Event-Driven Updates for Survival Endpoints

Why Trigger Updates by Events?

With a survival endpoint, each enrolled patient may contribute information over time: an event may occur, or the patient may still be event-free and under observation. An event-driven allocation rule uses observed endpoint events as occasions to refresh the probabilities for future assignments. This can be useful when the timing of events, rather than a prespecified calendar date, is the natural point at which new outcome information becomes available.

The trigger does not mean that only patients who experience events matter. At an update time, patients who remain at risk have accumulated event-free follow-up. Under a survival model, that follow-up contributes information about the event rate. The rule must distinguish no event observed so far from an event observed, and it must distinguish both from loss to follow-up or other censoring.

Model and Update Rule

Consider \(K\) treatment arms and a single time-to-event endpoint. For arm \(a\), assume event times follow an exponential model with constant hazard \(\lambda_a>0\): conditional on assignment to arm \(a\), \(T_i\sim \operatorname{Exp}(\lambda_a)\), with survival function \(P(T_i>t)=\exp(-\lambda_a t)\). The hazard is measured in events per unit of follow-up time. Assume patients’ event times are independent conditional on their assigned arms and the arm-specific hazards, and that censoring is noninformative for this model. The event definition and follow-up measurement must be consistent across arms. This simple rule is not a general method for arbitrary time-varying hazards or competing-risk outcomes.

At an update time \(\tau\), let \(U_i(\tau)\) be patient \(i\)'s observed follow-up time from enrollment through the earliest of an observed event, censoring, or \(\tau\). Let \(\delta_i(\tau)=1\) if the endpoint event has been observed by \(\tau\), and \(0\) otherwise. For arm \(a\), define \(D_a(\tau)\) as the number of observed events and \(Y_a(\tau)\) as the sum of observed follow-up times:

A patient still at risk at \(\tau\) contributes the time from enrollment through \(\tau\) to \(Y_a(\tau)\), and contributes zero to \(D_a(\tau)\). A patient censored before \(\tau\) contributes follow-up only through censoring. A patient with an observed event contributes follow-up through the event time and one event.

Why Both Event Counts and Exposure Matter

For one patient with follow-up \(u\), the exponential-model contribution to the likelihood is proportional to \(\lambda_a^\delta\exp(-\lambda_a u)\). If an event occurs, the density contributes the factor \(\lambda_a\); if the patient remains event-free through \(u\), the survival probability contributes \(\exp(-\lambda_a u)\). Multiplying these contributions within arm \(a\) gives:

Thus observed events increase the event count, while event-free follow-up increases exposure. A patient who is still at risk is not a completed “no event” outcome at a fixed horizon. The patient’s survival through the time observed is evidence about the hazard, but future event-free time has not yet been observed.

For a concrete Bayesian update, use independent Gamma priors on the arm-specific hazards. Here \(\operatorname{Gamma}(a_{0a},b_{0a})\) uses shape \(a_{0a}\) and rate \(b_{0a}\), so its mean is \(a_{0a}/b_{0a}\). Combining this prior with the likelihood gives a posterior with shape \(a_{0a}+D_a\) and rate \(b_{0a}+Y_a\).

The posterior mean is an estimate of the event hazard, not the probability of experiencing the event by a fixed time. To favor arms with lower estimated hazards, define a score \(s_a=1/E(\lambda_a\mid\text{data})\) and normalize it to \(r_a=s_a/\sum_{j=1}^K s_j\). A minimum randomization probability can protect continued exploration. For a specified floor \(\varepsilon\), with \(K\varepsilon\leq 1\), one rule is:

The scores define a particular allocation heuristic; they are not a claim that this rule is universally optimal. The floor prevents an arm’s probability from falling below \(\varepsilon\), while the normalized scores determine how the probability remaining above the floors is distributed.

What Happens to Patients Still at Risk?

A patient who has had no event by the trigger time contributes survival experience for the period observed. Under the stated model, that is exposure in the likelihood, not an extra event and not a definitive no-event outcome. The following example shows how that distinction enters an allocation calculation.

Event Timing, Refreshing, and Allocation

An event-driven rule has two distinct clocks. Follow-up exposure accumulates continuously as patients remain under observation, but the allocation probabilities are recalculated only when the prespecified event trigger occurs. Between triggers, incoming patients use the currently active probabilities. At a trigger, the analysis includes all eligible event and follow-up information accumulated through that time; it must not use information recorded later.

If several events are recorded at the same trigger time, process them together before calculating probabilities for the next assignment. This avoids making the result depend on an arbitrary ordering of simultaneous records. The precise trigger policy should be specified in advance—for example, update after every event or after a batch of a fixed number of events—because these choices produce different update schedules.

Assumptions, Diagnostics, and Alternatives

The calculation relies on a constant hazard within each arm over the period being modeled, a correctly defined endpoint, and appropriate handling of censoring. Independent censoring means, in this setting, that the censoring process does not carry unmodeled information about a patient’s event time after conditioning on the model information. A patient lost to follow-up cannot be treated as event-free for the unobserved period: their exposure ends at the last time their status is known. The model also assumes that observed events are recorded in time to serve as the trigger. If event recording is delayed, the rule triggers on recorded information, not necessarily on the underlying event time.

Before using an update, verify the event count and exposure total against the patient-level records and the risk set. Check that exposure units are consistent, that follow-up stops at events or censoring, and that each event is counted once. Review arm-specific event counts and rates \(D_a/Y_a\) alongside the posterior summaries. Very few events, extreme dependence of allocation on the prior, or probabilities repeatedly close to their floors indicate that the resulting assignments may be driven more by the modeling and safeguards than by stable evidence. A long interval without events can leave allocation probabilities unchanged even while exposure accumulates; this is a feature of the event-trigger policy, but it can make updates feel stale.

The main alternatives answer different operational needs. Calendar-time updates refresh probabilities at scheduled reviews, even if no event has occurred; a fixed-follow-up binary endpoint waits for a defined outcome window and does not use partial survival time in the same way. A time-to-event model with a nonconstant hazard can be more appropriate when the exponential assumption is implausible, but it requires a different model and update calculation. Event-driven updating is attractive when events are the natural information milestones, but it does not eliminate the need to examine the hazard model, censoring, or the consequences of infrequent triggers.

Common Mistake and Key Takeaway

Wrong reasoning: “A patient has not had an event by the update, so count that patient as a non-event and treat the outcome as complete.” This confuses being event-free through the observed follow-up with being known to remain event-free through a defined endpoint horizon.

Correction: Count only observed events in \(D_a\). Include each patient’s observed time at risk in \(Y_a\), ending at the trigger, event, or censoring, as appropriate. An event-free patient contributes survival exposure but may still experience the event later. If the event-driven rule has no trigger, keep the current allocation probabilities until the next specified trigger.