Statistics Glossary → Kaplan-Meier estimator (survival curve)
Kaplan-Meier estimator (survival curve)concept
A nonparametric method that estimates the survival function step by step from censored time-to-event data by multiplying the probabilities of surviving each event time.
At each event time tⱼ, nⱼ people are at risk and dⱼ experience the event; the probability of getting past that time is 1 - dⱼ/nⱼ. The Kaplan-Meier (product-limit) estimator multiplies these conditional probabilities: S(t) = Π (1 - dⱼ/nⱼ). Censored individuals (whose follow-up ended or who left the study) count among those at risk until the time they are censored and are then removed, so incomplete follow-up is used rather than discarded. The result is a step curve that drops only at event times; its standard error is usually computed with Greenwood's formula.
The median survival time is the first time the curve falls below 0.5; if it never reaches 0.5, the median cannot be estimated. The method assumes non-informative censoring (independent of the risk of the event); if, say, patients who are getting worse leave the study more often, the curve will be optimistic. At the right end of the curve few people remain at risk, so the estimate is very uncertain. When there are competing risks (e.g. death from another cause before the event of interest), 1 - KM overstates the cumulative probability of the event of interest, and the cumulative incidence function should be used instead.
An everyday example
A phone company follows how long 10 new subscribers keep their contract. In month 3, 1 of 10 leaves (S = 0.90). In month 5, one moves away and drops out of follow-up (censored). In month 7, 2 of the remaining 8 leave: S = 0.90 × (1 - 2/8) = 0.675. The subscriber who moved is counted neither as 'stayed' nor as 'left'; they contribute only up to month 5.
In MerQur
The core output of the Kaplan-Meier survival analysis; with the 'Time points' option, S(t), SE, CI and the number at risk are reported at the given times. In Cox and time-dependent Cox regression Kaplan-Meier is one of the time transforms of the proportional hazards test; competing risks analysis is used to avoid the bias Kaplan-Meier creates with competing events, and interval-censored survival uses the Turnbull NPMLE where Kaplan-Meier cannot be applied.
Analyses where this concept appears
Kaplan-Meier Survival AnalysisCox Regression (Proportional Hazards Model)Competing RisksTime-dependent Cox RegressionInterval-Censored Survival
Related concepts
Censoring (survival analysis)Log-rank test (and Breslow, Tarone-Ware)Cumulative hazard and log-minus-log plot