Recurrent Trial Events: Choose How the Endpoint Handles Death

Consider a heart-failure team deciding whether the primary endpoint should measure the first hospitalization, all hospitalizations, or hospitalizations plus cardiovascular death. The operational task is to keep those alternatives distinguishable before selecting the primary analysis. The synthetic ledger below shows how the same six records produce different summaries, including an early death with no hospitalization. These are endpoint-design decisions for trial teams, rather than conclusions about any particular treatment.

A time-to-first endpoint retains the timing of the first qualifying event. A recurrent-event endpoint also uses subsequent qualifying events. Two participants with the same first hospitalization date can therefore contribute identical first-event records but very different total counts. Neither summary is intrinsically the right one: delaying initial deterioration and reducing later episodes are distinct clinical objectives.

Gregson and colleagues’ 2023 review examines repeat-event methods in five cardiovascular trials while noting that time-to-first remains common in that setting. It compares negative binomial, Lin–Wei–Yang–Ying (LWYY), joint frailty, win-ratio and area-under-the-curve approaches. That comparison supports selecting a method around the clinical question; it does not establish that any recurrent-event analysis is automatically superior.

A team should write the proposed treatment-effect sentence before naming software: for example, “Compare the expected cumulative number of qualifying hospitalizations through month 24, with counts stopping at death.” A different sentence would compare hospitalizations per unit of time alive. A third would add cardiovascular death to the counted outcome. Each needs a specified population, treatment comparison, time origin, horizon and handling of other intercurrent events.

Death ends the possibility of further nonfatal events. Loss to follow-up ends observation while events may still occur. A common calendar cutoff ends the observation window by design. These situations can share an analysis stop date without being interchangeable. In particular, a relocation or withdrawal is not automatically independent administrative censoring merely because it ends contact.

ICH E9(R1) places intercurrent-event handling within the estimand. Its while-on-treatment strategy uses the term “while alive” when death is the relevant event. A composite-variable strategy can incorporate death into the outcome. A joint model is an estimation approach, rather than an additional ICH strategy. See the ICH E9(R1) addendum, sections A.3.1–A.3.3 and the site’s estimand framework guide.

The error to avoid is silently treating death as an ordinary loss of observations from a process that could continue afterward. A cause-specific analysis may use a censoring code at death for its risk set; the code alone does not determine whether the analysis is valid. What matters is the target quantity and its assumptions. The related competing-risks guide covers first-event probability and cause-specific questions; recurrent histories require the additional counting decisions developed here.

Ghosh and Lin’s 2000 paper defines the marginal mean frequency function as the expected cumulative number of recurrent events over time in the presence of death. Once someone dies, their count cannot increase. This is different from estimating a recurrence rate among time spent alive. The paper develops nonparametric estimation and comparisons, including joint consideration of recurrence and death. Interpreting the mean curve still requires attention to mortality and assumptions about incomplete follow-up.

The following decision map separates outcome definitions from models. “Events only” is not a complete estimand specification: the denominator and population summary still matter. Similarly, adding death changes the outcome, but leaves choices about weighting, observation time and interpretation. The distinctions below are a protocol-writing aid, not a qualification of a method for a particular trial.

The EMA qualification opinion adopted on 27 February 2020 discusses events-only and death-inclusive estimands, but its final statement is cautious about terminal events. It questions the interpretation of exposure-weighted effects when treatment affects mortality, notes limitations of the estimator that adds one for death, and envisages separate recurrence and mortality estimates with mortality assessed first. This is a methodological qualification opinion, not a general endorsement of total-event primaries.

The next table derives alternative endpoint records from that ledger. The first-hospitalization column identifies death before hospitalization explicitly; it does not assume that such a death can be ignored in an absolute-risk estimate. The first-composite column uses the earlier of admission or cardiovascular death. All counts are recorded counts up to the end of observed follow-up.

There are four observed first admissions, eight admissions in total, two cardiovascular deaths and ten admission-plus-death events. Five participants have a first composite event. Pt-102’s three later admissions affect the recurrent count but not their first-event time. Pt-103 illustrates why zero admissions must remain distinguishable from an event-free survivor. Pt-104 contributes two nonfatal events and one death without implying that these outcomes have equal clinical importance.

The ledger contains 97 observed months alive: 24 + 24 + 3 + 8 + 14 + 24. This is 97/12 patient-years. The crude recorded admission rate is 8 ÷ (97/12), approximately 0.99 admissions per patient-year. This pooled descriptive rate is not a model-adjusted effect, an average of participant-specific rates, or a mean cumulative count at month 24. Pt-105 has unobserved later follow-up, so neither eight divided by six nor the crude rate is automatically a complete 24-month burden estimate.

For actual data, retain an event identifier, onset and resolution dates, episode grouping, event type, source and adjudication status in addition to this compact display. Keep death date and cause, last known alive date, planned cutoff, actual observation end, and end reason as separate fields. A programmer can then derive first-event and recurrent records without reconstructing later episodes from a single total-count field.

The methods below produce different summaries. The 2026 practical overview by Duoerkongjiang and colleagues describes model assumptions and provides worked code. The table is a selection aid: a trial statistician still needs to verify that the implemented estimator targets the protocol’s question, including the handling of death and incomplete observation.

An offset is a definition of exposure in the fitted model, not a correction that automatically separates morbidity from survival. Likewise, fitting a joint frailty model changes the model and its interpretation; it does not retroactively define the clinical question. Ask the statistical reviewer to trace the proposed treatment-effect sentence to the exact estimator and to identify which assumptions support that trace.

Claggett and colleagues’ 2018 simulations varied risk heterogeneity and treatment discontinuation after a first event. They simulated 4,000 participants with a true relative risk of 0.80, assuming that discontinuers reverted to placebo-level risk. Recurrent-event methods had greater power when post-event discontinuation was at most 30%, less power at 60% or more, and comparable power between those settings. Increasing risk heterogeneity reduced power for both approaches. These are results under specified simulation assumptions, not universal discontinuation thresholds or a promised sample-size saving.

Hazewinkel and colleagues’ 2025 reanalysis used four SGLT2-inhibitor heart-failure trials. It found pronounced clustering of readmissions and cardiovascular deaths after hospitalization, especially early afterward. The authors reported that LWYY and negative binomial methods did not account for those time patterns. AUC and win-ratio approaches strengthened the composite treatment-effect estimates, yet time-to-first analyses tended to give the strongest evidence. These findings challenge automatic power claims; they do not establish that recurrent endpoints are unsuitable in every indication.

For planning, this means investigating plausible event timing, differences between participants, death and treatment discontinuation together. A justified design should keep the chosen clinical question when evaluating operating characteristics, rather than search across different endpoints for the largest apparent power. Continue collecting the events required by the estimand after treatment discontinuation when that strategy calls for them; stopping medication and losing outcome follow-up are distinct situations.

As checked on 31 October 2026, FDA’s June 2019 heart-failure endpoint guidance page still labels the document a draft with nonbinding recommendations, not for implementation. Its downloadable draft lists binary hospitalization outcomes, counts, time to initial hospitalization, time alive and out of hospital, and recurrent-hospitalization approaches. It also calls for attention to all-cause versus cause-specific hospitalization capture. Listing an approach does not establish its acceptability for an individual development program.

The EMA opinion describes established recurrent endpoints where terminal events are rare, including MS relapses, respiratory exacerbations, migraine attacks, hypoglycemia and seizures. Experience where mortality is substantial is more limited. Its concerns about mortality, weighting and interpretation remain central when using the qualification opinion. Neither this document nor the cited FDA draft requires a recurrent-event primary endpoint; transferring an analysis to another disease needs clinical and statistical justification.

The following practical crosswalk turns the preceding distinctions into reviewable decisions. These are recommended implementation controls. Set the clinical question and collection plan during design, then prespecify the analysis before unblinding. The SAP endpoint guide addresses the broader documentation process.

Test the derivation specification against edge cases before unblinding: no admissions followed by death; an admission ending in death; repeated records for one transfer; an unadjudicated episode; treatment discontinuation with continued follow-up; and loss to follow-up before the horizon. These checks reveal ambiguities in the chosen rules. They do not validate the clinical event definition or prove a model’s assumptions.

Under ICH E9(R1), sections A.5.2–A.5.3, sensitivity analyses examine assumptions while targeting the same estimand. Adding death to an admissions-only endpoint changes the variable and generally provides a supplementary answer. A joint-frailty contrast may also target a different quantity. Plan same-estimand checks for incomplete follow-up and model assumptions, and label changed questions explicitly. The missing-data and sensitivity-analysis guide develops that distinction.

Use hospital-free days when the objective is a day-based outcome with its own death-scoring convention; do not substitute that score for a recurrent-event history. For a recurrent endpoint, the reviewable deliverable is the connection between the clinical question, the complete ledger and the programmed summary. A team should be able to explain why an early death contributes its specified outcome, how later episodes enter, and what information is missing before interpreting any treatment contrast.