Competing Events Are Not Censoring

Two Different Reasons Follow-up Ends

In a first-event analysis, follow-up can end because an event happens or because observation ends before an event is seen. These are different data outcomes. A competing event is an observed event of another type; censoring means that observation ended without revealing which event would occur next. Confusing them can turn a question about observed event experience into a different question about a hypothetical event process.

Suppose the outcome of interest is the first hospitalization for a particular condition. A death before that hospitalization is a competing event: it is a known first event that prevents a later first hospitalization from occurring. By contrast, a participant who moves away while still alive and has no recorded hospitalization is censored at the last date their status is known. The participant's subsequent event history is unknown.

This definition assumes a first-event framework: event types refer to alternative types of the first event, and only the first event is being analyzed. When a competing event occurs, the first-event outcome is known even though no later event history is available or relevant to that endpoint. When censoring occurs, the first-event outcome may still be unknown.

Classifying Records Correctly

For a given cause of interest, distinguish three states: the cause of interest has occurred; a different event has occurred; or observation ended before any event was observed. The second and third states can both mean that no event of the cause of interest was recorded during the observed period. But they do not mean the same thing: in the second, another event is known to have happened; in the third, the later event history is unknown.

A crucial detail is that the first two outcomes are observed events, while the third is incomplete observation. A competing event ends a participant's time at risk for a later first event of the focal cause. That fact can make it look like censoring in a particular risk-set calculation. The reason the participant leaves the risk set, however, remains different: an observed competing event is not a loss of information at that time.

Why the Distinction Changes the Calculation

For a cause-specific hazard, the risk set at a time consists of individuals who have not yet had any first event and whose observation has not already ended. A competing event removes an individual from this risk set after the event, because that individual can no longer have a later first event of the focal cause. Censoring also removes an individual from subsequent risk sets, but because observation has ended and later event status is unknown. The risk-set arithmetic can therefore involve removal in both cases; the records and interpretation are not interchangeable.

A short discrete-time example shows why the distinction matters for event probabilities. Suppose all 100 hypothetical participants are followed completely through two intervals. In interval 1, 10 have the focal event and 20 have a competing event, leaving 70 event-free. In interval 2, among those 70, 14 have the focal event, 21 have a competing event, and 35 remain event-free through the end of follow-up.

The observed proportion with the focal event by the end of interval 2 is \(24/100=0.24\): 10 participants have the focal event in the first interval and 14 in the second. The 41 competing events are known outcomes, not missing follow-up. If, instead, the competing events are treated as censoring in a product-limit calculation for the focal event, the calculation is

The result, 0.28, is not the observed proportion with the focal event in this hypothetical cohort. It results from removing competing events as if they ended observation, then calculating as though those participants could remain relevant to later focal-event experience. In this example, the product-limit calculation is a net-risk calculation under a hypothetical removal of competing events, not the probability of the focal event in the world where those competing events occur.

A Justification: What the Data Identify

The distinction follows directly from the information recorded. For a participant with a competing event at time \(t\), the data establish that the first event occurred by \(t\), and that its type was not the focal cause. For a participant censored at time \(c\), the data establish only that no event had been observed through \(c\). Thus, even if both participants contribute no focal event after their respective times, only the competing-event record supplies a known first-event outcome.

This difference also explains why simply counting people who experienced the focal event and dividing by the original cohort size does not generally estimate the focal-event probability when follow-up is incomplete: some censored participants may have had an event later. In contrast, when follow-up is complete to a fixed horizon, every participant's first-event status by that horizon is known, and the observed focal-event proportion has a direct interpretation. In analyses with incomplete follow-up, handling censoring requires its own assumptions and methods; it does not make competing events into censored observations.

What Similarity Does—and Does Not—Mean

In a cause-specific risk-set calculation, both a competing event and censoring can end a participant's future contribution to the risk set. That limited procedural similarity sometimes leads to the mistaken label “censored for the other cause.” It is safer to say precisely what is meant:

• For a cause-specific hazard, a participant who has a competing event exits the risk set after that event because a later first event of the focal cause is no longer possible.
• A censored participant exits because observation ends, leaving the later first-event status unknown.
• For the probability of a particular cause as the observed first event, competing events must remain part of the event history. Treating them as censoring changes the target calculation.

The introductory material on cause-specific and subdistribution hazards distinguishes the corresponding hazard definitions. Here the practical point is narrower: a participant can be removed from one risk set without having been censored in the data. Conversely, a censoring label must not be used to conceal an observed event type.

Diagnostics, Limitations, and a Common Mistake

Check the data structure before interpreting a result. For each participant, ask whether there is a recorded event time and event type, or only a last-known event-free time. Confirm that the event type refers to the first event under study and that an event of another type truly precludes the focal first event. If records contain a code such as “other,” verify whether it means an observed event of another type or merely an unknown event type; the label alone is not enough.

A dataset can also be insufficient to distinguish competing event from censoring. For example, a record of “no focal event” without a follow-up end date or a known alternative event type does not establish either status. Such a record needs clarification rather than an assumed event classification. This tutorial concerns event classification and its interpretation; it does not establish that censoring is independent of event risk. That is a separate assumption, and the distinction here does not by itself validate any censoring adjustment.

The main limitation is that this framework describes a defined first-event endpoint. If the scientific question concerns recurrent events, events after the first event, or a sequence of transitions among states, a first-event label alone does not represent the full history. The event structure must match the question.