Event-Count Re-Estimation For Survival Endpoints

What Is Being Re-Estimated?

In an event-driven survival trial, the analysis is planned around the number of participants who experience the endpoint, not just the number enrolled. This distinction matters at an interim review: slower event incidence or more censoring can mean that the original enrollment will not produce the intended amount of information. Blinded estimates can help project how many participants would be needed, while the treatment comparison itself remains fixed.

A key distinction is that the required event count and the enrollment needed to reach it are not the same quantity. Under the usual log-rank planning approximation, the event target is determined by the planned hazard ratio, allocation, significance level, and power. Event incidence, accrual, and censoring primarily determine how many participants must be enrolled to attain that target. Updating an estimated event rate alone does not justify changing the planned treatment effect or the event target.

Required Events for the Planned Comparison

Let \(D\) be the total number of observed primary events, \(q\) the planned fraction allocated to treatment, and \(1-q\) the fraction allocated to control. Let \(HR\) be the planned treatment-to-control hazard ratio under proportional hazards. For a two-sided test with type I error rate \(\alpha\) and power \(1-\beta\), the standard large-sample log-rank approximation is

Here \(z_u\) is the \(u\)-quantile of the standard normal distribution. This approximation assumes proportional hazards, independent censoring, a sufficiently large event count for the normal approximation, and a risk-set allocation fraction close to the planned \(q\). It also assumes the stated hazard ratio is the effect the design intends to detect. If the allocation in the risk sets differs materially from the planned allocation, a design-specific information calculation may be needed rather than this simple formula.

The relation follows from the approximate log-rank information. With \(D\) events and treatment fraction \(q\), the information for the log hazard ratio is approximately \(Dq(1-q)\); thus the standard error of the estimated log hazard ratio is approximately \(1/\sqrt{Dq(1-q)}\). Under the planned alternative, the standardized statistic has approximate mean \(\log(HR)\sqrt{Dq(1-q)}\). To attain the specified power, the magnitude of this mean is set approximately to \(z_{1-\alpha/2}+z_{1-\beta}\). Squaring both sides and solving for \(D\) gives the formula above.

Turning an Event Target Into Enrollment

The next step is to estimate the proportion of enrolled participants expected to experience the event by the analysis time. A transparent model can be useful for this projection, provided its assumptions are plausible. Consider uniform accrual over \(A\) years, followed by an additional \(F\) years of follow-up after accrual ends. Assume each participant has a constant event hazard \(\lambda\) and an independent constant loss-to-follow-up hazard \(\mu\). The event and loss processes are modeled as independent exponential times. Administrative end of follow-up is determined by the study calendar.

For a participant followed for \(u\) years, the probability of an event before either loss to follow-up or the end of that follow-up is \(\frac{\lambda}{\lambda+\mu}[1-\exp\{-(\lambda+\mu)u\}]\). Under uniform accrual, \(u\) ranges from \(F\) to \(F+A\). Averaging that probability over recruitment time gives the expected event fraction:

This is a projection under the stated model, not an observed certainty. If the event fraction is \(\pi_{\text{event}}\), an initial enrollment estimate is \(N\approx D/\pi_{\text{event}}\), rounded up. For a constant event fraction, a lower incidence or greater loss to follow-up raises the projected enrollment without changing the required event count.

Updating the Projection Without Changing the Comparison

At an interim review, a blinded estimate may suggest that the planning rates are too optimistic or too pessimistic. A protocol can specify how to obtain the nuisance-rate estimates, when the estimate is made, and how the updated projection is translated into a permitted enrollment change. The planned hazard ratio and other comparison parameters remain the design inputs. Blinded event data inform the expected event yield; they do not by themselves supply an estimate of the treatment effect.

The update should also respect any prespecified enrollment floor or cap and the study’s operational constraints. If the projected enrollment exceeds the cap, the design cannot simply claim the original event target will still be reached. The shortfall should be made explicit for the planned review and governance process. Rules for changing enrollment, and the inferential consequences of the adaptation, must be specified in advance.

Allocation, Diagnostics, and Alternatives

Allocation affects the event requirement as well as the event yield. For example, keeping the same hypothetical hazard ratio of 0.80, \(\alpha=0.05\), and 80% power, but using \(q=0.40\), gives

Useful diagnostics at a blinded review include checking whether the observed pooled event timing is broadly compatible with the planning model, whether censoring is more frequent than expected, whether recruitment has followed the assumed schedule, and whether participants have had the follow-up time assumed by the projection. A projection can be unstable when few participants have mature follow-up. Interim nuisance estimates are uncertain, so a protocol may use conservative estimates, a prespecified range, or a rule that limits how much enrollment can change.

The exponential calculation is convenient but restrictive: it assumes constant event and loss hazards. It may be inappropriate when hazards vary over time, censoring is informative, or the trial has complex entry or follow-up patterns. In those settings, a prespecified survival model or simulation based on the planned calendar can provide a more suitable event-yield projection. A fixed-enrollment design is simpler but may miss its event target when incidence differs from assumptions. An event-driven design directly continues follow-up until its target is reached, but it does not remove the need to plan enrollment, accrual, and study duration.

Common Mistakes and Practical Interpretation

• Changing the event target because the event rate changed. The rate affects projected enrollment; under the stated log-rank approximation, the target event count is tied to the planned comparison. Recalculate the event target only when a prespecified design input that determines it changes.
• Using \(D/\pi_{\text{event}}\) as an exact enrollment guarantee. This is an expected-yield projection under a model. Actual event counts vary, and assumptions about accrual, event times, and censoring may be wrong.
• Treating censoring as irrelevant. Loss to follow-up can reduce the proportion with an observed event by the analysis time. State how censoring is modeled and assess whether the assumption is credible.
• Ignoring allocation in the information calculation. The factor \(q(1-q)\) matters for the event target, and unequal allocation can also make event yield differ between groups.
• Using a pooled interim estimate as though it were a precise prediction. Blinded nuisance estimates are uncertain. Report the assumptions and the uncertainty or conservatism built into the adaptation rule.

The practical value of this approach is disciplined separation: preserve the planned survival comparison, use blinded nuisance information to assess the event yield, and make any enrollment change only under a prespecified rule. This keeps the operational question—how many participants are likely to be needed—distinct from the inferential question of what treatment effect the trial is designed to detect.