Follow-Up Duration and Time-to-Event Estimands

Why Follow-Up Duration Is Part of the Target

In a time-to-event trial, extending follow-up can do more than provide additional observations. It can also change the question the treatment comparison answers. A contrast in the probability of remaining event-free through one year is not the same target as a contrast through two years. Restricted mean survival time (RMST) likewise depends on its chosen time horizon. A hazard ratio has a different interpretation: it compares instantaneous event rates among participants still event-free, and its meaning as a single summary depends on how those rates relate over time.

The central design distinction is between collecting more follow-up to estimate the same target and changing the target by extending its time horizon. In an adaptive trial, that distinction should be made explicit before follow-up is extended. The target also depends on the event definition, population, and treatment conditions; those are held fixed here so we can isolate the role of duration.

Definition: Let \(T^a\) be the potential time to the specified event under treatment condition \(a\), where \(a=1\) denotes the treatment condition and \(a=0\) the control condition. The survival probability at time \(t\) is \(S_a(t)=P(T^a>t)\). A fixed-time survival contrast at horizon \(\tau\) is \(S_1(\tau)-S_0(\tau)\). Positive values favor treatment when remaining event-free is desirable.

The contrast is an absolute difference in probabilities, measured in percentage points when expressed as a percentage. It answers how much the probability of being event-free at the specified time differs between treatment conditions. The horizon \(\tau\) is therefore part of the estimand, not merely a data-collection detail. The same two treatment conditions can have one contrast at one year and a different contrast at two years.

Three Summaries, Three Interpretations

Definition: The restricted mean survival time through horizon \(\tau\) under condition \(a\) is \(\operatorname{RMST}_a(\tau)=E[\min(T^a,\tau)]\). It is the expected event-free time accumulated from time zero up to \(\tau\). The treatment contrast is \(\operatorname{RMST}_1(\tau)-\operatorname{RMST}_0(\tau)\), measured in time units.

RMST does not count event-free time beyond \(\tau\): anyone who remains event-free longer contributes \(\tau\). Its treatment contrast is therefore an average difference in event-free time accumulated over the chosen period, rather than a difference in the probability of being event-free at the endpoint.

Definition: If \(f_a(t)\) is the event-time density under condition \(a\), the hazard at time \(t\) is \(h_a(t)=f_a(t)/S_a(t)\), where defined. It is the instantaneous event rate at \(t\), conditional on being event-free just before \(t\). The hazard ratio is \(h_1(t)/h_0(t)\). A constant hazard ratio over time is the proportional-hazards condition.

A hazard is a rate, not a probability, and a hazard ratio is not a ratio of survival probabilities. Its risk sets consist of participants who have remained event-free up to each time, so the people being compared within those sets change over time. Under proportional hazards the ratio is constant, but the absolute survival difference can still vary with time. If hazards are not proportional, a single-number hazard ratio can depend on the follow-up period and on how the analysis summarizes time-varying rates.

Assumptions for Interpretation and Estimation

The three summaries can be defined for a specified population, event, and pair of treatment conditions. Estimating them from trial data requires that event times and censoring be handled appropriately. Censoring means that a participant’s event time is not fully observed by the analysis; it is not itself the event. For example, a participant may leave the study event-free before the planned end of follow-up.

Consistent outcome ascertainment: the event has the same operational definition and is recorded comparably across treatment conditions and follow-up periods.
A defined time origin and time scale: time zero is specified consistently, such as randomization, and all horizons use the same units.
Adequate observation through the horizon: there must be participants at risk and sufficient event follow-up to estimate the target. A nominal two-year horizon cannot be supported merely by following the first participants for two years if most of the target population has little observation near that time.
Appropriate censoring assumptions for estimation: common survival estimators require censoring to be independent of event time, either unconditionally or conditional on variables included in the analysis, as appropriate. There must also be a positive chance of remaining observed over the times being estimated.

For example, the Kaplan–Meier estimator can estimate a survival curve with right-censored observations under independent censoring. Its estimated survival probability at \(\tau\) can be used for a fixed-time contrast, and the area under the estimated curve through \(\tau\) can estimate RMST. If loss to follow-up is related to unobserved event risk, those estimates may be biased unless the dependence is addressed. Extending follow-up does not, by itself, remove informative censoring or make a poorly supported late-time estimate reliable.

Why the RMST Formula Is an Area

The connection between RMST and the survival curve follows directly from a representation of a nonnegative time. For any individual with event time \(T^a\), the quantity \(\min(T^a,\tau)\) is the length of time in the interval from zero to \(\tau\) for which that individual remains event-free. It can be written as the area under an indicator that equals one before the event and zero afterward.

\min(T^a,\tau)=\int_0^\tau \mathbf{1}(T^a>t)\,dt

Taking expectations and exchanging expectation with integration gives the survival-curve identity:

\operatorname{RMST}_a(\tau) =E[\min(T^a,\tau)] =\int_0^\tau E[\mathbf{1}(T^a>t)]\,dt =\int_0^\tau S_a(t)\,dt.

Thus RMST is the area under the survival curve up to \(\tau\), not the area over the entire lifetime unless the horizon actually covers that lifetime. This identity also explains why changing \(\tau\) changes the RMST estimand: it changes the interval over which event-free time is accumulated.

Worked Example: A One-Year Survival Contrast and Hazard Ratio

Suppose, purely hypothetically, event times follow exponential distributions: the control hazard is constant at \(0.10\) events per person-year and the treatment hazard is constant at \(0.08\) events per person-year. Under this model \(S_a(t)=e^{-\lambda_a t}\), where \(\lambda_a\) is the constant hazard. The constant-hazard and exponential assumptions are deliberately strong; they make the example’s calculations transparent, not universally appropriate.

At one year, the control survival probability is \(S_0(1)=e^{-0.10}=0.904837\), and the treatment survival probability is \(S_1(1)=e^{-0.08}=0.923116\). Therefore, the treatment-minus-control contrast is

S_1(1)-S_0(1)=e^{-0.08}-e^{-0.10} =0.923116\ldots-0.904837\ldots \approx 0.01828.

The estimated probabilities in this hypothetical model differ by about 1.83 percentage points at one year. The hazard ratio is \(0.08/0.10=0.80\), meaning the treatment hazard is 20% lower at each time under the assumed constant-hazard model. It does not mean that treatment raises one-year survival by 20%; the one-year absolute difference is about 1.83 percentage points.

Worked Example: Extending the Horizon Changes Survival and RMST Targets

Use the same hypothetical exponential model. Compare treatment and control at two years, rather than one year. At two years, \(S_1(2)=e^{-0.16}=0.852144\) and \(S_0(2)=e^{-0.20}=0.818731\). The treatment-minus-control contrast, calculated from the unrounded probabilities, is

e^{-0.16}-e^{-0.20} =0.852143789\ldots-0.818730753\ldots \approx 0.0334130.

The two-year contrast is about 3.34 percentage points, compared with about 1.83 percentage points at one year. This is not an arithmetic inconsistency: the survival probabilities are evaluated at different times, so the contrasts are different estimands.

Under a constant hazard \(\lambda\), the survival curve is \(e^{-\lambda t}\), and its area through \(\tau\) gives \(\operatorname{RMST}(\tau)=(1-e^{-\lambda\tau})/\lambda\). At two years, the control RMST is \((1-e^{-0.20})/0.10=1.81269\) years and the treatment RMST is \((1-e^{-0.16})/0.08=1.84820\) years. The treatment-minus-control RMST contrast is \(1.84820-1.81269=0.03551\) years, or about \(0.03551\times365=12.96\) days of additional event-free time accumulated within the two-year window, on average.

The fixed-time contrast describes the probability of being event-free at year two; the RMST contrast describes average event-free time accumulated over the first two years. Neither is interchangeable with the hazard ratio of 0.80. The exponential model implies the same hazard ratio at both horizons, but it does not imply the same survival or RMST contrast.

When Effects Change Over Time

The constant-hazard example cannot show what happens when the relative treatment effect changes over time. Under nonproportional hazards, the hazard ratio may vary, survival curves can approach or cross, and a single hazard ratio may conceal clinically relevant differences at particular times. In that setting, the time horizon is especially consequential.

Worked Example: A Time-Varying Hazard Effect

Consider a hypothetical piecewise-exponential model. The control hazard is \(0.10\) per year throughout. The treatment hazard is \(0.05\) per year in year one and \(0.20\) per year in year two. Hazards are constant within each stated interval, and survival is continuous at the one-year boundary.

At one year, \(S_0(1)=e^{-0.10}=0.904837\), while \(S_1(1)=e^{-0.05}=0.951229\). The treatment-minus-control survival contrast is \(0.951229\ldots-0.904837\ldots\approx0.04639\), favoring treatment by about 4.64 percentage points at that time. During year one the hazard ratio is \(0.05/0.10=0.50\).

By year two, the control survival probability is \(e^{-0.10(2)}=e^{-0.20}=0.818731\). Treatment survival incorporates both intervals: \(S_1(2)=e^{-0.05(1)-0.20(1)}=e^{-0.25}=0.778801\). Thus the two-year treatment-minus-control contrast is \(0.778801\ldots-0.818731\ldots\approx-0.03993\), favoring control at two years. The hazard ratio in year two is \(0.20/0.10=2.00\), the opposite direction from year one.

For completeness, treatment RMST through two years is the sum of the area in year one and the area in year two:

\operatorname{RMST}_1(2) =\frac{1-e^{-0.05}}{0.05} +e^{-0.05}\frac{1-e^{-0.20}}{0.20} \approx 0.975412+0.862143 =1.83755\text{ years}.

Control RMST through two years is \((1-e^{-0.20})/0.10=1.81269\) years, so the treatment-minus-control RMST contrast is about \(0.02486\) years, or \(9.07\) days. This positive accumulated-time contrast can coexist with a negative two-year survival contrast: the latter focuses on event-free status at the endpoint, while RMST accounts for the full survival experience up to it. The hazard ratio is not constant here, so reporting a single hazard ratio without clarifying how it summarizes the time-varying effect would be inadequate.

Planning an Extension Without Silently Changing the Target

Suppose a protocol initially specifies survival at 18 months and later proposes collecting data through 24 months. If the primary estimand remains the 18-month contrast, the added observation period can provide more complete observation for that same target, especially for participants who otherwise would be censored before 18 months. The estimand stays fixed only if the population, treatment conditions, event definition, time origin, and 18-month horizon stay fixed.

If the revised primary comparison instead uses survival at 24 months, it targets a different probability contrast. If it uses RMST through 24 months rather than through 18 months, it targets a different accumulated-time contrast. Those may be sensible scientific choices, but they should be described as changes to the target, not as merely collecting additional data for the original target. A hazard-based target also needs careful specification: a constant proportional-hazards contrast has a distinct interpretation from a time-varying hazard comparison or a single fitted summary over a stated observation period.

Summary	Question answered	What changing the horizon can change
Survival contrast at \(\tau\)	How different are the probabilities of being event-free at this time?	The endpoint time, and therefore the target probability difference
RMST contrast through \(\tau\)	How different is average event-free time accumulated up to this time?	The interval over which event-free time is counted
Hazard ratio	How do instantaneous event rates compare among those still event-free?	Under nonproportional hazards, the single-number summary can depend on the period and risk sets represented

These summaries are alternatives when the scientific question differs, not interchangeable ways to report the same quantity. Fixed-time survival probabilities are directly interpretable at a clinically chosen time, but discard differences in event timing away from that point. RMST uses the curve over the full interval to the horizon and has time units, but still depends on the horizon and does not summarize events after it. A hazard ratio can be useful when proportional hazards is plausible and a relative rate comparison is relevant, but it is less direct as an absolute clinical effect and can obscure time variation.

Diagnostics and Common Pitfalls

Late estimates with little information: inspect numbers at risk, censoring, and uncertainty near the selected horizon. A smooth-looking curve does not make a tail estimate precise.
Unequal or informative censoring: compare follow-up patterns across treatment conditions and consider whether censoring can plausibly be treated as independent, possibly conditional on measured covariates. Longer planned follow-up does not solve bias from informative loss to follow-up.
Nonproportional hazards: inspect estimated survival curves and time-specific hazard patterns, and assess whether a constant hazard ratio is plausible. Crossing curves or clearly changing effects argue against an unqualified constant-ratio interpretation.
Changed event ascertainment: if assessment schedules or event definitions change during an extension, apparent differences may reflect measurement changes as well as event experience. Keep ascertainment consistent or account for the change explicitly.

Common mistake: “We collected more follow-up, so the estimand is unchanged.” This is true only when the target horizon and the other estimand attributes remain the same. If a one-year survival contrast is replaced by a two-year contrast, the treatment comparison now concerns event-free status at two years. The correction is to state the target horizon explicitly and distinguish additional observation for estimation from a revised time horizon for the clinical question.

Key takeaway: State the time horizon alongside every fixed-time survival or RMST target. More follow-up can improve estimation of a fixed target, but changing the horizon changes the probability being compared or the period over which event-free time is accumulated. Interpret hazard-based summaries separately, with attention to whether hazards are proportional.

Check Your Understanding

Answer each question using the distinction between the time horizon and the treatment-effect summary.

A trial changes its primary survival comparison from 12 to 18 months but keeps the same population, event, and treatment conditions. Has the primary estimand changed? Why?
In one sentence, distinguish a survival-probability difference at 18 months from an RMST difference through 18 months.
A hazard ratio is constant at 0.75 under a proportional-hazards model. Does that mean survival is 25 percentage points higher at every time? Explain.
Why might extending follow-up make an estimate at a fixed horizon more precise without changing its target?
Survival curves cross during follow-up. What does this imply for reporting a single hazard ratio, and name one alternative summary from this tutorial.

Solutions. (1) Yes. The fixed-time survival contrast now compares probabilities at 18 months rather than 12 months. (2) The survival difference compares probabilities of being event-free exactly at 18 months, whereas the RMST difference compares average event-free time accumulated from time zero through 18 months. (3) No. A hazard ratio of 0.75 is a relative instantaneous-rate comparison under the model, not a 25-percentage-point survival difference; the absolute survival difference can vary with time. (4) More participants may have observed event status through that same horizon rather than being censored earlier, improving information while the population, event, treatment conditions, and horizon stay fixed. (5) Crossing curves suggest time-varying effects, making a single hazard ratio potentially misleading; a fixed-time survival contrast or RMST contrast through a prespecified horizon is an alternative.