Can RMST exceed the chosen horizon?

No, for a valid survival distribution and nonnegative event time.

Is it the same as median survival?

No; it averages capped time rather than locating the point where survival reaches one-half.

Can it be useful when survival curves cross?

Yes, but the summary may depend materially on the chosen horizon.

AI Glossary / Statistics and probability / Intermediate

Restricted mean survival time

Definition

Restricted mean survival time, or RMST, is the expected time spent event-free up to a specified time horizon tau. For a nonnegative event time T, it equals E[min(T, tau)] and also the area under the survival curve from time zero to tau. It is measured in the same time units as the underlying event-time data.

What it means

The restriction places an upper limit on the time counted for every individual. Someone whose event occurs before the horizon contributes time until that event; someone whose event occurs later contributes the full horizon.
Survival analysis is not limited to death. The event can be component failure, service cancellation, or another clearly defined endpoint. Interpretation follows that endpoint: more event-free time can be desirable or undesirable depending on what event is being studied.

How it works

Define the start of follow-up, the event, and the target population. Specify a meaningful horizon before inspecting which choice gives the most attractive group difference. A comparison should ordinarily use the same horizon for both groups.
Let S(t) be the probability that the event has not yet occurred by time t. RMST is the integral of S(t) from zero to tau. The identity with E[min(T, tau)] explains why the result is an average time rather than a probability or hazard ratio.
With complete event times, calculate each person's time capped at tau and average those values. With right censoring before tau, the true capped time is not fully observed for everyone; simply averaging observed follow-up times is generally inappropriate.
A Kaplan - Meier survival estimate can provide a nonparametric RMST estimate by summing rectangular areas under its step function. This requires suitable censoring assumptions and adequate follow-up support through the chosen horizon. Covariate adjustment or more complex sampling may require other methods.
Censoring is not an event. A censored individual leaves the risk set, but censoring alone does not create a downward jump in the Kaplan - Meier curve. Treating every last observation as an event underestimates event-free time.
For two groups, an RMST difference expresses additional average event-free time within the horizon. Its sign depends on the order of subtraction. The difference does not require a constant hazard ratio, although valid estimation and causal interpretation still require their own assumptions.
Estimate uncertainty with an appropriate variance, interval, or resampling method. The amount of follow-up near tau matters: very sparse tail information can make estimates unstable, and extrapolating beyond observed support introduces model dependence.
Distinguish RMST from median survival, unrestricted mean survival, and restricted mean time lost. For a single nonnegative event time, restricted mean time lost equals tau minus RMST. Competing-event and multistate questions require clearly specified probabilities and estimands.

Three practical examples

Reliability analysis: engineers compare average failure-free operating time over a fixed warranty horizon.
Customer analytics: a team summarizes average subscription time retained within the first year.
AI-assisted research: analysts report an interpretable time-based summary alongside survival curves rather than relying only on a hazard ratio.

Think of it this way

Think of giving everyone the same eight-hour observation budget. The clock counts until the event or until eight hours, whichever comes first. RMST averages those capped clock readings, with survival methods accounting for people whose clocks become unobservable early.

Worked example

Consider five independent units followed for failure. One fails at time 2, one is censored at time 3, two fail at time 5, and the final unit is censored at time 8. Choose tau = 8 and assume the censoring mechanism permits Kaplan - Meier estimation.
Initially S(t) = 1. At time 2, one of five at-risk units fails, so the estimate drops to 4/5. The censoring at time 3 reduces the risk set but does not change survival. At time 5, two of the three remaining units fail, so survival becomes (4/5) × (1 minus 2/3) = 4/15.
The estimated RMST is the sum of three rectangle areas: 2 × 1 + 3 × (4/5) + 3 × (4/15) = 2 + 2.4 + 0.8 = 5.2 time units.
Restricted mean time lost is 8 minus 5.2 = 2.8. These are small-sample descriptive estimates under the stated censoring assumption, not evidence of a group effect or a precise population guarantee.

Illustrative case study

Fictional case: Sequoia Devices compares two maintenance programs using failure-free time through a common six-month horizon. The analysts report the survival curves, RMST difference, uncertainty, and numbers still under observation. They avoid extending the comparison into a poorly observed tail and distinguish a randomized assignment effect from a comparison based on customers choosing their own program.

Common misconceptions and mistakes

Do not interpret RMST as the mean event time only among individuals who experience the event. Those remaining event-free also contribute to the target quantity.
Do not treat censoring as failure or average raw observed times without accounting for incomplete follow-up.
A longer horizon naturally permits more counted time, so RMST values with different horizons are not directly comparable as though they measured the same quantity.
The absence of a proportional-hazards requirement does not remove censoring, confounding, or sampling assumptions.

Business use

RMST can make time-to-event comparisons easier to communicate in units such as days or months. For business decisions, choose a horizon connected to the operating question, assess follow-up quality, and relate the estimated time difference to costs and service consequences without overstating causal or long-term implications.

Check your understanding

Why does the curve not drop at time 3 in the example?

Answer: That observation is censored rather than a failure.

Why does the last interval contribute 0.8?

Answer: Its width is three time units and the survival estimate is 4/15.

Related terms

Survival function
Censoring
Hazard ratio
Estimand
Kaplan–Meier estimator
time-to-event analysis
restricted mean time lost
median survival

Learn next

Learn next how censoring assumptions support survival estimation, how common-horizon RMST differences are estimated, and how endpoint definitions change the meaning of event-free time.

Censoring

Sources

Sources: Frank Harrell, Glossary of Statistical Terms, https://hbiostat.org/glossary/ ; R survival package, print.survfit documentation, https://search.r-project.org/CRAN/refmans/survival/html/print.survfit.html . These primary author and developer references define the restricted area and describe horizon choices. The Kaplan–Meier arithmetic and fictional device case are original illustrations.

Previous: Restricted Boltzmann Machine Next: Restricted randomization