Why Two Times Produce One Survival Record
In survival data, the time an event would occur and the time observation ends are not always both visible. A participant may experience the event while being followed, or follow-up may end first because a study closes, the participant withdraws, or contact is lost. In the second case, the event time is not observed; the data establish only that the participant remained event-free through the observed follow-up.
This distinction matters when survival records are used in chained-equation imputation. The recorded follow-up duration is not always the event time, and the event indicator carries essential information about what that duration means. A censored duration is a bound on the unobserved event time, not a substitute value for it.
The time origin must be defined consistently, such as randomization, diagnosis, or study entry. Both \(T_i\) and \(C_i\) are measured from that origin in the same units. The event must also be specified: “time to relapse” and “time to death” are different event-time variables, even if they are observed in the same people.
For an event record, \(Y_i=T_i\): the event occurred before observation ended, so its time is observed. For a right-censored record, \(Y_i=C_i\) and \(T_i>C_i\): the event time is only known to be later than the end of observed follow-up. The record does not reveal how much later it occurs, or whether it occurs at all within some later period.
Reading an Observed Record
Suppose a study follows participants from enrollment until the first occurrence of a specified event, or until follow-up ends. A row containing \(Y_i=5\) months and \(\Delta_i=1\) says that the event occurred at 5 months. A row containing \(Y_i=5\) months and \(\Delta_i=0\) says that the participant was known to be event-free through 5 months; the actual event time, if it occurs, is later than 5 months.
The censoring time itself may have different practical causes. Administrative censoring occurs when a study ends while a participant remains event-free. Loss to follow-up occurs when observation ends earlier than planned. In either case, the record is right-censored if the event has not been observed by the end of follow-up. The reason for censoring is useful information for assessing assumptions, but the basic record still consists of the observed time and indicator.
Worked Example: Converting Follow-Up Records
Consider a hypothetical study measuring months from enrollment to first relapse. Each participant is followed until relapse or until observation ends. The table gives the event time \(T_i\), when it occurs during the study, and the non-event end-of-observation time \(C_i\). For this illustration, an unobserved event time is shown as “not observed.”
| Participant | Event time \(T_i\) (months) | End of observation \(C_i\) (months) | Observed \(Y_i\) (months) | \(\Delta_i\) | What the record establishes |
|---|---|---|---|---|---|
| A | 4 | 9 | 4 | 1 | Relapse occurred at 4 months |
| B | Not observed | 7 | 7 | 0 | Relapse time is later than 7 months |
| C | 5 | 3 | 3 | 0 | Relapse time is later than 3 months |
| D | 8 | 10 | 8 | 1 | Relapse occurred at 8 months |
For A, \(Y_A=\min(4,9)=4\) and \(\Delta_A=I(4\le9)=1\), so the event time is observed. For B, \(Y_B=\min(T_B,7)=7\) and \(\Delta_B=0\), so the available information is \(T_B>7\), not \(T_B=7\). For C, observation ends at 3 months before the event at 5 months; consequently \(Y_C=3\) and \(\Delta_C=0\), and the event time is only known to exceed 3 months. For D, \(Y_D=\min(8,10)=8\) and \(\Delta_D=1\), giving an observed event time of 8 months.
In an actual data set, the unobserved event time for B or C would not be entered as though it were known. The table’s unobserved values merely make the logic of the minimum explicit.
What Censoring Does and Does Not Tell You
A censored record contributes genuine information: the participant remained event-free up to the recorded time. It does not identify an exact event time after that point. If \(S(t)=P(T>t)\) denotes the event-free probability beyond time \(t\), then the information from a censored record at \(y\) is consistent with the event-free condition \(T>y\). A record with an observed event at \(y\), by contrast, identifies the event at that time, subject to the study’s measurement precision.
A short derivation shows why these records have different contributions to an event-time likelihood. Assume that \(T\) and \(C\) are independent continuous times, with event-time density \(f_T\) and censoring-time density \(f_C\). For an observed event at \(y\), the event time is at \(y\) and censoring occurs after \(y\), so the joint density contribution is \(f_T(y)P(C>y)=f_T(y)S_C(y)\). For censoring at \(y\), censoring occurs at \(y\) and the event time exceeds \(y\), giving \(f_C(y)P(T>y)=f_C(y)S_T(y)\). Therefore:
The factor involving the event-time distribution is \(f_T(y_i)\) for an event and \(S_T(y_i)\) for a censored record. If the censoring mechanism is independent of the event-time model, its factors can be separated when estimating the event-time distribution. This is why a censored observation contributes an event-free probability rather than an event density at \(y_i\). With discrete times or recorded ties, the corresponding likelihood uses probability masses and a consistent tie convention rather than continuous densities.
The independence assumption is substantive. In applications, a common version is that censoring is independent of the event time conditional on measured covariates used in the analysis. For example, if participants with worse measured health are more likely to leave follow-up, censoring may be associated with event time marginally but plausibly independent after conditioning on those health measures. If censoring still depends on the unobserved event time after conditioning, the simple event-time contributions may not adequately represent the data.
Worked Example: Interpreting a Censored Record as a Bound
Suppose a hypothetical participant is event-free through 8 months and is then lost to follow-up. The record is \(Y=8\) months and \(\Delta=0\). It establishes \(T>8\) months; it does not establish an event at 8 months or a particular later event time.
For illustration, suppose a specified event-time model gives \(S(8)=0.60\) and \(S(12)=0.36\). Conditional on being event-free at 8 months, the model-implied probability of an event after month 8 but by month 12 is
Thus, under this illustrative model, the conditional probability is 0.40, or 40%. This is a model-based probability for people known to be event-free at 8 months; it does not reveal whether this particular censored participant experienced the event by month 12. The values of \(S(8)\) and \(S(12)\) are assumptions for the example, not quantities determined by this one censored record.
Worked Likelihood Contributions
The likelihood distinction can also be made concrete. Suppose, only for this example, that event times follow an exponential model with rate \(\lambda=0.05\) per month. Its density and survival function are \(f_T(t)=\lambda e^{-\lambda t}\) and \(S_T(t)=e^{-\lambda t}\). This model assumes a constant event rate over time; additionally assume independent censoring so the event-time contributions can be considered separately from the censoring mechanism.
Worked Example: Event and Censoring Contributions at Six Months
For an observed event at 6 months, the event-time likelihood contribution is a density:
For a person censored at 6 months, the event-time contribution is instead the probability of remaining event-free beyond 6 months:
The event contribution has density units, while the censored contribution is a probability. They are not interchangeable. The observed event supports an event at month 6 under the model; the censored record supports survival beyond month 6. Neither calculation imputes an exact event time for the censored person.
Assumptions, Diagnostics, and Failure Modes
Interpreting the record requires more than calculating a minimum. The time origin and event definition must be clear; event status and follow-up time must refer to the same event and observation process; and the recorded time must represent when observation ended or the event occurred. The analysis also needs a defensible censoring assumption. Independent censoring, possibly conditional on observed covariates, allows the observed censoring to be handled without treating its timing as information about unobserved event times beyond what those covariates explain.
Independence cannot generally be verified from observed follow-up alone: the event times after censoring are unknown. Diagnostics therefore examine the process that generated censoring. Tabulate censoring proportions and reasons, compare measured covariates across event and censoring groups, and inspect whether censoring patterns vary across important groups or time periods. Strong differences are not proof that the assumption fails, but they prompt investigation of whether measured predictors should be represented. A high censoring rate late in follow-up also means that event-time information in that period is limited.
In MICE, preserve the paired meaning of \(Y_i\) and \(\Delta_i\) when survival records are used as predictors or otherwise included in an imputation procedure. A censored follow-up time is not an ordinary event-time value with a missing indicator; it gives the lower bound \(T_i>Y_i\). Treating it as an exact event time changes the data’s meaning. The earlier material on Survival Records for Chained-Equation Imputation establishes why this pair matters in survival-related imputation; here the key issue is the underlying event-time information encoded by each pair.
Two tempting alternatives are especially problematic. First, analyzing only participants with observed events discards censored participants and their known event-free time, potentially producing a selected group. Second, coding every censored duration as an event time asserts an event that was not observed. Neither approach respects the information in the record. A survival likelihood uses the event density for observed events and the survival probability for censored observations, under its censoring assumptions.
These records have limits. Right censoring gives a lower bound but no exact value beyond that bound; when many participants are censored early, late event-time behavior is weakly informed by the data. If follow-up ends for reasons related to unobserved event risk, the usual independent-censoring interpretation may be inadequate. Ties and coarse time recording also require a consistent convention for whether an event recorded at the same time as censoring counts as observed. The indicator definition above uses \(T_i\le C_i\); the study’s actual recording protocol should determine and document how ambiguous ties are resolved.
Specify what event is studied and the point from which both event and censoring durations are measured.
For each person, use the earlier of the event time and end of observation: \(Y_i=\min(T_i,C_i)\).
If \(\Delta_i=1\), the event time is \(Y_i\); if \(\Delta_i=0\), the event time exceeds \(Y_i\).
Check reasons and patterns for censoring and consider whether independence is plausible unconditionally or given measured covariates.
Check Your Understanding
Answer each question, then check the solutions below.
- A participant has an event time of 11 months and is observed until month 8 for a non-event reason. What are \(Y\), \(\Delta\), and the information about \(T\)?
- A record has \(Y=6\) months and \(\Delta=1\). What does it establish about the event time?
- Under an event-time model, why does a right-censored record at \(y\) contribute \(S_T(y)\), rather than \(f_T(y)\)?
- Suppose \(S(5)=0.75\) and \(S(9)=0.45\). What is the model-implied probability of an event between months 5 and 9, conditional on being event-free at month 5?
- Give one reason that independent censoring may be implausible and one diagnostic that could prompt further investigation.
Solutions. (1) \(Y=\min(11,8)=8\) months and \(\Delta=0\); the record establishes \(T>8\) months. (2) The event was observed at 6 months, so \(T=Y=6\) months. (3) Censoring at \(y\) establishes that the event time exceeds \(y\), whose probability under the event-time model is \(P(T>y)=S_T(y)\); it does not establish an event at \(y\). (4) \([S(5)-S(9)]/S(5)=(0.75-0.45)/0.75=0.40\), or 40%. (5) For example, participants with worse unmeasured health may be more likely to leave follow-up and also have earlier events. Comparing measured covariates or censoring reasons across observed event and censoring groups could reveal patterns needing investigation, although observed-data diagnostics cannot prove independence.