Why the Reason for Loss Matters
A participant can leave a study for many reasons: the planned observation period ends, a clinic closes, the participant moves, or worsening health makes study visits difficult. These events may all produce fewer observations, but they do not carry the same information about a participant’s unobserved future. The reason for loss helps assess whether the observed follow-up can represent the target group; the label alone does not settle that question.
A further distinction is essential: stopping longitudinal visits is not necessarily the same as losing information about the event time. A participant may stop returning for biomarker measurements while a registry continues to ascertain hospitalization or death. In that case, longitudinal follow-up has ended, but event follow-up has not. Conversely, a participant may be lost to all contact, so both measurements and event status after the last contact are unknown.
Administrative Censoring and Logistical Loss
Administrative censoring occurs because observation ends at a study-defined cutoff, such as the planned end of follow-up. If every participant is followed for 12 months after enrollment, the study-end censoring time is 12 months for those who have not yet had the event. This is not a participant’s decision to leave. With staggered enrollment and one common calendar end date, however, participants can have different follow-up durations; enrollment date and calendar time may then matter when assessing the censoring mechanism.
Logistical loss refers to follow-up ending for practical or organizational reasons, such as a clinic closing, a move, or difficulty scheduling visits. Such reasons are not, by themselves, evidence that loss is unrelated to event risk. For example, moving for a job might be unrelated to health in one setting, while moving to obtain care could be related to health in another. A neutral-sounding reason does not guarantee a neutral selection process.
A common working assumption in event-time analysis is conditional independent censoring. For baseline covariates \(Z_i\), its precise form is \(T_i\perp C_i\mid Z_i\): within levels of the recorded covariates, event time and censoring time are independent. This assumption permits censoring to be associated with event time marginally. If censoring depends on recorded follow-up history, a more relevant assumption must condition on the appropriate history up to each time; it is not enough to adjust for baseline variables automatically.
Worked Example: A Planned Study-End Cutoff
Hypothetical scenario. A study follows 100 participants for 12 months after enrollment. There is no loss of event ascertainment before month 12. By month 12, 24 participants have experienced the event; the remaining 76 have not. Assume that event times are recorded accurately and that study follow-up truly ends at the planned 12-month limit for everyone.
For each of the 76 participants without an event by month 12, \(Y_i=12\) months and \(\Delta_i=0\). Their event times are not known to equal 12 months; only \(T_i>12\) months is known. The 24 participants with events have \(\Delta_i=1\), and their event times are observed.
Because there is no earlier loss, the study can describe the proportion with an event by 12 months as \(24/100=0.240\), or 24.0%, for this hypothetical cohort. This calculation describes event occurrence during the observed 12-month period; it does not establish what happens after that period. In a time-to-event analysis, the 76 event-free participants contribute follow-up through month 12 rather than being treated as if their event times were observed.
The administrative cutoff is fixed by design, so it is not triggered by a participant’s worsening health. If the cutoff instead came from a common calendar end date with different enrollment dates, follow-up duration would vary. The analyst should then consider enrollment and calendar-time factors rather than assuming all censoring times are identical.
Conditional Independence Does Not Mean Marginal Independence
The distinction between conditional and marginal independence is not merely technical. Suppose a recorded characteristic predicts both event risk and the chance of being censored at a particular time. Within each covariate group, censoring can be unrelated to event time, while censored participants overall have a different mix of groups from the full cohort. Their overall event risk can therefore differ.
Worked Example: A Logistical Reason With Different Group Composition
Hypothetical scenario. Let \(Z_i\) indicate whether a participant belongs to group 0 or group 1, with \(P(Z_i=0)=P(Z_i=1)=0.5\). At time \(c\), suppose event-free probabilities are \(P(T_i>c\mid Z_i=0)=0.80\) and \(P(T_i>c\mid Z_i=1)=0.40\). A logistical event causes censoring at \(c\) with probabilities \(P(C_i=c\mid Z_i=0)=0.20\) and \(P(C_i=c\mid Z_i=1)=0.80\). Assume \(T_i\) and \(C_i\) are independent within each group.
The conditional independence assumption gives \(P(T_i>c\mid C_i=c,Z_i=0)=0.80\) and \(P(T_i>c\mid C_i=c,Z_i=1)=0.40\). The equality holds separately within each group. But among those censored at \(c\), group 1 is more common, so the marginal probability is different.
By the law of total probability,
In the full cohort, the event-free probability at \(c\) is \(0.5(0.80)+0.5(0.40)=0.60\). Thus \(P(T_i>c\mid C_i=c)=0.48\), not \(P(T_i>c)=0.60\). This difference does not contradict the assumed conditional independence; the groups have different event risks and different censoring probabilities. It shows why an unconditional equality must not be inferred from conditional independence.
Health-Related Attrition
Health-related attrition occurs when worsening health, symptoms, treatment burden, or care needs prompt a participant to stop study visits or make further contact difficult. If the health changes that drive attrition also predict the event or the unobserved longitudinal measurements, the participants who remain can become a selected subset. In that case, the observed follow-up may not represent the full cohort’s health trajectory or event experience.
A health-related reason does not automatically prove that censoring is informative after all relevant information is considered. If the health state driving loss is measured well and included in an appropriate conditional model, a conditional independence assumption might be plausible. If worsening health is unmeasured or poorly measured, conditioning on recorded baseline characteristics may not address the dependence. Whether the reason is recorded is not the same as whether the underlying health process is fully observed.
Worked Example: Retention Depends on Worsening Health
Hypothetical scenario. Consider 200 participants, half with a measured health state classified as higher risk and half with a lower-risk state. Suppose the probability of an event during a later one-year period is 0.60 in the higher-risk group and 0.20 in the lower-risk group. Before that period, the probability of health-related dropout is 0.50 for the higher-risk group and 0.10 for the lower-risk group. Assume dropout and the later event are independent within each measured health group.
The expected number retained from the higher-risk group is \(100(1-0.50)=50\), and the expected number of later events among them is \(50(0.60)=30\). The expected number retained from the lower-risk group is \(100(1-0.10)=90\), with \(90(0.20)=18\) expected events. Therefore, the expected event risk among retained participants is
The event risk in the full cohort before selection is \(0.5(0.60)+0.5(0.20)=0.40\). The retained group’s expected risk, about 34.3%, is lower because higher-risk participants are less likely to remain. This calculation is a selection illustration, not proof that a particular analysis will estimate exactly 34.3%. It also relies on the stated within-group independence assumption; if health changes within each group further predict dropout and events, that assumption would not hold.
How to Assess the Follow-Up Process
A reason for loss is useful information, but assessment should also consider when loss occurred, what was observed before it, and whether event ascertainment continued by another route. The following checks can reveal patterns that make a simple independent-censoring assumption less plausible:
- Tabulate losses by reason, timing, site, enrollment period, and relevant baseline characteristics.
- Compare recorded health and biomarker histories before loss with those of participants who remain under observation. In particular, check whether worsening measurements or increasing symptoms tend to precede dropout.
- Establish whether participants who stop visits can still have their event times ascertained through records or registries. If so, visit dropout and event censoring should not be treated as the same time.
- Check whether the administrative end date is genuinely set by the study schedule or varies with enrollment, site closure, or participant circumstances.
These checks diagnose observable patterns; they cannot establish independence from health information that was never measured. Similar recorded histories among retained and lost participants do not rule out different unobserved health states. Conversely, an association between recorded health and loss does not by itself prove that all censoring is unmanageable: the relevant question is whether the assumption is plausible after accounting for appropriate observed information.
Comparing the Main Follow-Up Situations
| Situation | What ends? | What to assess |
|---|---|---|
| Planned administrative cutoff | Study observation at a defined end time | Whether the cutoff is fixed as designed; whether varying follow-up relates to enrollment or calendar time |
| Logistical loss | Visits, event ascertainment, or both | Whether the practical reason or recorded history is associated with event risk after conditioning on relevant covariates |
| Health-related attrition | Often visits, sometimes event ascertainment too | Whether worsening health predicts loss and whether that health information is adequately observed |
Routine time-to-event methods can accommodate right-censoring under appropriate independent or conditionally independent censoring assumptions. Simply labeling a loss “administrative” or “logistical” does not verify those assumptions. When observed histories predict loss, analysts may consider methods that account for measured predictors; when important health information is unobserved, sensitivity analyses or models that explicitly represent dependence may be needed. Those approaches rely on their own assumptions and do not recover information without qualification.
The central limitation is that post-dropout outcomes are not directly observed merely because the reason for loss has been documented. Conclusions about those outcomes depend on assumptions. Administrative censoring at a planned endpoint limits the time horizon of inference; health-related attrition may additionally change which kinds of participants contribute later measurements or events.
Common Mistake
Wrong reasoning: “The participant left for a logistical reason, so censoring is independent; therefore the event-free probability among those censored equals the event-free probability in the full cohort.” A reason that is not explicitly health-related does not establish independence, and conditional independence given covariates does not imply marginal independence.
Correction: Specify the conditioning information and assess whether it is adequate. Under \(T_i\perp C_i\mid Z_i\), the valid equality is \(P(T_i>c\mid C_i=c,Z_i)=P(T_i>c\mid Z_i)\). The marginal equality without \(Z_i\) requires marginal independence, a stronger assumption. Also determine whether the participant lost visits, event ascertainment, or both.
Check Your Understanding
Answer each question, then compare with the brief solutions.
- A participant stops biomarker visits at month 8, but a registry records their event time through month 12. Is their event time necessarily censored at month 8?
- If \(T_i\perp C_i\mid Z_i\), must \(P(T_i>c\mid C_i=c)=P(T_i>c)\)?
- A study ends at a fixed 12 months after each participant enrolls. A participant remains event-free throughout. What is observed about their event time?
- Why might dropout prompted by worsening health make the observed group’s later health measurements unrepresentative of the original cohort?
Solutions. (1) No. Visit dropout does not end event ascertainment if the registry supplies later event information. (2) No. The equality is guaranteed conditional on \(Z_i\), not necessarily marginally; different covariate mixtures can induce marginal association. (3) The event time is right-censored at 12 months: \(T_i>12\) months is known, but its later value is not. (4) If worsening health predicts both dropout and later measurements, participants with worse health may be less likely to contribute those measurements, changing the composition of the observed group.