Joint Frailty Models for Recurrent and Terminal Events

Many clinical outcomes do not occur only once. Patients may experience repeated hospitalizations, disease relapses, infections, exacerbations, bleeding episodes, or treatment interruptions. Follow-up can then end because of death, transplantation, discharge, loss to follow-up, or another terminal event. When the terminal event is related to the patient’s underlying recurrence risk, analyzing recurrent events alone can misrepresent the event process.

A common shortcut is to analyze time to the first recurrence or death. That endpoint may be clinically useful, but it discards later recurrences and can make death look equivalent to a non-fatal recurrence. Another shortcut is to censor recurrent-event follow-up at death and assume that the censoring is non-informative. That assumption is questionable when patients with a high underlying recurrence burden are also more likely to die earlier.

A joint frailty model addresses this dependence by specifying separate hazard processes for recurrent and terminal events and connecting them through an unobserved subject-level frailty. The model does not create information that the study did not collect. Its value is that it makes the dependence structure explicit, estimates it under stated assumptions, and preserves the distinction between repeated non-fatal events and the terminal outcome.

What the model is designed to represent

Let a patient contribute recurrent event times and a terminal-event time. The recurrent process may be represented with a counting-process or gap-time formulation, while the terminal process is modeled with a survival hazard. A shared frailty term introduces patient-level heterogeneity that affects both hazards. A patient with a higher latent frailty may have more frequent recurrences and a higher terminal-event hazard, although the sign and strength of this association must be estimated rather than assumed.

The two processes need separate baseline hazards and covariate effects. A treatment may reduce recurrent events without changing mortality, affect both outcomes in different directions, or alter the terminal hazard while leaving recurrence unchanged. Reporting a single composite hazard ratio would conceal those possibilities. The recurrent-event effect, terminal-event effect, and dependence parameter answer different scientific questions.

Why censoring at death can be informative

For a recurrent-event analysis, death removes a patient’s opportunity to experience later events. If the probability of death is related to the unobserved recurrence propensity, the observed recurrent-event histories are selectively truncated. Treating death as independent censoring can then distort recurrence-rate estimates and standard errors. The issue is not simply that fewer observations remain; the remaining observation window is related to the event process.

The same logic applies to other terminal outcomes. A patient can be discharged from an intensive-care unit, receive a transplant, or enter a competing disease state. Whether that event should be treated as a terminal event, a competing event, or a censoring event depends on the estimand and the clinical question. A joint model is not automatically correct for every setting. The terminal event must have a scientifically meaningful relationship to the recurrent process and sufficient information for estimation.

Joint frailty modeling also differs from simply adding recurrent events to a composite endpoint. A composite endpoint may be appropriate when the estimand explicitly concerns time to any component. A joint model is useful when the investigator wants to retain component-specific processes and study their dependence. The two strategies should not be presented as interchangeable.

Choosing the recurrent-event time scale

The event-history definition changes the estimand. In a total-time model, the clock begins at baseline and continues through recurrent events. In a gap-time model, the clock may reset after each event, focusing on time until the next recurrence. Some analyses count events while others model transitions between states. The choice should reflect how clinical risk evolves after an event and whether the treatment effect is expected to change with event order.

Investigators should define whether recurrent events must be clinically distinct, whether events occurring within a short window are collapsed, how same-day events are handled, and whether an event resets follow-up. A model can be mathematically sophisticated while still answering the wrong clinical question if these operational definitions are left implicit.

Model assumptions and diagnostics

A joint frailty model commonly uses proportional-hazards structures, baseline hazard functions, and a specified frailty distribution. The frailty distribution is not a decorative technical choice. Misspecification can influence the estimated dependence between recurrent and terminal processes and, in turn, the component-specific effects. Sensitivity analyses should examine whether conclusions change under plausible alternative frailty or baseline-hazard specifications.

Diagnostics should include event counts by treatment group, numbers of patients with zero, one, and multiple recurrences, follow-up duration, terminal-event frequencies, and the timing of terminal events. Plot observed recurrence patterns and terminal-event-free survival by clinically important groups. Assess whether the model reproduces the broad empirical event-history features rather than relying only on a convergence message.

When the data include clustering by hospital, family, physician, or treatment center, the recurrent-terminal dependence may coexist with higher-level clustering. A subject-level shared frailty alone may not absorb site-level dependence. The analysis plan should distinguish patient heterogeneity from cluster correlation and explain how standard errors and baseline hazards address both.

Joint frailty models versus neighboring strategies

Researchers comparing these approaches should also read the site’s guides to restricted mean survival time, competing risks, joint longitudinal and survival models, and win ratio. Those methods can be complementary, but they do not answer the same estimand.

Five steps for a defensible analysis

Evidence Summary Table

Reporting checklist

A transparent manuscript should state why a joint model is needed rather than a first-event analysis, count model, or ordinary recurrent-event model. Define the recurrent and terminal outcomes, the observation window, the event-history rules, and the handling of zero-event participants. Report the number of recurrent events, terminal events, and patients at each stage of follow-up.

Describe the frailty distribution, baseline hazard representation, dependence parameter, covariate effects, estimation algorithm, convergence criteria, missing-data approach, and sensitivity analyses. Give component-specific effect estimates with uncertainty intervals and explain their clinical meaning. If competing terminal events are modeled, show how each terminal outcome enters the likelihood and whether any terminal outcome is treated as independent censoring.

Finally, distinguish statistical dependence from causation. A positive frailty association may indicate shared vulnerability, residual heterogeneity, or model structure; it is not automatically evidence that recurrent events cause death. The strength of a joint frailty analysis comes from an explicit estimand, a defensible event-history representation, and transparent sensitivity analysis.

Medical Disclaimer: This article is for research and educational purposes only. It does not provide medical advice, treatment recommendations, or a substitute for protocol-specific statistical review. Researchers must verify methods against the original sources, data structure, estimand, and clinical context before using them in a study or manuscript.