What a Common Study End Does
A study may plan to observe every participant until one prespecified date. A participant who experiences the event before that date has an observed event time; a participant who remains event-free when observation ends has a right-censored time. The latter is not an event at the study end: it records that the event time exceeds the participant’s observed follow-up.
The key condition for a fixed-end administrative censoring mechanism is that participants share a common time origin and the same prespecified follow-up duration. For example, if everyone enters follow-up on the same date and the study ends 12 months later, the censoring time is 12 months for each participant who has not already experienced the event. The date may be selected for logistical reasons, but it must not be selected or changed in response to participants’ unobserved event times.
The inclusive convention needs to match the study protocol. With continuous event times, exact equality to \(\tau\) has probability zero under an ordinary continuous-time model. With dates recorded only to the day, ties at the endpoint can occur; investigators must specify whether an event on the closing date is included and apply that rule consistently.
Worked Example: Recording Follow-Up at a 12-Month Endpoint
Suppose, hypothetically, that 40 participants begin follow-up on the same date, and the study ends 12 months later. Assume events are completely ascertained through the endpoint. Four illustrative participants have the following observed histories.
| Participant | Observed time \(\tilde T\) (months) | \(\Delta\) | Interpretation |
|---|---|---|---|
| A | 4 | 1 | Event observed at month 4 |
| B | 12 | 0 | Event-free through month 12; censored at the endpoint |
| C | 9 | 1 | Event observed at month 9 |
| D | 12 | 0 | Event-free through month 12; censored at the endpoint |
For participant B, the record establishes \(T>12\) months, not \(T=12\) months. The observed time is \(\min(T,12)=12\), and \(\Delta=I(T\leq12)=0\). For participant A, \(T=4\) months is observed, so \(\min(T,12)=4\) and \(\Delta=I(T\leq12)=1\). The same coding rule gives the entries for C and D. No event times beyond the endpoint are needed or observed for B or D.
The endpoint supports descriptions and inference about event occurrence through 12 months. It does not, by itself, reveal when a censored participant will experience an event after month 12.
Independence: Censoring Time Is Not the Censoring Indicator
Under a genuine common fixed endpoint, \(C=\tau\) is a constant random variable. A constant random variable is independent of every random variable: for any event-time set \(A\) and any set \(B\) of censoring-time values, \(P(T\in A,C\in B)=P(T\in A)P(C\in B)\). In particular, the censoring time is independent of \(T\). This is an unusually transparent case of independent censoring: the endpoint does not vary from person to person according to their event prospects.
That statement concerns \(C\), not \(\Delta\). The indicator \(\Delta=I(T\leq\tau)\) is determined by \(T\) and is generally dependent on \(T\). For example, knowing that \(\Delta=1\) tells us that \(T\leq\tau\); knowing that \(\Delta=0\) tells us that \(T>\tau\). There are exceptional cases: if \(T\leq\tau\) almost surely, then \(\Delta=1\) almost surely, and if \(T>\tau\) almost surely, then \(\Delta=0\) almost surely. In either case \(\Delta\) is constant and therefore independent of \(T\). So it is not correct to infer dependence merely because one variable is a function of another; the dependence is general, not logically inevitable.
Worked Example: Showing That the Indicator Is Generally Dependent on Event Time
For illustration, suppose hypothetical event times follow an exponential distribution with rate \(\lambda=0.05\) per month, and the fixed endpoint is \(\tau=12\) months. This distributional assumption is only for this calculation; fixed-end censoring itself does not require exponential event times.
The probability of an event by month 12 is
Now consider the event \(A=\{T\leq6\}\). Its unconditional probability is
But \(T\leq6\) implies \(T\leq12\), and hence implies \(\Delta=1\). Therefore,
The indicator and event time are dependent in this example. By contrast, \(C=12\) is constant and independent of \(T\). The independence assumption used for censoring refers to the censoring time, not to the event indicator.
Likelihood Contributions at a Fixed Endpoint
Suppose event times in the target cohort have density \(f(t)\) and survival function \(S(t)=P(T>t)\). Assume participants’ event histories are completely observed until \(\tau\), and the only reason follow-up ends before an event is the prespecified administrative endpoint. For an observed event at time \(t\leq\tau\), the contribution to the event-time likelihood is \(f(t)\). For a participant censored at \(\tau\), the information is \(T>\tau\), which contributes \(S(\tau)\). Thus, apart from factors involving the fixed endpoint that do not depend on the event-time parameters, an observation contributes
This expression follows directly from what each record reveals. When \(\Delta=1\), the event time is observed and its density is used. When \(\Delta=0\), the event time is not observed, but it is known to exceed \(\tau\), so the probability of that event-free follow-up is used. For independent participants, these individual contributions multiply. If there are distinct event causes, an observed event contribution must also reflect its recorded cause; a censored participant still contributes the probability of no event of any cause through \(\tau\).
Worked Example: Likelihood With Two Events and Two Administrative Censors
Consider hypothetical independent event times from an exponential model with rate \(\lambda\) per month. Four participants share a fixed endpoint of 12 months. Events occur at months 3 and 8; the other two participants remain event-free through month 12. Assume exact event times, complete ascertainment through month 12, and no other censoring. Under this model, \(f(t)=\lambda e^{-\lambda t}\) and \(S(t)=e^{-\lambda t}\).
The two event contributions are \(\lambda e^{-3\lambda}\) and \(\lambda e^{-8\lambda}\). Each administrative censor contributes \(S(12)=e^{-12\lambda}\). The likelihood is therefore
The log-likelihood is \(\ell(\lambda)=2\log(\lambda)-35\lambda\). Differentiating gives \(\ell'(\lambda)=2/\lambda-35\); setting this equal to zero yields \(\hat\lambda=2/35\approx0.05714\) per month. The censored participants contribute 24 participant-months of known event-free time, not two event times at month 12. This numerical estimate relies on the exponential model as well as the stated observation assumptions; the fixed-end censoring mechanism alone does not imply an exponential distribution.
When the Fixed-Endpoint Argument Is Credible
The claim \(C=\tau\) is credible only when the actual follow-up structure supports it. Before treating administrative censoring as independent, check the design and the records. The following conditions are central:
- A common time origin and duration: every participant’s follow-up starts from the same defined origin and the planned endpoint is the same elapsed time \(\tau\) later.
- A prespecified, outcome-independent closure: the endpoint is set for logistical or design reasons, not moved because of observed or anticipated event patterns.
- Complete event ascertainment through the endpoint: a participant not recorded as having an event is genuinely known to be event-free through their censoring time.
- No unrecorded earlier end to observation: loss to follow-up, withdrawal, or another reason for ending observation before \(\tau\) is not silently coded as administrative censoring at \(\tau\).
- A clearly defined analysis population: the independence claim applies within the population whose follow-up is being analysed. It does not remove selection issues arising from how that population was formed.
For a fixed endpoint, useful diagnostics are mainly design and data checks rather than a statistical test of independence. Compare recorded censoring dates with the planned closing date; check that censored participants have follow-up through the endpoint; inspect event dates, censoring reasons, and any protocol amendments; and verify the time origin and endpoint convention. A censoring date before \(\tau\), missing event ascertainment, or a reason such as withdrawal signals that the observation process is not described by \(C=\tau\) alone. The earlier material on censoring audits and independent censoring provides the broader framework for those checks.
Comparison, Scope, and Common Pitfalls
Fixed-end administrative censoring is simpler than censoring times that vary with participants’ circumstances. When everyone has the same planned duration, the censoring-time distribution is a point mass at \(\tau\), and the likelihood contribution for someone event-free at closure is \(S(\tau)\). With random loss to follow-up, censoring times vary and their relationship to event time must be assessed under the relevant independent-censoring assumptions. A common calendar closing date is not automatically the same as a common follow-up duration: if participants have different time origins, their elapsed times at closure differ. That setting requires a different representation of follow-up.
A fixed endpoint also limits the time range directly supported by observation. It can support estimation of event-time quantities through \(\tau\), subject to the event-time model and sampling assumptions, but it cannot identify the event-time distribution after \(\tau\) without additional assumptions or data. In competing-risks settings, the endpoint limits observation of all event causes: an event-free participant censored at \(\tau\) has no observed event cause, while an observed event before \(\tau\) must have its cause ascertained if cause-specific conclusions are intended.
A frequent mistake is to say, “The censoring indicator is independent of event time because the endpoint is fixed.” This confuses two variables. The fixed censoring time \(C\) is independent of \(T\) because it is constant; \(\Delta=I(T\leq\tau)\) is generally dependent on \(T\), since it reports whether \(T\) falls before the cutoff. The correction is to name the variable in the independence claim.
Another mistake is to conclude that any study with one common end date has fixed-end independent censoring. If participants have different origins, if event ascertainment stops early for some participants, or if the closure date is changed in response to outcome information, the mechanism is not captured by \(C=\tau\) for everyone. The correction is to verify the elapsed-time definition, the actual recorded follow-up, and the reason for each censoring record before using the fixed-end argument.
Check Your Understanding
Answer each question using the definitions and assumptions stated above. Solutions follow the questions.
- Everyone begins follow-up together and the study closes at 10 months. For someone event-free at closure, what do \(\tilde T\) and \(\Delta\) equal, and what does the record establish about \(T\)?
- Explain why \(C=10\) is independent of \(T\), but \(\Delta=I(T\leq10)\) is generally dependent on \(T\).
- Under an exponential model with rate \(\lambda=0.04\) per month, what is the likelihood contribution of a participant censored at a fixed endpoint of 10 months?
- What additional issue arises if every participant is followed until the same calendar closing date but participants begin follow-up on different dates?
- Suppose a participant’s event status is unknown from month 7 onward, even though the study’s planned endpoint is month 10. Can this record be coded as administrative censoring at month 10? Explain.
Solutions. (1) \(\tilde T=\min(T,10)=10\) and \(\Delta=I(T\leq10)=0\); the record establishes \(T>10\) months. (2) A constant random variable is independent of every random variable, so \(C=10\) is independent of \(T\). But \(\Delta\) indicates whether \(T\leq10\), which generally changes the distribution of \(T\); for instance, \(T\leq5\) implies \(\Delta=1\). (3) The contribution is \(S(10)=e^{-0.04(10)}=e^{-0.4}\approx0.6703\), the probability of remaining event-free beyond month 10 under the model. (4) Elapsed follow-up differs across participants, so a common calendar date does not mean \(C_i=10\) months for everyone; the time origin and resulting censoring times must be represented correctly. (5) No. Event status is not known through month 10, so the participant was not completely ascertained to the administrative endpoint. The record requires its actual observation end and reason to be represented rather than being treated as administrative censoring at month 10.