10 = 0.900 0.900
3 9 0 1 no event 0.900
4 8 2 0 6/8 = 0.750 0.900 × 0.750 = 0.675
6 6 0 1 no event 0.675
7 5 1 0 4/5 = 0.800 0.675 × 0.800 = 0.540
9 4 0 1 no event 0.540
10 3 1 0 2/3 = 0.667 0.540 × 2/3 = 0.360
12 2 0 2 no event 0.360
  1. Order the times. List every event and censoring time from shortest to longest.
  2. Count the risk set. At each event time, count the patients still followed: 10 at month 2, but only 8 at month 4, because one patient died at month 2 and one was censored at month 3.
  3. Multiply. Each drop multiplies the previous estimate by the share who survived that event time, giving 0.900, 0.675, 0.540 and 0.360.
  4. Read the median. The curve first falls to 0.5 or below at month 10, so median survival is 10 months.

Compare the naive answers. Counting all five censored patients as alive gives 5/10 = 50% survival at 12 months, although three of them were followed for only 3 to 9 months. Dropping those three leaves 2 alive out of 7, or 29%. The Kaplan-Meier estimate of 36% sits between the two because it uses each censored patient for exactly the time they were observed.

The area under this step curve up to 12 months is 2 × 1 + 2 × 0.900 + 3 × 0.675 + 3 × 0.540 + 2 × 0.360 = 8.165 months: the restricted mean survival time at 12 months, which becomes important later.

How to read a Kaplan-Meier curve: seven checks

To read a Kaplan-Meier curve, check what event is plotted, follow the steps and censoring marks, and judge every estimate against the numbers at risk below the plot. Seven checks cover most published curves.

  1. The axes and the event. The vertical axis is the estimated probability of being event-free, from 1 (or 100%) down; the horizontal axis is time since randomisation or diagnosis. Check whether the endpoint is OS, PFS or something else, and how each event is defined.
  2. The steps. Each vertical drop is one or more events. A drop is bigger when fewer patients remain at risk, because each event then represents a larger share of the risk set.
  3. Censoring ticks. Short vertical marks on the flat parts of the curve show censored observations. Many ticks early on mean heavy loss to follow-up or late recruitment, and a reason to ask whether censoring is non-informative.
  4. Numbers at risk. The table under the x-axis gives how many patients are still followed in each arm at set times. Where it falls into single figures, the curve rests on very few people and its shape is unreliable.
  5. Median survival. Draw a horizontal line at 0.5: the time at which the curve first reaches it is the median. If the curve never drops to 0.5 during follow-up, the median cannot be estimated and is reported as “not reached”.
  6. Landmark survival. Draw a vertical line at a chosen time, such as 12 or 24 months, and read the estimated proportion event-free. Landmark estimates should be quoted with their confidence intervals.
  7. Uncertainty. Confidence bands, where shown, widen to the right as the risk set shrinks. Clark and colleagues, citing Pocock and colleagues, suggest curtailing the time axis before the end of follow-up, because confidence limits are often wide in the tail of the curve.

Reporting standards for Kaplan-Meier plots

Pocock, Clayton and Altman (2002) urged trial reports to show the extent of follow-up and the statistical uncertainty around survival curves. The KMunicate study by Morris and colleagues (2019) surveyed more than 1,100 respondents, mostly statisticians and clinicians, on how Kaplan-Meier plots should look. In a sample of 50 trials published in 2013 in four major general medical journals, 47 plots showed numbers at risk, 20% showed censoring and only 10% showed confidence intervals. The authors suggest confidence intervals around each curve plus an extended table giving numbers at risk, censored and with events at each time point as a starting point.

Comparing Kaplan-Meier curves: log-rank test and hazard ratio

Two Kaplan-Meier curves are usually compared with the log-rank test, which asks whether they differ, and summarised with a hazard ratio, which estimates by how much. Neither is a reliable summary when hazards are not proportional.

The log-rank test compares the number of events observed in each group with the number expected if there were no difference, summed over all event times. It tests the null hypothesis that the survival curves are the same, which for two groups under proportional hazards is equivalent to a hazard ratio of 1. It gives a p-value but no effect size, which is why trials pair it with a hazard ratio from a Cox model; that article explains calculation, interpretation and the proportional hazards assumption in detail.

Summary What it tells you Main caveat
Log-rank test Whether the curves differ overall (p-value) No effect size; can lose power when hazards are not proportional
Hazard ratio Relative event rate, averaged over follow-up A single number misleads if the effect changes over time
Difference in medians Shift in the time by which half have had the event Ignores the tail; unavailable if a median is not reached
Landmark difference Difference in survival at one chosen time Depends on the time chosen; ignores the rest of the curve
Restricted mean survival time (RMST) Mean event-free time up to a set horizon; the difference between arms is the area between the curves Only as long as follow-up; not a lifetime mean

Crossing or separating late

When curves cross, or separate only after a delay as often happens with immunotherapy, hazards are not proportional. Royston and Parmar (2016) note that the log-rank and equivalent Cox tests may then lose power, sharply so when an early effect fades. RMST, the area under each curve up to a fixed time, stays interpretable whatever the shape of the hazards. In indirect comparisons the same problem carries into network meta-analysis of survival outcomes, covered in Module 4: Evidence Synthesis.

Limitations of the Kaplan-Meier method

The Kaplan-Meier method describes one group’s survival within follow-up under non-informative censoring. It does not adjust for covariates, handle competing events or say anything about the time after the last observation.

1

Informative censoring

If sicker patients drop out, the remaining risk set is healthier than the whole cohort and the curve is too optimistic.

2

Competing risks

When another event prevents the one of interest, treating it as censoring makes 1 minus Kaplan-Meier overstate incidence; a cumulative incidence function is used instead.

3

Unstable tail

Late steps rest on few patients, so a single event can produce a large drop and a long flat tail can look reassuring without much evidence behind it.

4

No adjustment

The estimator compares groups without adjusting for prognostic factors; that needs regression, such as the Cox model.

5

Interval-measured events

Bland and Altman note that when events are only detected at scheduled visits, the estimates are biased upwards, as can happen with progression assessed at scheduled scans.

6

Median, not mean

Survival data are skewed, so trials report medians. A mean cannot be calculated from a curve that has not reached zero.

Why Kaplan-Meier curves matter in HTA and economic models

In HTA, the Kaplan-Meier curve is the starting point, not the answer. Cost-effectiveness needs mean survival over a lifetime, so analysts fit parametric models to the observed curves, check them visually and statistically, and extrapolate beyond follow-up.

Input to partitioned survival models

In a partitioned survival model, the PFS and OS curves directly set the share of patients who are progression-free, progressed or dead over time, and the area under each curve becomes time in each health state. Those curves start from the trial’s Kaplan-Meier data. The same survival estimates feed transition probabilities in Markov models; see the overview of health economic modelling.

Extrapolation: NICE DSU TSD 14

NICE Decision Support Unit Technical Support Document 14 (Latimer, 2011, updated 2013) sets out how to extrapolate survival with patient-level trial data. Its main points for readers of Kaplan-Meier curves:

  • Means, not medians. Economic evaluation needs mean time to event, and a median should not be plugged into a model directly.
  • Parametric over restricted means. Fitted parametric models should be used rather than restricted means, unless the data are almost complete.
  • Visual fit is not enough. A model can track the Kaplan-Meier curve closely yet give an implausible tail, so visual inspection should not be relied on alone; if it is used, numbers at risk should be shown.
  • Proportional hazards must be justified. Log-cumulative hazard plots help decide whether one model with a treatment effect, or separate models per arm, is appropriate.

TSD 14 also describes approaches that use the Kaplan-Meier data directly and attach a fitted parametric tail, and warns that results can be sensitive to where the switch is made. Because the tail beyond follow-up can drive much of the QALY gain, these choices are tested in scenario and probabilistic sensitivity analysis; how QALYs are built is explained in what a QALY is. Modelling is taught in Module 8: Economic Modelling.

Reconstructing patient data from published curves

Analysts often lack individual patient data for the comparator trials. Guyot and colleagues (2012) published an algorithm that reconstructs approximate patient-level data from a digitised Kaplan-Meier curve, using the published numbers at risk and, where reported, the total number of events. In their validation, reconstructed survival probabilities and medians were very close to the originals; hazard ratios were reasonably accurate only when at least the numbers at risk or the total number of events were reported. They recommend that trials publish both alongside every curve.

For submissions, this means reviewers check where each curve came from (patient-level data or reconstruction), which data cut was used, and whether the extrapolation is plausible against external data. See our guide to the NICE technology appraisal process for how committees weigh these choices.

Key takeaways

  • The Kaplan-Meier curve estimates the probability of remaining event-free over time, using censored patients while they are followed.
  • The product-limit formula multiplies conditional survival probabilities at each event time.
  • Always read the curve with its numbers at risk; the tail is often based on few patients.
  • The log-rank test and hazard ratio summarise differences, but both can mislead when curves cross.
  • HTA models need mean survival, so Kaplan-Meier data are fitted and extrapolated following NICE DSU TSD 14.

Certificate Course in HTA

Learn to read survival evidence like an HTA reviewer

Module 3 covers the biostatistics behind survival curves, hazard ratios and confidence intervals; Module 4 brings them into evidence synthesis; Module 8 turns them into partitioned survival and Markov models. The course presents both the industry and payer perspective without assuming prior HTA knowledge.

15 structured learning modules
79 interactive lessons
100% online and self-paced
Global HTA and market access perspective

Frequently asked questions

Kaplan-Meier curve basics

What does a Kaplan-Meier curve show?

It shows the estimated proportion of patients who have not yet had an event, such as death or progression, at each time since the start of follow-up. The curve drops in steps at event times, and short tick marks show censored patients, whose event had not occurred when they were last observed.

How is the Kaplan-Meier estimator calculated?

At each event time, divide the number of patients who did not have the event by the number at risk just before it, then multiply this by the previous survival estimate. Censored patients leave the risk set without causing a drop. The result is also called the product-limit estimator.

How do you find median survival on a Kaplan-Meier curve?

Draw a horizontal line at 0.5 (50%) and find the time at which the curve first reaches it. If the curve stays above 0.5 for the whole follow-up, the median cannot be estimated and is reported as not reached.

What do the tick marks on a Kaplan-Meier curve mean?

Tick marks show censored observations: patients who left follow-up, or were still event-free at the data cut, at that time. They do not lower the curve, but they reduce the number at risk, so later events cause larger drops.

What is the difference between the log-rank test and the hazard ratio?

The log-rank test compares observed and expected events to test whether two Kaplan-Meier curves differ; it gives a p-value but no effect size. The hazard ratio, usually from a Cox model, estimates the relative event rate between groups. Both assume, or work best with, proportional hazards.

Why can HTA models not use the Kaplan-Meier curve alone?

Trials end before all patients have had the event, but cost-effectiveness needs mean survival over a lifetime. Analysts therefore fit parametric models to the Kaplan-Meier data and extrapolate them, as described in NICE DSU TSD 14, and test how the choice of model affects results.

Sources

  1. Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. Journal of the American Statistical Association 53(282):457–481, 1958.
  2. Bland JM, Altman DG. Statistics notes: survival probabilities (the Kaplan-Meier method). BMJ 317(7172):1572, 1998.
  3. Clark TG, Bradburn MJ, Love SB, Altman DG. Survival analysis part I: basic concepts and first analyses. British Journal of Cancer 89(2):232–238, 2003.
  4. Morris TP, Jarvis CI, Cragg W, et al. Proposals on Kaplan-Meier plots in medical research and a survey of stakeholder views: KMunicate. BMJ Open 9:e030215, 2019.
  5. Pocock SJ, Clayton TC, Altman DG. Survival plots of time-to-event outcomes in clinical trials: good practice and pitfalls. Lancet 359(9318):1686–1689, 2002.
  6. Royston P, Parmar MKB. Augmenting the logrank test in the design of clinical trials in which non-proportional hazards of the treatment effect may be anticipated. BMC Medical Research Methodology 16:16, 2016.
  7. Manzoor BS, Adimadhyam S, Walton SM. An introduction to competing risks. Value & Outcomes Spotlight, ISPOR, March/April 2017.
  8. Latimer N. NICE DSU Technical Support Document 14: Survival analysis for economic evaluations alongside clinical trials – extrapolation with patient-level data, June 2011, updated March 2013.
  9. Guyot P, Ades AE, Ouwens MJNM, Welton NJ. Enhanced secondary analysis of survival data: reconstructing the data from published Kaplan-Meier survival curves. BMC Medical Research Methodology 12:9, 2012.

Not ready to enrol yet?

Get the full course syllabus by email

  • All 15 modules, from evidence and statistics to modelling and reimbursement
  • Prices: 3 months €169, 6 months €289, 12 months €490
  • A short note to forward if your employer pays for training (VAT invoice available)

One email with the syllabus. See our privacy policy.

Article rating

Rate this article

Be the first to rate this article

Click a star to rate

Last updated: 16/11/2026