Right-Censoring Mechanisms and Their Assumptions

What Right-Censoring Assumptions Are About

Right-censoring means that the event of interest has not been observed by the time a subject’s observation ends. The recorded time is therefore a lower bound on the event time, not the event time itself. Knowing why observation ended matters because some mechanisms make censoring plausibly unrelated to event time, while others can preferentially remove people with different event risks.

Let \(T\) denote the event time of interest and \(C\) the time at which observation ends for a reason other than observing that event. The recorded time is \(Y=\min(T,C)\), and the event indicator is \(\Delta=I(T\leq C)\). Thus, \(\Delta=1\) records the event at \(Y=T\); \(\Delta=0\) records right-censoring at \(Y=C\). In a multi-state setting, \(T\) may be the time to a specified transition, and \(C\) marks the end of observation of the state process. It does not itself describe a transition.

Definition: A censoring mechanism is the process that determines when or why observation of the event time ends. Independent censoring is a sufficient assumption for standard event-time likelihood methods: the censoring time \(C\) is independent of the event time \(T\) in the analysis population. A common alternative is independence conditional on measured covariates; in that case, the analysis must account appropriately for those covariates.

“Independent” is a statement about the relationship between the censoring process and the event time, not a label assigned by the recorded reason. Administrative censoring is often plausibly independent under a well-defined design, but is not automatically so. Loss to follow-up and withdrawal raise concerns, but are not necessarily informative if their timing is unrelated to event time under the assumptions of the analysis. The mechanism and the justification must be considered together.

Four Common Ways Observation Ends

Mechanism	What happens	Key assumption to consider
Administrative censoring	Observation ends at a planned study or data-collection boundary.	The boundary and enrollment process must not select observation times in a way that depends on event time, unless that dependence is addressed.
Loss to follow-up	The study can no longer obtain outcome information from a participant who has not had the event.	Loss must not be related to future event time, at least after accounting for relevant observed information in the analysis.
Withdrawal	A participant or representative asks for study follow-up to stop.	The reason for withdrawal must not signal an unaccounted-for difference in future event risk.
Other observation-ending mechanisms	Examples include transfer of care, loss of a data feed, or a record-linkage boundary.	The circumstances causing observation to end must be evaluated for links to event time.

These categories describe how observation stopped; they do not change the definition of the event. For example, death may be an event of interest, a competing event, or—if the study’s endpoint is a different event—an event that prevents further observation of that endpoint. Whether it is a censoring event depends on the estimand and analysis, not on the fact that it occurred before the target event. This distinction is consistent with the earlier material on distinguishing censoring from event types.

Why Independence Supports the Event-Time Likelihood

For a continuous event time with density \(f_T(t)\) and survival function \(S_T(t)=P(T>t)\), the event-time contribution for one observation is \(f_T(Y)\) if the event is observed and \(S_T(Y)\) if the observation is censored. This contribution follows directly when the censoring process is independent of the event time.

To see why, suppose first that an event is observed at time \(y\). The joint contribution from observing the event at \(y\) and remaining under observation until then involves the event-time density and the chance that censoring has not occurred before \(y\). If a subject is censored at \(y\), the contribution involves the chance the event time exceeds \(y\) and the censoring-time density at \(y\). Under independence, these factors separate into an event-time part and a censoring-time part. When inference concerns the event-time distribution and the censoring distribution is treated as a nuisance, the event-time parts are \(f_T(y)\) for an event and \(S_T(y)\) for a censored observation.

L_T(y,\Delta) = \{f_T(y)\}^{\Delta}\{S_T(y)\}^{1-\Delta}

This expression is not a license to ignore the observation process. It relies on the stated independence condition and on correctly identifying the event and censoring times. If censoring depends on event time, a censored subject’s remaining event-time distribution may differ from the population distribution the analysis intends to describe. Conditional independence can sometimes provide a valid basis instead: if censoring is independent of event time within levels of measured covariates, an analysis must use those covariates in a way that supports the target inference. The precise marginal and conditional formulations require careful distinction.

Worked Example: A Planned End of Follow-Up

Suppose a hypothetical study observes participants for at most 12 months after enrollment. One participant has no recorded event by month 8, when their scheduled observation ends. Assume the study’s enrollment and event processes are stable over the relevant period, and that the scheduled end time is determined by the protocol rather than by this participant’s health or expected event time. For illustration, suppose a specified event-time model gives \(S_T(8)=0.72\).

The participant’s recorded values are \(Y=8\) months and \(\Delta=0\). Under the stated assumptions, the event-time likelihood contribution is

\{f_T(8)\}^{0}\{S_T(8)\}^{1} = S_T(8) = 0.72.

The value 0.72 is the model’s probability of remaining event-free beyond 8 months; it is not a probability that this participant will never have the event. The scheduled endpoint supports an independent-censoring argument only insofar as the assumptions are credible. For example, if enrollment date is related to calendar-time changes in event risk and determines follow-up length, the simple argument may not hold without accounting for those changes.

Mechanism-Specific Assumptions and Warning Signs

Administrative censoring. A protocol-defined study end is often a strong design-based reason for censoring, because the end date may be set in advance. The argument still depends on the study population and time scale. If participants enter at different calendar times, have different potential follow-up, or experience changing event risk across calendar time, censoring time can be associated with event time in the pooled data. The name “administrative” does not establish independence.

Loss to follow-up. A participant may become unreachable because of relocation, practical barriers, or disengagement. If these causes are unrelated to future event time, independent censoring may be reasonable. If worsening health, symptoms, or impending treatment changes make follow-up less likely, the event-time distribution among those lost may differ from that among those still observed. Relevant recorded history may help justify a conditional analysis, but unrecorded predictors can remain a concern.

Withdrawal. A request to stop follow-up can have many causes. A preference unrelated to prognosis may have little connection to event time. Withdrawal because of worsening symptoms, treatment burden related to disease severity, or dissatisfaction after an adverse change may be prognostic. The reason is useful evidence about plausibility, not proof of dependence or independence. A study should also distinguish stopping study visits from permission to continue outcome collection through records when that distinction is available.

Other mechanisms. A transfer to another health system, a change in data coverage, or a technical failure can end observation. A data-coverage boundary may be unrelated to an individual’s event time, while transfers prompted by worsening health may not be. The relevant question is always whether the process that stops observation is associated with the event-time information that would have been observed.

Worked Example: Loss to Follow-Up Related to Measured Risk

In a hypothetical cohort, 20% of participants are classified at baseline as high risk and 80% as low risk. Assume their probabilities of the target event by a fixed horizon are 0.50 and 0.10, respectively. Suppose the probability of being lost before that horizon is 0.40 for high-risk participants and 0.05 for low-risk participants. For this illustration, assume censoring is independent of event time within each risk group.

Among those lost, the probability of being high risk is

P(\text{high risk}\mid\text{lost}) = \frac{0.20(0.40)}{0.20(0.40)+0.80(0.05)} = \frac{0.08}{0.12} = 0.6667.

Thus, two-thirds of those lost are high risk, even though high-risk participants were only one-fifth of the original cohort. Applying the assumed within-group baseline horizon risks to the measured-risk composition among those lost gives the following mixture; it is not the event probability after loss, which requires loss times and conditional event risks beyond those times that are not provided.

0.6667(0.50)+0.3333(0.10) \approx 0.3667.

Among participants not lost, the high-risk proportion is

\frac{0.20(0.60)}{0.20(0.60)+0.80(0.95)} = \frac{0.12}{0.88} = 0.1364.

Their expected event probability is \(0.1364(0.50)+0.8636(0.10)\approx0.1545\). The lost and retained groups therefore differ in event risk marginally because they have different mixtures of measured risk groups. Under the example’s assumption of independence within each group, this does not by itself invalidate an analysis that properly accounts for risk group. It does show why treating all loss to follow-up as automatically independent can be misleading.

Worked Example: Withdrawal Associated With Unobserved Prognosis

Consider a hypothetical study in which 40 participants withdraw before the target event is observed. Suppose investigators later learn, for illustration, that 10 of the 40 withdrew after worsening symptoms. Assume the true probability of the target event within six months would have been 0.60 for those 10 participants and 0.20 for the other 30. These are hypothetical probabilities for demonstrating the consequence of an assumption violation, not estimates from a real study.

The six-month event probability among the 40 withdrawn participants would be

\frac{10(0.60)+30(0.20)}{40} = \frac{6+6}{40} = 0.30.

If an analysis effectively treated the withdrawn participants as having the same future risk as the 30 participants without worsening symptoms, it would use 0.20 for this group. The difference is \(0.30-0.20=0.10\), or 10 percentage points. The example shows how withdrawal linked to prognosis can make an independence assumption implausible and distort event-time inference. In practice, the unobserved outcomes of withdrawn participants are not known, so the relevant risk difference cannot be verified from the observed data alone.

Derivation: How Dependence Changes the Contribution

The factorization under independence can be made explicit. Write \(f_C\) for the censoring-time density and \(S_C\) for its survival function. If an event occurs at time \(y\) before censoring, the joint contribution is \(f_T(y)S_C(y)\). If censoring occurs at time \(y\) before the event, the joint contribution is \(S_T(y)f_C(y)\). Under independence, the first expression separates into an event-time term \(f_T(y)\) and a censoring term \(S_C(y)\); the second separates into \(S_T(y)\) and \(f_C(y)\).

\begin{aligned} \Delta=1 &: \quad f_T(y)S_C(y),\\ \Delta=0 &: \quad S_T(y)f_C(y). \end{aligned}

The censoring factors can then be treated as nuisance contributions when the target is the event-time distribution. If \(T\) and \(C\) are dependent, the corresponding joint event and censoring distributions generally do not factor this way. The usual event-time contribution need not describe the target population’s event-time distribution. This is the core reason that investigating the mechanism is part of the analysis rather than a clerical step.

Diagnostics, Alternatives, and Limits

No single diagnostic proves independent censoring, especially because the event times after censoring are unobserved. Useful checks include comparing baseline characteristics across observed censoring reasons, examining whether censoring rates vary with measured prognosis, and reviewing whether censoring clusters around clinical deterioration or treatment changes. Plotting or tabulating censoring times by enrollment period can also reveal patterns that a pooled summary hides. These checks can identify implausibility; they cannot establish independence from unmeasured information.

When censoring is plausibly independent by design, standard right-censored event-time methods are often appropriate. When censoring depends on measured covariates but is plausibly independent after conditioning on them, the analysis must account for that structure rather than rely on an unadjusted marginal argument. When dependence may remain after using observed information, possible responses include collecting more follow-up, modeling the observation process, using methods that address informative censoring under stated assumptions, and conducting sensitivity analyses. No method can recover unobserved outcomes without assumptions.

The main alternatives are not interchangeable labels. Administrative censoring describes a planned boundary; it does not guarantee noninformativeness. Loss to follow-up and withdrawal describe how contact ended; either may be independent or informative. A competing event is an observed event that changes the event process, not simply a participant who became unobservable. Choosing between treating an event as a competing event and censoring at its time changes the question being answered.

Wrong reasoning: “This person was censored administratively, so censoring is independent.”
Correction: Check whether the administrative time, enrollment pattern, and event process support independence for the target population.
Wrong reasoning: “Loss to follow-up is always informative.”
Correction: Loss can be independent, conditionally independent given relevant observed information, or informative even after adjustment; the reason and data context matter.
Wrong reasoning: “A participant without an observed event is censored.”
Correction: Determine whether observation actually ended before the event, or whether another event occurred that is part of the endpoint or a competing event.

Key takeaway: Record the event definition, censoring time, censoring reason, and information available before censoring. Then state the independence assumption needed by the analysis and explain why the observed mechanism makes it plausible—or why it may fail.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

A participant remains event-free when a protocol-defined observation period ends. What are \(Y\) and \(\Delta\), and what assumption is needed to use the standard censored event-time contribution?
Why does a higher loss-to-follow-up rate among high-risk participants not, by itself, prove that censoring is informative conditional on risk group?
In a hypothetical cohort, 30% of participants withdraw after symptom worsening. What should an analyst conclude about independent censoring?
A participant experiences a competing event before the target event. Why is it not automatically correct to treat this as ordinary right-censoring?
State one diagnostic that may reveal a concern about censoring and one reason it cannot prove independence.

Solutions. (1) If follow-up ends at \(c\) months without an observed event, \(Y=c\) and \(\Delta=0\). The analysis needs censoring to be independent of event time, or a suitable conditional independence assumption accounted for by the analysis. (2) The measured risk group can explain the difference in censoring rates; censoring may still be independent of event time within each group. The analysis must account for the group. (3) The analyst should investigate whether symptom worsening predicts future event time and whether that information is measured. The percentage alone neither proves nor rules out independence. (4) A competing event is an observed event that affects the possibility or meaning of the target event; treating it as censoring changes the event-time question and requires a distinct justification. (5) Comparing censoring rates across measured prognosis groups can expose a concern. It cannot prove independence because event times after censoring are unobserved and unmeasured predictors may remain.