Event-Free Survival

Event-Free Survival: Concept Architecture

Measure time until a defined event occurs

Event-free survival tracks the time from a stated starting point until the first event in a prespecified set, or censoring if no event is observed. This page explains why the event list and time origin matter, how the survival estimate is interpreted, and what an economic model can learn from it.

Define the composite endpoint before analysis

Specify the eligible population, starting time, each qualifying event, assessment schedule, and whether death from any cause is included. In an oncology trial, events might include disease progression preventing surgery, recurrence, or death; another disease setting may use a different set. An event-free survival result is only interpretable with its actual protocol definition.

The FDA guidance on oncology endpoints gives a context-specific definition including progression that precludes surgery, recurrence, and death. The NCI dictionary uses a broad patient-facing description, so the trial's event list and clock take precedence over an assumed universal definition.

Estimate the probability of remaining event-free

Let $T$ be the time from the specified origin to the first qualifying event. The event-free survival function is $S_{\mathrm{EFS}}(t)=P(T>t)$ under the study's endpoint and population. It gives the probability of no qualifying event by time $t$, not necessarily the probability of being cured or remaining alive indefinitely.

With incomplete follow-up, a Kaplan–Meier estimate can use observed event and censoring times when censoring is appropriately independent of future event times. At each event time, the estimate multiplies the fraction of those still at risk who avoid the event. Censoring rules for loss to follow-up, new therapy, missed assessments, or end of study should be reported, since inconsistent handling can bias comparisons. If death or another outcome prevents a qualifying event but is not itself included in the endpoint, treating it as ordinary independent censoring may answer a hypothetical question rather than the real-world probability of remaining free of that event; the competing-event estimand must be specified.

Work through a small cohort

Suppose 100 people enter a trial at randomization. By one year, 20 have experienced a qualifying first event and the other 80 have complete event-free follow-up through that year. The observed one-year event-free proportion is $80/100=0.80$, or 80%.

If 10 of the 80 had instead left observation at six months, the one-year estimate would depend on when the 20 events occurred and who remained at risk. Simply dividing 80 by 100 would no longer be a valid Kaplan–Meier calculation for that incomplete follow-up. The example is illustrative and assumes a common time origin and complete ascertainment for the first calculation.

Separate the endpoint from nearby measures

Overall survival counts death from any cause as its event, regardless of prior progression or recurrence. Progression-free survival usually counts disease progression or death, while an event-free survival definition can include additional failures such as inability to proceed to planned surgery or relapse after initial treatment. The exact distinctions vary by trial and disease.

A composite endpoint may improve event capture but combine components of unequal severity or different treatment effects. Show component event counts and timing when available and avoid assuming that a longer event-free interval proves longer overall survival. The FDA guidance treats interpretation and context of use as important for these endpoints.

Translate evidence into model outcomes

An economic model may use event-free survival to estimate time in a state before a first event, with associated utility and resource use. It also needs a coherent pathway after events, mortality, and subsequent treatment; the EFS curve alone does not reveal which component event occurred or determine post-event costs and health outcomes.

Report the endpoint definition, time origin, population, censoring, follow-up maturity, treatment comparison, and uncertainty. Check whether modelled event-free occupancy and overall survival can coexist without impossible state counts. Extrapolation beyond observed follow-up requires clinically plausible assumptions and external validation rather than extending a trial curve mechanically.