Why the Truncation Horizon Matters
Restricted mean survival time (RMST) summarizes average event-free time up to a specified time point. The truncation horizon, denoted by \(\tau\), is therefore part of the quantity being estimated, not merely a technical setting in a calculation. A horizon of 2 years asks about average event-free time accumulated during the first 2 years; a horizon of 5 years includes another 3 years of the survival experience.
The introductory material on RMST as an estimand defines the group-specific quantity. Here the central task is choosing \(\tau\). A defensible choice should have a clear clinical interpretation, lie within the follow-up support available in both randomized groups, and be specified without using the observed treatment contrast to select a favorable answer.
The horizon changes the question the contrast answers. Two studies, or two analyses of the same study, that use different horizons estimate different restricted means and potentially different treatment contrasts. There is no universally optimal \(\tau\): its choice depends on the clinical question and on whether the data can support estimating the survival curves through that time.
Three Requirements for a Defensible Horizon
First, tie \(\tau\) to a clinically interpretable period. It might correspond to a planned assessment time, a period relevant to treatment decisions, or a time by which a clinically important difference in event-free time could reasonably arise. State the horizon and its units explicitly. A horizon chosen only because it produces a larger estimated contrast does not have a sound clinical rationale.
Second, check follow-up support in each group. The Kaplan–Meier estimator can be calculated beyond the last observed event, but that does not mean the tail is informative. If very few participants remain under observation, or censoring has left little information about survival near \(\tau\), the estimated area there may be unstable. The maximum recorded follow-up time alone is not adequate evidence of support. Examine numbers at risk and censoring patterns over time, separately by group, and consider whether both groups provide useful information through the proposed horizon.
For the usual Kaplan–Meier-based estimate, event times and censoring must be recorded appropriately, and censoring must be independent of the event time within groups, or independent conditional on variables handled by the analysis. Randomization supports a treatment-group comparison but does not by itself guarantee independent censoring. A horizon cannot fix informative censoring or create follow-up information that was not collected.
Third, prespecify the horizon in the protocol or analysis plan, preferably before treatment-group outcomes are examined. This reduces the risk of choosing \(\tau\) after seeing which value gives the most favorable contrast or smallest p-value. If clinical or operational considerations require more than one horizon, specify their roles in advance—for example, one primary horizon and a limited set of sensitivity horizons—and interpret the additional results as such.
Worked Example: How the Contrast Accumulates With Time
Suppose a hypothetical randomized trial has estimated survival curves that are constant over the following intervals. Time is measured in years. For control, \(S_{\mathrm{ctrl}}(t)=1\) from 0 to 2, \(0.80\) from 2 to 4, and \(0.50\) from 4 onward. For treatment, the corresponding values are \(1\), \(0.90\), and \(0.70\). Assume, for this illustration, that follow-up supports both proposed horizons of 3 and 5 years.
At 3 years, the control RMST is the area from 0 to 3:
For treatment, the area is:
At 5 years, the additional interval from 4 to 5 contributes \(0.50\) years to control and \(0.70\) years to treatment. Thus:
The contrast is larger at 5 years because the treatment curve contributes more area than the control curve during years 2 to 5. This is not a contradiction: \(\hat{\Delta}(3)\) and \(\hat{\Delta}(5)\) summarize different time windows. The example illustrates why \(\tau\) must be chosen for the clinical question, not selected afterward because one estimate appears more favorable.
A Short Justification for Horizon Sensitivity
The reason the contrast can change with \(\tau\) follows directly from the area definition. If \(\tau_2>\tau_1\), the added area in each group is the integral of its survival curve from \(\tau_1\) to \(\tau_2\). Subtracting the two groups’ added areas gives:
This identity says that the change in the RMST contrast is determined by the difference between the survival curves over the added time interval. If treatment survival exceeds control survival over that interval, extending the horizon increases the contrast; if it is lower, the contrast decreases. If the curves are close, the added interval contributes little. The identity is about the estimand and applies whether the curves are known or estimated.
For estimated Kaplan–Meier curves, the corresponding calculation is the difference in estimated areas. Extending \(\tau\) does not automatically improve the analysis: it adds clinically relevant time only if that time is meaningful and adequately supported. It can also add an unstable tail contribution when few participants remain under observation.
Worked Example: Checking Follow-Up Support Before Choosing Five Years
Consider a hypothetical trial with a proposed 5-year horizon. The numbers at risk immediately before years 3, 4, and 5 are as follows:
| Time (years) | Control at risk | Treatment at risk |
|---|---|---|
| 3 | 120 | 118 |
| 4 | 45 | 42 |
| 5 | 8 | 6 |
Both groups have some participants under observation at 5 years, so the horizon is not beyond the observed follow-up. But only 8 control participants and 6 treatment participants remain at risk immediately before that time. A 5-year RMST estimate is therefore likely to depend on a sparsely supported tail. The fact that the largest observed time reaches 5 years is not, by itself, a convincing reason to choose \(\tau=5\).
Suppose the trial’s clinically relevant primary assessment is at 4 years, and the analysis plan was written before outcomes were reviewed. The risk sets at year 4 are larger than at year 5 in both groups, and the 4-year horizon directly answers the planned clinical question. On these facts, 4 years is a more defensible primary choice. The 5-year estimate could be reported as a prespecified sensitivity analysis if its uncertainty and limited tail support are made clear. The risk-set counts do not prove a particular horizon is correct; they inform whether a clinically motivated horizon is adequately supported.
Prespecification, Sensitivity Analyses, and Alternatives
Prespecification does not mean that investigators must ignore relevant information about follow-up. It means that the clinical rationale and the rule for selecting the analysis are set out before treatment effects are inspected. A protocol can identify a primary \(\tau\), explain why it matters clinically, and describe how follow-up support will be reviewed. If follow-up ends earlier than planned or a serious data-quality problem arises, any change should be transparently justified and distinguished from the original plan.
A limited sensitivity analysis across clinically plausible horizons can show whether the substantive interpretation depends on the time window. Report the horizons and estimates together, with units, and avoid presenting the most favorable result as if it had been the sole planned analysis. A sensitivity analysis is informative about the choice of time window; it does not erase the consequences of inspecting several outcome-dependent choices.
The main alternatives are not interchangeable. A single prespecified horizon gives a focused estimand and interpretation. Several prespecified horizons assess how conclusions vary over time, but require careful reporting and may raise multiplicity concerns if each is treated as a separate confirmatory test. Choosing a horizon after examining the estimated treatment contrasts is data-dependent and can bias the reported evidence in a favorable direction. Conversely, using the latest observed time simply because it is available may make the result depend heavily on sparse, unstable tails. Neither convenience nor a larger observed effect is an adequate substitute for clinical rationale and support.
A useful diagnostic is to inspect estimated survival curves, numbers at risk, and censoring patterns through the candidate horizon in each group. Look for an abrupt loss of follow-up, a markedly thin tail, or a horizon that extends beyond the period relevant to the decision. Also verify that the horizon and time units are consistent across the protocol, analysis, tables, and interpretation. These checks identify concerns; they do not establish independent censoring, and they cannot rescue a horizon unsupported by the study design.
Worked Example: Interpreting Two Prespecified Horizons
In a hypothetical analysis, investigators prespecified 3 years as the primary horizon because it matches the planned clinical decision point, and 5 years as a sensitivity horizon. The estimated treatment-minus-control contrasts and standard errors are:
| Horizon | Estimated contrast | Standard error |
|---|---|---|
| 3 years | 0.10 years | 0.18 years |
| 5 years | 0.40 years | 0.31 years |
Using a normal-approximation 95% interval, the 3-year result is:
For 5 years, the corresponding interval is:
The point estimate at 5 years is larger, but its standard error is also larger, and both intervals include zero. These estimates do not justify replacing the prespecified 3-year primary analysis with the 5-year result. The correct report identifies the 3-year estimate as primary and describes the 5-year estimate as a sensitivity result: at 5 years, the estimated difference is 0.40 additional years of event-free time for treatment, but it is less precise. A confidence interval describes uncertainty under the analysis assumptions; it is not the probability that the true contrast lies inside the particular interval.
Common Pitfall and Limits
Wrong reasoning: “We calculated RMST at several horizons and the treatment contrast was largest at 5 years, so 5 years is the best horizon.” This selects the estimand based on the observed result and can make the reported treatment effect look more favorable than a prespecified analysis would.
Correction: Choose the primary horizon based on clinical meaning and credible follow-up support, before examining treatment-group contrasts. If several horizons are reported, explain their prespecified roles and show the estimates transparently. A larger contrast at a longer horizon may reflect real differences in the survival curves over the added period, but it does not alone establish that the longer horizon is the right primary question.
Horizon choice cannot remove bias from informative censoring, correct poor endpoint measurement, or compensate for inadequate follow-up. Nor does one horizon summarize every clinically relevant feature of the survival curves. It defines a particular time-restricted summary; the interpretation should remain tied to that period and to the assumptions behind estimation.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In your own words, what question does a 4-year RMST contrast answer, and what are its units?
- Why is the maximum observed follow-up time alone insufficient to establish support for a proposed horizon?
- If the treatment survival curve is above the control curve from year 2 through year 4, what does the area identity imply about extending the horizon from 2 to 4 years?
- Why is selecting the horizon that gives the largest observed treatment contrast problematic?
- A clinically important decision occurs at 3 years, but follow-up is sparse after year 4. What would make a 3-year primary horizon more defensible than a 5-year horizon?
Solutions. (1) It compares the groups’ average event-free time accumulated through 4 years; the contrast is measured in years. (2) A few participants may have very long observed follow-up while risk sets near the proposed horizon are very small, making the estimated tail unstable. (3) The integral of \(S_{\mathrm{trt}}(t)-S_{\mathrm{ctrl}}(t)\) over years 2 to 4 is positive, so extending from 2 to 4 years increases the treatment-minus-control contrast. (4) It makes the estimand depend on observed outcomes and can favor an exaggerated or selectively reported result. (5) The 3-year horizon directly matches the clinical decision, was prespecified, and has better follow-up support in both groups; the 5-year tail is sparse. These considerations support the choice without proving that the estimates are free of other biases.