Censoring and Precision of RMST Estimates

Why Censoring Matters Near the RMST Horizon

Restricted mean survival time (RMST) through a horizon \(\tau\) is the area under the survival function from time zero to \(\tau\). It summarizes expected event-free time accumulated during that interval. Its estimate depends not only on observed event times but also on whether follow-up provides information about survival throughout the interval.

Censoring does not automatically invalidate a Kaplan–Meier estimate. With independent censoring and adequate follow-up, the product-limit estimator can estimate the survival function despite some participants leaving observation before an event. But if almost everyone has been censored before \(\tau\), the estimator can be highly imprecise near the horizon. If nobody remains under observation, the data do not identify survival beyond that point. A flat Kaplan–Meier tail can conceal this lack of support.

The earlier material on estimating RMST from Kaplan–Meier curves gives the step-function area calculation; the earlier material on Greenwood-based variance gives the variance framework. Here the focus is on how censoring and risk-set depletion determine where those calculations are informative.

Estimability Requires Follow-Up Support

Definition: The RMST through \(\tau\) is \(\mu(\tau)=\int_0^\tau S(t)\,dt\), where \(S(t)\) is the probability of remaining event-free beyond time \(t\). Its Kaplan–Meier estimate is \(\hat{\mu}(\tau)=\int_0^\tau \hat{S}(t)\,dt\), where \(\hat{S}(t)\) is the Kaplan–Meier estimate of survival. The estimate is supported through \(\tau\) only when follow-up provides information about survival over the interval being integrated.

The standard Kaplan–Meier conditions include independent participants and independent censoring of event times: within the population or analysis group, censoring must not selectively depend on a participant’s unobserved event time. In a covariate-adjusted setting, this condition may be formulated as censoring independent of event time conditional on the modeled covariates. The usual product-limit estimator also relies on correctly recorded event and censoring times and on sufficient follow-up at the times of interest.

A useful way to express the follow-up requirement is that the probability of remaining uncensored must be positive over the interval of interest. In a sample, this corresponds to having participants at risk as time advances. A positive count at risk is necessary for estimating subsequent event experience, but a very small count may still provide weak information. Thus, “someone remains at risk” is not a guarantee of good precision.

At an event time \(t_j\), let \(Y_j\) denote the number at risk immediately before that time and \(d_j\) the number of events. The Kaplan–Meier estimate is multiplied by \(1-d_j/Y_j\). Censoring reduces later risk sets \(Y_j\); with fewer people contributing event experience, each later event can have a larger effect on the estimate and its variance.

What the Greenwood Variance Says About Precision

For event times \(t_j\) up to \(\tau\), a commonly used Greenwood-based variance estimate for the Kaplan–Meier RMST is

\widehat{\operatorname{Var}}\{\hat{\mu}(\tau)\} = \sum_{t_j\leq \tau} \frac{d_j}{Y_j(Y_j-d_j)} \left\{\int_{t_j}^{\tau}\hat{S}(t)\,dt\right\}^{2}.

This expression is a large-sample approximation, used when the risk sets and event counts make the approximation reasonable. The integral inside each term is the estimated survival area remaining after event time \(t_j\). An event early in follow-up can affect a longer stretch of the area than an event close to \(\tau\). The factor \(d_j/[Y_j(Y_j-d_j)]\) reflects the event information at that time and grows as the risk set becomes small. If \(Y_j=d_j\), the expression is undefined; this is a warning of a terminal event pattern, not a license to treat the estimated tail as well supported.

The standard error is the square root of this estimated variance, and it has units of time, as does RMST. Larger standard errors indicate less precision. Censoring often increases uncertainty by shrinking later risk sets, although the amount depends on when censoring occurs, the event pattern, and the horizon. It does not follow that two studies with the same censoring percentage must have the same RMST precision.

Worked Example: A Flat Tail After the Last Participant Leaves Follow-Up

Consider a hypothetical group of eight participants. Two events occur at time 1. Two participants are censored at time 1.5. One event occurs at time 2, when four participants are at risk. One participant is censored at time 2.5. One event occurs at time 3, when two participants are at risk, and the final participant is censored at time 3.5. Use time units throughout.

At time 1, \(Y_1=8\) and \(d_1=2\), so \(\hat{S}(1)=1(1-2/8)=0.75\). At time 2, \(Y_2=4\) and \(d_2=1\), giving \(\hat{S}(2)=0.75(1-1/4)=0.5625\). At time 3, \(Y_3=2\) and \(d_3=1\), giving \(\hat{S}(3)=0.5625(1-1/2)=0.28125\).

The curve is supported through time 3.5: a participant remains under observation up to that censoring time. Integrating the step heights through 3.5 gives

\hat{\mu}(3.5) =1(1)+0.75(1)+0.5625(1)+0.28125(0.5) =2.453125\text{ time units}.

At time 3.5, the last participant leaves follow-up. If we mechanically integrate the plotted Kaplan–Meier curve to time 4, its flat tail adds \(0.28125(0.5)=0.140625\), giving \(2.59375\) time units. But no participant is observed beyond 3.5. That added tail area is not supported by observed follow-up; the curve’s flatness is a consequence of no further recorded events, not evidence that the true survival probability stayed constant.

For the supported horizon 3.5, the Greenwood-based variance terms use \(q_j=d_j/[Y_j(Y_j-d_j)]\), which are \(1/24\), \(1/12\), and \(1/2\) at times 1, 2, and 3. The remaining areas from those event times to 3.5 are \(1.453125\), \(0.703125\), and \(0.140625\). Therefore,

\widehat{\operatorname{Var}}\{\hat{\mu}(3.5)\} = \frac{1}{24}(1.453125)^2 +\frac{1}{12}(0.703125)^2 +\frac{1}{2}(0.140625)^2 \approx 0.1391\text{ time units}^2, \qquad SE\{\hat{\mu}(3.5)\}\approx 0.3729\text{ time units}.

This is an approximate standard error from a very small hypothetical sample, so it should not be treated as a precise measure of coverage. It illustrates that the late event has a large Greenwood factor because only two participants were at risk. Extending the calculation to time 4 would not fix the absence of follow-up support.

How Risk-Set Depletion Changes Precision

The same event can contribute very different amounts to estimated variance depending on the risk set. For an event with \(d=1\), the Greenwood factor is \(1/[Y(Y-1)]\). If a hypothetical event occurs with \(Y=10\), the factor is \(1/90\approx0.01111\). If the same event occurs with \(Y=4\), it is \(1/12\approx0.08333\), 7.5 times as large.

Suppose, for illustration, that the estimated survival area remaining after either event is 2 time units. The corresponding variance contributions are \(0.01111(2)^2\approx0.04444\) and \(0.08333(2)^2\approx0.3333\) time units squared. This comparison isolates the risk-set effect; it is not a complete variance calculation for a study. In a full analysis, contributions from all event times are added, and their remaining areas depend on the estimated survival curve.

This is why risk tables matter. A survival curve may look smooth while the number at risk near \(\tau\) is tiny. The estimated tail can then be sensitive to one or two events, with substantial uncertainty that is easy to miss if attention is limited to the plotted line or the RMST point estimate.

Worked Example: Comparing RMST Only Over a Shared Supported Horizon

Suppose a hypothetical randomized study has two groups with Kaplan–Meier estimates described as follows. In Group A, the estimated survival probabilities are 1 from time 0 to 1, 0.8 from 1 to 2, and 0.6 from 2 to 4, with participants remaining under observation through time 4. In Group B, they are 1 from 0 to 1, 0.7 from 1 to 2, 0.5 from 2 to 3, and 0.4 from 3 to 3.2. The last Group B participant is censored at 3.2, leaving no follow-up after that time.

Both groups have support through time 3. Their RMST estimates at this shared horizon are

\hat{\mu}_A(3)=1(1)+0.8(1)+0.6(1)=2.4\text{ time units}, \qquad \hat{\mu}_B(3)=1(1)+0.7(1)+0.5(1)=2.2\text{ time units}.

The estimated contrast, Group A minus Group B, is \(2.4-2.2=0.2\) time units through time 3. This arithmetic does not by itself establish statistical significance; uncertainty must also be assessed. It does show that the contrast is computed over a horizon with follow-up support in both groups.

At time 4, Group A’s curve remains supported, but Group B’s does not. A mechanical flat-tail calculation would add \(0.4(4-3.2)=0.32\) to Group B’s area through time 3.2, producing \(\hat{\mu}_B(4)=2.2+0.4(0.2)+0.4(0.8)=2.60\) time units. Group A’s area through time 4 is \(1+0.8+0.6(2)=3.0\) time units. The resulting mechanical contrast, \(3.0-2.60=0.40\) time units, includes an unsupported Group B tail from 3.2 to 4. It should not be presented as a data-supported nonparametric RMST contrast through time 4.

Diagnostics, Alternatives, and Limitations

Before interpreting an RMST estimate, check the following features of the data and analysis:

Numbers at risk near the horizon. Inspect the risk table alongside the curve. A very small risk set signals that late estimates may be unstable; a risk set of zero means later survival is not empirically estimated.
Last event and last follow-up times. Distinguish the last observed event from the last time anyone remains under observation. A curve can be flat after the last event while censoring continues; it cannot gain observed support after the last follow-up time.
How the horizon was chosen. Select \(\tau\) using clinical relevance and expected follow-up support. Choosing it after inspecting where the curves look most favorable can make the comparison difficult to interpret and may introduce bias.
Censoring patterns across groups. A common RMST contrast requires support through the chosen horizon in each group. Unequal follow-up can leave one group’s tail less estimable even if the other group has adequate follow-up.
Uncertainty as well as the point estimate. Report an appropriate standard error or confidence interval and interpret it in time units. Greenwood-based approximations can be unreliable with very sparse late risk sets; a numerically computed interval does not compensate for unsupported follow-up.

The main nonparametric remedy for an unsupported tail is to use a shorter horizon that is clinically meaningful and supported in all groups. This changes the estimand: it asks about event-free time only up to the shorter \(\tau\). If the scientific question requires a longer horizon, model-based extrapolation may be considered, but its tail depends on assumptions beyond the observed data and should be identified and examined through sensitivity analyses. Neither option makes the unsupported nonparametric tail observed.

Independent censoring is an assumption, not a consequence of using Kaplan–Meier. If loss to follow-up is related to event risk, the Kaplan–Meier estimate may be biased even where risk sets remain nonzero. Conversely, independent censoring does not guarantee precision: extensive censoring can leave very little information near the horizon. A trustworthy analysis therefore needs both a defensible censoring assumption and adequate follow-up support.

Common mistake: “The Kaplan–Meier curve is flat after the last event, so the survival probability is known to stay at that level through the chosen horizon.” The correction is that a flat curve after follow-up ends is a computational continuation, not new evidence. Check the risk table and last follow-up time; if the interval is unsupported, shorten the horizon or clearly distinguish model-based extrapolation from the nonparametric estimate.

Set the horizon.
Specify the clinical time interval of interest before interpreting the estimated area.

Check support in each group.
Review risk sets and last follow-up times through the full interval, not just the plotted survival values.

Calculate and qualify.
Integrate the Kaplan–Meier steps only over the supported interval, assess precision, and state any assumption or limitation that affects interpretation.

Check Your Understanding

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

In a hypothetical study, the last participant is censored at time 5, and the Kaplan–Meier curve remains flat to time 8. What does the flat segment establish about survival from 5 to 8?
For a single event with \(d=1\), compare the Greenwood factors when \(Y=8\) and \(Y=3\). Which event contributes more to estimated variance, all else equal?
Both groups have follow-up support through time 2. One group has participants at risk through time 4, while the other has none after time 2.5. What is the maximum horizon for a nonparametric contrast supported in both groups based on the information given?
Does independent censoring guarantee a precise RMST estimate near the horizon? Explain.

Solutions. (1) Nothing about actual survival from 5 to 8 is established by the flat segment; it reflects no observed follow-up after 5. (2) The factors are \(1/[8(7)]=1/56\approx0.01786\) and \(1/[3(2)]=1/6\approx0.1667\). The event with \(Y=3\) contributes more, assuming its remaining-area term is the same. (3) Time 2.5 is the latest horizon explicitly known to have support in both groups; the group with no follow-up after time 2.5 has stated follow-up support through that time. (4) No. Independent censoring supports the validity of the estimator under its assumptions, but sparse risk sets can still produce large variance, and zero follow-up leaves later survival unsupported.