Why the Follow-Up Horizon Is Part of the Question
A causal question about an intervention is incomplete if it says only what outcome matters. “Does treatment reduce hospitalization?” could mean hospitalization within 30 days, within one year, or over whatever time happens to be available in the records. Those are different questions, even when the endpoint is named identically. The protocol must specify the start of follow-up and how long outcomes are sought for each endpoint.
The time origin, or time zero, is the point from which the protocol measures follow-up. It should be aligned with eligibility and assignment, as established in the earlier material on aligning eligibility, assignment, and time zero. A horizon is then a duration measured from that origin. The calendar date on which a study’s records close is not itself a participant’s horizon: it may leave some participants with less observation than the protocol planned.
A protocol may use a different horizon for each endpoint. For example, it could assess an acute adverse event through 30 days and a chronic outcome at one year. The endpoint and its horizon should be stated together: an outcome without a time frame leaves the target underspecified. The exact clinical or data-based definition of an endpoint is a separate protocol decision; here, the focus is when it is assessed and how long it is observed.
Defining the Planned Window
For an endpoint assessed at a fixed horizon \(\tau\), the target quantity may be expressed as a potential outcome \(Y^a(\tau)\), the value that would be observed at that horizon under strategy \(a\). For a time-to-event endpoint with event time \(T^a\), the corresponding horizon-specific event indicator is \(I(T^a \leq \tau)\). These expressions make the time frame explicit; they do not imply that every participant’s outcome will in fact be observed through \(\tau\).
The start of the window must match the protocol’s time zero. Measuring from treatment initiation for one group but from eligibility for another can give the groups different amounts of time at risk, even if both analyses label the follow-up “one year.” The protocol should also specify whether the boundary is inclusive—for example, whether an event on day 365 counts in a 365-day horizon—and apply that convention consistently.
A fixed horizon is not the same as a fixed calendar closing date. If the study ends data collection on a common date, participants who enter earlier can contribute more observation than those who enter later. To guarantee a full \(\tau\)-length window for every participant, the last time zero must occur no later than the closing date minus \(\tau\), allowing for the protocol’s date-counting convention.
Worked Example: Checking an Administrative Cutoff
Suppose a hypothetical protocol assesses a 365-day endpoint after each participant’s time zero. Data collection closes at the end of December 31, 2025. A participant enters on January 10, 2025. Does the administrative cutoff provide the full planned window?
The planned endpoint date is January 10, 2025 plus 365 days, which is January 10, 2026. The data cutoff precedes that date. In 2025, January 10 is day 10 and December 31 is day 365, so the participant has \(365-10=355\) days from time zero through the cutoff under an elapsed-day convention. The planned window is short by \(365-355=10\) days.
Thus, this participant does not have a complete observed 365-day window based on the stated data. The protocol’s horizon remains 365 days; the study cutoff does not turn it into a 355-day endpoint. To provide a full year of potential observation for all participants, the latest time zero would need to be at least 365 days before the cutoff.
Administrative Dates and Participant-Specific Events
An administrative end date applies at the study or data-source level. Examples include the end of record linkage, a prespecified database extract date, or the planned end of a trial. Participant-specific events can end usable observation earlier. Examples include withdrawal from data collection, loss of insurance coverage in a claims source, or emigration from a registry’s coverage area. These events do not necessarily mean that the endpoint occurred; they may only mean that later endpoint information is unavailable.
For planning purposes, let \(A\) denote the administrative closing date and \(L_i\) the last date on which participant \(i\)'s endpoint information is available because of a participant-specific end of observation. Under a fixed horizon, the latest available date is the earliest of the planned endpoint date, the administrative cutoff, and the participant-specific end date:
This calculation describes the available window, not how to analyze incomplete observations. If \(E_i\) occurs before \(t_{0i}+\tau\), the endpoint may not be known at the planned horizon. A protocol must not silently treat “no event recorded before observation ended” as “no event through the horizon.”
The meaning of a participant-specific event depends on the endpoint. For all-cause mortality, death is itself the outcome event, not a reason to regard that endpoint as unobserved. For a nonfatal endpoint, death may prevent the event from occurring later and may require a distinct protocol decision. Similarly, withdrawal can stop collection without being an endpoint. The protocol should state how such events relate to each endpoint rather than applying one generic rule to all outcomes.
Worked Example: Several Endpoints, Several Horizons
In a hypothetical study, a participant’s time zero is March 1, 2025. The protocol specifies a 30-day acute-event endpoint and a 180-day functional-status endpoint. Data are available through a participant-specific coverage end date of July 15, 2025; the overall administrative cutoff is December 31, 2025. Assume elapsed-day horizons and that the time-zero date is day 0.
For the acute endpoint, the planned date is March 1 plus 30 days, or March 31. Both the coverage end and administrative cutoff occur later, so the full 30-day window is available.
For functional status, the planned date is March 1 plus 180 days, or August 28, 2025. The administrative cutoff is later, but coverage ends July 15. From March 1 to July 15 is 136 elapsed days, so the observed window is 136 days, which is \(180-136=44\) days short of the planned horizon.
The two endpoints therefore have different observation situations for this participant. The acute endpoint has a complete planned window, while the 180-day status is not observed at its planned date in the available source. A single label such as “follow-up ended July 15” would conceal this endpoint-specific distinction.
Why Incomplete Follow-Up Can Change What Is Compared
A fixed-horizon estimand concerns outcomes at the same prespecified duration under each strategy. If an analysis instead includes only participants whose outcomes are observed at that duration, it may compare a selected subset rather than the protocol’s full target population. This is not automatically biased in every setting, but the interpretation changes unless the analysis appropriately accounts for who lacks observation and why.
A short justification makes the issue clear. Let \(R_i(\tau)=1\) indicate that participant \(i\)'s outcome is observed through horizon \(\tau\). The average among observed participants is \(E[Y^a(\tau)\mid R(\tau)=1]\), whereas the target average for the population is \(E[Y^a(\tau)]\). These are equal only under additional conditions, such as observation being unrelated to the potential outcome within the relevant group, or after a valid adjustment that recovers the target. If observation depends on prognosis or treatment, the observed subset can differ systematically.
For example, participants with worsening health might be more likely to leave a data source. Their observed outcomes could then differ from those of otherwise similar participants who remain observable. The appropriate response is not to redefine the horizon after seeing the available follow-up. Instead, the protocol keeps the target horizon explicit, documents the observation limits, and specifies how incomplete outcome information will be addressed.
Worked Example: Different Availability Does Not Mean Different Targets
Consider a hypothetical protocol with a 90-day endpoint and a common data cutoff of June 30. Participant A has time zero on April 1, while participant B has time zero on May 15. Assume elapsed-day counting.
For A, the planned endpoint date is April 1 plus 90 days, or June 30. A has a full 90-day window. For B, the planned date is May 15 plus 90 days, or August 13. The June 30 cutoff provides only 46 elapsed days after May 15, leaving 44 days before the planned horizon.
It would be incorrect to describe A’s endpoint as “90-day risk” and B’s as “46-day risk” and then treat them as measurements of the same outcome. The protocol has one target horizon, 90 days; B’s outcome at that horizon is not available from this data cutoff. Nor should the study label the overall endpoint “up to 90 days” without explaining how the actual observation window varies. That phrase could refer to quite different durations for different participants.
Choosing Among Follow-Up Designs
The horizon should follow the scientific question and the expected timing of the endpoint, not merely the longest period present in the database. A short horizon can miss delayed benefits or harms; a long one can increase the amount of incomplete observation and may make the result less directly relevant to a near-term decision. Protocols should justify each endpoint’s horizon and make clear whether it is a single assessment time or an event-counting period.
| Design choice | What it specifies | Main consideration |
|---|---|---|
| Common fixed horizon | Each participant is assessed at the same duration after time zero. | Answers a clearly comparable time-specific question, but later entrants may lack complete observation under a fixed data cutoff. |
| Different horizons by endpoint | Each endpoint has its own prespecified duration. | Allows the time frame to match the endpoint, but each endpoint needs an explicit window and availability check. |
| Common calendar closing date | All observation ends on a shared date, so duration varies by entry date. | Useful as an administrative rule, but does not by itself define a common-duration causal question. |
| Event-triggered follow-up | Observation ends at a prespecified event or transition rather than at a fixed duration alone. | Can suit questions naturally tied to that event, but requires a precise trigger and a clear account of what outcome is measured afterward. |
A fixed horizon is often easiest to interpret when the question is explicitly time-specific. A calendar-based end can still be used operationally, but the protocol must say how variable durations relate to the target question. Event-triggered follow-up can be appropriate—for example, assessing an outcome after discharge—but the trigger must be defined in a way that is comparable across participants and consistent with time zero. A trigger that occurs only after treatment begins also needs careful consideration, since it may create groups with different entry conditions.
Diagnostics and Common Mistakes
Before finalizing a protocol, make a participant-level window table or equivalent check using the planned enrollment range, endpoint horizons, administrative dates, and known observation-ending events. At minimum, calculate the amount of potential follow-up for the earliest and latest eligible time zeros. If the closing date leaves late entrants short of the target horizon, revise the enrollment period, data collection period, or stated target—before analyzing outcomes.
- Calling the last available date the horizon. This confuses the data limit with the protocol’s target duration. State both separately.
- Using different origins across strategies. If time zero is not aligned, nominally equal horizons may cover different periods at risk. Use the protocol’s common time origin.
- Treating absence of a recorded event as proof of no event. A participant may simply have left observation before the horizon. Record that as an availability limitation, not automatically as an event-free outcome.
- Using one horizon for every endpoint by habit. Acute and long-term endpoints may require different, scientifically justified windows.
- Assuming death, withdrawal, and database closure mean the same thing. Their role depends on the endpoint and data source; the protocol must classify each event appropriately.
The key comparison is between the planned horizon and the available window for each endpoint and participant. This design-stage check makes gaps visible without pretending that every gap can be resolved by changing the question after the data are collected. It also leaves the later analysis plan to specify how censoring and missing outcomes are handled.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- A protocol targets a 120-day endpoint from time zero. A participant enters on February 1, and the data cutoff is May 31 in a non-leap year. How many elapsed days are available, and is the horizon complete?
- Why is an outcome observed at a participant’s last available record date not necessarily the outcome at the protocol’s fixed horizon?
- For a 30-day endpoint and a 180-day endpoint, coverage ends 60 days after time zero. Which endpoint has a complete planned window?
- Give one reason a common calendar cutoff can create different observation durations for participants.
Solutions. (1) February 1 to May 31 is \(28+31+30=89\) elapsed days in a non-leap year, so the 120-day window is incomplete by \(120-89=31\) days. (2) The record date may precede the planned assessment, so it gives no automatic information about the outcome at the later horizon. (3) The 30-day endpoint has a complete window; the 180-day endpoint is observed for only 60 of its planned days. (4) Participants enter on different calendar dates, so those entering later have less time before the same closing date.