Restricted Mean Survival Time as an Area

Why Restrict Mean Survival Time?

A mean survival time can be difficult to interpret when some event times may be very long or when the full mean is not finite. A fixed horizon provides a clear alternative: count each subject’s time up to that horizon, but do not count additional time beyond it. The resulting quantity has a direct interpretation in time units and can be represented as an area under the survival function.

This tutorial develops the definition and its interpretation. It treats a survival function as given; it does not address how to calculate an area from a Kaplan-Meier curve. The focus is the quantity being measured and why the area formula gives that quantity.

Definition: Let $T$ be a nonnegative event time, with survival function $S(t)=P(T>t)$. For a fixed horizon $\tau>0$, the restricted mean survival time through $\tau$, also called the restricted mean, is $$ \mu(\tau)=E[\min(T,\tau)]=\int_0^\tau S(t)\,dt. $$ It is the expected amount of time survived within the interval from time 0 through time $\tau$. Its units are the same as the units used for $T$ and $\tau$.

The definition concerns time from a specified origin, such as enrollment, diagnosis, or the start of coverage. The horizon $\tau$ is measured from that same origin. Unless discounting is explicitly introduced, this is an undiscounted expected duration: no interest rate is used.

Why the Area Equals Expected Time

For any individual whose event time is $T$, the quantity $\min(T,\tau)$ is the time survived before either the event occurs or the horizon is reached. At each time $t$ between 0 and $\tau$, that individual contributes one unit of time if they have survived beyond $t$, and contributes no time otherwise. Adding those contributions across the interval gives the individual’s restricted duration.

Write restricted duration as an integral.
For a fixed event time $T$, the indicator $\mathbf{1}(T>t)$ equals 1 while the individual is still event-free at time $t$, and 0 after the event. Thus $\min(T,\tau)=\int_0^\tau \mathbf{1}(T>t)\,dt$, apart from endpoint values that do not affect the integral.

Take expectations.
Because the integrand is nonnegative, expectation and integration can be interchanged. At each $t$, the expected value of the indicator is $P(T>t)=S(t)$.

Obtain the survival-area identity.
Combining the preceding statements gives $E[\min(T,\tau)]=\int_0^\tau E[\mathbf{1}(T>t)]\,dt=\int_0^\tau S(t)\,dt$.

\begin{aligned} \min(T,\tau) &=\int_0^\tau \mathbf{1}(T>t)\,dt,\\ E[\min(T,\tau)] &=\int_0^\tau E[\mathbf{1}(T>t)]\,dt\\ &=\int_0^\tau P(T>t)\,dt\\ &=\int_0^\tau S(t)\,dt. \end{aligned}

This argument requires a nonnegative event time and a fixed, finite horizon. It does not require $E[T]$ to be finite: $\min(T,\tau)$ is bounded above by $\tau$, so its expectation exists. A survival function may have jumps; the integral remains well-defined, and the value assigned at isolated endpoints does not change the area.

What the Restricted Mean Does—and Does Not—Measure

The restricted mean is an average of capped times. A person who experiences the event before $\tau$ contributes their event time; a person who remains event-free through $\tau$ contributes $\tau$. Consequently, $0\leq\mu(\tau)\leq\tau$. Increasing the horizon cannot decrease the restricted mean, because it adds a nonnegative area under the survival curve.

It is important not to confuse $\mu(\tau)$ with $E[T\mid T\leq\tau]$. The latter is a conditional average among those whose event occurs by $\tau$, and excludes those who survive past the horizon. The restricted mean includes everyone in the target population, capping the duration of those who survive longer than $\tau$. Nor is it generally the unrestricted mean $E[T]$; it equals that mean only when the full mean exists and the chosen horizon captures all possible event times.

Quantity	What is averaged?	Role of the horizon
$\mu(\tau)=E[\min(T,\tau)]$	Capped event-free time for the full population	Times beyond $\tau$ count as $\tau$
$E[T\mid T\leq\tau]$	Event time among those with events by $\tau$	Selects a subgroup; does not include those surviving beyond $\tau$
$E[T]$	Unrestricted event time for the full population	No cap; may be infinite or unavailable

For a population survival function, the area is the target quantity. If survival is estimated from censored data, an estimated survival function can be used to estimate this target. Censoring is not itself an event, and the validity of an empirical estimate depends on the assumptions behind the survival estimator.

Worked Example: Uniform Event Time With a Six-Year Horizon

A hypothetical time to event $T$ is uniformly distributed from 0 to 10 years, so $T\sim\operatorname{Uniform}(0,10)$. Find the restricted mean through $\tau=6$ years and the probability of remaining event-free through that horizon. Assume no discounting; no interest rate or separate mortality assumption is used.

Step 1: Write the survival function on the integration interval. For $0\leq t\leq 10$, the uniform distribution has $F_T(t)=t/10$, so $S(t)=1-t/10$. The full interval from 0 to 6 lies within that range.

Step 2: Integrate through the horizon.

\begin{aligned} \mu(6) &=\int_0^6 \left(1-\frac{t}{10}\right)\,dt\\ &=\left[t-\frac{t^2}{20}\right]_0^6\\ &=6-\frac{36}{20}\\ &=6-1.8\\ &=4.2\text{ years}. \end{aligned}

Step 3: Find survival at the horizon. $S(6)=1-6/10=0.4$, so the probability of remaining event-free through six years is 0.4000.

Check. The restricted mean, 4.2 years, is between 0 and the six-year cap. The area calculation also has the expected form: survival falls linearly from 1 to 0.4, so the area is the trapezoid’s average height, $(1+0.4)/2=0.7$, times 6 years, or 4.2 years. The method is direct integration of the survival function; the common mistake to avoid is using an incorrect area or treating the restricted mean as the conditional event time among those with events by year 6.

Worked Example: Exponential Survival With a Five-Year Horizon

For a hypothetical event time, assume an exponential distribution with constant hazard $\lambda=0.20$ per year. Find the restricted mean through $\tau=5$ years and compare it with the unrestricted mean. Assume no discounting; there is no interest-rate assumption.

Step 1: State the survival function. Under the exponential model, $S(t)=e^{-\lambda t}=e^{-0.20t}$, and the unrestricted mean is $1/\lambda=5$ years.

Step 2: Integrate to five years.

\begin{aligned} \mu(5) &=\int_0^5 e^{-0.20t}\,dt\\ &=\left[-\frac{1}{0.20}e^{-0.20t}\right]_0^5\\ &=\frac{1-e^{-1}}{0.20}\\ &=5(1-0.367879\ldots)\\ &=3.1606\text{ years}. \end{aligned}

Step 3: Check the horizon probability and scale. $S(5)=e^{-1}=0.3679$. The restricted mean, 3.1606 years, is less than both the five-year horizon and the unrestricted mean of five years, as it must be. The method is integration under the assumed exponential survival model. The common mistake is to report $1/\lambda=5$ as the restricted mean; that is the unrestricted mean, not the mean of time capped at five years.

Worked Example: Area Under a Piecewise Linear Survival Function

A hypothetical event-time model has the following survival function, with time measured in years:

S(t)= \begin{cases} 1-0.10t, & 0\leq t\leq 4,\\ 0.60-0.15(t-4), & 4<t\leq 8,\\ 0, & t>8. \end{cases}

Find the restricted mean through $\tau=6$ years. Assume no discounting; the stated survival function is the complete distributional assumption needed.

Step 1: Check the pieces and their boundary values. At $t=0$, $S(0)=1$. At $t=4$, the first formula gives $1-0.10(4)=0.60$, matching the value that starts the second piece. At $t=6$, the second formula gives $0.60-0.15(2)=0.30$. The survival function is nonincreasing and nonnegative through the requested horizon.

Step 2: Integrate the first segment.

\begin{aligned} \int_0^4 (1-0.10t)\,dt &=\left[t-0.05t^2\right]_0^4\\ &=4-0.80\\ &=3.20\text{ years}. \end{aligned}

Step 3: Integrate from year 4 to year 6. Set $u=t-4$, so $u$ ranges from 0 to 2 and the survival function on this segment is $0.60-0.15u$.

\begin{aligned} \int_4^6 \left(0.60-0.15(t-4)\right)\,dt &=\int_0^2 (0.60-0.15u)\,du\\ &=\left[0.60u-0.075u^2\right]_0^2\\ &=1.20-0.30\\ &=0.90\text{ years}. \end{aligned}

Step 4: Add the areas. $\mu(6)=3.20+0.90=4.10$ years. As a check, this is below the six-year cap. On the second segment the survival declines linearly from 0.60 to 0.30, so its area is also $2(0.60+0.30)/2=0.90$ years. The method is to integrate each formula over the portion of its interval included in the horizon. A common mistake is to integrate the second piece all the way to year 8 even though the requested horizon is year 6.

Interpretation and Practical Checks

A restricted mean is naturally interpreted in the same time units as the horizon. For example, a value of 4.1 years means that the expected time survived during the first six years is 4.1 years; it does not mean that every individual survives for 4.1 years. It is an average across the full population, including those who remain event-free to the horizon.

When interpreting or calculating a restricted mean, check that the survival function corresponds to the event and population of interest, that its time scale matches $\tau$, and that the integration stops exactly at $\tau$. The result must lie between zero and $\tau$, and increasing $\tau$ cannot reduce it. These checks catch common errors such as using a conditional survival probability as though it were the unconditional $S(t)$, confusing years with months, or reporting an unrestricted mean instead of the capped expectation.

Key takeaway: Restricted mean survival time through $\tau$ is $E[\min(T,\tau)]$, the expected time survived before the event or the horizon, whichever comes first. The survival-area formula $\mu(\tau)=\int_0^\tau S(t)\,dt$ follows by adding the probability of being event-free at each instant through the horizon.

Check Your Understanding

For each question, use the stated survival model and horizon. Unless a question says otherwise, assume no discounting.

A hypothetical event time is uniform on 0 to 12 years. Find the restricted mean through 3 years. (Answer: 2.6250 years)
A hypothetical event time is exponential with constant hazard $\lambda=0.10$ per year. Find the restricted mean through 4 years. (Answer: 3.2968 years)
A survival function is $S(t)=1-0.05t$ for $0\leq t\leq 10$ years. Find the restricted mean through 8 years and check it against the horizon. (Answer: 6.4000 years; it is below the 8-year horizon)
A hypothetical survival function is $S(t)=1-0.20t$ for $0\leq t\leq2$, and $S(t)=0.60-0.10(t-2)$ for $2<t\leq5$. Find the restricted mean through 5 years. (Answer: 2.9500 years)
Compare two hypothetical exponential event-time models through $\tau=4$ years: Model A has $\lambda=0.25$ per year, and Model B has $\lambda=0.15$ per year. Find each restricted mean and the difference, Model B minus Model A. (Answer: Model A = 2.5285 years; Model B = 3.0079 years; difference = 0.4794 years)

Quantity	What is averaged?	Role of the horizon
\(\mu(\tau)=E[\min(T,\tau)]\)	Capped event-free time for the full population	Times beyond \(\tau\) count as \(\tau\)
\(E[T\mid T\leq\tau]\)	Event time among those with events by \(\tau\)	Selects a subgroup; does not include those surviving beyond \(\tau\)
\(E[T]\)	Unrestricted event time for the full population	No cap; may be infinite or unavailable

What You'll Learn