Recurrent Event Survival Model

Standard Cox regression stops the clock for a subject after the first event and cannot use information from subsequent recurrences. A recurrent event model keeps each subject in the dataset across multiple event episodes by adopting counting-process format, so every interval from start to stop contributes to the partial likelihood. The within-subject correlation introduced by repeated observations is handled via robust standard errors or frailty terms.

Choose PWP when the risk of the next event is likely to depend on how many events the subject has already experienced — for example, successive cancer relapses or repeated hospitalisations in a deteriorating patient. PWP stratifies the baseline hazard by event number, allowing it to change shape across recurrences, and restricts the risk set for the k-th event to subjects who have already experienced k-1 events. AG is preferable when events are exchangeable and a single hazard ratio across all recurrences is a satisfactory summary.

A frailty is a subject-specific random effect — typically a Gamma or log-normal multiplier on the baseline hazard — that captures unobserved heterogeneity between subjects. If some individuals are intrinsically more prone to events than others for reasons not captured by covariates, ignoring this leads to under-dispersed event counts and biased hazard ratios. Adding a frailty term accounts for this clustering and yields more honest uncertainty estimates.

The registry recommendation is at least 50 recurrent events (not subjects). With fewer events, the partial likelihood surface may be flat for later recurrence strata in PWP, causing convergence problems. If the total event count is below this threshold, a standard Cox regression on the first event is a safer fallback.