Estimands for Recurrent Events and Event Burden

Why Recurrent-Event Estimands Need More Than an Event Count

For many chronic conditions, a patient may experience several events during a trial: exacerbations, hospital admissions, infections, or episodes of severe pain. A single “event or no event” outcome discards how often events recur. Yet a total count alone does not describe every relevant aspect of burden: patients may have different lengths of observed follow-up, and an event rate or the amount of time spent symptomatic may answer a different clinical question.

An estimand makes that question precise. For recurrent events, it must specify the treatment strategies being compared, the target population, the event definition, the follow-up horizon, how intercurrent events are handled, and the summary measure of burden. The choice between counts, rates, and time with events is not merely a choice of statistical model. It changes what treatment effect is being asked about.

The potential outcome \(N^a(\tau)\) describes the event history under a treatment strategy, not simply the observed count for someone who happened to receive that treatment. As with other estimands, the target population and treatment strategies must be stated; the observed-data analysis is a separate step. In a randomized trial, randomization supports comparison of assigned strategies, but the endpoint definition and intercurrent-event strategy still determine what outcome is being compared.

Three Ways to Summarize Recurrent-Event Burden

A mean count answers, “How many events, on average, occur per patient by the horizon?” It gives each patient a count over the same stated horizon and does not require dividing by observed time. A count difference is often direct to explain, while a count ratio expresses relative burden. Neither is universally preferable: a difference emphasizes absolute events avoided, whereas a ratio emphasizes proportional change.

An event rate relates event frequency to time at risk or another prespecified exposure duration. A rate-based estimand must define the exposure time as carefully as the event count. In a population, one possible marginal rate is the ratio of expected event count to expected exposure time. This is not generally the same as the average of each patient’s individual rate.

This definition makes the denominator part of the estimand. Exposure might be time alive, time on treatment, or time under observation, depending on the clinical question and the chosen intercurrent-event strategy. These choices are not interchangeable. For instance, an on-treatment rate may describe event frequency while patients remain on assigned treatment, but it does not by itself describe the effect of assignment over the full trial horizon. The earlier material on while-on-treatment estimands and treatment discontinuation is relevant when selecting such a strategy.

A third summary measures time with events. For a symptom that can be present on multiple days, define \(I^a(t)=1\) when the patient has the symptom at time \(t\) under strategy \(a\), and \(0\) otherwise. The accumulated burden is the integral of this indicator over follow-up. For day-level data, it is the number of days with the symptom. If a day can contain several episodes, counting days with any symptom differs from counting episodes.

These measures can lead to different conclusions because they answer different questions. A reduction in event counts does not necessarily imply fewer days affected: one strategy might reduce the number of episodes while making the remaining episodes longer. A rate can differ from a fixed-horizon count comparison when exposure time differs. The estimand should therefore match the clinical meaning of burden that matters to patients and decision-makers.

Worked Examples

Why the Burden Definitions Matter

For a fixed horizon with complete follow-up, the mean count gives every patient the same calendar window. For a rate, patients may contribute different exposure times. Dividing by exposure is useful when the target is event frequency per unit time, but it also changes the question. In particular, a group’s total events divided by total exposure estimates a ratio of expected counts to expected exposure in the target population; it is not automatically an estimate of the average patient-specific rate.

The relationship between time with events and a daily symptom indicator follows directly from accumulation over time. If \(I(t)\) is one on symptom days and zero otherwise, then adding the daily indicators gives the number of symptom-days. Dividing by the fixed number of days \(\tau\) gives the fraction of days affected:

Thus a mean difference in symptom-days and a mean difference in the fraction of days affected differ only by the fixed scale factor \(\tau\), if everyone has the same complete horizon and the day-level definition is unchanged. If follow-up lengths vary, that simple equivalence no longer holds without specifying how unobserved days and differing exposure are handled.

Death, Discontinuation, and Incomplete Follow-Up

A patient who dies before the planned horizon cannot simply be assigned zero events for the remaining time without changing the meaning of the outcome. Zero events after death may make a strategy that causes earlier death appear to reduce event burden. Nor is it always appropriate to divide only by time alive: that produces a rate among time alive and may conceal differences in survival. The earlier material on death as an intercurrent event or endpoint component helps frame the choices.

Possible targets include a count by a fixed horizon under a specified strategy for death, a rate per time alive, a composite outcome that incorporates death and recurrent events, or a burden summary defined only while alive. Each has a distinct interpretation. Similarly, stopping assigned treatment does not automatically end outcome collection: a treatment-policy question may seek events through the horizon regardless of discontinuation, whereas an on-treatment question may stop or redefine follow-up at discontinuation. The chosen strategy must be stated rather than inferred from the analysis method.

Incomplete event observation is also not itself an estimand choice. If events are missed, or follow-up ends early for reasons related to a patient’s event history, the observed count may not represent the intended count. Defining the target does not eliminate the need for assumptions or suitable data collection to identify it.

Comparisons, Diagnostics, and Limitations

A mean-count contrast is straightforward when a common follow-up horizon and complete event collection are central to the question. A rate contrast is more natural when exposure time is central, but it relies on a meaningful and consistently defined denominator. Time-with-events summaries capture duration or frequency of symptomatic days, but depend on reliable repeated measurement and a clear rule for overlapping episodes. These summaries are complementary, not competing estimates of one universal “burden.”

Before interpreting an estimate, check whether the follow-up horizon is comparable across arms; whether event definitions and ascertainment are consistent; whether patients with early death or discontinuation are handled as specified; and whether the reported denominator matches the stated estimand. A rate based on time on treatment should not be described as the effect of assignment over the full follow-up period unless those targets coincide by design. A count without its horizon is incomplete, and a symptom-day proportion without its definition of a symptom day is ambiguous.

A frequent mistake is to treat a lower observed event count as evidence of lower patient burden even when follow-up differs substantially. The correction is to inspect the target: if the question is events per fixed period, define the fixed-horizon count and address incomplete follow-up; if it is event frequency per exposure time, define the exposure-based rate and interpret it as such. Neither adjustment nor a particular model can repair a mismatch between the clinical question and the estimand.

All these summaries compress an event history. Two patients with the same count may differ in event timing, severity, clustering, or duration. A rate may hide whether events occur early or late; total symptom-days may hide whether burden is continuous or intermittent. If those features matter clinically, the outcome definition or additional summaries must capture them. The estimand should remain explicit about which aspects of burden it does and does not represent.