Kaplan-Meier Survival Curves

The Kaplan-Meier (product-limit) estimator is the standard nonparametric method for estimating a survival function from censored time-to-event data. It produces the familiar stepwise survival curve, which falls at each observed event time and stays flat in between, and lets investigators read off survival probabilities and median survival without assuming any particular distribution for event times.

Definition

The Kaplan-Meier estimator is a nonparametric estimate of the survival function obtained as a running product over event times of the conditional probability of surviving each event time given survival up to it, with censored observations removed from the risk set at their censoring time.

Scope

This topic covers how the Kaplan-Meier estimator is constructed from the risk set at each event time, how censored observations are accommodated, how survival curves and median survival are read, and how groups are compared using the log-rank test. It is methodological reference material and not clinical guidance.

Core questions

• How is the survival curve estimated from event times and the risk set without assuming a distribution?
• How do censored observations enter the calculation?
• How are survival probabilities, median survival, and their confidence intervals read from the curve?
• How are two or more survival curves compared statistically?

Key concepts

• Product-limit estimator
• Risk set at each event time
• Conditional survival probability
• Stepwise survival curve
• Median survival
• Greenwood's formula (variance)
• Log-rank test
• Number at risk

Mechanisms

At each distinct event time the estimator computes the conditional probability of surviving that instant — one minus the number of events divided by the number at risk just before — and multiplies these conditional probabilities together to give the cumulative survival probability, producing a step down at each event time. Subjects censored before an event time leave the risk set and therefore do not pull the curve down, but do reduce the denominator for later steps. The variance of the estimate is commonly obtained from Greenwood's formula, supporting confidence intervals around the curve. Because it assumes no parametric form, the estimator is robust and widely applicable; group comparison is typically done with the log-rank test, which contrasts observed and expected events across groups over time (Kaplan & Meier, 1958; Bland & Altman, 1998).

Clinical relevance

Kaplan-Meier curves are the most common way prognosis and treatment effects on survival are displayed in clinical literature, and reading them — including the numbers at risk and the median survival — is a core appraisal skill. This entry explains the method descriptively and is not a basis for individual prognostic or treatment decisions.

Epidemiology

The estimator is used across essentially all medical fields that study time to an event, from oncology trials to cohort studies; its 1958 paper is among the most highly cited in all of science, reflecting how routine the method has become (Kaplan & Meier, 1958).

Evidence & guidelines

There are no clinical guidelines for the estimator itself; the methodological reference standard is Kaplan and Meier's 1958 paper, with widely used tutorials (Bland & Altman, 1998; Clark et al., 2003) and texts (Collett, 2015) describing best practice, including reporting numbers at risk and confidence intervals.

History

Kaplan and Meier introduced the product-limit estimator in 1958, unifying earlier actuarial life-table ideas into a rigorous nonparametric estimate that handles censoring exactly; their independent work was merged into a single landmark paper. The log-rank test for comparing curves and Greenwood's earlier variance formula complete the standard toolkit that accompanies the estimator (Schoenfeld, 1981).

Debates

When is the log-rank test the right comparison?

The log-rank test is most powerful under proportional hazards; when hazards cross or survival curves diverge non-proportionally it can lose power, motivating weighted or alternative tests, an issue tied to the asymptotic theory of these nonparametric comparisons.

Key figures

• Edward L. Kaplan
• Paul Meier
• Major Greenwood
• Douglas Altman

Seminal works

• kaplan-meier-1958

Frequently asked questions

Why does the Kaplan-Meier curve look like a staircase?

It changes only at observed event times, stepping down at each event and remaining flat in between, because survival probability is updated only when an event is seen, not while subjects are merely under observation.

How do censored subjects affect the curve?

A censored subject leaves the risk set at their censoring time without causing a step down, but reduces the number at risk used to compute later steps, so the curve reflects only the information actually observed.