Administrative End of Follow-Up and Loss to Follow-Up

Why the End of Observation Needs a Precise Definition

A follow-up period can end because the study database closes, or because a particular participant can no longer be observed. These are different events. The first is a common administrative end on the calendar; the second is loss to follow-up, which occurs at a participant-specific time. Both can prevent outcome ascertainment, but they imply different observation windows and may require different assumptions for analysis.

The distinction matters when emulating a target trial with a fixed outcome horizon. Someone who enters shortly before a database cutoff may have only a short opportunity for observation, even if they remain in the health system. Someone who leaves the observable system much earlier has a different reason for incomplete follow-up. Treating both as if everyone had the same complete observation period can misclassify unknown outcomes as event-free outcomes.

Definition: Let $E_i$ be participant $i$'s calendar date of time zero, $A$ the common calendar date when administrative data collection ends, $L_i$ the elapsed time from time zero to participant-specific loss to follow-up (infinite if none occurs), and $T_i$ the elapsed time to the outcome event. The maximum observable follow-up, before applying a target horizon $\tau$, is $$ K_i=\min(A-E_i,L_i). $$ For a horizon $\tau$, the observation window ends at $C_i=\min(K_i,\tau)$. With the convention that an outcome event recorded exactly at the censoring time is observed, the observed follow-up duration is $\widetilde T_i=\min(T_i,C_i)$, and the event indicator is $\Delta_i=I(T_i\le C_i)$.

Here $A-E_i$ is the time available before the common administrative end, expressed as elapsed time since participant $i$'s time zero. It varies with entry date, even though $A$ is shared. By contrast, $L_i$ varies because participants may stop contributing observable information at different times. In practice, the protocol may impose another planned end, such as five years after time zero; that limit is represented by $\tau$ in $C_i$.

A fixed-horizon outcome can be written $Y_i(\tau)=I(T_i\le\tau)$. If an event is observed by $\tau$, this outcome is known even if observation later ends. If no event is observed, the outcome is known to be zero only when the participant was observed event-free through $\tau$. If observation ends earlier, the fixed-horizon outcome is unknown, not automatically zero.

Calculating the Observation Window

Worked Example: One Calendar Cutoff, Different Follow-Up Durations

Suppose a database closes on December 31, 2024. Participant A has time zero on January 1, 2022, and remains observable. Participant B has time zero on July 1, 2023, and remains observable. Participant C has time zero on January 1, 2022, but is lost to follow-up on January 1, 2023. Find each participant’s maximum observable follow-up before applying a target horizon.

For A, the calendar interval from January 1, 2022, to December 31, 2024, is just under three years. Using elapsed days, the interval is 1,095 days, so $A-E_A=1{,}095$ days. A has no earlier loss, and therefore $K_A=1{,}095$ days.

For B, the interval from July 1, 2023, to December 31, 2024, is 549 days. Thus $A-E_B=549$ days and, with no earlier loss, $K_B=549$ days. The administrative date is the same as for A, but B has a shorter observation window because B entered later.

For C, the administrative interval is again 1,095 days, but $L_C=365$ days. Therefore $K_C=\min(1{,}095,365)=365$ days. C’s observation ends because of participant-specific loss, well before the database closes.

For a two-year horizon of 730 days, A can be observed through the horizon, B cannot, and C cannot. A no-event record for B or C does not establish that the participant remained event-free through two years.

The dates and event times must use the study’s actual time convention. If follow-up is measured in days, define how partial days and calendar boundaries are handled. If an outcome is ascertained only at scheduled visits, the last date with valid outcome information—not merely the administrative database date—may determine the observable period.

Worked Example: Is the Fixed-Horizon Outcome Known?

Consider a one-year outcome horizon. Three participants have the following records: A has $K_A=300$ days and an event at day 120; B has $K_B=300$ days and no recorded event; C has $K_C=365$ days and no recorded event. Determine $Y_i(365)$ where possible.

For A, the event occurs by day 365 and is observed before follow-up ends. Hence $Y_A(365)=1$. The fact that A was subsequently lost does not erase an already ascertained event.

For B, no event was recorded before day 300, but observation stopped before day 365. B may have had an event during days 301 through 365. Therefore $Y_B(365)$ is unknown, not zero.

For C, the full 365-day period was observed without an event. Hence $Y_C(365)=0$. This example shows why a no-event record and a known event-free outcome at the target horizon are not interchangeable.

What Can Be Estimated When Follow-Up Is Incomplete?

A target-trial protocol specifies the outcome horizon and the population whose risk is of interest. If every participant can be observed through $\tau$, the empirical proportion with $Y_i(\tau)=1$ directly estimates the horizon-specific risk, subject to the usual sampling and causal-identification requirements. When some outcomes are unknown because observation ended early, simply treating those participants as event-free changes the estimator and can bias the result.

One simple option is to restrict the analysis to participants with complete observation through $\tau$. This gives a direct estimate for that complete-follow-up subset. It estimates the target population’s risk only under additional conditions—for example, if completion is independent of the fixed-horizon outcome, possibly conditional on measured baseline covariates, and the target population is represented among completers. If completion depends on prognosis, that restriction can select a different-risk group.

A common alternative is inverse-probability-of-censoring weighting. The following result gives a precise version for complete-horizon outcomes. Let $X$ denote measured baseline covariates, $O=I(K\ge\tau)$ indicate complete observable follow-up, and $G(X)=P(O=1\mid X)$. Suppose consistency holds for the observed fixed-horizon outcome, $Y(\tau)$ is observed whenever $O=1$, and $Y(\tau)$ is conditionally independent of $O$ given $X$. Also require positivity: $G(X)>0$ for covariate values in the target population. Then weighting complete cases by $1/G(X)$ recovers the marginal risk in expectation.

Result: Under the assumptions above, $$ E\left[\frac{O\,Y(\tau)}{G(X)}\right]=E[Y(\tau)]. $$ For a sample, the corresponding ratio estimator is $$ \widehat R(\tau)= \frac{\sum_{i=1}^n O_iY_i(\tau)/\widehat G(X_i)} {\sum_{i=1}^n O_i/\widehat G(X_i)}. $$ The estimated denominator normalizes the weights; consistent estimation requires an appropriate model for $G(X)$ and the relevant regularity conditions.

To prove the expectation identity, first condition on $X$. Conditional independence gives $E[O\,Y(\tau)\mid X]=E[O\mid X]E[Y(\tau)\mid X]$. Since $E[O\mid X]=G(X)$, dividing by the positive $G(X)$ yields $$ E\left[\left.\frac{O\,Y(\tau)}{G(X)}\right|X\right]=E[Y(\tau)\mid X]. $$ Taking expectations over $X$ on both sides gives the stated result by iterated expectation. The argument requires that $Y(\tau)$ be well-defined for the target population and that the conditional independence and positivity assumptions hold; weighting does not make those assumptions true. $\square$

Worked Example: A Complete-Follow-Up Weighted Risk

Suppose a cohort has 100 participants in each of two baseline strata, $X=0$ and $X=1$. In stratum 0, 50 participants have complete follow-up to $\tau$, and 10 of them experience the event by $\tau$. In stratum 1, 80 have complete follow-up, and 16 of them experience the event. Assume the completion probabilities are $G(0)=0.50$ and $G(1)=0.80$. Calculate the normalized inverse-probability-weighted risk.

Each complete case in stratum 0 receives weight $1/0.50=2$; each complete case in stratum 1 receives weight $1/0.80=1.25$. The weighted event total is $$ 10(2)+16(1.25)=20+20=40. $$ The weighted complete-case total is $$ 50(2)+80(1.25)=100+100=200. $$ Thus the estimated risk is $40/200=0.20$, or 20%. As a check, the observed event proportions are $10/50=0.20$ and $16/80=0.20$ in both strata, so a 20% overall risk is consistent with these numbers. The calculation illustrates the weighting; it does not verify the conditional independence assumption from the data.

Assumptions, Diagnostics, and Alternatives

Administrative censoring is sometimes described as noninformative because the database closes on a date rather than in response to an individual’s health. That description is not automatically sufficient. Entry date can be associated with treatment practices, prognosis, or available care, and calendar time can therefore relate to the outcome. Moreover, if a late entrant has no possibility of being observed through $\tau$, then $G(X)=0$ for that participant’s covariate pattern. No weighting method can recover that person’s horizon-specific risk from the observed data without additional information or assumptions.

Loss to follow-up is often more directly related to health, treatment, or social circumstances. Conditional independent censoring requires that, after conditioning on the covariates used in the analysis, the probability of being observed through the relevant time does not depend on the outcome that would otherwise be observed. For time-varying loss processes, the required conditioning information may need to include measured histories up to each time. A poorly measured or misspecified censoring model can leave substantial bias.

Useful diagnostics include comparing entry dates and observed follow-up across treatment groups and baseline risk factors; tabulating reasons and timing of loss; checking whether many participants have little chance of reaching $\tau$; and examining the distribution of estimated completion probabilities and weights. Very small estimated $G(X)$ values produce large weights, making results sensitive to a few observations. Weight truncation can reduce variability but changes the estimator and does not repair a structural lack of positivity. Diagnostics can reveal warning signs, but cannot prove conditional independence.

The main alternatives answer different questions. A complete-case analysis is transparent but may target a selected subset. Restriction to an earlier-entry cohort can ensure a common opportunity for follow-up, but changes the population to which the result applies. Inverse-probability weighting can use incomplete-follow-up data under explicit assumptions, but depends on positivity and a sufficiently well-specified censoring model. Collecting outcomes through linkage or extended follow-up can reduce missing outcome information, though it does not by itself resolve every source of selection.

Common pitfall: “No event was recorded before the study ended, so this participant had no event by the target horizon.” This is wrong whenever observation ended before that horizon. The correct classification is: the horizon outcome is zero only if event-free observation continued through the horizon; an event observed before the end gives outcome one; otherwise the horizon outcome is unknown. Record the reason for the observation end as well as its date.

Putting the Observation Period Into the Protocol

A reproducible emulation should state the time-zero date, target horizon, common administrative end date, participant-specific loss rule, and event-at-cutoff convention. It should also say how incomplete outcomes will be handled and which population the analysis targets. These choices make the observation period auditable rather than an implicit consequence of the available data.

Fix the time origin and horizon.
Define each participant’s time zero and the elapsed-time horizon $\tau$ for the target outcome.

Identify both observation limits.
Calculate $A-E_i$ from the common calendar cutoff and $L_i$ from participant-specific follow-up records.

Classify endpoint status.
An event by $\tau$ is observed if ascertained before the observation window ends; an event-free outcome requires observation through $\tau$.

Assess identification and estimation.
Check whether the target population has a chance of complete observation and state the assumptions behind any restriction or censoring adjustment.

The practical lesson is not that one censoring type is always harmless and the other always harmful. It is that the study must represent each observation end correctly, then justify how the resulting incomplete outcomes inform the specified estimand.

Check Your Understanding

Answer each question, then compare with the solutions below.

A study closes on June 30, 2025. A participant enters on June 30, 2023, and has no recorded loss. What is the maximum administrative follow-up, and can a three-year event-free outcome be known?
A participant is lost at day 200, with no event recorded, for a one-year horizon. Is the one-year outcome zero or unknown? What additional observation would establish zero?
State the conditional independence and positivity assumptions needed for the complete-follow-up weighting identity.
Why can weighting fail for a late-entry participant who has no possibility of completing the target horizon?

Solutions. (1) The administrative window is two years, or approximately 731 days depending on the calendar interval convention. It is shorter than three years, so an event-free three-year outcome cannot be established. (2) The outcome is unknown because 200 days is less than one year. Documented event-free observation through day 365 would establish zero. (3) Conditional on measured $X$, complete observation $O$ must be independent of the fixed-horizon outcome $Y(\tau)$, and the completion probability $G(X)=P(O=1\mid X)$ must be positive in the target population. (4) The completion probability is zero, so the inverse weight $1/G(X)$ is undefined; observed data contain no complete cases from which to learn that horizon outcome without further assumptions or data.