Restricted mean survival time, or RMST, is the average time without the event of interest from study entry through a chosen horizon. If the endpoint is death, it is average survival time within that window. If the endpoint is recurrence, it is average recurrence-free time. The estimate is the area under the survival curve from time zero to the horizon, usually written as tau.
Unlike a median, RMST can be estimated even when fewer than half the participants have had the event, and unlike a hazard ratio, it does not require the relative event hazards to remain proportional over time. It gives an effect in days, months, or years.
Turn the survival curve into time#
A Kaplan-Meier curve shows the estimated probability of remaining event-free over time. RMST adds up the height of that curve across a fixed interval. A curve that stays higher contributes more area and therefore more average event-free time.
Imagine a fictional two-year trial. The estimated RMST is 20.4 months in the treatment group and 18.9 months in the comparison group. The difference is 1.5 months through two years. That means random assignment to treatment produced an estimated average of 1.5 additional event-free months per participant during the first 24 months, subject to the confidence interval and trial assumptions.
The statement does not mean every participant personally gained 1.5 months. Some had no event in either group, some had early events, and some were censored. It is a group average over a bounded period.
Researchers can also report an RMST ratio, here 20.4 divided by 18.9. The difference is usually easier to communicate because it preserves time units. Either measure needs its horizon.
Why hazard ratios can be hard to read#
A hazard is an instantaneous event rate among people still at risk. A hazard ratio compares that rate between groups. In a Cox model, the familiar single hazard-ratio estimate is commonly interpreted under a proportional-hazards assumption: the relative hazards are stable enough over follow-up for one ratio to summarize them.
When survival curves cross, a treatment may cause early harm and later benefit. When they separate only after a delay, the early and late hazard ratios differ; one overall ratio then compresses changing effects into a number that may have no simple clinical meaning.
RMST does not need proportional hazards because it compares the accumulated survival experience directly, and it remains defined when curves cross. None of that makes it automatically superior, because a net time difference can also average early loss against later gain, and the only way to see which pattern you are looking at is to read the curves themselves. The summary number will not tell you.
Even when proportional hazards are plausible, RMST can be useful as an absolute measure. “An average of 1.5 additional event-free months through two years” often communicates more directly than “a hazard ratio of 0.78.” The two summaries address related but not identical aspects of the data.
The horizon is part of the endpoint#
RMST through one year and RMST through five years are different outcomes. The longer window includes later events and can change the sign or size of a difference, and tau should therefore come from clinical relevance and expected follow-up, not from whichever choice gives the smallest p-value.
A defensible horizon usually lies within the period where both groups have adequate observation, because extending tau into a tail with few participants at risk makes the estimate depend heavily on a small, selected remainder. Choosing each group's last event as a different limit would also undermine comparability.
Prespecification is strongest. If the horizon is chosen after viewing the curves, the analysis becomes data-driven and uncertainty must reflect that selection. Methodological work has proposed data-adaptive choices, but those methods need a stated procedure and valid inference, not informal trial and error.
Read the number-at-risk table before you take a horizon at face value. A reported five-year RMST is difficult to trust if almost everyone was censored before year five. Administrative censoring at a common study end can be suitable, while differential loss to follow-up needs closer examination.
Censoring assumptions still matter#
Standard Kaplan-Meier RMST estimation assumes censoring is noninformative within the analysis framework. Loosely, after accounting for the model and design, participants censored at a time should not have systematically different future event prospects from comparable participants who remain observed.
Loss to follow-up caused by deteriorating health can violate that assumption. So can treatment discontinuation if outcome follow-up stops and discontinuation is related to prognosis. Randomization does not fix missing outcomes after assignment.
Reports should distinguish administrative censoring from loss to follow-up, provide reasons by group, and describe sensitivity analyses. Competing events also need an endpoint-specific plan. For example, “time to disease recurrence” cannot treat death casually as ordinary noninformative censoring if death prevents recurrence from being observed.
Estimation and uncertainty#
In a randomized trial, unadjusted RMST is commonly calculated from each group's Kaplan-Meier curve. The difference receives a standard error, confidence interval, and test. A 1.5-month estimate with a 95% confidence interval from 0.1 to 2.9 months supports a different degree of certainty than the same estimate with an interval from minus 2.0 to 5.0 months.
Prespecified adjustment for strong baseline prognostic variables can improve precision while preserving the randomized comparison when done appropriately. Several regression and augmentation methods are available. What you want to know is whether the adjustment was planned, whether the covariates were measured before randomization, and whether the model was fitted across the full trial process without post hoc selection. For observational studies, adjustment does not confer the protection of randomization. Confounding, treatment selection, immortal-time errors, and informative censoring can bias an RMST contrast just as they can other effect estimates.
RMST can reveal absolute scale#
A relative effect can sound dramatic when the event is rare. RMST incorporates the survival curve's absolute level. If almost everyone remains event-free through 12 months, even a substantial relative hazard reduction may translate into a small average time difference within that short window.
The reverse can occur when events are common or follow-up is long. The same hazard ratio can correspond to very different RMST differences in populations with different baseline risk. That is one reason absolute and relative measures belong together.
RMST also avoids describing a hazard ratio as a reduction in the probability of ever experiencing an event. A hazard ratio is not a risk ratio. The RMST difference is still bounded by tau and should not be extrapolated beyond it.
What RMST does not solve#
RMST cannot rescue a poorly defined endpoint, biased outcome assessment, informative missingness, or a flawed comparison group. It does not explain biological mechanism. It does not automatically identify a subgroup that benefits, nor does it prove that a statistically detectable time difference is important to patients.
A single average can hide heterogeneity. Some participants may receive large benefit and others harm, yet the mean is modest. Subgroup analyses need prespecification, interaction testing, adequate sample size, and caution about multiplicity.
Competing risks may require a different estimand, such as restricted mean time lost to a particular cause or time spent in health states. Multi-state methods can separate time before and after events, but their added assumptions and complexity should serve a clear question.
A practical reading sequence#
Name the event and time origin. Find tau, and ask why that horizon and not another. Then work through the survival curves, the number at risk, the event counts, and the censoring by group before the RMST values, the difference, the confidence interval, and the adjustment method, and only when you have all of that should you set the result beside the hazard ratio and the other absolute measures.
Then translate the estimate in full for yourself: outcome, population, contrast, units, and horizon. “An average of six additional weeks without recurrence through three years” is meaningful. “RMST improved by 0.12” is not.
References#
- Royston and Parmar, RMST for randomized time-to-event trials
- Uno and colleagues, moving beyond the hazard ratio
- Tian and colleagues, choosing an RMST time window
- Karrison and Kocherginsky, covariate-adjusted RMST
- Royston and colleagues, scaling and interpreting RMST effects
Questions and answers
Is RMST the same as median survival?
No. Median survival is the time by which the estimated survival curve reaches 50%. RMST averages the entire curve up to a chosen horizon and can be estimated before the median is reached.
Does RMST require proportional hazards?
No. It remains defined when relative hazards change over time. Other assumptions, especially about censoring and endpoint definition, still apply.
Can researchers choose tau after seeing the data?
That weakens a conventional primary inference unless a valid adaptive method was prespecified. A clinically meaningful, data-supported horizon should usually be set before unblinded analysis.
Can an RMST difference be projected beyond follow-up?
No. The estimate describes average time through tau. Extrapolation requires additional modeling assumptions and should be labeled clearly.
Should a trial report RMST instead of a hazard ratio?
It depends on the question and design. When proportional hazards are doubtful, RMST can be a strong primary or supportive summary. Reporting curves and complementary absolute and relative measures is often most informative.