Kaplan–Meier Plug-in Estimator Overview

• Kaplan–Meier plug-in estimator is a nonparametric method that substitutes the empirical survival curve into functionals like average hazard and RMST under censoring.
• It extends the classical Kaplan–Meier approach by integrating modifications like differential privacy, weighting, and expert augmentation to address complex data structures.
• The method underpins asymptotic theory and bootstrap inference, ensuring robust statistical properties even in stratified or censored data settings.

The Kaplan–Meier plug-in estimator denotes a family of nonparametric estimators built from the Kaplan–Meier (KM) product-limit estimator by substitution into a target functional or into a broader estimating formula. In its classical form, the KM estimator targets the survival function $S(t)=P(T>t)=1-F(t)$ under right censoring by replacing the unknown survival law with the empirical product-limit curve (Dobler, 2016). In the wider literature, the same plug-in idea appears in several closely related senses: functionals such as the average hazard, restricted mean survival time, mean residual life, Lorenz curve, and Gini index are estimated by replacing $S$ with $\widehat S$; differentially private, weighted, inverse-probability-weighted, and expert-augmented estimators are obtained by replacing the classical KM inputs with modified counterparts; and asymptotic theory is often derived by treating KM as a smooth functional or as an M-estimator (Gondara et al., 2019, Heuser et al., 2017, Talbot et al., 2022, Uno et al., 21 Jul 2025, Gu et al., 2020).

1. Classical product-limit construction

For right-censored data, the KM estimator is the canonical nonparametric estimator of the survival function. If $t_1,\dots,t_k$ are the distinct failure times, $d_j$ is the number of failures at time $t_j$, and $r_j$ is the number at risk just before $t_j$, then

$$\hat S(t) = \prod_{j:t_j \le t} \frac{r_j - d_j}{r_j}.$$

This yields a step function that drops only at event times and remains constant through censoring times (Gondara et al., 2019).

Equivalent product-limit representations recur throughout the literature. One form writes

$$\widehat S_n(t) = \prod_{i: X_{i:n}\le t}\left(1-\frac{\delta_{[i:n]}}{n-i+1}\right),$$

for ordered observed times $S$0 and concomitant event indicators $S$1 (Dobler, 2016). Another equivalent notation is

$S$2

with $S$3 the at-risk denominator (Anevski, 2016). In all of these formulations, censoring affects future denominators rather than causing immediate downward jumps.

A recurring interpretive point is that the plug-in principle begins already at the level of the survival curve itself. In “Bootstrapping the Kaplan-Meier Estimator on the Whole Line,” the KM curve is described as the canonical plug-in estimator for the survival function, because later functionals of the full survival distribution are estimated by substituting $S$4 for $S$5 (Dobler, 2016). This usage is the most classical one.

2. Plug-in estimation of functionals of the survival curve

A central use of the KM plug-in principle is to estimate functionals of the full survival law by replacing $S$6 with $S$7. This includes both pointwise and whole-curve functionals.

For the average hazard over $S$8, the target is

$S$9

and the KM plug-in estimator is

$\widehat S$0

The numerator estimates cumulative incidence by $\widehat S$1, the denominator estimates restricted mean survival time up to $\widehat S$2, and the ratio estimates the average hazard (Uno et al., 21 Jul 2025).

For the attributable fraction with time-to-event outcomes, the target at time $\widehat S$3 is

$\widehat S$4

where $\widehat S$5 is the observed cumulative incidence and $\widehat S$6 is the cumulative incidence that would be observed if everyone were unexposed. The proposed estimator uses KM twice: once in the full sample to estimate observed survival, and once as a weighted KM among the unexposed after inverse probability of exposure weighting. The resulting plug-in estimator is

$\widehat S$7

This construction is explicitly designed to handle censoring and potentially non-proportional hazards (Talbot et al., 2022).

For whole-curve functionals, “Bootstrapping the Kaplan-Meier Estimator on the Whole Line” treats mean residual life, Lorenz curves, and the Gini index as direct KM plug-ins. The mean residual life-time function is

$\widehat S$8

so its estimator is obtained by replacing $\widehat S$9 with $t_1,\dots,t_k$0 (Dobler, 2016). The Lorenz curve and Gini index are likewise estimated by plugging $t_1,\dots,t_k$1 into the defining functionals (Dobler, 2016).

This plug-in logic also underlies restricted mean survival time estimation. In randomized trials, the RMST difference is the area under the arm-specific survival curves up to $t_1,\dots,t_k$2, and the unadjusted KM-based estimator is

$t_1,\dots,t_k$3

The covariate-adjusted TMLE proposed in (Díaz et al., 2015) targets the same marginal treatment effect as Kaplan–Meier, but replaces the unadjusted arm-specific survival estimates by a targeted, covariate-adjusted substitution estimator. The paper explicitly positions this as a precision-improved Kaplan–Meier estimator (Díaz et al., 2015).

3. Plug-in estimators based on modified Kaplan–Meier ingredients

A second major meaning of “Kaplan–Meier plug-in estimator” is obtained by replacing the classical KM ingredients with altered versions and then applying the ordinary product-limit formula.

In differentially private survival analysis, the sensitive sufficient counts $t_1,\dots,t_k$4 are replaced by private counts $t_1,\dots,t_k$5. The method constructs a partial count matrix $t_1,\dots,t_k$6, adds Laplace noise

$t_1,\dots,t_k$7

with global $t_1,\dots,t_k$8 sensitivity $t_1,\dots,t_k$9, reconstructs the risk sets recursively by

$d_j$0

and then defines

$d_j$1

The paper explicitly frames this as a Kaplan–Meier plug-in estimator: the release is obtained by replacing the sensitive sufficient counts $d_j$2 in the standard KM formula with their differentially private counterparts $d_j$3 (Gondara et al., 2019). The same private-count mechanism is then used, by post-processing, for Greenwood intervals, logrank tests, competing-risk cumulative incidence, and Nelson–Aalen estimation (Gondara et al., 2019).

In stratified population comparisons, the plug-in idea applies at the cohort level. For population $d_j$4, the marginal survival curve is

$d_j$5

and the weighted survival estimator is

$d_j$6

Here the cohort-specific $d_j$7 are the usual Kaplan–Meier estimates computed within each cohort (Heuser et al., 2017). The paper emphasizes that this is not the same as applying Kaplan–Meier once to the pooled sample, because the weighted sum of cohortwise Kaplan–Meier curves is not a linear functional of the pooled lifetimes (Heuser et al., 2017). The principal inferential object is the area between two such weighted curves: $d_j$8 with asymptotic Gaussian theory under the null (Heuser et al., 2017).

Inverse-probability weighting yields another plug-in generalization. For time-to-event attributable fractions, the counterfactual survival under no exposure is estimated by a weighted KM estimator among the unexposed, with weights

$d_j$9

The AF estimator then plugs that weighted survival estimate into the AF formula (Talbot et al., 2022). Closely related causal survival estimators arise in the inverse probability of treatment weighted Kaplan–Meier literature, where the treatment-specific survival estimator is

$t_j$0

with weighted event and risk-set counts $t_j$1 and $t_j$2 built from estimated propensity-score weights (Zhang et al., 2 Nov 2025).

4. Theoretical reformulations and asymptotic structure

The KM plug-in estimator has been analyzed through several complementary theoretical frameworks.

One line of work treats Kaplan–Meier as a smooth functional of the Nelson–Aalen estimator. Under dependent stationary data, the cumulative hazard estimator is

$t_j$3

and the survival estimator is recovered by the product-integral map

$t_j$4

Because the product-integral map is compactly Hadamard differentiable with derivative

$t_j$5

functional central limit theorems for Nelson–Aalen transfer directly to KM via the functional delta method (Anevski, 2016). In this setting, KM is explicitly handled as a plug-in functional of the hazard estimator.

A second line reconstructs KM as an M-estimator. “Reconstruct Kaplan--Meier Estimator as M-estimator and Its Confidence Band” defines a quadratic empirical criterion based on concordance and shows that, under right censoring, an EM algorithm maximizing the incomplete-data M-function converges to

$t_j$6

which is exactly the classical KM estimator (Gu et al., 2020). This yields an alternative route to consistency, asymptotic distribution, and confidence bands (Gu et al., 2020).

A third line studies bootstrap inference on the full support rather than on intervals bounded away from the endpoint. Under the censoring condition

$t_j$7

Efron’s bootstrap satisfies

$t_j$8

where the limit matches the Gaussian limit of the original KM estimator (Dobler, 2016). This supports time-simultaneous confidence bands for the entire survival curve and, via the functional delta method, for functionals such as mean residual life, Lorenz curves, and the Gini index (Dobler, 2016).

A later reformulation expresses KM as a sum over units rather than only as a product over time points. For ordered times $t_j$9 and indicators $r_j$0, the cumulative failure estimator

$r_j$1

is shown to admit the representation

$r_j$2

where failed units contribute empirical step functions and censored units contribute KM-imputed conditional distributions based on the remaining risk set (Tichy, 6 Nov 2025). This suggests a self-consistent plug-in interpretation at the level of individual units.

5. Specialized extensions and domain-specific variants

Several recent extensions preserve the product-limit logic while altering the observed event process or the estimand.

In contaminated right-censored samples, “Expert Kaplan--Meier estimation” replaces the observed failure indicator by expert information. The crude expert estimator is the product-limit estimator with $r_j$3 replaced by $r_j$4, intended to estimate the unobservable true closure indicator. Under contaminated right-censoring being entirely random and suitable conditions on expert quality, the estimator is strongly uniformly consistent on compact intervals (2206.13120). The sophisticated expert estimator instead uses an expert-provided distribution $r_j$5 on the latent event time and yields an IPCW-based plug-in estimator (2206.13120).

The conditional version introduces covariates through kernel smoothing. The conditional expert Kaplan–Meier estimator first forms a conditional expert Nelson–Aalen estimator

$r_j$6

and then applies the product-integral map

$r_j$7

The paper establishes functional consistency and weak convergence under asymptotically unbiased expert judgments, and derives an explicit deterministic asymptotic bias under systematic expert deviation (Bladt et al., 16 Dec 2025).

Privacy-preserving collaboration supplies another extension. “Private and Collaborative Kaplan-Meier Estimators” introduces differentially private methods that perturb either a sampled KM curve or a discretized event-probability vector, then reconstruct surrogate datasets or directly aggregate private curves across sites (Rahimian et al., 2023). A separate distributed approach updates ordinary and inverse propensity weighted KM curves sequentially across sites via the KM influence function, with the goal of obtaining multi-center survival analysis without sharing individual-level data (Risk et al., 18 Jul 2025).

These developments preserve the role of KM as the nonparametric backbone, but the object being plugged in may be a private count table, a kernel-smoothed conditional hazard, an expert-corrected event process, or an influence-function update.

6. Limitations, disputed interpretations, and unsuitable analogies

The term “Kaplan–Meier plug-in estimator” has also appeared in methodological disputes about when the substitution principle is appropriate.

One dispute concerns average hazard estimation when the truncation time $r_j$8 does not coincide with an observed event time. The KM plug-in estimator

$r_j$9

was criticized in a recent article, but the subsequent commentary argues that the criticism rests on undefined hazard quantities at non-event times and on an incorrect expectation that the average hazard should be flat between event times (Uno et al., 21 Jul 2025). The commentary emphasizes that while $t_j$0 is flat between failures, $t_j$1 continues to increase, so $t_j$2 typically declines between event times; this is presented as a consequence of the definition rather than a defect (Uno et al., 21 Jul 2025).

A stronger objection appears in stochastic highway capacity estimation. There the Kaplan–Meier estimator, or product-limit method, had been used by mapping time to traffic flow intensity and failure to traffic breakdown. “Stochastic highway capacity: Unsuitable Kaplan-Meier estimator, revised maximum likelihood estimator, and impact of speed harmonisation” argues that this analogy is fundamentally unsuitable because traffic intensity is not a continuously accumulating age variable, capacity is latent rather than directly observed, and the KM risk set is misdefined in that setting (Mikolášek, 1 Jul 2025). The paper reports that the product-limit method systematically underestimates breakdown probability at low intensities and overestimates breakdown probability at high intensities, and recommends a corrected parametric likelihood instead (Mikolášek, 1 Jul 2025).

These episodes delineate an important boundary. Within standard right-censored survival analysis and closely aligned extensions, the plug-in principle is technically natural: substitute $t_j$3, or modified KM ingredients, into a target functional and derive inference from the resulting estimator. Outside that setting, especially when the data-generating process does not have the survival-analysis risk-set structure, the same substitution can become conceptually invalid (Mikolášek, 1 Jul 2025).

In the broader literature, then, “Kaplan–Meier plug-in estimator” is not a single estimator but a methodological pattern. Its most classical form is the substitution of the KM survival curve into functionals of interest; its modern variants replace the KM inputs by private, weighted, distributed, or expert-corrected quantities while preserving the product-limit logic; and its validity depends on whether the underlying risk-set and censoring structure matches the survival framework in which Kaplan–Meier is defined (Gondara et al., 2019, Heuser et al., 2017, Dobler, 2016, Talbot et al., 2022).