Kaplan-Meier Estimator

What the Kaplan–Meier estimator does

It updates the estimated probability of remaining event-free at each observed event time while retaining censored participants in the risk set until their censoring time. This page explains the product-limit calculation, censoring assumptions, uncertainty, comparison and use in health-economic modelling.

“Survival” can mean freedom from any clearly defined event, not only death. The event definition, time origin and analysis population must be specified before the curve can be interpreted.

The survival function describes time without the event

For event time (T), the survival function is:

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

It gives the probability that an individual remains event-free beyond time (t). The corresponding cumulative distribution function is:

$$F(t)=P(T\le t)=1-S(t)$$

The Kaplan–Meier estimator represents (S(t)) as a right-continuous step function. It changes only at observed event times; censoring reduces later risk sets but does not itself create a downward step.

Risk sets drive the product-limit calculation

Let (t_1<t_2<\cdots<t_k) be the distinct observed event times. At event time (t_j), let (n_j) be the number at risk immediately before the time and (d_j) the number experiencing the event at that time. The conditional probability of surviving through that event time is estimated as:

$$1-\frac{d_j}{n_j}$$

The Kaplan–Meier estimator multiplies those conditional probabilities through time:

$$\widehat{S}(t) = \prod_{t_j\le t} \left(1-\frac{d_j}{n_j}\right)$$

This product-limit form allows participants with different follow-up lengths to contribute all information observed before their event or censoring time.

A worked Kaplan–Meier calculation

Suppose six participants enter a study. Events occur at months 2, 4 and 7; one participant is censored at month 3, one at month 6 and one remains event-free through the end of follow-up. The risk sets just before the event times are 6, 4 and 2.

The event-time calculation is:

The estimated survival probability just after month 7 is therefore:

$$\widehat{S}(7)=0.3125$$

The censoring at months 3 and 6 changes the later risk sets but does not reduce the curve at those times. The participant remaining event-free at study end is censored at the administrative end of follow-up.

Censoring contributes partial follow-up information

Right censoring occurs when the event has not been observed by a participant's last known follow-up time. The participant's exact event time is then unknown, but it is known to exceed the censoring time. Kaplan–Meier estimation uses that partial information without treating censoring as the event.

Common reasons include:

• Administrative end of study before the event occurs.
• Loss to follow-up.
• Withdrawal from observation.
• Entry into another state that ends observation under the analysis definition.

These mechanisms are not automatically equivalent. Administrative censoring may be relatively benign, while loss to follow-up related to prognosis can create informative censoring and bias the curve.

Non-informative censoring is a central assumption

The standard estimator assumes that, conditional on any structure used in the analysis, those censored at a time have the same subsequent survival prospects as comparable participants who remain under observation. This is often described as independent or non-informative censoring. It cannot be verified from the observed censored data alone.

Credibility should be examined through:

• Reasons and timing of censoring by treatment group.
• Baseline and time-varying characteristics associated with loss to follow-up.
• Differences between participants with complete and incomplete follow-up.
• Sensitivity analyses for plausible informative-censoring mechanisms.
• Whether treatment switching, competing events or withdrawal are being misclassified as censoring.

Heavy or differential censoring makes the tail of a curve particularly uncertain, even when the estimator is calculated correctly.

Delayed entry requires the correct risk set

Some participants become observable only after the chosen time origin. This left truncation, or delayed entry, requires them to join the risk set at their entry time rather than at time zero. Including them from the origin would give them risk time during a period in which they could not have been observed.

With delayed entry time (L_i) and exit time (Y_i), participant (i) contributes to the risk set at (t_j) only when:

$$L_i<t_j\le Y_i$$

The exact boundary convention should match the event and data definitions. Left censoring, interval censoring and left truncation are different data structures and do not all use the ordinary right-censored Kaplan–Meier estimator.

Tied events are handled at the same time point

When several events share the same recorded time, they are grouped in (d_j) and enter one survival step. Participants censored at that recorded time require an explicit ordering convention because the available data may not show whether censoring occurred before or after the events.

A standard implementation treats individuals with an event at (t_j) as at risk immediately before (t_j). Software conventions for censoring ties and time intervals should be checked, especially when time is recorded coarsely and ties are common.

Greenwood's formula estimates sampling variance

Greenwood's formula provides a commonly used variance estimate for the Kaplan–Meier curve:

$$\widehat{\operatorname{Var}}\!\left[\widehat{S}(t)\right] = \widehat{S}(t)^2 \sum_{t_j\le t} \frac{d_j}{n_j(n_j-d_j)}$$

The formula becomes unstable or undefined at a time where all remaining participants experience the event. Pointwise confidence intervals often use transformations, such as the complementary log–log transformation, to respect the probability range and improve behaviour near zero or one.

A transformed interval is not a simultaneous confidence band for the entire curve. Pointwise intervals answer uncertainty at selected times, while simultaneous bands address coverage over a range of times.

The curve should show uncertainty and numbers at risk

A useful Kaplan–Meier plot displays time on the horizontal axis and estimated survival on the vertical axis. Downward steps occur at events, censoring marks show censored observations and a table reports numbers at risk at selected times. Confidence intervals reveal increasing uncertainty as the risk set becomes smaller.

Every graph should identify:

• The event and time origin.
• The analysis group or treatment strategy.
• Axis units and follow-up range.
• Censoring marks or an explanation if they are omitted.
• Numbers at risk at clinically meaningful times.
• The confidence-interval method when intervals are displayed.
• Whether the curve is unadjusted, weighted or otherwise modified.

Displaying a long horizontal tail after few participants remain can give a false impression of precision. The number-at-risk table is essential context for interpreting that region.

Median survival is a quantile of the curve

Median survival is the earliest time at which estimated survival is at or below 0.5:

$$\widehat{t}_{0.5} = \inf\left\{t:\widehat{S}(t)\le0.5\right\}$$

In the worked example, the curve is 0.625 after month 4 and 0.3125 after month 7, so estimated median survival is 7 months. If the curve never reaches 0.5 during observed follow-up, the median is not reached and should not be invented by extending the final step.

Other quantiles can be estimated similarly when the curve crosses the required probability. Confidence intervals for survival quantiles require methods appropriate to the inverted step function.

Mean survival requires integration and sufficient follow-up

Mean survival is the area under the survival curve:

$$E[T]=\int_0^{\infty}S(t)\,dt$$

The full mean is not identifiable non-parametrically when the Kaplan–Meier curve ends above zero because survival beyond the observed period is unknown. The restricted mean survival time, or RMST, avoids this problem by selecting a finite horizon (\tau):

$$\operatorname{RMST}(\tau) =\int_0^{\tau}S(t)\,dt$$

RMST is interpretable as expected event-free time up to (\tau). Treatment comparisons must use the same clinically justified horizon, chosen without selecting the value that produces the preferred result.

Comparing curves requires more than visual separation

The log-rank test evaluates a null hypothesis of equal survival functions by comparing observed and expected event counts over event times. It is most sensitive under proportional-hazards-type alternatives and may have limited power when curves cross or treatment effects occur mainly early or late.

A statistically significant log-rank test does not quantify the size or clinical importance of a difference. Useful summaries can include:

• Survival probabilities at prespecified times.
• Differences or ratios of those probabilities.
• Median survival when estimable.
• RMST and its between-group difference at a common horizon.
• A hazard ratio from a separate regression model when its assumptions and interpretation are suitable.

Kaplan–Meier curves are unadjusted unless weighting, stratification or model-based standardisation has been applied. Apparent group differences may reflect baseline imbalance in observational data.

Competing risks require a different probability estimator

A competing event prevents the event of interest from occurring or changes its interpretation. Treating a competing event as ordinary independent censoring and using (1-\widehat{S}(t)) generally overestimates the cumulative probability of the event of interest because it imagines that those experiencing the competing event remain capable of the event later.

For competing risks, the cumulative incidence function estimates the probability of a particular event type while accounting for other event types. Cause-specific Kaplan–Meier methods can address a hypothetical world where competing events are removed, but that estimand should not be confused with real-world cumulative incidence.

The choice depends on the question:

Death should not be censored casually when estimating the probability of a non-fatal outcome.

Kaplan–Meier estimates are not hazard estimates

The survival function and hazard function describe different aspects of event-time data. The hazard is an instantaneous event rate among those still at risk, while survival is a cumulative probability. Similar survival probabilities at one time do not imply equal hazards over the full follow-up period.

The cumulative hazard (H(t)) and survival function satisfy:

$$S(t)=e^{-H(t)}$$

The transformation (-\log\widehat{S}(t)) is related to cumulative hazard, but the Nelson–Aalen estimator directly estimates cumulative hazard using a sum of hazard increments. These estimators are closely related but not algebraically identical in finite samples.

Health-economic models use curves as evidence, not complete extrapolations

Kaplan–Meier curves provide an empirical summary during observed follow-up. Economic evaluations often need lifetime costs and health outcomes extending far beyond that period, so the curve alone is insufficient when substantial survival remains at study end. Parametric or flexible extrapolation then requires clinical plausibility, external evidence and sensitivity analysis.

Health-economic uses include:

• Estimating observed progression-free and overall survival in trial follow-up.
• Calculating restricted mean event-free time over a supported horizon.
• Informing health-state occupancy in partitioned-survival models.
• Assessing proportional-hazards assumptions and candidate extrapolation models.
• Comparing reconstructed and reported trial survival when individual data are unavailable.
• Benchmarking modelled survival against empirical evidence.

A fitted extrapolation should join the observed evidence coherently and should not be selected solely by in-sample statistical fit. Long-term predictions depend strongly on tail assumptions where the Kaplan–Meier estimate is weakest.

Reconstructing individual data from a published curve adds uncertainty

When individual participant data are unavailable, analysts may digitise published Kaplan–Meier coordinates and combine them with numbers at risk and event totals to approximate individual event and censoring times. This can support secondary modelling, but the reconstructed data are not the original trial records.

Potential errors arise from:

• Image resolution and digitisation precision.
• Incomplete or rounded numbers-at-risk tables.
• Unknown censoring distribution within intervals.
• Ambiguous curve overlap or censoring marks.
• Inconsistent published totals and follow-up summaries.
• Algorithmic assumptions used to allocate events and censoring.

Reconstructed data should be labelled clearly, validated by reproducing the published curve and key summaries, and examined in sensitivity analyses.

Common errors distort survival estimates

The Kaplan–Meier method is mechanically simple, but incorrect event definitions, risk sets or censoring assumptions can produce a convincing yet invalid curve.

• Counting censored observations as events produces artificial survival drops.
• Removing a censored participant before their censoring time discards valid risk time.
• Treating competing events as ordinary censoring overstates cumulative incidence.
• Reading a horizontal tail as evidence of long-term cure ignores sparse follow-up.
• Reporting median survival when the curve never reaches 0.5 invents an unsupported estimate.
• Comparing curves visually without uncertainty or numbers at risk hides imprecision.
• Using a log-rank p-value as the treatment-effect magnitude confuses testing with estimation.
• Ignoring delayed entry creates incorrect risk sets and possible immortal-time bias.
• Assuming censoring is independent without investigating reasons for loss to follow-up can bias results.
• Extending the Kaplan–Meier curve unchanged to a lifetime horizon creates an unjustified extrapolation.

A defensible Kaplan–Meier workflow

A high-quality analysis preserves the connection between the clinical question, event-time data, risk-set calculation and downstream decision use. Every curve should be reproducible from a documented record of time origin, events, censoring and inclusion rules.

1. Define the estimand: State the population, time origin, event, competing events, treatment strategy and follow-up horizon.
2. Prepare event-time data: Verify entry times, exit times, event indicators, duplicates, impossible dates and time units.
3. Classify censoring: Document administrative censoring, withdrawal, loss to follow-up and treatment changes separately.
4. Construct risk sets: Include each participant only while observed and at risk, with delayed entry handled explicitly.
5. Calculate survival: Apply the product-limit estimator at each distinct event time and verify monotonicity.
6. Quantify uncertainty: Use an appropriate variance estimator, confidence transformation and number-at-risk display.
7. Check assumptions: Investigate censoring patterns, competing risks, sparse tails and group comparability.
8. Summarise appropriately: Report survival at prespecified times, median when reached and RMST at a justified horizon.
9. Support modelling cautiously: Separate observed Kaplan–Meier evidence from parametric extrapolation and test long-term assumptions.
10. Report transparently: Provide event definitions, analysis rules, software, uncertainty methods, risk tables and unresolved limitations.

The Kaplan–Meier estimator is most informative within the region where participants remain under observation and at risk. Its step curve is an evidence summary, not a guarantee about unobserved survival beyond the data.