Restricted Mean Survival with a Horizon Between Event Times

Why the Horizon’s Position Matters

Restricted mean survival time up to a horizon \(\tau\) is the area under the survival curve from time 0 to \(\tau\). When \(\tau\) is an event time, the endpoint coincides with a Kaplan–Meier step. When it falls between two event times, the final area is only part of a step. The key is to use the survival level that applies over the interval containing \(\tau\), not the level after the next event.

This calculation is useful when a requested horizon does not line up with observed failures. It also tests whether you can distinguish a failure, which changes the Kaplan–Meier estimate, from a censoring time, which does not create a drop in the curve.

Definition: Let the distinct observed event times be \(0<t_1<t_2<\cdots\), and let \(\hat S(t)\) be the Kaplan–Meier survival estimate. If \(t_k<\tau<t_{k+1}\), then \(\hat S(t)\) is constant at \(\hat S(t_k)\) throughout the interval from \(t_k\) up to, but not including, \(t_{k+1}\). The restricted mean survival estimate is \(\widehat{\operatorname{RMST}}(\tau)=\int_0^\tau \hat S(t)\,dt\), so the final partial area is \((\tau-t_k)\hat S(t_k)\).

Formula for a Horizon Between Events

At an event time \(t_j\), write \(n_j\) for the number at risk just before that time and \(d_j\) for the number of events at that time. The Kaplan–Meier estimate after that event time is the product of the conditional survival factors up to and including the event:

\hat S(t_j)=\prod_{r=1}^{j}\left(1-\frac{d_r}{n_r}\right).

This product is the Kaplan–Meier estimator for right-censored time-to-event data. Its standard interpretation relies on independent observations and censoring that is non-informative for the event time, conditional on the information used in the analysis. Censoring changes later risk sets, but it does not itself multiply the survival estimate by a failure factor.

Set \(t_0=0\) and \(\hat S(t_0)=1\). If the horizon lies between event times \(t_k\) and \(t_{k+1}\), sum the areas of the completed steps and then add the part of the current step up to \(\tau\):

\widehat{\operatorname{RMST}}(\tau) = \sum_{j=0}^{k-1}(t_{j+1}-t_j)\hat S(t_j) + (\tau-t_k)\hat S(t_k).

Here the completed intervals are \([t_j,t_{j+1})\) for \(j=0,\ldots,k-1\). The last term is the partial interval \([t_k,\tau]\). The value at a single endpoint does not change an area, so it is immaterial whether an interval is written with an open or closed endpoint.

Why the Final Partial-Area Formula Holds

The Kaplan–Meier curve is a step function: between event times it has no slope and remains at the level set by the most recent event. Therefore its integral over a constant interval is simply the interval’s width multiplied by its height. Splitting the integral at the event times gives the full-interval areas plus the final partial area:

\begin{aligned} \int_0^\tau \hat S(t)\,dt &=\sum_{j=0}^{k-1}\int_{t_j}^{t_{j+1}}\hat S(t_j)\,dt +\int_{t_k}^{\tau}\hat S(t_k)\,dt\\ &=\sum_{j=0}^{k-1}(t_{j+1}-t_j)\hat S(t_j) +(\tau-t_k)\hat S(t_k). \end{aligned}

This also explains why linear interpolation is not appropriate. The Kaplan–Meier estimate does not gradually move from one event-time survival level to the next; it stays flat, then drops at the next event time.

Worked Calculations

Worked Example: A Horizon After an Intermediate Censoring Time

A hypothetical follow-up study records event times at 2, 5, and 9 years. The risk sets and event counts are: at 2 years, \(n_1=10\) and \(d_1=2\); at 5 years, \(n_2=8\) and \(d_2=2\); at 9 years, \(n_3=5\) and \(d_3=1\). One participant is censored at year 6. Calculate the restricted mean survival estimate through \(\tau=7\) years.

Model and method. Use the nonparametric Kaplan–Meier step estimate for right-censored time-to-event data, assuming independent, non-informative censoring. The time unit is years. No parametric event-time distribution is assumed; interest discounting is not part of this survival-area calculation.

Step 1: Calculate the event-time survival levels. At year 2, the conditional survival factor is \(1-2/10=8/10=0.8\), so \(\hat S(2)=0.8\). At year 5, the factor is \(1-2/8=6/8=0.75\), giving \(\hat S(5)=0.8(0.75)=0.6\). The year-6 censoring mark does not cause a drop. It reduces the risk set from 6 to 5 before the event at year 9, but year 9 is beyond the requested horizon.

Step 2: Identify the final partial interval. Since \(5<7<9\), the curve is at level \(\hat S(5)=0.6\) throughout the interval containing the horizon. The partial width is \(7-5=2\) years, giving area \(2(0.6)=1.2\) years.

Step 3: Add the completed areas. From year 0 to year 2, the area is \(2(1)=2\) years. From year 2 to year 5, it is \(3(0.8)=2.4\) years. Thus

\widehat{\operatorname{RMST}}(7) =2+2.4+1.2 =5.6\text{ years}.

Check and mistake avoided. The widths \(2+3+2=7\) cover the full horizon, and all three survival heights are the levels applying over those intervals. Using the post-year-9 level for the final interval would wrongly treat a later failure as if it had already occurred.

Worked Example: A Noninteger Horizon and a Censoring Mark

In a hypothetical time-to-event study measured in years, the data at event times are \(n=12,d=3\) at year 1; \(n=9,d=2\) at year 4; and \(n=6,d=2\) at year 8. One individual is censored at year 5. Find the Kaplan–Meier restricted mean estimate through \(\tau=6.5\) years.

Model and method. Apply the Kaplan–Meier product-limit estimator under independent, non-informative censoring. No parametric distribution is specified or required, and no interest-rate adjustment is made.

Step 1: Calculate survival after year 1. The factor at year 1 is \(1-3/12=9/12=3/4\), so \(\hat S(1)=3/4\).

Step 2: Calculate survival after year 4. The factor at year 4 is \(1-2/9=7/9\). Therefore \(\hat S(4)=(3/4)(7/9)=21/36=7/12\). The censoring at year 5 does not lower this estimate. It changes the number at risk for later events, but it does not alter the survival level between years 4 and 8.

Step 3: Add the full and partial areas. The first interval contributes \(1(1)=1\) year. The interval from year 1 to year 4 contributes \(3(3/4)=9/4\) years. Since \(4<6.5<8\), the final partial interval has width \(6.5-4=2.5=5/2\) years and height \(7/12\), so its area is \((5/2)(7/12)=35/24\) years. Hence

\widehat{\operatorname{RMST}}(6.5) =1+\frac{9}{4}+\frac{35}{24} =\frac{24+54+35}{24} =\frac{113}{24} =4.7083\text{ years}.

Check and mistake avoided. The interval widths \(1+3+2.5\) total \(6.5\) years. The answer is rounded to four decimal places only at the final step. Using the year-8 event to lower the survival height over years 4 to 6.5 would incorrectly include information from after the horizon.

Worked Example: Reconstructing Risk Sets Before the Horizon

A hypothetical study begins with 15 individuals. At year 2, three events occur. At year 4, two of those still under observation are censored. At year 6, four events occur. At year 7, one remaining individual is censored. At year 10, two events occur. Calculate the restricted mean survival estimate through \(\tau=8.5\) years.

Model and method. Use the Kaplan–Meier estimator for right-censored event times, with independent, non-informative censoring. The event-time distribution is left unspecified; this is not a parametric model. The time unit is years, and interest discounting is not applied.

Step 1: Determine the risk sets and survival levels. At year 2, all 15 are at risk, so the factor is \(1-3/15=12/15=0.8\), giving \(\hat S(2)=0.8\). Three events leave 12 individuals. The two censorings at year 4 leave 10 at risk at year 6. The factor at year 6 is \(1-4/10=6/10=0.6\), so \(\hat S(6)=0.8(0.6)=0.48\). After those four events, 6 remain; the censoring at year 7 leaves 5 at risk for the year-10 event. This later risk-set detail confirms the sequence, but the year-10 event occurs after the horizon.

Step 2: Integrate each interval through year 8.5. From year 0 to 2, the area is \(2(1)=2\). From year 2 to 6, it is \(4(0.8)=3.2\). From year 6 to 8.5, the width is \(8.5-6=2.5\) years and the applicable level is 0.48, giving \(2.5(0.48)=1.2\) years. Therefore

\widehat{\operatorname{RMST}}(8.5) =2+3.2+1.2 =6.4\text{ years}.

Check and mistake avoided. The interval lengths \(2+4+2.5=8.5\) match the horizon. The year-7 censoring has no direct effect on the curve’s height from year 6 to 8.5; it affects the later risk set. Treating it as a survival drop would produce an incorrect area.

Common Exam Traps

The most frequent error is to use the survival level at the next event rather than the level that holds before it. In the first worked example, the incorrect calculation would use \(\hat S(9)=0.48\) over years 5 to 7:

2+3(0.8)+2(0.48)=5.36\text{ years}.

That calculation is wrong because the event at year 9 has not happened by year 7. The correct final interval uses \(\hat S(5)=0.6\), giving \(2+3(0.8)+2(0.6)=5.6\) years.

Other checks help prevent related errors:

Use the partial width. The final width is \(\tau-t_k\), not the distance from \(t_k\) to the next event time.
Do not interpolate survival. A Kaplan–Meier curve stays at the previous event-time level between events; it does not slope toward the next level.
Do not count censoring as a failure. A censoring mark does not create a downward step, though it can change a later event’s risk set.
Keep units consistent. If event times are in months, each interval width and the reported restricted mean are in months.

The final answer should also be plausible: it is an area over a time horizon, so it has units of time and cannot exceed \(\tau\), because \(0\leq \hat S(t)\leq1\). This provides a quick arithmetic check, not a replacement for checking the step levels.

Key takeaway: When \(t_k<\tau<t_{k+1}\), the final Kaplan–Meier area is \((\tau-t_k)\hat S(t_k)\). Add it to the completed step areas; do not use the next event’s survival level, interpolate, or treat censoring as a failure drop.

Check Your Understanding

For each question, identify the step containing the horizon before calculating its area.

A hypothetical study has an event at year 2 with \(n=8,d=2\), followed by an event at year 6 with \(n=6,d=3\). Calculate the restricted mean survival estimate through year 4. (Answer: 3 years)
At year 1, a risk set of 10 has one event; at year 5, a risk set of 9 has three events. Find the restricted mean estimate through year 3.5. (Answer: 3.25 years)
At year 2, \(n=20,d=4\); at year 5, \(n=16,d=4\); at year 9, \(n=10,d=2\). Two individuals are censored at year 7. Calculate the estimate through year 7.5. (Answer: 5.9 years)
A curve has event times at years 2, 5, and 8. Its survival levels are 0.9 after year 2 and 0.6 after year 5. If the restricted mean through an unknown \(\tau\), where \(5<\tau<8\), is 6.2 years, find \(\tau\). (Answer: 7.5 years)
Curve A has survival 0.8 after year 2 and 0.6 after year 5, with no further event before year 8. Curve B has survival 0.9 after year 3 and 0.65 after year 6, with no further event before year 8. Calculate both restricted mean estimates through year 8 and their difference, B minus A. (Answer: A = 6.2 years, B = 7.0 years, difference = 0.8 years)

What You'll Learn