The Kaplan–Meier estimator is the workhorse statistic for estimating how long things—or people—last when some of them drop out of view before their fate is known. A new open-access paper in the Springer journal Discover Data, published on 14 July 2026, now offers a fresh way to look inside this venerable method. Its author, Malte C. Tichy, shows that the familiar product-limit formula can be rewritten exactly as a simple sum of contributions from each individual unit in the dataset, a decomposition that promises to make survival curves far easier to explain, visualize, and critically assess.
The problem the estimator solves is ubiquitous. In medicine, engineering, business, and science, researchers track a non-negative continuous quantity such as a patient’s survival time or the functional lifetime of a technical component. Because studies end and subjects leave datasets, much of the data is right-censored: a unit is observed only until some age, after which its eventual failure time is unknown. The Kaplan–Meier estimator, formalized by Kaplan and Meier after long use in the demographic and actuarial sciences, handles this gracefully and has become the de facto textbook standard for estimating cumulative failure probability in partially right-censored populations.
In its textbook form, the estimator for the cumulative failure distribution at age τ is a product over all event ages up to τ, with each factor equal to one minus the ratio of failures to units still at risk at that age. When no censoring occurs, this product collapses into something much more intuitive: the empirical cumulative distribution function, which is simply a sum of step functions, one per unit. Each unit’s step function is zero before its observed failure age and one afterwards, so the population curve is just the fraction of units that have failed by that age. Every unit’s fate is known, so every contribution is unambiguous.
The new paper asks whether the same unit-level clarity can be recovered when censoring obscures some fates. The answer is yes. For a unit whose failure was observed, the contribution remains the empirical step function jumping to one at the failure age. For a right-censored unit, the author assigns the Kaplan–Meier estimator of the risk set consisting of all units that are still functional and uncensored immediately after the censoring age. Averaging these individual contributions over all units reproduces the population-level Kaplan–Meier estimator exactly, a result the paper proves by induction.
The construction has a pleasing self-consistency. A censored unit’s estimator matches the Kaplan–Meier estimator of the remaining, later-observed units, which is precisely the maximum-likelihood best guess for that unit’s fate given the available information. The paper also shows this unit-wise form is equivalent to an imputation approach in which the unknown post-censoring behavior of each censored unit is filled in using the Kaplan–Meier curve itself as the imputation function. The result connects to a long line of alternative interpretations, including redistribution-to-the-right heuristics introduced by Efron and implemented by Dinse, and a representation by Satten and Datta as an average of step functions weighted by the probability of censoring.
The decomposition yields some surprising insights. One is an idempotence property: adding new units that are censored before the first observed failure leaves the Kaplan–Meier estimator completely unchanged. Such units neither survive a previously observed failure nor contribute a failure of their own, so they add no information. But the picture changes subtly if many such units are added. The numerical value of the curve stays the same, yet the contributions shift: the estimator becomes less a description of what actually happened and more a prediction of what will happen, borne increasingly by imputed rather than observed contributions.
To make this concrete, the paper visualizes a small dataset of six units with observed ages one through six and failure markers indicating which were censored. Units whose failures were observed contribute constant steps to the overall curve, while censored units contribute rising segments reflecting assumed failure likelihoods. The final unit, which never failed within the observation window, contributes nothing at all—in fact, the existence of such an aged survivor pushes the overall estimate downward. The first unit, censored early with no prior failures, has an individual estimator identical to the full population curve, since everything that can be said about it derives from the other units’ behavior.
Larger simulated examples drive the point home. The author simulates 100 units failing under a Weibull process with shape parameter 1.4 and scale parameter 1, then applies different Weibull censoring processes. With heavy censoring, only a few failures are observed after early ages, and the estimator beyond that point is driven almost entirely by imputation, with larger steps reflecting the small remaining risk population. With weak censoring, observed failures dominate for much longer. Under strong infant mortality, many units are censored early, while a censoring process with high shape leaves the curve almost purely empirical until late ages. The paper introduces an “imputation fraction” quantifying what share of the curve at each age is borne by censored units, and shows how this fraction evolves differently under each censoring regime.
The practical upshot is a diagnostic tool. Analysts can now assess whether a Kaplan–Meier curve at a given age is primarily data-driven or imputation-driven, which matters because artifacts of Kaplan–Meier plots may be misread as evidence for or against alignment with a predictive model. The author emphasizes that although the estimator is parameter-free, it is not assumption-free: its stepwise form keeps the estimated failure probability for a censored unit constant between observed failure ages, however far into the future the next failure lies, which can produce paradoxical situations where other imputation assumptions would be more appropriate. The decomposition makes these mechanics transparent and helps judge when the estimator fits the data at hand.
The paper closes with open questions. The unit-wise representation is a deterministic decomposition of the point estimate and does not directly yield confidence intervals, though resampling observed failures and censored units separately might distinguish statistically induced from censoring-induced uncertainty. The analysis covers only right-censored data, where a closed formula exists; extending the approach to left-, right-, and interval-censored populations, where no closed form of the Turnbull estimator is available, remains speculative and demanding. An iterative algorithm whose fixed point is the unit-level estimator, and an analogous representation for the Nelson–Aalen hazard estimator, are sketched as promising directions for the many fields that rely on finite, partially censored survival datasets.
Subject of Research: A sum-over-units decomposition of the Kaplan–Meier estimator for right-censored survival data
Article Title: The Kaplan–Meier estimator as a sum over units
Article References: Tichy, M. C. (2026). The Kaplan–Meier estimator as a sum over units. Discover Data, 4(1), Article 11. https://doi.org/10.1007/s44248-026-00111-4
Image Credits: AI Generated
DOI: 10.1007/s44248-026-00111-4
Keywords: Kaplan–Meier estimator, survival analysis, right-censoring, cumulative failure distribution, imputation, maximum likelihood, Weibull process, risk set, empirical CDF, Nelson–Aalen estimator, statistical visualization, Discover Data
News Source: Denise Maddox. (October 29, 2026). Survival Curve Secrets: Every Unit Tells Its Own Story in the Kaplan–Meier Estimator. Scienmag.
