Skip to content
Back to search results

Kaplan-Meier Analysis — Nonparametric Survival Estimation

Introduced by
Edward L. Kaplan and Paul Meier
Year
1958
Type
Nonparametric survival estimator
Family
Clinical / epidemiology
Font size

What does it do?

Kaplan-Meier (KM) analysis is a nonparametric method for estimating the survival function from time-to-event data.

Introduced by Kaplan and Meier in 1958, it produces the classic step-function survival curve that shows the probability of surviving beyond each observed event time, correctly accounting for censored observations — participants who left the study or had not yet experienced the event by the end of follow-up. It is one of the most widely used techniques in clinical and epidemiological research.

The intuition behind the method

Imagine following a group of cancer patients to see how long they survive after treatment. Some die during the study (observed events); others are still alive when the study ends or drop out (censored). A simple average ignores censoring and understates survival. The Kaplan-Meier estimator instead multiplies a series of conditional survival probabilities — the chance of surviving each successive event time given survival up to that point — yielding a step-function curve that drops only when an actual event occurs. Censored observations reduce the risk set but do not count as events, so their information is used until the moment they exit.

History

Kaplan-Meier analysis originates from a landmark 1958 paper by Edward L. Kaplan and Paul Meier in the Journal of the American Statistical Association. Meier had derived the product-limit estimator independently; a journal editor recognised the parallel work and encouraged a joint publication. The paper was initially met with modest interest but grew to become one of the most cited scientific articles ever published. The accompanying log-rank test was developed by Nathan Mantel in 1966. The method became central to clinical trial reporting from the 1970s onward, and regulatory agencies worldwide now expect KM curves as part of any time-to-event analysis submission.

When to use it

Use Kaplan-Meier analysis whenever the outcome is time to a single well-defined event (death, relapse, discharge, treatment failure) and follow-up times vary across participants. It is the standard first step in any survival analysis and is appropriate for both randomised trials and observational cohort studies. Prefer KM curves for descriptive display and group comparison when the proportional hazards assumption holds; move to Cox regression when you need to adjust for multiple covariates simultaneously. Do not use KM when the event cannot occur more than once per person and competing events (e.g., death from a different cause before the event of interest) are present — use competing-risks analysis instead. KM is also unsuitable when censoring is likely to be informative, when continuous covariate adjustment is needed, or when the outcome is binary rather than time-to-event.

Limitations and considerations

  • Cannot adjust for multiple confounders simultaneously — use Cox proportional hazards for covariate-adjusted estimates.
  • Assumes non-informative censoring; biased estimates result if censoring is related to prognosis.
  • Does not handle recurrent events or competing risks without extension to more specialised methods.
  • The log-rank test comparing curves has reduced power when the hazard ratio changes over time (crossing or converging curves).
  • Estimates become unstable at long follow-up times when few participants remain at risk.

Source conditions do not replace independent scientific review or validation by an analysis engine.

Strengths

  • Completely nonparametric — makes no assumption about the shape of the survival distribution.
  • Correctly incorporates censored observations, maximising use of all available follow-up time.
  • Produces an intuitive, immediately interpretable visual (the step-function survival curve).
  • Median survival with confidence intervals is easily derived as a summary statistic.
  • Widely understood across clinical, epidemiological, and regulatory audiences.

Applications

  • Comparing overall survival between treatment arms in oncology randomised controlled trials.
  • Estimating time to hospital readmission after cardiac surgery in cardiovascular cohort studies.
  • Reporting progression-free survival in regulatory submissions for new oncology drugs.
  • Analysing time to treatment discontinuation or adverse event onset in pharmacovigilance studies.
  • Estimating graft survival in transplant registries and organ rejection studies.
  • Describing time to recurrence after surgical resection in colorectal or breast cancer cohorts.

How it works

How it works

1

Specify the event of interest (e.g., death, disease recurrence, hospital readmission), the time origin (e.g., randomisation date, diagnosis date), and the censoring mechanism. Every participant contributes a survival time and an event indicator (1 = event occurred, 0 = censored). Ensure that censoring is non-informative — that is, participants who are censored should not systematically differ in prognosis from those who remain at risk.

2

Sort all observed event times from smallest to largest. Ties (multiple events at the same time point) are resolved by standard convention. At each distinct event time t_i, record n_i (the number of participants at risk just before t_i) and d_i (the number of events at t_i). Censored observations that occur between consecutive event times reduce n_i for subsequent steps but are not counted as events.

3

Compute the survival probability at each event time using the product-limit formula: S(t) = product over all t_i <= t of (1 - d_i / n_i). The estimate starts at S(0) = 1 and decreases in steps only at event times, remaining constant between events. This makes no assumption about the underlying distribution of survival times.

Formula
S(t) = \prod_{t_i \le t} \left(1 - \frac{d_i}{n_i}\right)
4

Censored observations — participants lost to follow-up, withdrawn, or event-free at the study cutoff — are incorporated by removing them from the risk set at their censoring time, after contributing to the denominator n_i up to that point. Proper handling of censoring is critical: if censoring is informative (correlated with prognosis), the KM estimate will be biased and more sophisticated methods are needed.

5

When two or more groups (e.g., treatment vs. control) are compared, a separate KM curve is estimated for each group and the curves are displayed together. The log-rank test (or Mantel-Cox test) provides a formal significance test for the null hypothesis of equal survival functions. The log-rank test is most powerful when the hazard ratio between groups is approximately constant over time (proportional hazards assumption); weighted variants such as Wilcoxon-Breslow are available when early differences are of greater interest.

6

Present the KM curves graphically with a numbers-at-risk table below the x-axis. Report median survival time (the time at which S(t) crosses 0.5) with 95% confidence intervals (typically Greenwood's formula). If a comparison was made, report the log-rank p-value and the hazard ratio from a Cox model. Specify the total sample size, number of events, and follow-up duration.

Common pitfalls

  • Ignoring competing risks — using KM when participants can experience a competing event (e.g., death before relapse) leads to overestimation of the cumulative incidence of the event of interest; use the Fine-Gray or cause-specific hazard approach instead.
  • Presenting curves without a numbers-at-risk table — late portions of the curve based on very few individuals are unreliable and misleading without context.
  • Informative censoring — if patients are censored because they deteriorated and switched treatment, the surviving cohort is artificially selected; sensitivity analyses or inverse-probability-of-censoring weighting should be considered.
  • Reporting only the log-rank p-value without an effect size — the p-value conveys significance but not the magnitude of the difference; always accompany it with a hazard ratio and confidence interval from Cox regression.
  • Crossing survival curves with a single log-rank test — when curves cross, the log-rank test may be non-significant even when there is a real early or late difference; restricted mean survival time or weighted log-rank tests are more informative.

Frequently asked questions

What is the difference between the Kaplan-Meier curve and the cumulative hazard curve?

The KM curve shows S(t), the probability of surviving (not experiencing the event) beyond time t, which decreases from 1 toward 0. The cumulative hazard H(t) = -log S(t) increases over time and represents the accumulated instantaneous risk. Both are equivalent representations of the same underlying data; the cumulative hazard is linear on a log scale under an exponential distribution, making it useful for model checking.

When should I use the log-rank test vs. a Cox model to compare groups?

Use the log-rank test when you want a simple, nonparametric significance test comparing two or more unadjusted KM curves. Use Cox regression when you need to adjust for confounders, include continuous predictors, or obtain a hazard ratio as a summary effect size. The two approaches assume proportional hazards; when that assumption is violated, consider restricted mean survival time analysis.

How do I calculate the confidence interval for the median survival?

The standard approach applies Greenwood's formula to estimate the variance of S(t), then constructs pointwise confidence bands around the curve. The confidence interval for the median is the set of time points where the confidence band crosses the 0.5 probability line. Most statistical software (R survival package, SAS LIFETEST, Stata stci) computes this automatically.

Can I use Kaplan-Meier when I have competing risks?

Standard KM analysis treats competing events (e.g., death from another cause before the event of interest) as censored observations, which leads to overestimation of cumulative incidence because censored individuals are assumed to remain at risk. When competing risks are present, use the cumulative incidence function (CIF) from Fine-Gray regression or cause-specific hazard models instead of the KM estimator.

What sample size do I need for a reliable Kaplan-Meier estimate?

Reliability depends primarily on the number of observed events rather than total sample size. As a rough rule, at least 10–20 events per group are needed for stable curve estimates, and formal power calculations for the log-rank test should be conducted based on the expected event rate, median survival, and desired power. At long follow-up times where few participants remain at risk (fewer than ~5–10), the curve becomes unreliable and should be truncated or presented with appropriate uncertainty.

Related methods

Sources

  1. Kaplan, E. L., & Meier, P. (1958). Nonparametric estimation from incomplete observations. Journal of the American Statistical Association, 53(282), 457–481.

    Open DOI reference
    Copied.
  2. Kaplan–Meier estimator. Wikipedia.

    Open publication reference
    Copied.