Cause Labels And The Meaning Of A Competing Event

Why the Cause Label Depends on the Question

A clinical outcome does not carry an intrinsic label of “competing event.” Its role depends on the endpoint being studied. A death before a planned hospitalization can compete with that hospitalization as a first event, while the same death may itself be part of the focal endpoint in an analysis of a broader clinical composite. The patient’s outcome has not changed; the question and its event classification have.

This distinction matters because cause labels determine which observations count as the event of interest and which count as other first events. Those classifications are part of the estimand: they specify what event the analysis is about. They are not merely cosmetic labels applied after the analysis.

The definition concerns the event being analyzed, not every event that may eventually happen to a patient. If the endpoint is time to first heart-failure (HF) hospitalization, a death before HF hospitalization is a different first event. An HF hospitalization after that death cannot occur; an HF hospitalization after a transplant may be possible clinically, but it is no longer the first event in this particular endpoint. A first-event analysis therefore treats the endpoint’s event categories as mutually exclusive: each person can have at most one observed first-event label.

This framework assumes that the endpoint, time origin, and eligible first-event categories have been stated clearly; that the event categories do not overlap under the coding rule; and that each qualifying first event receives one label. If the data cannot support a unique label, the classification rule needs to be clarified before interpreting cause-specific results. Questions about simultaneous labels and event misclassification require separate treatment.

One Set of Outcomes, Two Clinical Questions

Consider a hypothetical cohort of 100 patients followed for 12 months. Suppose each patient’s first relevant outcome by 12 months, if any, is classified as follows. Assume for this illustration that follow-up and outcome ascertainment are complete through the stated horizon.

For the endpoint “time to first HF hospitalization,” HF hospitalization is the focal event. Each of the other three event types is competing with that focal first event: a cardiac death, noncardiac death, or transplant was recorded first. The 70 patients with none of these outcomes by 12 months have no recorded first event in the table by that horizon.

Now consider a different question: “time to first cardiac deterioration,” defined in advance as either HF hospitalization or cardiac death. Under this definition, the 8 cardiac deaths are no longer competing events. They are part of the focal endpoint. Noncardiac death and transplant remain other first-event types. The same patients and outcomes produce different event counts because the endpoint has changed.

Cause Labels and Cause-Specific Hazards

A cause-specific hazard describes the instantaneous rate of a specified event among people who have not yet had any event in the defined first-event system. For cause \(k\), one standard definition is

Here \(T\ge t\) denotes being event-free just before time \(t\), and \(\Delta\) is a short time interval. The cause label \(k\) is meaningful only after the endpoint and the categories defining \(J\) have been specified. In the HF-hospitalization endpoint, cardiac death is one of the alternative first-event types, not a focal HF hospitalization. In the cardiac-deterioration endpoint, cardiac death is included in the focal label.

Changing a classification can change the hazard for the focal event even when the underlying clinical records are identical. With the composite definition, a cardiac death occurring in a short interval is counted as a focal event alongside an HF hospitalization. The denominator is also defined by the endpoint’s first-event system: after a first event of any included type, the person is no longer event-free for that system. Consequently, hazard comparisons are interpretable only when the event definitions and risk-set meanings are aligned.

A useful special case is a broader cause-specific category formed by combining disjoint event types. Suppose \(A\) and \(B\) are mutually exclusive labels under the same first-event definition. For a person event-free at \(t\), the events \(J=A\) and \(J=B\) cannot occur as two different first events in the same short interval. Therefore,

Divide both sides by \(\Delta\) and take the limit as \(\Delta\to0^+\). The cause-specific hazard for the grouped category is then

The equality depends on using the same event-free condition and combining disjoint labels. It does not mean that any two quantities called “hazards” can be added; their definitions and denominators must match. In particular, a composite endpoint should be constructed by defining its categories first, not by combining reported results without checking their underlying definitions.

Choosing a Classification That Matches the Clinical Aim

A detailed classification and a composite endpoint answer different questions. Keeping HF hospitalization and cardiac death separate can help describe each type of first event. Combining them can instead answer a broader question about first cardiac deterioration. Neither is automatically preferable: the choice depends on the clinical aim, and it should be made before examining the results.

The following checks help determine whether a cause scheme is usable and interpretable:

• State the focal endpoint in words. “Cardiac events” is not precise enough unless the included event types are listed.
• Specify the first-event rule. Say whether the endpoint is the first occurrence of any listed event, and name the time origin.
• Check the alternative categories. For the chosen focal event, list which observed first outcomes fall outside it and whether they prevent that focal first event from being observed.
• Check that labels are disjoint and complete. Under the coding rule, every qualifying first event should receive one and only one category.
• Keep the question fixed in comparisons. A cause-specific hazard for first HF hospitalization is not the same target as a hazard for a composite of HF hospitalization and cardiac death.

One limitation is that a first-event classification does not describe the full subsequent clinical course. A patient may experience several events over time, but a first-event endpoint records only the first qualifying event. A different scientific question may require a different endpoint or model. Also, a clinically sensible composite can conceal differences among its components: a higher composite event count alone does not show which component accounts for the difference.

Common Mistake: Treating “Competing” as an Intrinsic Property

Wrong reasoning: “Cardiac death is always a competing event for HF hospitalization, so it stays a competing event in every analysis.”

Correction: Cardiac death is a competing first event when the focal endpoint is first HF hospitalization. If the focal endpoint is defined as first HF hospitalization or cardiac death, cardiac death belongs to the focal category instead. The role follows from the endpoint definition, not from the event’s name alone.

A related diagnostic warning is a table or report that gives a focal-cause label but does not define which outcomes were grouped into it. Without that information, the competing-event classification—and therefore the meaning of the reported cause-specific result—cannot be reconstructed. Ask for the operational event definitions and their mutually exclusive coding rules before interpreting the analysis.