Reporting Cohort Flow and Cause-Specific Event Counts

Why Cohort Flow and Event Counts Matter

A competing-risk result is only interpretable if readers can tell who was analyzed, how much follow-up was observed, and what happened first. A concise flow account connects the people initially considered with the analysis population. A count of each first-event type then shows how that population’s follow-up ended by the data cutoff. These are descriptive accounting tasks: they do not, by themselves, estimate a risk or compare groups.

The account should be consistent with the endpoint and analysis population defined in the methods. In a first-event competing-risks analysis, a person contributes one terminal classification during the stated follow-up: the event of interest, one of the competing event types, or censoring before any event. A competing event is an observed event, not a censored observation. This distinction is essential to an auditable report.

Set the Accounting Rules Before Reporting Counts

Before making a flow diagram or table, specify the population, time origin, event definition, event types, censoring rule, and data cutoff. The time origin might be diagnosis, randomization, or another prespecified point. Follow-up duration is measured from that origin to the first event or censoring time. Where entry occurs after the origin, the report should make clear who enters the analysis population and when their at-risk follow-up begins; do not imply they were observed before entry.

Eligibility exclusions should be reported with enough detail to explain the difference between the initial group and the analysis population. If exclusions are applied in stages, report the number remaining at each stage. If categories are intended to be mutually exclusive, define the assignment rule so that one person is counted only once. Alternatively, if reasons can overlap and the counts are not mutually exclusive, say so explicitly; their sum will not then equal the number excluded.

For the analysis population, report an exhaustive set of terminal statuses at the cutoff. Under a first-event endpoint with \(K\) mutually exclusive event types, let \(N\) be the number of analyzed people, \(d_k\) the number whose first event was type \(k\), and \(c\) the number censored before any event. If every person has an observed status and each person is counted once, the accounting identity is:

Here, censoring includes follow-up ending without an observed first event, whether because of loss to follow-up, withdrawal, or the administrative cutoff, as defined in the study. Reports may show these reasons separately, but they should still reconcile to the total censored count. A known competing event belongs in its event-type count, not in \(c\). If event status or cause is genuinely unknown, describe that uncertainty rather than quietly assigning the person to a known cause or to ordinary censoring.

Worked Example: Reconciling Screening, Exclusions, and Follow-Up

Report Follow-Up and Censoring Transparently

The number censored alone does not show how long people were observed. Report the follow-up period or data cutoff and summarize observed follow-up duration with a measure appropriate to its distribution, commonly the median and interquartile range. State the units and the definition of the duration being summarized. If the report separates censoring due to loss to follow-up from censoring at the administrative cutoff, provide those counts and clarify whether each person is assigned to one reason only.

For example, “Median follow-up was 18 months (interquartile range, 11–26); follow-up was administratively closed on 30 June 2025” gives a duration summary and cutoff. This sentence should correspond to the actual time variables used. A reader should not have to infer whether the stated follow-up is measured from enrollment, diagnosis, or another origin.

A flow diagram is often useful when there are several stages, but a table or concise text can work for a simple cohort. Whichever format is chosen, use consistent labels and put the denominator beside percentages. Counts are usually the central information: a percentage without its denominator can obscure how many people were actually observed.

Checking the Arithmetic and the Meaning

The accounting identity follows by classifying each analyzed person once and only once. Let \(I_{ik}\) equal 1 if person \(i\)'s first event is type \(k\), and 0 otherwise; let \(C_i\) equal 1 if that person is censored before an event, and 0 otherwise. Under exhaustive, mutually exclusive terminal classification, each person satisfies:

Summing over the \(N\) analyzed people gives \(\sum_{i=1}^{N}\sum_{k=1}^{K} I_{ik} + \sum_{i=1}^{N} C_i = N\). The first term is the total number of first events across causes, and the second is the number censored. This yields \(N = \sum_{k=1}^{K}d_k + c\). The identity is a check on classification and data handling, not an assumption that people have equal follow-up or equal event probabilities.

If the counts fail to sum, investigate before reporting. Common explanations include a person counted under two event types, an event recorded after censoring, an unreported unknown status, or a mismatch between the analysis dataset and the flow diagram. A discrepancy should be resolved or explicitly accounted for; it should not be hidden by changing a denominator.

Common Pitfalls and Limits

• Calling a competing event “censoring.” This wrongly treats an observed event as if its occurrence were unknown. Record each specified competing cause as an event type.
• Giving percentages without denominators. State whether a percentage uses the screened group, eligible group, or analysis population. For terminal statuses, the analysis population is a natural denominator when clearly labeled.
• Mixing exclusions with censoring. A person excluded before entering the analysis population is not part of its censored count. Keep pre-analysis eligibility decisions separate from follow-up outcomes.
• Counting recurrent events as first events. If the endpoint is the first event, later events do not add to the first-event cause counts. Recurrent-event analyses require a different accounting unit and should be labeled accordingly.
• Interpreting crude percentages as estimated risk. A proportion classified by the cutoff does not account for differing follow-up times. Use the appropriate competing-risk estimand when reporting time-specific event probabilities.
• Assuming counts establish comparability. Similar cohort sizes or event counts do not show that groups had comparable follow-up, baseline characteristics, or censoring patterns.

These reports are descriptive and depend on correct event classification, consistent time origins, and complete accounting of included participants. They do not replace analysis of event probabilities or associations. Their value is that readers can see the path from the starting cohort to the data behind those analyses.