Greenwood-Based Variance for RMST

Why RMST Needs a Variance Calculation

The restricted mean survival time through a fixed horizon \(\tau\) is the area under the survival curve from time zero to \(\tau\). When that curve is estimated by Kaplan–Meier, its area is also estimated, and the estimate has sampling variability. The variance is not obtained by treating the separate areas between event times as independent: each later Kaplan–Meier estimate reuses the earlier survival factors. Greenwood’s formula, together with the covariance it implies across times, accounts for this shared uncertainty.

This tutorial develops the single-group, Greenwood-based plug-in variance for RMST. It builds on the definition of RMST and its Kaplan–Meier area estimate in the earlier material on estimating RMST from Kaplan–Meier curves. The focus here is how uncertainty in that estimated curve propagates to uncertainty in its area.

Definition: Let \(T\) be an event time, and let \(C\) be a censoring time. For a prespecified horizon \(\tau\), the restricted mean survival time is \(\mu(\tau)=\int_0^\tau S(t)\,dt\), where \(S(t)=P(T>t)\). Given distinct event times \(t_1<\cdots<t_K<\tau\), let \(Y_j\) be the number at risk just before \(t_j\), and \(d_j\) the number of events at \(t_j\). The Kaplan–Meier estimate is \(\hat S(t)=\prod_{t_j\le t}(1-d_j/Y_j)\). The RMST estimate is \(\hat\mu(\tau)=\int_0^\tau\hat S(t)\,dt\).

Assumptions and the Greenwood Covariance

The usual Greenwood-based variance estimate relies on independent subjects and non-informative censoring: an individual’s censoring process must not reveal that individual’s unobserved event time. Administrative censoring at a fixed study end is a familiar example under which this assumption may be reasonable. The event-time and censoring-time data must be recorded consistently, and the risk set \(Y_j\) must include the individuals who are event-free and uncensored immediately before \(t_j\). With tied events, \(d_j\) counts all events at that time.

For an event time \(t_j\), define the Greenwood increment \(g_j\) and cumulative quantity \(G_j\) as

g_j=\frac{d_j}{Y_j(Y_j-d_j)}, \qquad G_j=\sum_{r=1}^{j}g_r.

The standard Greenwood estimate of the variance of \(\hat S(t)\) is \(\hat S(t)^2\) multiplied by the sum of \(g_j\) over event times no later than \(t\). For two times \(t\) and \(u\), the corresponding plug-in covariance estimate is

\widehat{\operatorname{Cov}}\{\hat S(t),\hat S(u)\} = \hat S(t)\hat S(u) \sum_{t_j\le \min(t,u)}g_j.

The sum stops at the earlier time because the Kaplan–Meier estimates at \(t\) and \(u\) share the survival factors through \(\min(t,u)\); later factors affect only the estimate at the later time. This covariance is nonnegative. In particular, for \(t\le u\), it is \(\hat S(t)\hat S(u)\) times the Greenwood sum through \(t\).

From Survival Covariance to RMST Variance

Because \(\hat S(t)\) is constant between event times, the estimated area can be written as a finite sum. Put \(t_0=0\), \(t_{K+1}=\tau\), and define interval lengths \(a_0=t_1\) and \(a_j=t_{j+1}-t_j\) for \(1\le j\le K\). Let \(\hat S_0=1\), and let \(\hat S_j=\hat S(t_j)\). Then

\hat\mu(\tau)=a_0+\sum_{j=1}^{K}a_j\hat S_j.

This expression assumes the event times included in the sum are strictly before \(\tau\). An event exactly at \(\tau\) does not change the area over \([0,\tau]\), so it adds no post-event interval. The constant \(a_0\) has no sampling variance. Applying the covariance formula to the remaining terms gives the Greenwood-based plug-in variance estimator:

\widehat{\operatorname{Var}}\{\hat\mu(\tau)\} = \sum_{j=1}^{K}\sum_{\ell=1}^{K} a_j a_\ell\hat S_j\hat S_\ell G_{\min(j,\ell)}.

A useful equivalent form groups the contributions by Greenwood increment:

\widehat{\operatorname{Var}}\{\hat\mu(\tau)\} = \sum_{r=1}^{K}g_r \left(\sum_{j=r}^{K}a_j\hat S_j\right)^2.

Justification: In the double sum, the covariance for \(\hat S_j\) and \(\hat S_\ell\) contains \(G_{\min(j,\ell)}=\sum_{r=1}^{\min(j,\ell)}g_r\). Substitute this sum and interchange the finite summations. For a fixed \(r\), exactly the pairs with \(j\ge r\) and \(\ell\ge r\) include \(g_r\). Their combined contribution is \(g_r(\sum_{j=r}^{K}a_j\hat S_j)(\sum_{\ell=r}^{K}a_\ell\hat S_\ell)\), which is the squared term above.

This derivation shows why the interval contributions cannot generally be treated as independent. An increment \(g_r\) contributes to the uncertainty of every later survival estimate, and therefore to every later area segment. The squared tail sum in the equivalent formula collects those effects.

Worked Example: One Event Time and One Greenwood Increment

Suppose a hypothetical group has 10 independent participants, all under observation through time 2. At time 2, two participants experience the event, and the remaining eight are followed without further events or censoring through \(\tau=5\). Assume censoring, if present, is independent of event time. Calculate the Kaplan–Meier RMST estimate and its Greenwood-based variance.

At time 2, \(Y_1=10\), \(d_1=2\), and \(\hat S_1=1-2/10=0.8\). The estimated curve is 1 until time 2 and 0.8 from time 2 through time 5. Thus

\hat\mu(5)=2(1)+(5-2)(0.8)=2+2.4=4.4\text{ time units}.

The Greenwood increment is \(g_1=2/\{10(10-2)\}=2/80=0.025\). There is one post-event interval of length 3, so the variance is

\widehat{\operatorname{Var}}\{\hat\mu(5)\} =0.025\{3(0.8)\}^2 =0.025(2.4)^2 =0.144\text{ time units}^2.

The estimated standard error is \(\sqrt{0.144}=0.3795\) time units. This is a plug-in estimate of sampling uncertainty, not a guarantee that a normal approximation will be accurate in a sample of only 10.

A Complete Calculation from Event-Time Data

Worked Example: RMST Variance with Events and Censoring

Consider six hypothetical participants followed to \(\tau=5\) time units. Their observed times and outcomes are: an event at 1, censoring at 2, an event at 3, an event at 4, censoring at 5, and an event at 6. Suppose participants are independent and their censoring times are independent of their event times. The event at 6 is after the horizon and does not change the estimated area through 5.

At time 1, all six are at risk, so \(Y_1=6\), \(d_1=1\), and \(\hat S_1=5/6\). After the censoring at 2, four remain at risk just before time 3. Thus \(Y_2=4\), \(d_2=1\), and \(\hat S_2=(5/6)(3/4)=5/8\). At time 4, \(Y_3=3\), \(d_3=1\), and \(\hat S_3=(5/8)(2/3)=5/12\). The interval lengths after these event times are 2, 1, and 1; the initial interval has length 1. Therefore

\hat\mu(5) =1(1)+2(5/6)+1(5/8)+1(5/12) =\frac{89}{24} =3.708\text{ time units}.

The Greenwood increments at the three event times are

g_1=\frac{1}{6(5)}=\frac{1}{30},\qquad g_2=\frac{1}{4(3)}=\frac{1}{12},\qquad g_3=\frac{1}{3(2)}=\frac{1}{6}.

Use the tail-sum variance form. For the first increment, the relevant tail area is \(2(5/6)+1(5/8)+1(5/12)=65/24\). For the second it is \(5/8+5/12=25/24\); for the third it is \(5/12\). Consequently,

\begin{aligned} \widehat{\operatorname{Var}}\{\hat\mu(5)\} &=\frac{1}{30}\left(\frac{65}{24}\right)^2 +\frac{1}{12}\left(\frac{25}{24}\right)^2 +\frac{1}{6}\left(\frac{5}{12}\right)^2\\ &=0.2445+0.09042+0.02894\\ &=0.3639\text{ time units}^2. \end{aligned}

The estimated standard error is \(\sqrt{0.3639}=0.6032\) time units. The censoring at time 2 affects the risk set at time 3, but it is not itself an event and does not create a Greenwood increment. The censoring at \(\tau=5\) does not change the curve over the integration interval.

Worked Example: Why Ignoring Covariances Understates Uncertainty

Use the three post-event intervals in the preceding calculation to see the role of covariance. The separate interval contributions are \(2\hat S_1\), \(\hat S_2\), and \(\hat S_3\). If one incorrectly treated these survival estimates as independent, one would add only their marginal variances:

\begin{aligned} V_{\text{independent}} &=2^2\hat S_1^2G_1+\hat S_2^2G_2+\hat S_3^2G_3\\ &=4(5/6)^2(1/30)+(5/8)^2(1/30+1/12)\\ &\quad +(5/12)^2(1/30+1/12+1/6)\\ &=0.1874\text{ time units}^2. \end{aligned}

This is not the Greenwood-based RMST variance. The interval estimates share earlier event-time increments. For example, \(2\hat S_1\) and \(\hat S_2\) both depend on the first survival factor, giving covariance \(2\hat S_1\hat S_2G_1>0\). Including all such shared contributions gives \(0.3639\) time units squared, as calculated above. The discrepancy illustrates that the area segments are not independent observations.

Diagnostics, Alternatives, and Limitations

Before interpreting the standard error, check that the event-time risk sets and horizon support the calculation. In particular:

Verify that each \(Y_j\) counts individuals at risk just before the event time and that \(d_j\le Y_j\). A coding error in censoring or tied-event handling changes both the Kaplan–Meier estimate and its variance.
Check that the study provides meaningful follow-up through \(\tau\). If few people remain under observation near the horizon, the estimated area and its variance can be sensitive to the last observed events and censorings.
Inspect the Greenwood increments. Large increments indicate that survival estimation is becoming unstable, often because risk sets are small.
Consider the censoring assumption in the study context. If censoring is related to an individual’s event risk after accounting for the design or measured information, ordinary Kaplan–Meier and Greenwood calculations may not provide valid inference.
Check whether every event time used has \(Y_j-d_j>0\). If all individuals at risk fail at one time, the Greenwood increment is undefined; standard plug-in inference near that boundary requires special care and should not be repaired by silently inserting a finite value.

A bootstrap that resamples independent participants and recomputes the Kaplan–Meier area can provide an alternative estimate of uncertainty, subject to appropriate handling of censoring and adequate sample size. A parametric survival model can also produce model-based uncertainty, but it depends on the chosen event-time model. The Greenwood plug-in method is attractive because it follows directly from the Kaplan–Meier estimator and does not require choosing a parametric survival distribution. Its limitations are the same practical concerns revealed by the diagnostics: sparse risk sets, heavy censoring near \(\tau\), small samples, and violations of independent censoring can make the approximation unreliable.

A common mistake is to calculate a Greenwood variance for survival at each event time and then add only the variances of the interval areas. That calculation drops the covariances between survival estimates. The correction is to use the full double sum or, more efficiently, the equivalent Greenwood-increment tail-sum formula. Both retain the uncertainty shared across later parts of the area.

Key takeaway: Greenwood-based RMST variance is the integrated covariance of the Kaplan–Meier curve. Each event-time Greenwood increment affects all subsequent area segments, so a correct calculation must preserve that shared uncertainty.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

At an event time, \(Y_j=20\) and \(d_j=2\). Calculate \(g_j\).
Why does an event exactly at \(\tau\) contribute no post-event interval to the RMST variance sum?
For \(t\le u\), which Greenwood increments enter \(\widehat{\operatorname{Cov}}\{\hat S(t),\hat S(u)\}\), and why?
What is the main error in treating RMST interval contributions as independent?

Solutions. (1) \(g_j=2/\{20(20-2)\}=2/360=0.005556\). (2) The area ends at \(\tau\), so there is no interval after that event within the integration range. (3) The increments at event times no later than \(t\) enter, because both survival estimates share those factors; increments after \(t\) affect only the estimate at \(u\). (4) That approach omits the positive covariance created by shared earlier Kaplan–Meier factors, and can therefore understate the variance.