RMST Under Delayed and Crossing Effects

Why Timing Matters for an RMST Contrast

A single summary can combine very different patterns over time. A treatment might have little effect early and a benefit later; alternatively, one group’s survival curve might initially be higher and then fall below the other’s. The restricted mean survival time (RMST) contrast captures the net difference over a specified time window. Its value depends on the whole survival curves within that window, not just on which group has a higher curve at the end.

Write \(S_g(t)\) for the survival function in group \(g\), where \(g=1\) denotes treatment and \(g=0\) denotes control. For a common truncation horizon \(\tau\), the RMST in group \(g\) is the expected event-free time accumulated from time zero through \(\tau\). The contrast used here is treatment minus control, so a positive value indicates more restricted mean survival time in the treatment group.

Definition: The RMST in group \(g\) through horizon \(\tau\) is \(\mu_g(\tau)=\int_0^\tau S_g(t)\,dt\). The RMST contrast is \(\Delta(\tau)=\mu_1(\tau)-\mu_0(\tau)=\int_0^\tau [S_1(t)-S_0(t)]\,dt\). Its units are units of time, such as months or years.

The curve-area interpretation is especially useful under delayed or crossing effects. At each time, \(S_1(t)-S_0(t)\) is the vertical difference between the survival curves. Integrating those differences adds the areas where the treatment curve is higher and subtracts areas where it is lower. The result is a signed net area, not the total unsigned area between the curves.

Justification: RMST as Area Under a Survival Curve

Let \(T_g\) be the event time in group \(g\). The restricted time is \(\min(T_g,\tau)\): a person contributes their event time if the event occurs by \(\tau\), and otherwise contributes \(\tau\). For any nonnegative random variable \(Y\), its expectation equals the area under its tail probability, \(E(Y)=\int_0^\infty P(Y>t)\,dt\). Applying that identity to \(Y=\min(T_g,\tau)\), whose tail probability is zero after \(\tau\), gives the RMST formula.

\begin{aligned} \mu_g(\tau) &=E[\min(T_g,\tau)]\\ &=\int_0^\infty P\{\min(T_g,\tau)>t\}\,dt\\ &=\int_0^\tau P(T_g>t)\,dt\\ &=\int_0^\tau S_g(t)\,dt. \end{aligned}

Subtracting the two group-specific integrals gives \(\Delta(\tau)=\int_0^\tau[S_1(t)-S_0(t)]\,dt\). This derivation does not require proportional hazards. It does require that both groups’ survival curves and the common horizon refer to the same time origin and event definition.

Assumptions and Practical Conditions

The estimand itself is defined by the survival distributions through \(\tau\). In a randomized trial, interpreting the contrast as an effect of assignment also relies on valid randomization, consistent outcome measurement, and appropriate handling of loss to follow-up. Estimating the curves with ordinary group-specific Kaplan–Meier methods ordinarily requires censoring to be independent of the event time within each group marginally, or conditions that make the marginal estimate valid. If censoring is independent only conditional on covariates, an appropriate conditional, stratified, or weighted method is needed; for a marginal target, specify how the result is standardized over covariates. The horizon should be chosen so both groups have meaningful follow-up and the curves are estimable through that time.

For an estimated contrast, estimate each RMST from its group’s survival curve and subtract in the stated order. With Kaplan–Meier curves, the estimate is the area under each step function through \(\tau\). The earlier material on estimating RMST from Kaplan–Meier curves covers that integration; the examples below focus on how the areas combine under particular effect patterns.

Worked Example: A Delayed Benefit in Hypothetical Survival-Curve Areas

Suppose a hypothetical randomized study follows participants for 12 months. The two estimated survival curves coincide through month 6. Integrating the common curve from month 0 through month 6 gives 5.183636 months. After month 6, the hypothetical estimated treatment curve has a conditional remaining area of 4.277065 months, and the estimated survival probability at month 6 is 0.740818. The control RMST through month 12, obtained by integrating its full estimated curve, is 7.356723 months. These are hypothetical curve summaries, not results from a real study.

Because the curves coincide before month 6, that shared area contributes equally to both RMSTs. The treatment-group area after month 6 is the survival probability at month 6 multiplied by the remaining area conditional on being alive at month 6. Using the unrounded curve-area calculations gives:

\begin{aligned} \hat{\mu}_1(12) &\approx 5.183636\ldots+0.740818\ldots(4.277065\ldots)\\ &\approx 8.352162925\text{ months}\\ &\approx 8.352163\text{ months}. \end{aligned}

The control estimate is \(\hat{\mu}_0(12)\approx7.356723165\) months, or 7.356723 months to six decimal places. Therefore:

\hat{\Delta}(12)=\hat{\mu}_1(12)-\hat{\mu}_0(12) \approx 8.352162925-7.356723165 =0.995439760\text{ months} \approx 0.995440\text{ months}.

The contrast is positive: over the first 12 months, the treatment group has an estimated 0.9954 additional months of restricted mean survival time. The extra digits above show the subtraction before rounding; rounded intermediate curve summaries need not reproduce every final decimal. This contrast is a net area over 12 months, not a claim that each participant gains that amount or that the benefit begins exactly at month 6.

Crossing Curves: Positive and Negative Areas Can Cancel

When curves cross, the RMST contrast can be small even when the groups differ substantially at particular times. The early area favoring one group may be offset by a later area favoring the other. The direction of the contrast alone therefore does not describe the timing or shape of the differences; it describes their signed total through \(\tau\).

Worked Example: A Contrast When Survival Curves Cross

Consider hypothetical piecewise-linear survival curves for two groups, A and B, with time measured in months. The curves have the following values at selected times, with straight-line interpolation between them.

Time (months)	Group A survival	Group B survival
0	1.00	1.00
2	0.90	0.80
5	0.60	0.65
10	0.30	0.35

The curves cross between months 2 and 5. Over that interval, the difference falls linearly from \(0.90-0.80=0.10\) to \(0.60-0.65=-0.05\). It reaches zero after \(0.10/(0.15/3)=2\) months, at month 4. To calculate the RMSTs through month 10, use trapezoid areas for each straight-line segment:

\begin{aligned} \hat{\mu}_A(10) &=2\frac{1.00+0.90}{2} +3\frac{0.90+0.60}{2} +5\frac{0.60+0.30}{2}\\ &=1.90+2.25+2.25 =6.400\text{ months},\\[4pt] \hat{\mu}_B(10) &=2\frac{1.00+0.80}{2} +3\frac{0.80+0.65}{2} +5\frac{0.65+0.35}{2}\\ &=1.80+2.175+2.50 =6.475\text{ months}. \end{aligned}

Thus the A-minus-B contrast is \(6.400-6.475=-0.075\) months. The signed areas make the cancellation clear: from month 0 to 2, A has an area advantage of 0.100 months; from month 2 to 5, the net advantage is 0.075 months; and from month 5 to 10, B has an area advantage of 0.250 months. The total is \(0.100+0.075-0.250=-0.075\) months. Although A’s curve is initially higher, B has slightly greater RMST through month 10. These hypothetical interpolated curves illustrate the geometry; actual Kaplan–Meier curves are step functions.

Delayed Effects and the Choice of Horizon

A delayed effect contributes little or nothing to the contrast before the groups’ survival curves begin to separate. After separation, the contrast accumulates according to the vertical gap. For example, if treatment and control have the same hazard until a prespecified delay and treatment has a lower hazard afterward, the RMST contrast through a horizon beyond that delay will generally be positive. Its size still depends on how much follow-up remains after the delay and on the survival probabilities at that time.

Worked Example: A Simple Delayed-Hazard Model

Suppose, hypothetically, that both groups follow a piecewise-exponential model: both hazards are 0.10 per month through month 3. After month 3, the control hazard remains 0.10 per month and the treatment hazard is 0.05 per month. Assume complete follow-up for the calculation and choose \(\tau=9\) months. The survival curves coincide through month 3, and the treatment curve is higher afterward.

For control, the hazard is 0.10 throughout, so its RMST is the integral of \(\exp(-0.10t)\) from 0 to 9. For treatment, split the integral at month 3. The first part is shared; the second part starts at survival probability \(\exp(-0.30)\) and follows the post-delay hazard:

\begin{aligned} \mu_0(9) &=\frac{1-e^{-0.10(9)}}{0.10} \approx 5.9343\text{ months},\\ \mu_1(9) &=\frac{1-e^{-0.10(3)}}{0.10} +e^{-0.10(3)}\frac{1-e^{-0.05(6)}}{0.05}\\ &\approx 2.5918+0.7408(5.1836) \approx 6.4319\text{ months},\\ \Delta(9)&\approx 6.4319-5.9343 =0.4976\text{ months}. \end{aligned}

The model-based treatment advantage is about 0.498 restricted mean months through month 9. The shared area before month 3 cancels in the contrast; the positive contrast comes from the treatment curve being higher after the hazard changes. This is a model illustration, not a claim that an exponential model is necessary for RMST comparisons.

Changing \(\tau\) changes the estimand, not merely the precision of a fixed estimand. A horizon that ends before a delayed separation will produce little or no contrast even if the curves diverge later. Extending the horizon can add a positive or negative area, depending on which curve is higher. The horizon should therefore be clinically interpretable, selected with adequate follow-up in mind, and applied equally to both groups. A contrast should always be reported with its horizon and time units.

Interpretation, Diagnostics, and Alternatives

A confidence interval for \(\Delta(\tau)\), when reported, quantifies uncertainty in the contrast under the analysis assumptions. It is not a probability statement about the fixed parameter. A point estimate’s sign does not establish statistical significance, and a small contrast does not prove the curves are similar at all times: sizable positive and negative areas may have nearly canceled.

Before interpreting the result, inspect the estimated survival curves and check that each group has adequate support near \(\tau\). Heavy censoring late in follow-up can make the tail area unstable; informative censoring can bias the estimated curves. Also check that the event definition, time origin, and analysis population are consistent between groups. If the apparent crossing is driven by very few participants at risk, its late portion may be imprecise even when the point estimate is calculable.

RMST is an absolute-time summary and does not require a constant hazard ratio. A hazard-ratio summary or a logrank-type test addresses a different aspect of the comparison and can emphasize event-time differences differently. Neither a single hazard ratio nor a single RMST contrast fully describes a time-varying pattern. The RMST contrast’s distinctive limitation is its dependence on \(\tau\): two analyses with different horizons answer different questions. The introductory material on RMST as an estimand defines that target; here, the essential practical point is to connect its value to the signed area over the chosen window.

Common mistake: “The treatment curve is higher early, so the treatment RMST must be larger.” This ignores any later area where the treatment curve is lower. The correction is to integrate the signed difference over the full common horizon; under crossing effects, positive and negative areas both count.

Check Your Understanding

Use the definition and area interpretation to answer each question. Solutions follow.

If the treatment and control curves coincide until month 4 and treatment is higher afterward through month 10, what can be said about the sign of the RMST contrast through month 10?
In the crossing-curves example, why is the treatment-minus-control contrast negative even though group A’s curve is higher at month 2?
Suppose a delayed treatment effect begins after the selected horizon. What does the RMST contrast measure, and what does it not establish?
Why must the same truncation horizon be used for both groups in a direct RMST contrast?
Name one assumption or diagnostic concern relevant to estimating RMST from censored time-to-event data.

Solutions. (1) The contrast is positive if treatment is higher over a nonzero interval after month 4 and never lower through month 10, because the signed area is then positive. (2) Group A’s early positive area is outweighed by its later negative area; the net contrast through month 10 is \(-0.075\) months. (3) It measures the signed area only through the selected horizon, so it may be zero or small; it does not establish the absence of a later effect. (4) Each group’s area must cover the same time window for the subtraction to represent a direct contrast over a common estimand. (5) Kaplan–Meier estimation relies on independent censoring within groups; heavy censoring near the horizon is also a warning that the tail area may be unstable.