Defining Causes, Competing Events, and Censoring in Methods

Why Definitions Belong in the Methods

A competing-risk analysis is only interpretable if readers can tell what counted as an event, which events were treated as competing, and when follow-up ended. Those decisions are not clerical details. For example, “cardiovascular event” could mean a first myocardial infarction, any hospitalization for a cardiovascular diagnosis, or cardiovascular death. Those definitions produce different observed event histories. A reproducible methods description lets another analyst classify the same records in the same way.

The aim here is to write the event and follow-up rules clearly, not to choose an estimand or report its estimate. The earlier material on choosing and naming the reported competing-risk estimand addresses that separate decision. Here, the practical task is to state exactly how the records were converted into event types and observed follow-up.

For a simple first-event analysis, let \(T_i\) be subject \(i\)'s time from the stated time origin to the first qualifying event, and let \(C_i\) be the time from that origin to censoring. The observed time is \(Y_i=\min(T_i,C_i)\). If an event is observed, its cause label \(J_i\) identifies the event type; if follow-up is censored first, \(J_i=0\) denotes censoring. With \(K\) defined event types, an observed cause label is one of \(1,\ldots,K\). This notation is useful only when the event categories and the censoring rule are clearly specified.

A competing event is therefore not simply “an event other than the one of interest.” It must be an observed outcome that prevents the target first event from occurring later in the defined event process. Death before a nonfatal relapse is a common example: after death, a later relapse cannot occur. By contrast, an administrative end of observation does not establish that relapse will never occur; it is censoring. If events can recur, or one event does not preclude another, the study endpoint and event-history structure need to say so rather than forcing every outcome into a first-event classification.

Write the Classification Rules Before Applying Them

A useful methods description makes the classification reproducible by answering four questions: what qualified, when it counted, how its cause was assigned, and what happened when assignment was uncertain or more than one cause seemed applicable. Define the event in observable terms, such as a prespecified diagnostic criterion, a laboratory threshold, or a documented clinical event. State the data sources used to establish it and any adjudication process. A broad label without operational criteria leaves analysts to make unstated judgments.

For first-event categories, the recorded cause labels should ordinarily be mutually exclusive and collectively cover all observed qualifying first events. If one record meets criteria for two categories, a hierarchy, adjudication rule, or explicit “multiple/uncertain cause” category is needed. Do not silently assign the event to whichever category is most convenient. If a genuinely overlapping event process is scientifically important, a multi-state representation may be more appropriate than a forced single cause label.

Cause assignment can be based on a prespecified source, such as a blinded adjudication committee, a registry field, or a defined combination of clinical records. State how disagreements were resolved and how unclassifiable events were handled. “Unknown cause” is not automatically censoring: the event itself may be known to have occurred even when its cause is not. The analysis plan should distinguish missing cause information from absence of an event and explain whether such cases are retained as a separate category, excluded from cause-specific classification, or addressed through a sensitivity analysis.

Time Origin, Entry, and End of Follow-Up

“Follow-up began at enrollment” is not sufficiently precise when enrollment can occur long after diagnosis, treatment, or another clinically meaningful date. State the time origin and the rule that determines each participant’s starting point. Examples include randomization, diagnosis, surgery, or a fixed eligibility date. Also state the time scale and units, such as days since surgery. If participants become eligible after the origin and enter observation later, report both the origin and the entry rule; otherwise, readers may mistake delayed observation for absence of earlier events.

Define the end of follow-up as a rule, not just a date label. Common endpoints include the first qualifying event, death, withdrawal of permission for further follow-up, loss to follow-up, or a common administrative closing date. When multiple endpoints could apply, say which earliest date determines the observed time. Specify whether the event date itself is included and how dates with limited precision are handled. These details matter when event dates are recorded only by month or when several events share a calendar date.

Censoring should describe the reason observation stopped. Administrative censoring is generally set by a common study closing date; loss to follow-up occurs when continued event ascertainment becomes unavailable; withdrawal may stop data collection under the applicable rules. These mechanisms are not interchangeable descriptions. For inference, censoring assumptions also matter: commonly, the analysis requires censoring to be independent of future event experience, at least conditional on variables included in the analysis. A clear definition helps assess that assumption but does not prove it.

Ambiguous, Multiple, and Simultaneous Causes

Clinical records may use inconsistent terminology, list several diagnoses, or fail to identify a single underlying cause. The report should say whether causes were determined by a hierarchy of coded diagnoses, adjudication, or another prespecified rule. If a hierarchy is used, state its ordering and justify it clinically. If cases may remain unresolved, state the label assigned to them and how they enter the analysis. The rule should be applied consistently without access to information that would make classification depend on the desired result.

When two qualifying events have the same recorded date, do not imply an ordering that the data cannot support. Specify whether more precise timestamps were used, whether a prespecified priority rule was applied, or whether the record was classified as simultaneous/ambiguous. If the analysis requires one first event, the chosen resolution rule must be transparent and its potential effect considered. A tie rule should not be invented after inspecting which classification gives more favorable results.

Justification: Why a Competing Event Is Not Censoring

The distinction follows directly from what each recorded outcome says about the future event process. A competing event is an observed outcome that rules out the target event as a later first event. Censoring says only that event observation stopped at a particular time; the target event may occur later, but it is no longer observed. Treating a known competing event as censoring discards a real outcome and changes the interpretation of the data.

The first line records an event and its type; the second records incomplete event-time observation. If death is a competing event for relapse, replacing its event label with \(J_i=0\) would falsely describe death as mere loss of observation. Conversely, assigning a competing-event label to a participant whose records simply stop would claim an outcome that was not observed. This is a classification distinction, not a claim that censoring is harmless: inference still depends on appropriate assumptions about how censoring relates to future outcomes.

Methods Wording and Diagnostic Checks

A concise methods paragraph can be assembled from explicit rules rather than generic labels. For example: “Time zero was the date of [origin]. The event of interest was [operational criteria], dated at [date rule] and identified from [source]. [Named event] before the event of interest was treated as a competing event. Participants without either event were censored at the earliest of [specified endpoints]. Cause assignment used [source/adjudication rule]; simultaneous or unclassifiable events were [rule]. Follow-up was measured in [units] from time zero.” Replace each bracketed phrase with a study-specific definition.

Before finalizing the methods, check that every observed record can be assigned by the stated rules and that the assignment does not rely on unstated judgment. Useful diagnostics include:

• Look for events with no cause label, multiple cause labels, or a date that precedes the stated follow-up origin.
• Check whether “unknown cause,” “no event,” and “lost to follow-up” have been mistakenly combined.
• Review records with same-day events and verify that the stated tie rule matches the available date precision.
• Confirm that the stated endpoint actually precludes the target event as a later first event; if not, a competing-event label may be inappropriate.
• Compare the written censoring rule with the dates used to create the analysis records, including administrative closure and last-known-status dates.

The main alternatives should be distinguished by their scientific meaning. A first-event competing-risk definition assigns one earliest event type or censoring status. A composite endpoint instead counts any event meeting one or more stated criteria as the same outcome, sacrificing separate cause labels. A multi-state description can retain sequences or transitions when one event does not end the possibility of another. These are not interchangeable coding conveniences: choose the representation that matches the question, then report its event definitions and rules.

A common mistake is to write, “Non-target events were censored at their occurrence.” That is wrong when a non-target event is an observed event that precludes the target first event: it is a competing event, not censoring. The correction is to name the event and its classification explicitly. A second mistake is to treat a missing cause as if no event occurred. State separately that an event was observed but its cause was unresolved, and document the prespecified handling rule.