Restricted Mean Survival Time as an Estimand

Why Use a Restricted Survival-Time Estimand?

When treatment effects change over time, a single hazard ratio can be difficult to interpret as an overall summary. The hazard ratio compares instantaneous event rates among people still event-free at each time; those at-risk groups change as events occur. Restricted mean survival time (RMST) instead summarizes the expected event-free time accumulated up to a specified horizon. It is measured in time units and does not require proportional hazards.

The horizon is essential: RMST is not simply the unrestricted mean survival time. It is the mean of survival time after capping every event time at a chosen time $\tau$. Thus the estimand answers a bounded question: how much event-free time, on average, is accumulated during the first $\tau$ time units?

Definition: Let $T$ be a nonnegative event time and let $\tau>0$ be a fixed truncation horizon. The restricted mean survival time through $\tau$ is $$ \mu(\tau)=E[\min(T,\tau)]. $$ If $S(t)=P(T>t)$ is the survival function, then $$ \mu(\tau)=\int_0^\tau S(t)\,dt. $$ It has the same units as $T$, such as months or years.

The definition applies to an event-time distribution in a specified population or treatment group. The horizon must be stated whenever the RMST is reported: for example, “RMST through 18 months,” not just “the RMST.” A change in $\tau$ changes the quantity being summarized, even when the underlying population and event-time distribution stay fixed.

Why the Survival-Curve Area Is an Expectation

The area identity follows from a direct representation of the truncated event time. For any realized $T\geq 0$, the indicator $I(T>t)$ equals one from time zero until $T$, and zero afterward. Integrating it only to $\tau$ therefore gives the length of time in the interval $[0,\tau]$ before the event, which is $\min(T,\tau)$.

\min(T,\tau)=\int_0^\tau I(T>t)\,dt.

Taking expectations and interchanging expectation and integration is valid because the integrand is nonnegative and bounded on a finite interval. Since $E[I(T>t)]=P(T>t)=S(t)$, the result is

E[\min(T,\tau)] =E\left[\int_0^\tau I(T>t)\,dt\right] =\int_0^\tau E[I(T>t)]\,dt =\int_0^\tau S(t)\,dt.

This identity is a population definition, not an estimation recipe. It requires a well-defined nonnegative event time and finite $\tau$; it does not require a particular hazard model, a constant treatment effect, or proportional hazards. Estimating the area from observed, possibly censored data requires additional methods and assumptions.

Worked Example: Constant Hazard Through 12 Months

Suppose, hypothetically, that event times in a population follow an exponential distribution with constant hazard $\lambda=0.10$ events per person-month. The survival function is $S(t)=e^{-0.10t}$. Find the RMST through $\tau=12$ months.

The model implies nonnegative event times and a valid survival function. Its exponential form is used only for this example; the RMST definition itself does not require constant hazard. Integrating the survival function gives

\mu(12)=\int_0^{12}e^{-0.10t}\,dt =\left[-\frac{1}{0.10}e^{-0.10t}\right]_0^{12} =\frac{1-e^{-1.2}}{0.10} \approx 6.988\text{ months}.

Thus, the expected event-free time accumulated during the first 12 months is about 6.988 months. This is not an estimate of the unrestricted mean survival time: the latter, under this exponential model, is $1/\lambda=10$ months. The distinction arises because RMST caps any event time beyond 12 months at 12.

RMST and the Hazard Ratio Summarize Different Things

For a group with survival function $S(t)$, the hazard at time $t$ is the instantaneous event rate among individuals who have remained event-free just before $t$. A hazard ratio compares hazards between groups at the same time. Under proportional hazards, this ratio is constant over time; when hazards are nonproportional, it can vary with time.

RMST instead averages event-free time over a specified interval. It is based on the area under $S(t)$ from zero to $\tau$, not on the ratio of hazards at any one time. Its group-specific value has units of time. A difference between two groups’ RMST values, if defined, also has units of time. Neither the RMST nor its time-based interpretation requires a constant hazard ratio.

These summaries can lead to different conclusions without contradicting one another. A hazard ratio describes a comparison of instantaneous rates among survivors at each time. RMST describes the average amount of event-free time accumulated up to a horizon. In particular, the hazard ratio alone does not state how many months of event-free time separate groups through that horizon.

Summary	What it describes	Units	Key condition or choice
RMST through $\tau$	Expected event-free time accumulated through a fixed horizon	Time	The horizon $\tau$ must be specified
Hazard ratio	Ratio of instantaneous event rates among those still event-free	None	A single constant ratio requires proportional hazards
Survival probability at $\tau$	Probability of remaining event-free beyond the horizon	None	Describes status at one time, not accumulated time
Median survival time	Time by which survival probability reaches one-half	Time	May not be reached during observed follow-up

Worked Example: A Delayed Hazard Reduction

Consider hypothetical event-time distributions through 12 months. In a control group, the hazard is $0.10$ per month throughout. In a treatment group, the hazard is $0.10$ per month for the first 4 months and $0.05$ per month from month 4 through month 12. Assume these piecewise-constant hazards fully specify the event-time distributions over this interval. Find each group’s RMST through 12 months.

The control survival function is $S_C(t)=e^{-0.10t}$, so

\mu_C(12)=\int_0^{12}e^{-0.10t}\,dt =\frac{1-e^{-1.2}}{0.10} \approx 6.988\text{ months}.

For treatment, survival is $e^{-0.10t}$ up to month 4. After month 4, survival starts at $e^{-0.4}$ and declines at hazard $0.05$ per month. Therefore

\begin{aligned} \mu_T(12) &=\int_0^4 e^{-0.10t}\,dt +\int_4^{12}e^{-0.4}e^{-0.05(t-4)}\,dt\\ &=\frac{1-e^{-0.4}}{0.10} +e^{-0.4}\frac{1-e^{-0.4}}{0.05}\\ &\approx 3.297+4.420 =7.717\text{ months}. \end{aligned}

Through 12 months, the treatment-group event-time distribution has an RMST about 0.729 months greater than the control-group distribution. The hazards are equal during the first 4 months and the treatment-to-control hazard ratio is $0.05/0.10=0.5$ afterward. The RMST calculation summarizes the resulting survival curves’ area; it does not treat 0.5 as a constant hazard ratio over the entire period.

The Horizon Is Part of the Estimand

A horizon does more than set a reporting window. It determines which part of each survival curve contributes to the RMST. Area accumulated after $\tau$ is excluded, while survival through $\tau$ contributes at every time before the horizon. If survival curves cross, the difference in their accumulated areas can depend on where the horizon falls. The RMST is therefore a horizon-specific summary, not a universal ranking of event-time distributions.

Worked Example: Crossing Curves and Two Horizons

Suppose two hypothetical populations have discrete event times. In group A, $P(T=2)=0.5$ and $P(T=10)=0.5$. In group B, $P(T=1)=0.2$ and $P(T=5)=0.8$. These probabilities sum to one in each group, and all event times are nonnegative. Calculate the RMSTs through 2 and 4 time units.

At $\tau=2$, group A’s event times are both at least 2, so $\min(T,2)=2$ with probability one. In group B, the truncated time is 1 with probability 0.2 and 2 with probability 0.8. Thus

\mu_A(2)=2,\qquad \mu_B(2)=0.2(1)+0.8(2)=1.8.

At $\tau=4$, group A’s truncated times are 2 and 4, each with probability 0.5. Group B’s are 1 with probability 0.2 and 4 with probability 0.8. Therefore

\mu_A(4)=0.5(2)+0.5(4)=3.0,\qquad \mu_B(4)=0.2(1)+0.8(4)=3.4.

Group A has the larger RMST through 2 time units, whereas group B has the larger RMST through 4. This is possible because the survival curves cross: A retains more event-free time over part of the earlier interval, while B’s pattern yields more area over the longer interval. The two results answer different horizon-specific questions.

Interpretation, Assumptions, and Failure Modes

A careful report identifies the event, population or group, horizon, and time units. For example: “The RMST for time to hospitalization through 12 months was 7.7 months in the treatment group.” This describes an expectation under the group’s event-time distribution; it does not say that every participant remains event-free for 7.7 months or that the treatment caused a particular individual’s outcome.

For a causal interpretation of group-specific RMSTs in a randomized trial, the event time must be defined under each assigned treatment strategy, and the comparison must preserve the randomized groups’ meaning. Consistency (the observed outcome corresponds to the strategy received), random assignment, and no interference between participants support a causal reading of the group distributions. The RMST definition itself is descriptive and does not establish causation.

Horizon not stated. An RMST without $\tau$ is incomplete. The value can differ materially across horizons.
Unsupported follow-up range. If the event-time distribution is not adequately observed near $\tau$, estimating area through $\tau$ may depend heavily on extrapolation or on weakly supported data. The mathematical estimand remains defined, but the data may not identify it precisely.
Censoring treated as an event. Censoring is not the event time in the definition. Treating a censored observation as if the event occurred at censoring changes the data analysis and can bias an estimate.
RMST described as a hazard ratio. RMST is a mean time, while a hazard ratio is a relative instantaneous rate. They have different units and interpretations.
Assuming the horizon is innocuous. With nonproportional hazards or crossing curves, changing $\tau$ can change the magnitude, and sometimes the direction, of a group contrast.

The main alternative summaries answer different questions. A hazard ratio is compact when proportional hazards is a reasonable working description, but a single ratio can hide time-varying effects. Survival probability at $\tau$ describes the chance of remaining event-free beyond one time point, not the accumulated time before it. Median survival time focuses on the halfway point of the distribution and may be unavailable within follow-up. RMST provides a bounded average-time summary, but it depends on a chosen horizon and does not describe what happens after that horizon.

Key takeaway: RMST through $\tau$ is $E[\min(T,\tau)]$, equivalently the area under the survival curve from 0 to $\tau$. It is an interpretable, time-based estimand that does not require proportional hazards, but its horizon must be specified and supported by the study’s follow-up.

Check Your Understanding

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

In one sentence, define RMST through a horizon $\tau$, and state its units.
If $S(t)=e^{-0.20t}$ for time measured in months, calculate RMST through 5 months.
Why is an RMST through 12 months not interchangeable with a hazard ratio?
For the two populations in the crossing-curves example, what are the RMSTs through 2 time units, and which is larger?
Does independent censoring define RMST, or is it relevant to estimating RMST from observed data? Explain briefly.

Solutions. (1) RMST through $\tau$ is $E[\min(T,\tau)]$, the expected event-free time accumulated up to $\tau$; its units are units of time. (2) $\mu(5)=\int_0^5 e^{-0.20t}dt=(1-e^{-1})/0.20\approx 3.161$ months. (3) RMST is an expected time in months (or the relevant time unit), while a hazard ratio is a unitless comparison of instantaneous event rates; they summarize different aspects of the event-time distributions. (4) $\mu_A(2)=2.0$ and $\mu_B(2)=1.8$ time units, so group A’s RMST is larger. (5) RMST is defined by the event-time distribution and horizon. Censoring assumptions matter when estimating it from observed follow-up, because censoring can prevent event times from being observed.

Summary	What it describes	Units	Key condition or choice
RMST through \(\tau\)	Expected event-free time accumulated through a fixed horizon	Time	The horizon \(\tau\) must be specified
Hazard ratio	Ratio of instantaneous event rates among those still event-free	None	A single constant ratio requires proportional hazards
Survival probability at \(\tau\)	Probability of remaining event-free beyond the horizon	None	Describes status at one time, not accumulated time
Median survival time	Time by which survival probability reaches one-half	Time	May not be reached during observed follow-up