Restricted Mean Survival Time at an Interim Landmark

Why Use Restricted Mean Survival Time at an Interim Look?

At an interim selection look, an arm’s event-free survival can be summarized in several ways. A hazard ratio from a Cox model is one option, but it may be difficult to interpret as an amount of time gained or lost. A survival probability at one time point is another option, but it does not distinguish arms whose event patterns differ earlier in follow-up. Restricted mean survival time (RMST) instead summarizes the expected event-free time accumulated over a prespecified interval.

Here, the landmark is a fixed duration $\tau$ after randomization, selected before examining the interim arm results. The interim data cut is the calendar time at which information is reviewed; it is not itself the RMST horizon. This distinction matters when participants have different amounts of follow-up at the data cut. An RMST contrast can help describe intermediate evidence when considering whether an arm should continue, but it does not by itself establish benefit on a later endpoint or determine a calibrated retention rule.

Definition: Let $T^a$ be the event time that would be observed under assignment to arm $a$, and let $S_a(t)=P(T^a>t)$ be that arm’s event-free survival function. The restricted mean survival time through the prespecified horizon $\tau$ is $$ \mu_a(\tau)=E\{\min(T^a,\tau)\}=\int_0^\tau S_a(t)\,dt. $$ For a treatment arm $a$ compared with control arm $0$, define the RMST contrast as $\Delta_a(\tau)=\mu_a(\tau)-\mu_0(\tau)$. When the event is undesirable, a positive contrast means that the treatment arm has more expected event-free time, restricted to the first $\tau$ time units.

Why the Integral Is an Expected Time

For any nonnegative event time $T$, the quantity $\min(T,\tau)$ is the amount of time from zero to $\tau$ during which the person remains event-free. It can be written as the area under an indicator:

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

Taking expectations and exchanging expectation and integration gives

E\{\min(T,\tau)\} =\int_0^\tau E\{I(T>t)\}\,dt =\int_0^\tau P(T>t)\,dt =\int_0^\tau S(t)\,dt.

Thus RMST is the area under the event-free survival curve through $\tau$. It is expressed in the same units as time, such as days or months. It is not the unrestricted mean event time: events after $\tau$ do not contribute additional time to this estimand.

Estimation From Interim Survival Data

For each arm, the usual unadjusted estimator substitutes the arm-specific Kaplan–Meier estimate $\hat S_a(t)$ for $S_a(t)$ and integrates through $\tau$:

\hat\mu_a(\tau)=\int_0^\tau \hat S_a(t)\,dt.

Because the Kaplan–Meier curve is a step function, this integral is a sum of rectangles. If there are event times $t_1,\ldots,t_J$ before $\tau$, then the survival estimate after each event is held constant until the next event or $\tau$. Censoring times change the risk set, but do not themselves create a downward step in the estimated survival curve.

The ordinary, unadjusted arm-specific Kaplan–Meier estimator requires censoring to be independent of event time marginally within each arm, as well as adequate follow-up support through $\tau$. Independence only conditional on covariates is not enough to justify a pooled, unadjusted Kaplan–Meier curve for marginal survival. If censoring is independent only conditional on covariates, use an appropriate adjusted or stratified estimator and marginalize it to the target population. The estimand, covariate adjustment, and target population should be specified consistently.

At an interim data cut, administrative censoring is common because participants entered the trial at different calendar times. That censoring may be compatible with the unadjusted estimator if it is marginally independent of event time within arm and there is adequate support through $\tau$. But if calendar time or entry patterns are related to prognosis, marginal independence may fail. A long horizon with only a few participants still at risk is also weakly supported, even if a curve can technically be drawn to $\tau$.

Worked Example: Calculating RMST From One Arm’s Kaplan–Meier Curve

Suppose a hypothetical arm has six participants and the prespecified horizon is $\tau=6$ months. One event occurs at month 2, one participant is censored at month 3, one event occurs at month 4, and one event occurs at month 5. The two remaining participants are followed to month 6 without an event. Assume, for this illustration, that censoring is marginally independent of event time within the arm and follow-up is adequate through six months.

At month 2, $Y_1=6$ participants are at risk and $d_1=1$ event occurs, so the Kaplan–Meier survival estimate becomes $5/6$. Censoring at month 3 does not change that estimate. At month 4, $Y_2=4$ are at risk and one event occurs, giving $(5/6)(3/4)=5/8$. At month 5, $Y_3=3$ are at risk and one event occurs, giving $(5/8)(2/3)=5/12$.

The curve is therefore 1 from month 0 to 2, $5/6$ from month 2 to 4, $5/8$ from month 4 to 5, and $5/12$ from month 5 to 6. Its area is

\hat\mu(6) =(2)(1)+(2)\left(\frac{5}{6}\right) +(1)\left(\frac{5}{8}\right)+(1)\left(\frac{5}{12}\right) =4.708\text{ months}.

The estimate is 4.708 months of event-free time per participant, averaged over the first six months under this arm. It does not mean that each participant remained event-free for 4.708 months, nor does it estimate average event-free time beyond month 6.

Comparing Arms and Quantifying Uncertainty

For two randomized arms, estimate each arm’s RMST using its survival curve, then subtract in the stated direction. A common approximate variance estimator for an arm’s RMST uses the Greenwood increments. Let $Y_j$ be the number at risk just before event time $t_j$, and $d_j$ the number of events at that time. Then

\widehat{\mathrm{Var}}\{\hat\mu_a(\tau)\} = \sum_{j:t_j<\tau} \left\{\int_{t_j}^{\tau}\hat S_a(t)\,dt\right\}^{2} \frac{d_j}{Y_j(Y_j-d_j)}.

The integral in each term is the remaining area under the estimated survival curve after that event time. Under independent arm samples, the estimated variance of the treatment-minus-control contrast is the sum of the two arm-specific variance estimates. The square root is the contrast’s standard error, denoted $SE$. A normal-approximation interval is the estimate plus or minus 1.96 standard errors for a nominal 95% confidence interval. With small samples or sparse late risk sets, this approximation can be unreliable; the calculation should not be mistaken for a guarantee of precise inference.

Worked Example: A Two-Arm RMST Contrast With Approximate Uncertainty

Use the hypothetical six-month data from the preceding example as arm A. In a separate hypothetical control arm with six participants, events occur at months 3 and 5, and the other four participants are followed to month 6 without an event. Assume marginally independent censoring within each arm and adequate follow-up through the horizon.

For control, the survival estimate is 1 until month 3, $5/6$ from month 3 to 5, and $2/3$ from month 5 to 6. Therefore

\hat\mu_0(6) =(3)(1)+(2)\left(\frac{5}{6}\right)+(1)\left(\frac{2}{3}\right) =5.333\text{ months}.

The estimated treatment-minus-control contrast is

\hat\Delta_A(6)=4.708-5.333=-0.625\text{ months}.

Using the Greenwood-based formula, the remaining areas after arm A’s event times at months 2, 4, and 5 are 2.708, 1.042, and 0.4167 months. The corresponding risk sets are 6, 4, and 3, with one event at each time. Thus

\widehat{\mathrm{Var}}\{\hat\mu_A(6)\} =(2.708)^2\frac{1}{6(5)} +(1.042)^2\frac{1}{4(3)} +(0.4167)^2\frac{1}{3(2)} =0.3639\text{ months}^2.

For control, the remaining areas after months 3 and 5 are 2.333 and 0.6667 months. The risk sets are 6 and 5, with one event at each time, so

\widehat{\mathrm{Var}}\{\hat\mu_0(6)\} =(2.333)^2\frac{1}{6(5)} +(0.6667)^2\frac{1}{5(4)} =0.2037\text{ months}^2.

The estimated standard error of the contrast is $\sqrt{0.3639+0.2037}=0.7534$ months. The approximate 95% confidence interval is

-0.625\ \mathbin{\pm}\ 1.96(0.7534) =[-2.102,\ 0.852]\text{ months}.

In this small hypothetical example, the estimated arm A contrast is 0.625 fewer event-free months through month 6 than control. The interval is wide and includes zero, so these data do not provide a precise estimate of the direction or size of the contrast. The normal approximation is especially fragile with only six participants per arm; the numerical interval illustrates the calculation, not a reliable trial analysis.

Using the Contrast to Inform Arm Retention

An RMST contrast is useful for an intermediate comparison because its direction and magnitude are stated in time units. For an undesirable event, an estimate of $0.6$ months means an estimated 0.6 additional event-free months per participant, on average, over the chosen horizon. The meaning is tied to that horizon: changing $\tau$ can change the contrast, including its magnitude and potentially its direction.

Worked Example: Interpreting RMST Contrasts for Three Arms

Suppose a hypothetical interim analysis estimates event-free survival at the start of each three-month interval through a 12-month horizon. The estimates are treated as constant within each interval for this illustration:

Interval (months)	Control survival	Arm X survival	Arm Y survival
0 to 3	0.95	0.96	0.93
3 to 6	0.85	0.90	0.82
6 to 9	0.70	0.78	0.68
9 to 12	0.60	0.66	0.55

The area for control is $3(0.95+0.85+0.70+0.60)=9.30$ months. For arm X it is $3(0.96+0.90+0.78+0.66)=9.90$ months, and for arm Y it is $3(0.93+0.82+0.68+0.55)=8.94$ months. Thus the estimated contrasts are $9.90-9.30=0.60$ months for X versus control and $8.94-9.30=-0.36$ months for Y versus control.

Over the first 12 months, arm X has an estimated 0.60 more event-free months per participant than control, whereas arm Y has an estimated 0.36 fewer. If the selection process is considering which arms to retain for further study, this pattern may support considering X more favorably than Y. It is not, by itself, a retention rule: uncertainty, the prespecified selection procedure, and the consequences of an incorrect decision also matter. Nor does the contrast prove that X improves the final endpoint.

Assumptions, Diagnostics, and Alternatives

The interpretation depends on a few choices and conditions that should be checked before an arm contrast is used:

Event and time origin: Define the event, time origin, and whether the analysis follows randomized assignment. If the event is undesirable, greater RMST means more event-free time; the sign convention should be reported explicitly.
Horizon chosen in advance: Specify $\tau$ before reviewing interim arm contrasts. The horizon determines the estimand and should be clinically and operationally meaningful.
Censoring: For an unadjusted Kaplan–Meier estimator of marginal arm survival, censoring must be independent of event time marginally within arm. Conditional independence alone does not suffice. If only conditional independence is credible, use an appropriate adjusted or stratified estimator and average over the target population.
Follow-up support: Inspect numbers at risk and censoring patterns near $\tau$. If few participants remain under observation, the tail of the curve—and hence the area near the horizon—can be unstable. Do not extend the interpretation beyond the supported follow-up.
Interim maturity: Staggered entry means many participants may not yet have had the opportunity to reach $\tau$. Check whether the data cut provides useful information over the full interval rather than relying on a nominal sample size alone.
Selection variability: A noisy estimate can change arm rankings. When several arms are compared, the apparently most favorable estimate may partly reflect random variation. Report uncertainty and use the selection procedure established for the design, rather than treating a point estimate as conclusive.

RMST differs from a Cox-model hazard ratio: RMST directly estimates accumulated event-free time through $\tau$ and does not require proportional hazards for its definition. A hazard ratio summarizes relative event rates and may be difficult to interpret as time gained, while RMST depends explicitly on the chosen horizon. RMST also differs from survival probability at $\tau$: that single probability records event-free status at one time, whereas RMST incorporates the survival curve across the whole interval. Neither alternative is universally superior; the appropriate intermediate summary depends on the scientific question and the selection design.

Important limitations remain. The contrast does not describe what happens after $\tau$, and a different horizon can yield a different comparison. It can be imprecise when follow-up is immature or censoring is heavy. It also does not establish that an intermediate event-time contrast predicts a final endpoint or that retaining an arm will lead to final success. Those are separate questions from estimating the restricted mean through a landmark.

Common Mistake

Common mistake: “Censoring is independent of event time after adjusting for a prognostic covariate, so the ordinary pooled Kaplan–Meier curve estimates marginal survival.” This conclusion does not generally follow. If the covariate affects survival and censoring, the mix of covariate values among those still at risk can change over time, biasing the pooled curve for marginal survival. Use an estimator appropriate to the conditional censoring assumption and marginalize it to the target population. For the unadjusted arm-specific Kaplan–Meier estimator, the required independence is marginal within arm.

Check Your Understanding

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

In one sentence, interpret an RMST contrast of $-0.4$ months for a treatment arm versus control when the event is undesirable and $\tau=9$ months.
Why can a pooled unadjusted Kaplan–Meier curve be inappropriate when censoring is independent of event time only conditional on a prognostic covariate?
For a curve with survival estimates 1.00 from month 0 to 2, 0.80 from month 2 to 5, and 0.50 from month 5 to 6, calculate the estimated RMST through month 6.
Give one reason an RMST contrast at an interim look does not establish that an arm will succeed on the final endpoint.

Solutions. (1) Through the first nine months, the treatment arm has an estimated 0.4 fewer event-free months per participant than control. (2) Changing covariate composition in the pooled risk set can make the marginal censoring process informative for the pooled curve, even when censoring is independent within covariate levels. An appropriate adjusted or stratified estimator must be marginalized to the target population. (3) The area is $(2)(1.00)+(3)(0.80)+(1)(0.50)=4.90$ months. (4) RMST only summarizes event-free time through its chosen horizon; it does not, on its own, validate the intermediate measure as a predictor of the final endpoint.