03 · What You Need to Know
How to Determine and Justify Follow-Up Duration in Prognostic Research
What Does Follow-Up Duration Mean?
Follow-up duration refers to the period during which participants are observed for outcomes after a defined starting point.
In prognostic research, this period is meaningful only when the starting point and outcome are specified.
For example, a study may investigate mortality during the first year after diagnosis, recurrence within five years after treatment, or academic withdrawal during the first two semesters after enrollment.
The starting point determines when the clock begins, while the follow-up horizon determines how far into the future the outcome is evaluated.
These decisions should be coordinated rather than made independently.
The importance of defining time zero in prognostic research is therefore directly connected to selecting follow-up duration.
What Is the Difference Between Follow-Up Duration and Prediction Horizon?
The prediction horizon is the future period over which a prognostic estimate is intended to apply.
For example, a model may estimate the probability of an event within 12 months after diagnosis.
Observed follow-up is the actual amount of information available for each participant.
Some participants may be observed for the entire 12 months. Others may experience the event earlier, withdraw, die from a competing cause, or reach the administrative end of data collection before 12 months.
Consequently, a study can have a 12-month prognostic horizon without every participant contributing 12 months of event-free observation.
The analysis must account for these differences according to the outcome and censoring structure.
Why Should the Follow-Up Horizon Match the Intended Decision?
Prognostic estimates are most useful when their time horizon corresponds to a decision that researchers, clinicians, or other stakeholders need to make.
For example, a clinician deciding whether to intensify monitoring during the first month after discharge may need short-term risk estimates.
A patient considering long-term treatment may be more interested in recurrence over several years.
In education, early withdrawal risk may be relevant to first-semester support, while degree completion requires a longer horizon.
A technically precise estimate at an irrelevant horizon may have limited practical value.
Researchers should therefore justify the duration by explaining what the estimated outcome probability will be used to understand or decide.
How Does the Natural History of the Outcome Affect Follow-Up?
Different outcomes occur over different timescales.
Acute complications may develop within hours or days. Functional recovery may take months. Some recurrences and chronic complications may emerge years later.
A follow-up period that ends before most relevant events can occur will provide limited information about their longer-term probability.
However, a short horizon may still be entirely appropriate if the objective concerns early events.
Researchers should examine prior evidence about outcome timing, the underlying disease or process, and the intended decision.
Where event timing is uncertain, pilot data or existing cohorts may help determine whether the proposed observation period is realistic.
How Does Outcome Frequency Affect the Required Duration?
When the outcome is uncommon, a longer observation period may allow more events to accrue, potentially improving the precision of certain estimates.
However, extending follow-up does not necessarily solve an inadequate sample size.
If the research question concerns six-month prognosis, observing additional events after six months does not increase the number of events occurring within the original six-month horizon.
Likewise, if the outcome becomes less frequent over time, extending observation may produce relatively few additional events.
Sample size planning should therefore consider expected event probabilities or rates at the intended horizon, rather than assuming that longer follow-up will always compensate for limited recruitment.
The number of events needed also depends on whether the objective is to estimate overall prognosis, examine prognostic factors, or develop a prediction model.
Researchers should evaluate outcome-event requirements for prognostic studies alongside the proposed follow-up period.
Should Every Participant Be Followed for Exactly the Same Duration?
Not necessarily.
In some studies, participants are recruited over time and followed until a common administrative study end date. Participants recruited later consequently have less potential follow-up.
In other studies, each participant is scheduled for a fixed duration after their own entry date.
Both approaches can be appropriate.
For time-to-event outcomes, survival methods may accommodate different observation durations under appropriate censoring assumptions.
However, if the study seeks a fixed-horizon risk estimate, it must have sufficient information to estimate that horizon reliably.
Researchers should not assume that unequal follow-up is inherently invalid or that statistical methods automatically overcome inadequate observation.
What Is Administrative Censoring?
Administrative censoring occurs when follow-up ends because of a predetermined study end date or another planned observation limit rather than because the participant experienced the event.
For example, a participant enrolled late in a registry may have only eight months of observation before the database closes, even though the study is interested in one-year outcomes.
Time-to-event methods can accommodate right-censoring under appropriate assumptions.
However, the resulting estimates depend on the availability of information across the timescale and on the censoring mechanism.
Administrative censoring may be relatively straightforward when it is unrelated to future event risk conditional on relevant factors, but calendar-time changes and recruitment patterns can complicate this assumption.
Researchers should document how censoring occurs and whether the data support the intended horizon.
How Does Loss to Follow-Up Differ From Administrative Censoring?
Loss to follow-up occurs when outcome information becomes unavailable before the planned observation period ends.
Participants may withdraw, relocate, transfer care, or become unreachable.
Unlike a predetermined administrative endpoint, loss to follow-up may be related to health status, risk factors, or the outcome itself.
For example, participants experiencing deterioration may be more likely to leave a study or seek care elsewhere.
If this relationship is not adequately addressed, the observed outcomes among retained participants may not represent the intended population.
Researchers should record reasons for loss, compare relevant baseline characteristics, and consider sensitivity analyses or suitable missing-data methods.
Merely reporting the percentage lost does not establish that censoring is noninformative.
What Does Noninformative Censoring Mean?
Many standard survival methods rely on assumptions concerning the relationship between censoring and the event process.
Informally, noninformative censoring means that the censoring process does not provide additional information about future event risk beyond what is appropriately accounted for in the analysis.
This assumption is not guaranteed by the use of Kaplan-Meier curves or Cox regression.
For example, if participants at high risk of deterioration systematically disappear from observation, ordinary survival estimates may be biased.
Methods such as inverse-probability-of-censoring weighting may be considered under suitable assumptions, including adequate measurement of variables related to censoring and outcome risk.
Researchers should distinguish the statistical capacity to handle censoring from evidence that the required assumptions are plausible.
How Do Competing Events Affect Follow-Up Duration?
A competing event prevents the outcome of interest from occurring in the form being studied.
For example, death may prevent subsequent nonfatal recurrence, and academic completion may prevent later withdrawal from the same program.
As follow-up extends, competing events may become increasingly important.
When the objective is to estimate the real-world cumulative probability of an outcome, treating competing events as ordinary noninformative censoring can produce misleading estimates.
Competing-risk methods, including cumulative incidence functions, may be appropriate depending on the estimand.
Researchers should identify the relevant event structure before selecting the horizon and analysis.
More detailed decisions about competing events in prognostic study design depend on whether the interest is absolute event probability, cause-specific hazards, or another outcome quantity.
Does Longer Follow-Up Always Produce Stronger Evidence?
No. Longer observation may capture later outcomes, but it can also introduce additional complexities.
Participants may receive new treatments, experience changes in exposure, or move into different risk states.
Outcome measurement procedures may change, and attrition may increase.
These developments do not automatically invalidate long-term prognosis research. They may be part of the natural course the study intends to describe.
However, the interpretation must reflect the evolving conditions under which outcomes occur.
Researchers should avoid treating longer follow-up as inherently superior when it changes the scientific question or reduces the credibility of outcome ascertainment.
How Should Treatment Changes During Follow-Up Be Handled?
Prognosis is often observed under real-world treatment patterns rather than in the absence of treatment.
Patients may begin, discontinue, or change treatments during follow-up, and those changes may influence outcomes.
A study estimating prognosis under usual care may legitimately include these changes as part of the observed care pathway.
However, researchers should not interpret the resulting prognosis as the natural untreated course of disease unless the design supports that interpretation.
When treatment changes are central to the research question, investigators may need to define treatment strategies and consider causal methods rather than ordinary prognostic analysis alone.
Similarly, a prediction model intended for use at baseline should be interpreted in relation to the treatment context represented in its development and validation data.
How Frequently Should Outcomes Be Measured?
Follow-up duration and assessment frequency are separate decisions.
A study may follow participants for one year while collecting outcomes continuously through administrative records, at monthly visits, or at one final assessment.
The appropriate schedule depends on the outcome.
For precisely dated events such as death or hospitalization, reliable records may establish event times without frequent participant visits.
For fluctuating symptoms or functional outcomes, repeated assessments may be necessary to characterize change adequately.
If events are known only to have occurred between assessments, the data may be interval-censored rather than precisely timed.
Researchers should choose measurement intervals that support the intended outcome definition and statistical analysis.
What Is Median Follow-Up, and How Should It Be Interpreted?
Median follow-up is a summary of the observation time available in a study, but its calculation requires care when participants experience events or censoring.
Simply taking the median of observed times to event or censoring may not provide a satisfactory description of potential follow-up.
The reverse Kaplan-Meier method is commonly used to summarize follow-up in time-to-event studies by reversing the roles of events and censoring for the follow-up distribution.
Researchers should specify how median follow-up was calculated and avoid treating it as evidence that every participant was observed for that duration.
For fixed-horizon prognostic questions, the proportion with adequate outcome ascertainment at the target horizon may be more directly informative.
How Can Researchers Assess Whether Follow-Up Is Adequate?
Adequacy should be judged according to the estimand, not merely the total number of years in the dataset.
For a fixed-time outcome, researchers should examine how many participants have known outcome status at the specified horizon and whether missingness can be addressed credibly.
For time-to-event analyses, they should inspect numbers at risk, censoring patterns, event counts, and uncertainty across follow-up time.
Estimates near the end of follow-up may be unstable when few participants remain at risk.
Confidence intervals and risk tables can help readers recognize when apparent long-term precision is limited.
Researchers should avoid emphasizing late estimates supported by very small risk sets without appropriate qualification.
Should Prognostic Studies Include More Than One Horizon?
Multiple horizons may be appropriate when short-term and long-term outcomes support different decisions.
For example, 30-day and one-year mortality may describe distinct aspects of prognosis. First-semester withdrawal and four-year completion likewise concern different educational processes.
However, adding many horizons without a clear rationale can increase analytical complexity and encourage selective emphasis on favorable findings.
Researchers should prespecify the primary horizon and identify secondary horizons that have substantive relevance.
Each horizon should have adequate outcome information and an appropriate analysis.
Results should not be presented as interchangeable merely because they concern the same underlying outcome.
What If the Available Dataset Has Shorter Follow-Up Than the Intended Horizon?
Researchers should first determine whether the intended outcome probability can be estimated reliably from the available observations.
For example, a dataset containing only six months of follow-up cannot directly establish five-year prognosis without additional modeling assumptions or external information.
Extrapolation beyond observed follow-up may be possible in certain model-based settings, but it introduces assumptions that cannot be verified solely from the observed later outcomes.
Researchers should distinguish observed estimates from extrapolated predictions and report uncertainty appropriately.
If the data cannot support the intended horizon, narrowing the research question may be more defensible than presenting unsupported long-term estimates.
How Should Follow-Up Be Reported?
Researchers should report the intended outcome horizon, recruitment period, administrative study end date, and the actual observation available for participants.
For time-to-event studies, reporting may include numbers at risk, event counts, censoring, and an appropriately calculated summary of follow-up.
For fixed-horizon outcomes, researchers should report completeness of outcome ascertainment at the relevant time.
Reasons for missing outcomes and the methods used to address them should be described.
STROBE provides general reporting guidance for observational cohorts, while prediction-model studies should also follow applicable TRIPOD guidance.
Transparent reporting allows readers to judge whether the available follow-up supports the stated prognostic conclusions.
