Kaplan–Meier Estimator

Kaplan–Meier estimator (also known as the product-limit estimator) is the standard non-parametric statistical method in biostatistics, clinical trials, and longitudinal epidemiology for estimating the survival function S(t)—the probability that an individual survives beyond elapsed time t without experiencing a defined failure event—in the presence of incomplete, right-censored observations. Formulated independently in the 1950s by Edward L. Kaplan and Paul Meier and published jointly in the Journal of the American Statistical Association in 1958 after journal editor John Tukey persuaded them to combine their manuscripts, the product-limit estimator solved a long-standing analytical problem in medical follow-up studies and prospective cohort studies: how to utilize every day of person-time contributed by patients who withdraw early, are lost to follow-up, or remain event-free when the study ends, without discarding incomplete observations or assuming a parametric distribution for survival times.

Survival Function and Conditional Probability Chain Rule

Let T denote a non-negative random variable representing the elapsed time from a well-defined origin (t = 0, such as randomization in a clinical trial, surgical resection, or cohort enrollment mapped along a Lexis diagram) until a binary event of interest occurs (death, cancer recurrence, or infection). The true population survival function is S(t) = Pr(T > t). Suppose distinct observed failure events in a sample occur at ordered times t₁ < t₂ < ··· < t_k. For each ordered event time t_i:

• Let n_i denote the number at risk just prior to time t_i (all individuals who have neither experienced the event nor been right-censored before t_i).
• Let d_i denote the number of events (deaths or failures) occurring at time t_i.
• Let c_i denote the number of observations right-censored in the half-open interval [t_i, t_i+1), so that the risk set at the next failure time updates recursively via n_i+1 = n_i − d_i − c_i.

By the law of conditional probability, surviving past t_i requires surviving past t_i−1 and then surviving through the instant t_i. The empirical conditional probability of failing at t_i is d_i / n_i, and the conditional probability of surviving through t_i is 1 − (d_i / n_i). Multiplying these conditional probabilities across all observed event times t_i ≤ t yields the Kaplan–Meier product-limit estimator:

Graphically, Ŝ(t) is a right-continuous, monotonically non-increasing step function that starts at Ŝ(0) = 1.0, stays flat between event times, drops vertically by factor (1 − d_i / n_i) at each observed failure time t_i, and marks right-censored subjects with vertical tick marks (+ or |) along the horizontal treads where they reduce n_i+1 without causing Ŝ(t) to drop.

Step-by-Step Numerical Calculation Table

Greenwood's Formula and Log–Log Transformation

Applying the delta method to ln Ŝ(t) = ∑ ln(1 − d_i / n_i) yields Greenwood's formula (originally derived by Major Greenwood in 1926 for actuarial life tables) for the asymptotic variance of the Kaplan–Meier estimator:

Because naive linear confidence intervals Ŝ(t) ± 1.96 × SE[Ŝ(t)] can exceed 1.0 near t = 0 or fall below 0.0 in the right tail, biostatisticians construct confidence bands on the complementary log–log scale, g(S) = ln[ −ln Ŝ(t) ], which maps (0, 1) onto (−∞, +∞) and guarantees that back-transformed 95% confidence limits remain inside [0, 1]. The non-parametric median survival time is read directly as the smallest time t_0.50 at which the step function crosses Ŝ(t) ≤ 0.50.

The Mantel–Cox Log-Rank Test and Non-Informative Censoring

To test the null hypothesis H₀ : S₁(t) = S₂(t) across two or more treatment arms, investigators construct a 2 × 2 contingency table at every distinct failure time t_i and compute the expected number of events E_1i = d_i (n_1i / n_i) under H₀. Summing observed minus expected events across all k event times gives the Mantel–Cox log-rank statistic:

where V_1i is the hypergeometric variance at time t_i. The log-rank test is mathematically equivalent to the score test of a univariate proportional hazards model and attains maximum statistical power when hazards in the two groups are proportional over time.

Unbiased Kaplan–Meier estimation requires independent (non-informative) right-censoring: at any time t, individuals who are censored must have the same conditional probability distribution of future failure as individuals who remain under observation in the risk set n_i.

How does the Kaplan–Meier estimator relate to the Nelson–Aalen cumulative hazard estimator?

The Nelson–Aalen estimator non-parametrically estimates the cumulative hazard function Ĥ(t) = ∑_{ti ≤ t} (d_i / n_i). Because the theoretical survival and cumulative hazard functions are linked by S(t) = exp[−H(t)] and the first-order Taylor expansion of exp(−d_i / n_i) is 1 − d_i / n_i, the Kaplan–Meier estimator is the product-limit first-order approximation to the Breslow–Fleming–Harrington survival estimator exp[−Ĥ(t)].

Why does a Kaplan–Meier curve overestimate cumulative incidence in the presence of competing risks?

When patients can experience a competing risk that precludes the primary event (for example, dying of myocardial infarction before experiencing local cancer recurrence), treating competing-event deaths as ordinary right-censoring violates non-informative censoring because dead patients can never experience recurrence. Standard 1 − Ŝ(t) then overestimates true event probability, requiring the Aalen–Johansen cumulative incidence function (CIF) instead.

1. Greenwood, M. (1926). "The natural duration of cancer." Reports on Public Health and Medical Subjects, No. 33, 1–26. Her Majesty's Stationery Office.
2. Kaplan, E. L., & Meier, P. (1958). "Nonparametric estimation from incomplete observations." Journal of the American Statistical Association, 53(282), 457–481. https://doi.org/10.1080/01621459.1958.10501452
3. Mantel, N. (1966). "Evaluation of survival data and two new rank order statistics arising in its consideration." Cancer Chemotherapy Reports, 50(3), 163–170. https://pubmed.ncbi.nlm.nih.gov/5910392/
4. Peto, R., & Peto, J. (1972). "Asymptotically efficient rank invariant test procedures." Journal of the Royal Statistical Society: Series A, 135(2), 185–207. https://doi.org/10.2307/2344317
5. Kalbfleisch, J. D., & Prentice, R. L. (2002). The Statistical Analysis of Failure Time Data (2nd ed.). John Wiley & Sons. https://doi.org/10.1002/9781118032985