Define Competing Events Before Estimating Event Probability

Start with the event that can no longer occur

A competing event prevents the target event from occurring under the specified endpoint definition. Death before a first nonfatal outcome is a common example, but the classification depends on the scientific question and observation process. Treatment discontinuation or transfer is not automatically a competing event merely because follow-up becomes difficult.

Write the initial state, target event, possible competing events, and ordinary censoring reasons. Decide whether the question concerns the first event, repeated events, or transitions between states. A person can be eligible for a different analysis after an event even when the first-event analysis has ended.

Preserve event type and time

Keep a status field that distinguishes target events, competing events, and censoring, with documented codes. Store the relevant dates or elapsed times separately from the status. If two events are recorded at the same time, resolve their ordering or tie rule from the protocol and source information rather than choosing the category that produces a preferred result.

The survival package authors' multi-state and competing-risks tutorial explains probability-in-state estimation and the connection with first-event competing-risk analyses. It provides an appropriate framework for deciding what a curve represents.

Inspect a fictional denominator contrast

Imagine ten people followed completely through day five. One dies without the target event on day one; two others experience the target event on days two and three. The remaining seven have neither event by day five. The observed target-event probability in this complete, fixed-horizon example is 2/10 = 20%.

If the day-one death were instead treated as ordinary censoring in a target-only Kaplan–Meier calculation, the later factors would be 8/9 and 7/8. One minus their product is 2/9, approximately 22.22%. This illustrates how a different treatment of the competing event changes the quantity being estimated; it is not a real clinical result.

Match interpretation to the method

For a target-event probability in the presence of competing events, consider an appropriate cumulative-incidence analysis. A cause-specific hazard model can answer a useful question, but its hazard ratio is not itself the cumulative probability of the target event. Avoid relabeling a target-only survival complement as that probability without checking the required assumptions and estimand.

The risk-set worksheet can support a timing audit, while the survival-risk-set tool explicitly does not compute competing-risk cumulative incidence. Keep those boundaries clear when using simple examples to explain a full analysis.

Finish with the event taxonomy, time origin, handling of simultaneous events, and the probability or hazard quantity to be reported. Review plausible event histories manually before fitting software. A correctly coded numeric status variable still needs a scientifically coherent definition of what each state means.

Source checked on 4 November 2026; fictional probabilities independently recalculated.