Why the Classification Matters
An event-history record does not simply say whether a person “had an event.” It says what happened first, when it happened, and whether observation ended before an event was seen. Those distinctions matter because a person who experiences a different event is not in the same position as a person who remains event-free but is no longer observed.
The classification also depends on the question. Death may be the target event in one analysis and a competing event in another. A transition to a new health state may be a competing event for a first-event endpoint, but not necessarily for an endpoint that continues to count outcomes after that transition. Therefore, define the target event and the follow-up rule before coding records.
For example, if the endpoint is time to first hospital admission, death before admission is usually a competing event: after death, a later first admission cannot occur. A person alive and admission-free at the final contact has a right-censored record. In contrast, merely not observing an admission by the final contact does not establish that an admission never occurred afterward; it establishes only that none was observed by that time.
A Record-Level Representation
In a simple right-censored first-event setting, let \(C\) be the time at which observation ends, and let \(T\) be the time of the first event of any relevant type. The recorded follow-up time is \(Y=\min(T,C)\). Define \(\Delta=1\) if an event is observed by the end of follow-up and \(\Delta=0\) if observation ends first. When \(\Delta=1\), the event type \(J\) is recorded as well.
Thus, \(\Delta=1\) is not enough to identify the target event: the type \(J\) must also be checked. The record is a target event when \(\Delta=1\) and \(J=k\); it is a competing event when \(\Delta=1\) and \(J\ne k\); and it is right-censored when \(\Delta=0\). A competing event is observed information about what happened, not an unknown event status.
The basic representation assumes a first-event endpoint, event times measured from a defined origin, and event type ascertainable when an event is recorded. If two event types share a recorded time, the protocol must specify how ties are adjudicated or whether they are represented as a joint event; silently assigning one type is not justified. If onset is known only to have occurred between visits, the event may be interval-censored rather than right-censored. Uncertainty about the exact event date does not turn an observed event into censoring.
Worked Example: Admission, Death, and End of Follow-Up
Consider a hypothetical study whose endpoint is time from enrollment to first unplanned hospital admission. Death before admission is defined as a competing event. Follow-up ends at 365 days. For each record, calculate \(Y=\min(T,C)\) when the event or censoring time is known, and classify the record.
Record A: Admission occurs on day 120, with no earlier event. The event time is \(T=120\), and follow-up would otherwise continue to day 365. Therefore \(Y=\min(120,365)=120\), \(\Delta=1\), and \(J=\text{admission}\). This is a target event.
Record B: The participant dies on day 80 without a prior admission. Here the first event time is \(T=80\), so \(Y=\min(80,365)=80\), \(\Delta=1\), and \(J=\text{death}\). Death is a competing event for this endpoint, not right-censoring: it was observed and prevents a later first admission under the endpoint definition.
Record C: The participant is alive and admission-free at the last verified contact on day 200, after which follow-up is unavailable. The record is right-censored at \(C=200\): \(Y=200\) and \(\Delta=0\). The data support “known admission-free through day 200,” not “never admitted.”
Record D: An admission is first reported at a visit on day 150, and records establish that it occurred after the day-100 visit but do not establish its exact date. Its type is a target event, but its timing is known only to lie in the interval \((100,150]\). It is not a right-censored record: an admission is known to have occurred. An analysis must retain the interval information rather than inventing an exact event time.
Why a Competing Event Is Not Ordinary Censoring
For a target event \(k\), the cumulative incidence by time \(t\) is \(F_k(t)=P(T\le t,J=k)\): the probability that the target event is the first event by time \(t\). A person with a competing event has a known first-event outcome, but cannot subsequently experience the target event as the first event. A right-censored person, in contrast, has no observed first event through the censoring time; what happens later is unknown.
This distinction can be seen by partitioning people by their first outcome by time \(t\). The target-event probability counts those whose first event is the target by \(t\). Competing-event cases belong to another observed first-event category, while people with no observed event by \(t\) remain event-free through that time (or may have incomplete follow-up). Treating the competing category as if its members simply disappeared from observation erases a real outcome category.
A short numerical calculation shows the consequence. Suppose, hypothetically, 100 people are followed completely to one year. Twenty experience a competing event on day 100, then 20 of the remaining 80 experience the target event on day 200; the other 60 have neither event by one year. The observed target-event proportion is \(20/100=0.20\), or 20%.
If the 20 competing events are incorrectly treated as ordinary censoring for a product-limit calculation of target-event risk, the first 20 are removed from the risk set before day 200. The calculation then uses \(20/80=0.25\) as the target-event fraction at day 200, yielding a target-event estimate of 25% in this simplified example. The 25% is not the observed one-year probability of first experiencing the target event. It results from treating competing outcomes as though their later target-event status were simply unobserved.
This does not mean a competing event can never be removed from a risk set in another analysis. For example, a cause-specific hazard analysis counts individuals as at risk for a cause up to their first event, then no longer at risk afterward. That bookkeeping does not make the competing event ordinary censoring, and it does not by itself turn a cause-specific hazard into the target-event cumulative incidence. The estimand and the method must match.
Worked Example: The Same Death, Two Different Endpoints
In a hypothetical cohort, a participant dies from cardiovascular disease on day 300. Consider two analyses, both starting at enrollment.
Endpoint 1: time to cardiovascular death. Cardiovascular death is the target event. If no other event in the endpoint definition occurs first, this record is a target event at \(Y=300\), with \(\Delta=1\) and \(J=\text{cardiovascular death}\).
Endpoint 2: time to cancer death. Cancer death is the target event, and cardiovascular death is a competing event because it prevents a later cancer death from being the first death. For this endpoint, the same observed death is a competing event at \(Y=300\), with \(\Delta=1\) and \(J=\text{cardiovascular death}\).
The observed facts have not changed; the target endpoint has. Coding “death” as a single undifferentiated event would therefore be insufficient for either analysis. The cause and endpoint definition together determine the classification.
Endpoint Decisions and Practical Checks
Before assigning a status code, verify that the recorded event matches the endpoint definition and that the follow-up time has the intended origin and units. A person’s status at the last contact must be checked against the event history: a last-contact date is a censoring time only if the person is known to be free of the relevant events through that date. An event recorded after the nominal end of follow-up must not be pulled into the analysis without an explicit follow-up rule.
The following comparison is useful when reviewing data records:
| Record at the observed time | What is known? | Classification for a target-event endpoint |
|---|---|---|
| Target event occurs first | The target event and its time are observed. | Target event |
| Different qualifying event occurs first | A non-target first event and its time are observed; the target can no longer occur as defined. | Competing event |
| Follow-up ends with no qualifying event observed | The person is known to be event-free through the recorded time; later status is unknown. | Right-censoring |
| Event is established between visits, exact time unknown | The event type is known, but its time is only bounded by observations. | Interval-censored event, not right-censoring |
A common mistake is to code every non-target outcome as censored because “the target did not happen.” That reasoning confuses “not observed by the end of follow-up” with “a different event was observed first.” The correction is to record event status and type separately: an observed competing event is an event, while right-censoring records an event-free observation limit. Whether censoring can be handled as noninformative is a separate assumption; classification alone does not establish it.
Diagnostics for this basic coding task are concrete. Tabulate event counts by type and by follow-up status; inspect records where an event date equals the last-contact date; verify that all event codes correspond to the written endpoint definition; and check for impossible combinations such as \(\Delta=0\) paired with a recorded event type. Investigate same-day event ties and cases with unknown cause rather than silently recoding them. If the cause is unknown, the record is not automatically right-censored: an event may be known to have occurred even when its type is missing.
The main alternatives answer different questions. A single-event analysis may distinguish only target events from event-free censoring when other event types truly are outside the endpoint or cannot occur. A competing-risks formulation retains distinct first-event types and describes their probabilities. A multi-state formulation can represent transitions among several states, including transitions after an initial event, when those later transitions remain part of the scientific question. None of these choices can be made from the status code alone; the endpoint and observation structure determine what the record means.
Check Your Understanding
Classify each record for the endpoint stated. Then check the concise solutions below.
- For time to first stroke, a participant has a stroke on day 90. What is the record type?
- For time to first stroke, a participant dies without a stroke on day 70. What is the record type, assuming death prevents a later first stroke?
- For time to first stroke, a participant is alive and stroke-free at the last contact on day 250. What is the record type?
- A participant’s stroke is known to have occurred between visits on days 40 and 60, but its exact date is unavailable. Is this ordinary right-censoring?
- Explain why an observed competing event cannot be described as merely “a person whose target-event status is unknown.”
Solutions. (1) Target event: the stroke is the endpoint event and is observed on day 90. (2) Competing event: death is observed first and precludes a later first stroke under the stated definition. (3) Right-censoring at day 250: the participant is known to be event-free through the last contact. (4) No. The event type is known, while its time is interval-censored between days 40 and 60. (5) The competing event is a known first outcome that prevents the target from occurring as defined; unknown later status characterizes censoring, not this observed event.