AI Glossary / Statistics and probability / Intermediate
Kaplan-Meier estimator
Definition
The Kaplan - Meier estimator is a nonparametric estimate of a survival function from time-to-event observations, including right-censored observations. Also called the product-limit estimator, it multiplies estimated conditional probabilities of remaining event-free at successive observed event times. Its step-shaped curve estimates the probability of an event time exceeding a specified duration. It is widely useful beyond medical survival, including equipment lifetime and customer-duration analysis. Its ability to use incomplete follow-up depends on assumptions about censoring and does not make all missing outcomes harmless.
What it means
A right-censored observation tells us that the event had not occurred by the last observed time; it does not reveal the eventual event time. Kaplan - Meier uses that observation while it remains under follow-up and removes it from later risk sets. Censoring itself does not cause a downward jump in the curve. The method avoids choosing a parametric lifetime distribution, but still needs a clearly defined event, time origin and appropriate observation process. In a basic analysis, censoring should be independent of the event time, with extensions or conditioning needed when that assumption is only plausible within groups.
How it works
Order distinct observed event times. At time tᵢ, count nᵢ observations at risk immediately before that time and dᵢ events at the time. Compute S-hat(t) = product over tᵢ ≤ t of (1 − dᵢ/nᵢ). Between event times the estimate remains constant. Right-censored observations contribute to risk sets up to their censoring time. Handle tied events and censoring according to a stated convention, and account appropriately for delayed entry if present. Report uncertainty intervals and numbers at risk because late portions of a curve may depend on very few observations.
Three practical examples
- Equipment reliability: estimate the proportion of devices still operating when some devices have not failed by the study's end.
- Customer duration: summarize time to a defined cancellation event while recognizing customers whose follow-up is incomplete.
- Operational studies: describe time until a specified process milestone without assuming an exponential or Weibull distribution. In each setting, reasons for loss to follow-up need scrutiny before interpreting the curve.
Think of it this way
Think of a sequence of checkpoints. At each observed event time, calculate what fraction of those still eligible to experience the event avoid it at that checkpoint, then multiply those fractions. An observation that leaves follow-up has contributed information up to departure but cannot be counted as known event-free forever. The approach preserves partial information without pretending to know the unseen remainder of each duration.
Worked example
Follow six units from time zero. Events occur at times 2, 4 and 5; other units are right-censored at 3, 6 and 7. At time 2, all six are at risk, so survival becomes 5/6. The censoring at 3 produces no jump, but only four units remain at risk just before the event at 4. Survival then becomes (5/6)(3/4) = 5/8. At time 5, three units are at risk, giving (5/8)(2/3) = 5/12, approximately 0.4167. Later censoring does not lower the estimate. The estimated median event time is 5, the first time the curve reaches or falls below 0.5. The area under the curve through time 6 is 2 + 2(5/6) + 1(5/8) + 1(5/12) = 113/24, approximately 4.7083. This is a restricted mean through 6, not an unrestricted lifetime mean.
Illustrative case study
Fictional case study: Maple Support estimates time until customers close an onboarding ticket. Several customers disappear from tracking after difficult cases are transferred to another system. An initial Kaplan - Meier report treats those departures as ordinary censoring. The analyst recognizes that transfer may relate to longer resolution times and investigates the missing follow-up before trusting the curve. The team also displays the number of unresolved cases still observed at each milestone. The revised report distinguishes useful partial information from an observation process that may bias the estimate.
Common misconceptions and mistakes
Common mistakes include treating censored durations as events, discarding them entirely, or counting them as indefinitely event-free. Nonparametric does not mean assumption-free. A curve based on very few late observations can be unstable, and a median may be unavailable if the curve never falls to 0.5. The unrestricted mean is not identified merely by extending a positive final plateau indefinitely. With competing events, one minus Kaplan - Meier is generally not the cumulative incidence of a particular event type when competing events are simply censored. Group differences also need not have a causal interpretation.
Business use
Businesses can use the estimator to summarize durations when follow-up lengths differ, provided event definitions and censoring mechanisms are defensible. Show the curve with uncertainty and risk-set information, and select reporting horizons supported by actual follow-up. Keep entry rules, exclusions and event coding reproducible. When comparing groups, consider differences in customer mix, calendar time and observation practices. If a decision needs extrapolation beyond the observed range, introduce and evaluate an additional model rather than claiming that the Kaplan - Meier steps supply that forecast automatically.
People also ask
Does a censored case count as an event?
No. It indicates that the event time exceeds the observed follow-up duration.
Does censoring reduce the curve immediately?
No. It changes later risk sets rather than producing an event jump.
Is a median always available?
No. It may not be reached during observed follow-up.
Check your understanding
Why are four units at risk at time 4 in the example?
Answer: One of the original six had an event at time 2 and another was censored at time 3. Both leave the later risk set, although only the event changed the survival estimate at its time.
Related terms
- Survival function
- Risk set
- Restricted mean survival time
- right censoring
- hazard
- competing risks
- log-rank test
Learn next
build a risk-set table manually and compare it with software output. Then examine confidence intervals, censoring assumptions and the difference between ordinary survival probability and event-specific cumulative incidence when competing events exist.
More in Statistics and probability
- Kaniadakis Gamma distribution - The Kaniadakis Gamma distribution is a κ-deformed family for positive random variables, built by replacing the exponential component of a generalized Gamma...
- Kaniadakis Gaussian distribution - The Kaniadakis Gaussian distribution, or κ-Gaussian, is a symmetric probability family obtained by normalizing a κ-exponential of a negative quadratic...
- Kaniadakis Weibull distribution - The Kaniadakis Weibull distribution, or κ-Weibull, is a positive-valued probability model commonly defined by the survival function S(x) = expκ(−βx^α), with...
- Kendall's coefficient of concordance ( W ) - Kendall's coefficient of concordance, written W, measures agreement among several rankings of the same set of objects.
- Kendall's tau - Kendall's tau is a rank-based measure of association between two variables, calculated by comparing the ordering of pairs of observations.
- Kernel perceptron - A kernel perceptron is a perceptron classifier expressed through kernel evaluations between examples.
Sources
Sources: Frank Harrell, Glossary of Statistical Terms, https://hbiostat.org/glossary/. NIST/SEMATECH e-Handbook, Empirical model fitting—distribution free approach, https://www.itl.nist.gov/div898/handbook/apr/section2/apr215.htm. NIST, Censoring, https://www.itl.nist.gov/div898/handbook/apr/section1/apr131.htm. Worked example and fictional case are original.
Previous: Kaniadakis Weibull distribution Next: Karmarkar's algorithm