Comparing RMST Between Randomized Groups

Why Compare Restricted Mean Survival Times?

When treatment effects may change over time, a single hazard ratio can be difficult to summarize. Restricted mean survival time (RMST) gives a time-scale summary: the area under a survival curve up to a specified time. Comparing the areas between randomized groups yields a difference with a direct interpretation in units of time.

The comparison requires a common truncation time, denoted by \(\tau\), for both groups. The focus here is the estimated difference and its uncertainty—not how to select \(\tau\). The introductory material on RMST as an estimand defines the target; the essential point for this calculation is that each group’s RMST is its mean event-free time restricted to the same interval from zero to \(\tau\).

Definition: Let \(S_1(t)\) and \(S_0(t)\) be the survival functions for the randomized treatment and control groups. Their RMSTs through \(\tau\) are \(\mu_1(\tau)=\int_0^\tau S_1(t)\,dt\) and \(\mu_0(\tau)=\int_0^\tau S_0(t)\,dt\). Define the treatment-minus-control contrast as \(\Delta(\tau)=\mu_1(\tau)-\mu_0(\tau)\). A positive value means greater restricted mean event-free time in the treatment group; a negative value means less.

The contrast is measured in the same time units as follow-up. For example, a difference of 1.5 months means that, over the interval from zero to \(\tau\), the estimated treatment-group area is 1.5 months greater. It is an average contrast in restricted time, not a statement that every individual gains that amount.

Estimating the Difference and Its Standard Error

In each group, estimate the survival function with the Kaplan–Meier estimator, then integrate it up to \(\tau\). For a step-function Kaplan–Meier curve, this integral is a sum of rectangles: each interval’s length multiplied by the estimated survival level over that interval. The earlier material on estimating RMST from Kaplan–Meier curves develops this calculation.

For the variance, use the Greenwood-based RMST estimator. In one group, let \(t_j\) be an event time no later than \(\tau\), \(Y_j\) the number at risk immediately before that time, and \(d_j\) the number of events at that time. Let \(\hat S(t)\) be the group’s Kaplan–Meier estimate and define the remaining area after the event time as \(A_j=\int_{t_j}^{\tau}\hat S(t)\,dt\). The event-time contribution depends on this remaining area and on the Greenwood risk-set term.

Variance formula: With event times \(t_j\leq\tau\), the Greenwood-based variance estimate for the group’s RMST is \(\widehat{\operatorname{Var}}\{\hat\mu(\tau)\}=\sum_j \frac{d_j}{Y_j(Y_j-d_j)}A_j^2\). For two independent randomized groups, estimate the contrast variance by adding the two group-specific variance estimates. Its standard error is the square root of that sum.

The addition follows because the treatment and control estimates are calculated from separate groups and are treated as independent under the usual independent-arm analysis. The estimated covariance between them is therefore zero. This is not a rule for subtracting variances: the uncertainty in a difference of independent estimates is the sum of their variances.

A useful justification for the single-group formula comes from the large-sample covariance of Kaplan–Meier estimates. In its Greenwood form, the estimated covariance at times \(s\) and \(t\) is \(\hat S(s)\hat S(t)\) multiplied by the sum of Greenwood terms for event times no later than \(\min(s,t)\). Integrating this covariance over \(s\) and \(t\) from zero to \(\tau\) gives the RMST variance. For a given event time \(t_j\), its term appears only when both integration times are at least \(t_j\). Thus its double integral is

\frac{d_j}{Y_j(Y_j-d_j)} \left\{\int_{t_j}^{\tau}\hat S(t)\,dt\right\}^{2} = \frac{d_j}{Y_j(Y_j-d_j)}A_j^2.

Summing over event times gives the stated variance estimator. This large-sample calculation relies on the Kaplan–Meier estimator being appropriate for the observed censoring and on enough information being available to estimate survival through the chosen horizon.

A Wald confidence interval for the contrast is

\hat\Delta(\tau)\ \pm\ z_{1-\alpha/2}\, \widehat{SE}\{\hat\Delta(\tau)\}, \qquad \widehat{SE}\{\hat\Delta(\tau)\} = \sqrt{\widehat{\operatorname{Var}}\{\hat\mu_1(\tau)\} +\widehat{\operatorname{Var}}\{\hat\mu_0(\tau)\}}.

Here \(z_{1-\alpha/2}\) is the standard normal critical value; for a two-sided 95% interval it is approximately 1.96. The interval is an approximate repeated-sampling confidence interval: under the assumptions and approximation, procedures of this kind cover the true contrast about 95% of the time over repeated studies. It does not assign a 95% probability to the fixed contrast being in this particular interval.

Worked Example: Calculating a Contrast from Kaplan–Meier Curves

Suppose a hypothetical randomized trial follows participants for up to 6 months. The event and risk-set summaries are as follows. In the control group, 2 of 10 at risk have events at month 2, and 2 of the 8 then at risk have events at month 4. In the treatment group, 1 of 10 at risk has an event at month 2; one more participant is censored before month 4, when 2 of the 8 then at risk have events. Assume there are no further events through month 6. Calculate the treatment-minus-control RMST difference and a 95% confidence interval.

Control estimate. The Kaplan–Meier survival levels are 1 from month 0 to 2, \(8/10=0.8\) from month 2 to 4, and \(0.8(6/8)=0.6\) from month 4 to 6. Therefore,

\hat\mu_0(6)=2(1)+2(0.8)+2(0.6)=4.80\text{ months}.

The remaining areas are \(A_2=2(0.8)+2(0.6)=2.8\) months and \(A_4=2(0.6)=1.2\) months. The Greenwood contributions are

\widehat{\operatorname{Var}}\{\hat\mu_0(6)\} = \frac{2}{10(8)}(2.8)^2 + \frac{2}{8(6)}(1.2)^2 = 0.196+0.060 = 0.256\text{ months}^2.

Treatment estimate. Survival is 1 before month 2, \(9/10=0.9\) from month 2 to 4, and \(0.9(6/8)=0.675\) from month 4 to 6. The censoring before month 4 changes the risk set but does not itself reduce the Kaplan–Meier survival estimate. Thus,

\hat\mu_1(6)=2(1)+2(0.9)+2(0.675)=5.15\text{ months}.

Here \(A_2=2(0.9)+2(0.675)=3.15\) months and \(A_4=2(0.675)=1.35\) months. The variance estimate is

\widehat{\operatorname{Var}}\{\hat\mu_1(6)\} = \frac{1}{10(9)}(3.15)^2 + \frac{2}{8(6)}(1.35)^2 = 0.11025+0.0759375 = 0.1861875\text{ months}^2.

The estimated contrast and standard error are

\hat\Delta(6)=5.15-4.80=0.35\text{ months},\qquad \widehat{SE}=\sqrt{0.1861875+0.256}\approx0.6650\text{ months}.

The 95% Wald interval is \(0.35\pm1.96(0.6650)\), or approximately \((-0.953,\ 1.653)\) months. The point estimate favors treatment by 0.35 months, but the interval includes zero and also allows a modest negative contrast. These data do not establish a difference in restricted mean event-free time at the 5% two-sided level.

Inference and Interpretation in Context

A two-sided Wald test of \(H_0:\Delta(\tau)=0\) uses \(Z=\hat\Delta(\tau)/\widehat{SE}\{\hat\Delta(\tau)\}\), compared with the standard normal distribution under the null. The test and confidence interval are two presentations of the same approximate inference: a 95% interval excluding zero corresponds to a two-sided test with \(p<0.05\), subject to rounding.

Worked Example: Testing a Hypothetical Contrast

In a separate hypothetical trial, the estimated treatment-minus-control RMST difference through a common horizon is 1.10 days. The estimated RMST variances are 0.36 days squared for treatment and 0.25 days squared for control. Test \(H_0:\Delta=0\) against a two-sided alternative.

The variance and standard error of the difference are

\widehat{\operatorname{Var}}(\hat\Delta)=0.36+0.25=0.61\text{ days}^2,\qquad \widehat{SE}=\sqrt{0.61}\approx0.7810\text{ days}.

Thus \(Z=1.10/0.7810\approx1.409\). The two-sided normal-approximation p-value is \(2\{1-\Phi(1.409)\}\approx0.159\). The corresponding 95% interval is \(1.10\pm1.96(0.7810)\), approximately \((-0.431,\ 2.631)\) days. The result is not statistically significant at the 5% level; the p-value is not the probability that the null hypothesis is true.

Worked Example: Distinguishing a Positive Difference from a Meaningful One

Suppose a hypothetical analysis reports a treatment-minus-control RMST difference of 2.40 months through the prespecified horizon, with estimated variances of 0.64 and 0.81 months squared. For illustration, suppose investigators regard 1 month as a clinically important difference. Does the interval show that the benefit exceeds that threshold?

The standard error is \(\sqrt{0.64+0.81}=\sqrt{1.45}\approx1.204\) months. A 95% interval is

2.40\pm1.96(1.204)\approx(0.040,\ 4.760)\text{ months}.

The interval lies just above zero, but it includes differences below 1 month. Thus it provides approximate evidence of a positive contrast at the 5% level, but it does not establish that the benefit exceeds the stated 1-month threshold. Statistical evidence against a zero difference and evidence that an effect is clinically important are different questions.

Assumptions, Diagnostics, and Alternatives

The calculation has a clear interpretation only when its design and estimation assumptions are plausible. In a randomized study, group assignment supports a causal comparison of assigned treatment strategies when outcomes are analyzed consistently with that assignment and relevant trial conduct assumptions hold. The Kaplan–Meier estimator also requires censoring to be non-informative for event time within each group: roughly, among participants still at risk, those censored should not have systematically different future event prospects from those remaining under observation. Administrative end of follow-up can be compatible with this condition; dropout related to unobserved worsening may not be.

Check the event and censoring counts over time, the numbers at risk near \(\tau\), and whether each group has adequate follow-up to that horizon. A curve supported by very few people near its end can make the area and especially its estimated uncertainty unstable. The Greenwood expression also requires \(Y_j>d_j\) for event contributions as written; if everyone at risk fails at a time, the denominator is zero and this standard formula cannot be used naively. Sparse events, heavy censoring, or a horizon beyond well-supported follow-up undermine the normal approximation and can produce unreliable Wald intervals. No variance formula repairs informative censoring or absent follow-up support.

RMST differences answer a different question from several familiar alternatives. A hazard ratio summarizes a relative instantaneous event rate under its model and interpretation; it is not a difference in average time. A logrank or weighted logrank test assesses evidence through event-time comparisons, rather than directly estimating the area difference. Survival probabilities at a particular time compare the chance of remaining event-free at that point, not accumulated event-free time. RMST is attractive when a time-scale summary is useful, but the contrast depends on the chosen horizon, discards event-free time beyond that horizon, and does not describe the full shape of either survival curve. Reporting the horizon and the group-specific estimates alongside the difference helps make the result transparent.

Common mistake: Treating the variance of a difference as the difference of the two variances, or reporting the point estimate without its sign convention and time units. For independent randomized groups, add the variance estimates, take the square root for the standard error, and state explicitly whether the contrast is treatment minus control. A positive estimate then has a consistent meaning.

Check Your Understanding

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

At a common horizon, treatment RMST is estimated as 7.2 months and control RMST as 6.5 months. What is the treatment-minus-control contrast, and how should its sign be interpreted?
Two independent groups have estimated RMST variances of 0.16 and 0.25 years squared. Calculate the standard error of their difference.
A 95% confidence interval for a treatment-minus-control RMST difference is \((-0.4,\ 1.8)\) days. What does it say about a zero difference, and what does it not say about the null hypothesis?
Name one diagnostic concern that could make a Greenwood-based Wald interval unreliable near the truncation horizon.

Solutions. (1) The contrast is \(7.2-6.5=0.7\) months, favoring treatment in restricted mean event-free time through the stated horizon. (2) The variance is \(0.16+0.25=0.41\) years squared, so the standard error is \(\sqrt{0.41}\approx0.6403\) years. (3) Because the interval includes zero, it does not exclude a zero contrast at the corresponding approximate two-sided 5% level. It does not give the probability that the null is true. (4) Examples include very few participants remaining at risk, heavy censoring near the horizon, or lack of follow-up support through that horizon.

Comparing RMST Between Randomized Groups

What You'll Learn