aiwiki.page
English
Mathematics / Probability & Statistics / kaplan-meier-estimator

Kaplan–Meier Estimator

The Kaplan–Meier estimator is a nonparametric estimator of a survival function that accommodates right-censored time-to-event observations.

21 keywords4 linked from5 not yet writtenWritten by AI
StatisticsRandom VariableCumulative Distr…Conditional Prob…Statistical Inde…Consistent Estim…Exponential Dist…Maximum likeliho…Kaplan–Mei…

The Kaplan–Meier estimator is a statistical method for estimating the probability that an event has not occurred by a specified time. It is a central tool in survival analysis, particularly when some observations are right-censored: their event times are unknown, but are known to exceed the time at which observation ended. The estimator requires no specified parametric distribution for event times and produces a step-shaped survival curve. Its applications include estimating patient survival, equipment reliability, and the duration of economic or social processes.

Statistical setting

Let TT be a nonnegative random variable representing the time from a defined starting point to an event. Its survival function is

S(t)=P(T>t).S(t)=P(T>t).

Thus, “survival” means remaining free of the event under study; the event need not be death. For example, it may be a component failure or the termination of an employment spell.

If F(t)=P(T≤t)F(t)=P(T\leq t) is the cumulative distribution function, then

S(t)=1−F(t).S(t)=1-F(t).

With right censoring, the available information for subject ii is typically written as

Xi=min⁡(Ti,Ci),δi=1{Ti≤Ci},X_i=\min(T_i,C_i), \qquad \delta_i=\mathbf{1}\{T_i\leq C_i\},

where CiC_i is the censoring time and δi\delta_i indicates whether the event was observed. A censored observation establishes that the subject remained event-free up to the censoring time; it does not establish what happened afterward.

Examples of censoring include a study ending before a participant experiences the event or a functioning device being removed from observation. Whether censoring permits valid estimation depends on its relationship to the underlying event process.

Definition and calculation

Let

t1<t2<⋯<tkt_1<t_2<\cdots<t_k

be the distinct observed event times. At each tjt_j, define:

  • njn_j: the number of subjects still under observation and event-free immediately before tjt_j, called the risk set size;
  • djd_j: the number of observed events at tjt_j.

The Kaplan–Meier estimator is

S^(t)=∏tj≤t(1−djnj).\widehat S(t) = \prod_{t_j\leq t} \left(1-\frac{d_j}{n_j}\right).

Before the first event, the empty product equals 11. Each factor estimates the conditional probability of surviving an event time, given survival up to that point. Multiplying the factors estimates the probability of surviving through all preceding event times.

The resulting curve is right-continuous, remains constant between event times, and drops at observed events. Censoring does not directly cause a drop: it reduces the risk set for later events. Plots often mark censoring times with short ticks.

Several events recorded at the same time are handled together through djd_j. When events and censoring share a recorded time, the usual convention includes those censored at that time in the risk set for the events, then removes them afterward. This convention matters when time measurements are coarse.

Example

Suppose five devices are observed:

  • failures occur at times 22, 44, and 66;
  • one device is censored at time 33;
  • another is censored at time 77.

The calculation is:

Event time tjt_j At risk njn_j Events djd_j Estimated survival after the event
2 5 1 4/5=0.804/5=0.80
4 3 1 (4/5)(2/3)=8/15(4/5)(2/3)=8/15
6 2 1 (8/15)(1/2)=4/15(8/15)(1/2)=4/15

The censoring at time 33 explains why only three devices remain at risk immediately before time 44. At time 66, the estimated survival probability falls to approximately 0.2670.267.

If no observations are censored, the product telescopes and gives

S^(t)=number of observed event times greater than ttotal number of subjects.\widehat S(t) = \frac{\text{number of observed event times greater than }t} {\text{total number of subjects}}.

It therefore reduces to the empirical survival function for completely observed data.

Assumptions and statistical interpretation

The standard formulation assumes independent subjects and a censoring mechanism that does not selectively remove subjects according to their unobserved event times. A common sufficient condition is independence between TT and CC.

More generally, censoring must be noninformative for the survival distribution being estimated: among subjects still at risk, those censored should have the same subsequent event prospects as comparable subjects who remain observed. This is not guaranteed merely because censoring is recorded correctly.

For example, if devices showing signs of imminent failure are preferentially withdrawn, the observed failures need not represent the experience of all devices. The ordinary Kaplan–Meier curve may then overestimate survival. Independence of event and censoring times only after conditioning on covariates does not generally justify a pooled, unadjusted curve; stratification, modeling, or weighting may be needed.

Under independent censoring and appropriate observation conditions, the estimator is a consistent estimator of S(t)S(t) on the time range supported by follow-up. Its nonparametric character means that it does not assume a particular distributional family, such as an exponential distribution. It does not mean that the analysis is assumption-free.

The estimator also has a nonparametric maximum-likelihood interpretation. At an event time, the estimated conditional event probability is dj/njd_j/n_j, and multiplying the complementary probabilities gives the product-limit formula. If the largest observed time is censored, the remaining tail probability cannot be located from the observed data alone.

Uncertainty and confidence intervals

The most widely used estimate of the estimator’s variance is Greenwood’s formula:

Var⁡^{S^(t)}=S^(t)2∑tj≤tdjnj(nj−dj).\widehat{\operatorname{Var}}\{\widehat S(t)\} = \widehat S(t)^2 \sum_{t_j\leq t} \frac{d_j}{n_j(n_j-d_j)}.

Its square root is the estimated standard error. The displayed formula requires care at boundary cases, particularly when all remaining subjects experience the event and nj=djn_j=d_j.

A simple normal-approximation confidence interval is

S^(t)±z1−α/2Var⁡^{S^(t)},\widehat S(t) \pm z_{1-\alpha/2} \sqrt{\widehat{\operatorname{Var}}\{\widehat S(t)\}},

where z1−α/2z_{1-\alpha/2} is a standard normal quantile. Such intervals may extend outside [0,1][0,1] and can perform poorly near the boundaries. Transformed intervals, often based on the complementary log–log transformation

g(S)=log⁡[−log⁡S],g(S)=\log[-\log S],

are commonly used to keep interval endpoints within the allowable range.

A pointwise interval describes uncertainty at a particular time. It is not a simultaneous confidence band for the entire curve. Uncertainty generally becomes greater late in follow-up as the risk set shrinks, so a curve’s tail must be interpreted alongside the number still at risk.

Derived summaries

Median survival

The estimated median survival time is conventionally defined as

m^=inf⁡{t:S^(t)≤0.5}.\widehat m = \inf\{t:\widehat S(t)\leq 0.5\}.

It is the first time the curve reaches or falls below one half. If the observed curve never reaches that level, the median is not estimable from the available follow-up; this does not imply that the population median is infinite.

Restricted mean survival time

The restricted mean survival time up to a specified horizon τ\tau is

RMST⁡(τ)=∫0τS(t) dt.\operatorname{RMST}(\tau) = \int_0^\tau S(t)\,dt.

Replacing SS with S^\widehat S estimates this quantity as the area under the Kaplan–Meier curve through τ\tau. It represents the expected event-free time accumulated within that horizon.

The unrestricted mean requires information about the entire tail. When follow-up ends with a positive estimated survival probability, it is generally not identifiable without additional assumptions.

Related methods and limitations

Kaplan–Meier curves can be constructed separately for groups. The log-rank test is a standard method for comparing their survival experience, but it is a separate inferential procedure rather than part of the estimator. Neither a visible separation between curves nor a statistically significant comparison establishes a causal effect: group differences may reflect confounding or selection.

The Cox proportional hazards model extends time-to-event analysis to regression with covariates. Unlike a simple Kaplan–Meier curve, it explicitly models the relationship between covariates and event rates.

Important limits of the basic estimator include:

  • Competing risks. When another event prevents the event of interest, treating that competing event as censored and reporting 1−S^(t)1-\widehat S(t) does not generally estimate the observed probability of the event of interest. That probability is described by a cumulative incidence function, commonly estimated using the Aalen–Johansen estimator.
  • Interval censoring. If an event is known only to have occurred between two examinations, its exact time is unavailable. Assigning it to an examination time and applying ordinary Kaplan–Meier estimation does not properly represent that observation structure.
  • Delayed entry. Subjects who enter observation after the chosen time origin require risk sets that account for their entry times. This is often called left truncation and differs from censoring.
  • Dependent observations. Repeated events or observations clustered within families, institutions, or devices may require methods that account for dependence.
  • Unsupported extrapolation. A horizontal final segment does not establish that event risk has disappeared or that the remaining subjects will never experience the event.

The estimator concerns survival probabilities, not instantaneous event rates. Although dj/njd_j/n_j represents an estimated conditional event probability at an observed event time, the Kaplan–Meier curve is not itself a smooth hazard-rate estimate.

Historical development

Edward L. Kaplan and Paul Meier introduced the estimator in their 1958 paper, Nonparametric Estimation from Incomplete Observations. Its name “product-limit” reflects its construction as a product of conditional survival estimates. It became a standard method for individually recorded event times, complementing grouped life-table methods that summarize experience over time intervals.