Choosing a Restricted-Mean Horizon Under Censoring

Why the Horizon Matters

The restricted mean survival time summarizes survival over a specified time window. Its value depends on the window’s endpoint, so choosing that endpoint is part of the analysis—not a cosmetic detail. When follow-up is incomplete, the horizon must also be supported by the observed data. Otherwise, a calculation may look precise while relying on survival behavior that was never observed.

The Kaplan-Meier (KM) curve estimates survival using observed event times and the people still at risk at those times. Between observed events, the estimated curve is horizontal. That convention lets us calculate an area, but it does not mean we have observed what happens indefinitely after the last person leaves follow-up. In particular, a flat tail after the last follow-up time is not evidence that no later failures would occur.

The calculation uses the KM curve as a step function. To interpret that curve as an estimate of the event-time survival distribution, the usual non-informative censoring condition is required: censoring must not systematically hide different event risks among people who remain under observation and those who are censored. In the simplest formulation, event times and censoring times are independent. If censoring is informative, the ordinary KM estimator may not represent the target survival distribution.

Choosing a Supported Horizon

A defensible horizon is one for which there is follow-up information through the end of the interval. For one group, the largest observed follow-up time is a practical upper bound on a data-supported horizon. For a comparison of groups, a common horizon should be supported in every group; a conservative choice is no later than the shortest maximum follow-up among them. The horizon should also be chosen for the question being asked, preferably before examining which value produces the most favorable result.

Being within the observed follow-up range is necessary, but it does not guarantee a stable estimate. If only a few people remain under observation near \(\tau\), the estimated curve near the endpoint can be sensitive to individual events. Thus horizon selection should consider both the follow-up limit and the number at risk near that limit. A shorter horizon with broader support may be more interpretable than a longer one based on very sparse follow-up.

In practice, state the selected horizon and its units, explain why it is supported, and report the RMST for that horizon. When comparing groups, use the same \(\tau\) for each group; otherwise, the difference combines survival differences with differences in the time window being summarized.

Why the RMST Is an Area

The area formula follows from the definition of a nonnegative event time \(T\). For each individual, the time contributed before the horizon is \(\min(T,\tau)\). At any time \(t\) between 0 and \(\tau\), this quantity includes a small interval of length \(dt\) exactly when the person survives beyond \(t\). Averaging those interval contributions gives the integral of the probability of surviving beyond \(t\).

The first equality can be seen by writing \(\min(T,\tau)\) as the accumulated indicator of survival through time \(t\): it equals \(\int_0^\tau I(T>t)\,dt\), apart from endpoint conventions that do not affect the integral. Taking expectations gives the displayed result. Replacing \(S(t)\) with the estimated KM step function gives the estimated RMST.

Because the KM estimate is piecewise constant, each part of the integral is a rectangle. If the estimated survival is \(s_j\) over an interval of length \(L_j\), that interval contributes \(L_js_j\). For a horizon between event times, the final rectangle ends at \(\tau\), not at the next event time.

Common Horizons for Group Comparisons

A group-specific RMST answers how much survival time, on average, is accumulated within a given window for that group. To compare groups fairly, calculate both areas through the same \(\tau\). If one group has follow-up only through time 4 and another through time 6, a comparison through time 6 is not supported for both groups. One option is a common horizon no later than time 4, subject to adequate follow-up near that endpoint.

The difference \(\widehat{\mu}_A(\tau)-\widehat{\mu}_B(\tau)\) is interpreted in time units. A negative value means the estimated accumulated survival time through \(\tau\) is smaller for group A. It does not, by itself, establish why the groups differ or describe survival after \(\tau\).

Why Extrapolation Is Not Supported

The KM estimator is built from the observed risk sets and event counts. In the usual product-limit form, each observed failure time contributes a conditional survival multiplier based on the number at risk just before that time. Once everyone has left observation, there is no later risk set and no later multiplier to estimate. Thus the calculation supplies no information about the probability of survival after the last observed follow-up.

A software plot may visually continue the final horizontal segment. That display is useful for showing the last estimated level, but it does not turn an unobserved time interval into observed evidence. A parametric survival model could produce predictions beyond follow-up by imposing a distributional form; that is a different method and relies on additional assumptions. The nonparametric KM curve alone does not provide that extrapolation.

Practical Checks Before Reporting

Before reporting an RMST, check the horizon and the interpretation as carefully as the arithmetic. A defensible report states \(\tau\), its units, the estimated area, and how the horizon relates to observed follow-up. For a group comparison, it also states that the same horizon was used in both groups.

• Check the follow-up limit. Do not integrate a KM curve beyond the largest observed follow-up time for the population being summarized.
• Check support near the endpoint. A horizon near the end of follow-up may be based on very few people still under observation, even if it is technically within the observed range.
• Use a common comparison window. Different horizons produce summaries over different periods and can make a group difference hard to interpret.
• Choose for the question, not the preferred result. Selecting \(\tau\) after seeing which horizon makes one group look better can distort the comparison.
• Keep the meaning restricted. The RMST is expected survival accumulated up to \(\tau\); it is not the unrestricted mean unless additional conditions support that interpretation.