Define recurrent events before choosing a model

Decide what counts as a new episode

Repeated events can describe readmissions, infections, device failures or other recurring outcomes. Before considering a model, define when one episode ends and another begins. Multiple records for the same episode should not become multiple events simply because they appear on different dates or in different tables.

The survival package multistate vignette shows how different state and transition definitions represent different questions. Draw the process as states or episodes and decide which transitions contribute to the outcome of interest.

Define time at risk after an event

State whether participants return to risk immediately, after discharge, after a washout period or only after recovery. Decide whether time is measured from study entry or reset after each event. Record periods during which another episode is impossible or cannot be observed.

Death or another terminal event may prevent further recurrent events. Treating such an event as ordinary independent censoring can be inappropriate for the intended question. Specify whether the target is a rate while alive, a mean cumulative number, a first-event outcome or another summary. These are not interchangeable descriptions of disease burden.

Hypothetical example

A fictional participant has three emergency-department records on one day and is admitted overnight. The protocol defines the entire continuous care episode as one event. Counting the three records as three recurrences would inflate the event total and create zero-length intervals.

Later, the participant has a separate qualifying episode after discharge. That event can be counted according to the prespecified return-to-risk rule. The example illustrates record-to-episode derivation; it does not establish an appropriate clinical washout period.

Test the derived event history

Create synthetic timelines with overlapping encounters, same-day events, transfers, missing discharge dates and a terminal event. Check that the derivation respects the episode definition and produces valid intervals. Preserve person identifiers so repeated observations are not treated as independent people.

Choose the model and variance method only after defining the target quantity and dependence structure. Report both people and events, follow-up exposure, episode rules and terminal-event handling. Use the endpoint adjudication worksheet to document ambiguous cases. A sophisticated recurrent-event model cannot repair an outcome count that changes whenever a source table is joined.

Source verification: 6 November 2026. Examples are hypothetical.