Why an Unresolved Event Is Not a Non-Event
In a trial with a time-to-event endpoint, a patient who has not yet experienced the event may still experience it later. If an allocation decision is made before that patient’s follow-up is complete, the event status is unresolved at the decision time. Recording that patient as a non-event would convert “no event observed yet” into “no event by the endpoint,” which is a different statement.
The appropriate record preserves how long the patient has been observed and whether the event was observed during that time. These are the essential ingredients for representing a right-censored observation. They allow a later analysis to distinguish a short period without an event from a longer one, rather than treating both as the same binary response.
The pair \((Y_i(d),\Delta_i(d))\) is a compact description of the event-time information available at the cutoff. A complete operational record should also retain relevant dates and the reason observation ended. For example, a patient still being followed at the cutoff and a patient lost to follow-up earlier may both have \(\Delta_i(d)=0\), but their observation status and implications for future data collection differ.
Three Statuses That Must Not Be Confused
At a particular cutoff, “no event recorded” can describe several situations. If the patient remains under observation and has not yet reached the endpoint’s required follow-up, the outcome is pending. If observation ended before the event, the event time is right-censored; the patient is known to have remained event-free through the observed time, but later event status is unknown. If the patient has completed a prespecified follow-up horizon without the event, the patient is a non-event for that horizon-specific binary endpoint. That last classification does not mean that the patient can never experience the event after the horizon.
| Situation at the data cutoff | What is known | Appropriate interpretation |
|---|---|---|
| Follow-up ongoing; horizon not reached | No event through the observed time | Pending; right-censored at the cutoff for the current record |
| Observation ended before the event | No event through the last observed time | Right-censored; later event status is unknown |
| Prespecified horizon reached without an event | No event through the full horizon | Resolved non-event for the horizon-specific endpoint |
| Event observed during follow-up | Event time is known | Event indicator equals one |
For adaptive randomization, this distinction matters because the trial can only use information available by the allocation decision time. A patient enrolled recently may contribute a short event-free follow-up, but not the same evidence as a patient who has remained event-free through a long, completed follow-up period. A zero event indicator records censoring at an observed time; it is not a resolved binary response.
Worked Example: Building Records at a Calendar-Time Cutoff
Suppose a hypothetical trial freezes its data on day 50. All times below are measured in days from randomization. Patient A was randomized on day 10 and experienced the event 18 days later. Patient B was randomized on day 20, remains under observation, and has not experienced the event by the cutoff. Patient C was randomized on day 5, was lost to follow-up after 22 days, and had no observed event before loss. Construct the records available at day 50.
For patient A, \(\tau_A(50)=50-10=40\). The event occurred at \(T_A=18\), before the cutoff and before any censoring. Therefore \(Y_A(50)=18\) days and \(\Delta_A(50)=1\). This is an observed event, not a censored record.
For patient B, \(\tau_B(50)=50-20=30\). The patient has no observed event and is still under follow-up, so the current record is \(Y_B(50)=30\) days and \(\Delta_B(50)=0\). The event time is pending: the data show event-free follow-up through day 30, not a completed non-event.
For patient C, the available follow-up ends at loss on day 22 after randomization, before the data cutoff. Thus \(Y_C(50)=22\) days and \(\Delta_C(50)=0\). This record is right-censored at loss to follow-up. It differs operationally from patient B, whose follow-up is ongoing, even though both event indicators are zero.
A concise data table would therefore include \((18,1)\), \((30,0)\), and \((22,0)\) for \((Y,\Delta)\), together with status or censoring-reason fields. Coding B and C as “no event” without their follow-up times would discard information and misstate what is known.
Fixed Horizons and Resolved Binary Outcomes
A time-to-event endpoint can also induce a binary endpoint at a prespecified horizon \(H\): whether the event occurred by time \(H\). A patient whose event occurs at or before \(H\) is an event for that binary endpoint. A patient observed event-free through \(H\) is a non-event by \(H\). A patient with only \(u<H\) days of event-free follow-up has not yet resolved the binary endpoint at \(H\).
The horizon must be specified in advance and the patient must actually have event ascertainment through that horizon to call the outcome resolved. If the event does not occur by \(H\), the exact event time after \(H\) is still unknown; the binary statement is only that the event did not occur by \(H\). This distinction prevents a horizon-specific outcome from being confused with a claim about lifetime event status.
Worked Example: Determining Whether a 90-Day Outcome Is Resolved
Consider a hypothetical endpoint of first hospitalization within 90 days after randomization. At a data cutoff, three patients have the following records: patient D had a hospitalization on day 63; patient E remains under observation and is event-free through day 90; patient F remains under observation and is event-free through day 41.
Patient D experienced the event before day 90, so the 90-day binary outcome is a resolved event. Its observed event-time record is \(Y_D=63\) days and \(\Delta_D=1\).
Patient E completed 90 days of event-free observation. The 90-day binary outcome is a resolved non-event. For a record truncated at the horizon, \(Y_E=90\) days and \(\Delta_E=0\). This says “no hospitalization by day 90,” not “no hospitalization ever.”
Patient F has only 41 days of follow-up. Its current record is \(Y_F=41\) days and \(\Delta_F=0\), but the 90-day binary outcome is unresolved. It is incorrect to label F a 90-day non-event because the remaining 49 days of the specified risk period have not been observed.
Why Follow-Up Time Carries Information
A censored record is not an observed event time, but it is not devoid of information. Under a model with event-time density \(f(t)\) and survival function \(S(t)=P(T_i>t)\), an observed event at time \(y\) contributes event-time density \(f(y)\). A patient known to remain event-free through \(y\), but censored at \(y\), contributes the probability \(S(y)\) of surviving beyond that observed time.
This likelihood expression relies on an assumption about censoring: the censoring process must be independent of the event time, either unconditionally or after conditioning on variables included in the model. If patients are more likely to leave observation because their event risk has changed, treating censoring as independent may give misleading conclusions. The pair \((Y_i,\Delta_i)\) remains a correct description of what was observed, but a model for the censoring process or other methods may be needed for valid inference.
Worked Example: Likelihood Contributions for Event and Censored Records
Suppose, hypothetically, that event times follow an exponential distribution with rate \(\lambda=0.02\) per day, so \(f(y)=\lambda e^{-\lambda y}\) and \(S(y)=e^{-\lambda y}\). Three patients have records: an event at day 10, a right-censoring time of day 20, and an event at day 5. Assume censoring is independent of event time for this calculation.
The event at day 10 contributes \(f(10)=0.02e^{-0.02(10)}=0.02e^{-0.2}\). The censored observation at day 20 contributes \(S(20)=e^{-0.02(20)}=e^{-0.4}\). The event at day 5 contributes \(f(5)=0.02e^{-0.02(5)}=0.02e^{-0.1}\).
Multiplying the contributions gives
The two observed events contribute rate-density factors, while the censored patient contributes evidence of remaining event-free for 20 days. The likelihood is a density-based quantity, not the probability that the null hypothesis is true or the probability that a particular patient will experience the event.
Recording and Checking Unresolved Outcomes
A practical record should make it possible to reconstruct what was known at each allocation decision. At minimum, preserve the randomization date, event date if observed, last date with reliable event ascertainment, data cutoff, and censoring reason. Derive follow-up duration and the event indicator from those dates using a stated convention. If the endpoint is defined at a fixed horizon, also record whether that horizon has been fully observed.
Several checks can reveal errors before the data are used:
- Verify that observed follow-up is nonnegative and does not extend beyond the data cutoff or a prespecified endpoint horizon.
- For every record with \(\Delta_i=1\), confirm that an event date is available and falls within the observed follow-up period.
- For every record with \(\Delta_i=0\), confirm that the patient was known to be event-free through the recorded censoring time.
- Separate ongoing administrative censoring at the data cutoff from loss to follow-up or another reason observation ended.
- For a horizon-specific binary endpoint, flag patients with less than \(H\) days of event-free observation as unresolved, not as non-events.
These checks do not establish that censoring is statistically independent of event risk. They establish that the event-time record matches the actual observation history. Independence, when needed for a particular analysis, is a separate modeling assumption that should be assessed from the study context and addressed explicitly.
Common Mistake and Alternatives
A common mistake is to code every patient with \(\Delta_i=0\) as a binary non-event. The mistaken reasoning is: “No event is recorded, so the patient had no event.” The correction is that the record supports only “no event was observed through \(Y_i\).” A resolved non-event at horizon \(H\) requires ascertainment through \(H\); otherwise the patient’s horizon outcome remains pending.
The appropriate representation depends on the question being answered. A fixed-horizon binary response is suitable once each patient’s horizon status is resolved. A time-to-event record retains both event timing and the amount of event-free follow-up, including information from patients who have not completed follow-up. Complete-case analysis, which simply excludes unresolved patients, avoids labeling them non-events but discards their observed follow-up and can be misleading when inclusion depends on enrollment time or follow-up length. None of these representations makes incomplete event information disappear; the data structure should match the endpoint and decision being considered.
| Approach | What it represents | Main limitation before all follow-up is complete |
|---|---|---|
| Time-to-event record \((Y,\Delta)\) | Observed event time or event-free follow-up through censoring | Requires appropriate survival analysis assumptions for inference |
| Fixed-horizon binary outcome | Event by \(H\) versus no event by \(H\) | Unresolved until event or complete observation through \(H\) |
| Complete-case analysis | Only outcomes already classified as complete | Can discard pending patients and their partial follow-up information |
The central limitation is that censoring does not reveal what happened afterward. A carefully recorded \(Y_i\) and \(\Delta_i\) preserve the information actually observed; they do not justify predicting or imputing a future event status without additional assumptions. For an adaptive trial, maintaining that distinction prevents an incomplete follow-up record from being mistaken for a completed response.
Check Your Understanding
Answer each question, then compare with the solutions below.
- A patient is randomized on day 12 and remains under observation without an event through a data cutoff on day 37. What are the observed follow-up time and event indicator, and is a 60-day outcome resolved?
- A patient has no event during 25 days of follow-up, then is lost to follow-up. Explain what is known and why this is not automatically a non-event at a 60-day horizon.
- For a right-censored record with \(Y_i=14\) days and \(\Delta_i=0\), what event-time statement is supported by the record?
- Under the independent-censoring model, write the event-time likelihood contribution for an observed event at time \(y\), and for censoring at time \(y\).
- Give one reason that two records with \(\Delta_i=0\) may need different status labels in a trial database.
Solutions. (1) Follow-up is \(37-12=25\) days and \(\Delta_i=0\). The 60-day outcome is unresolved because only 25 of the required 60 days have been observed. (2) The patient is known to be event-free through day 25, so the record is right-censored at 25 days. The patient’s status through day 60 is unknown, so the 60-day binary outcome is not resolved. (3) The supported statement is \(T_i>14\) days: the patient remained event-free through the observed 14 days. No claim about later event status follows. (4) An observed event contributes \(f(y)\); censoring at \(y\) contributes \(S(y)=P(T_i>y)\). (5) One patient may still be under follow-up and censored at the current data cutoff, whereas another may have been lost to follow-up earlier. Both have a zero event indicator, but the censoring time and reason differ.