Blinded Event-Rate Re-Estimation

Why Re-Estimate a Blinded Event Rate?

In a time-to-event trial, the number of events observed by a planned analysis date depends on both the underlying event process and how long participants are followed. If events accrue more slowly than anticipated, an event-driven analysis may take longer to reach its target. A blinded interim review can estimate the event rate across the trial without separating participants by treatment assignment, then use that estimate to assess expected event accrual under continued follow-up.

This is a planning quantity, not a treatment-effect estimate. The pooled rate can help answer questions such as “How many events might accrue during the remaining planned person-time?” It does not, by itself, identify either treatment arm’s rate or the contrast between arms. The validity of an update therefore depends on the intended use, the assumptions about future follow-up, and the information made available to reviewers.

The basic calculation is simple, but its interpretation needs care. It estimates an average rate over the person-time represented in the interim data. It is not automatically the hazard at every time since randomization, and it is not the event rate that would occur in either treatment arm alone.

Model, Estimator, and Assumptions

A useful starting model assumes a constant event hazard \(\lambda\) over the relevant observation period. Under this model, each participant contributes time at risk until the first event, censoring, or the end of observation. For a pooled dataset with \(D\) observed events and \(T\) total person-time at risk, the event count has a Poisson likelihood proportional to \(\lambda^D \exp(-\lambda T)\). This form applies when event occurrences follow the constant-rate process and censoring is non-informative for the event process.

The likelihood gives a short justification for this estimator. Taking its logarithm gives \(\ell(\lambda)=D\log(\lambda)-\lambda T+\text{constant}\). Differentiating and setting the derivative to zero yields \(D/\lambda-T=0\), so the maximum-likelihood estimate is \(D/T\). Under the Poisson model, \(\operatorname{Var}(D)=\lambda T\); therefore the estimated variance of \(\widehat{\lambda}\) is approximately \(\widehat{\lambda}/T=D/T^2\), and its estimated standard error is \(\sqrt{D}/T\).

This standard error describes sampling uncertainty under the model; it does not account for a future change in the participant mix, event definition, follow-up process, or hazard pattern. A confidence interval can summarize uncertainty in the pooled rate, but an interim rate should not be treated as known exactly when projecting future events.

The constant-hazard model is a working assumption, not a universal description of survival data. If hazards vary over time, \(D/T\) remains an observed events-per-person-time summary for the interim population, but a single rate may be a poor predictor of future accrual when the remaining participants will contribute time at different points in their risk trajectories. A blinded update should therefore check whether the interim data and planned future follow-up are sufficiently comparable for the intended projection.

Calculating and Interpreting the Pooled Rate

For an event-driven trial, a pooled rate may be used to project how quickly events could accrue if future person-time resembles the person-time already observed. The projection is conditional: it uses the observed rate as a planning estimate and does not guarantee the eventual event count. It is also distinct from calculating the event target required to detect a specified treatment effect. The target and its assumptions must already be specified; the pooled rate concerns the expected pace of accrual.

What Blinding Does—and Does Not—Protect

A blinded pooled calculation uses aggregate event information without identifying treatment groups. Reviewers may need the number of events, total person-time at risk, the snapshot cutoff, and relevant aggregate follow-up summaries. The data system and review process should prevent access to arm-specific event counts or rates if the adaptation is intended to remain blinded. Treatment codes alone are not enough to establish blinding if other information effectively reveals which observations belong to which arm.

Pooling removes the direct arm-by-arm comparison from the calculation, but it does not remove treatment effects from the pooled quantity. To see why, suppose equal amounts of person-time are contributed by two arms with constant hazards \(\lambda_C\) and \(\lambda_T\). The pooled rate is approximately their average, \((\lambda_C+\lambda_T)/2\). It therefore depends on both arm rates. Knowing the average does not identify either rate separately.

Consequently, a blinded rate update can support a forecast about overall event accrual, but it should not be described as estimating the control rate, treatment rate, hazard ratio, or treatment benefit. If the rate will be inserted into a planning formula that specifically requires an arm-specific rate, that use needs an explicit justification; the pooled estimate does not supply the missing arm-specific information.

Diagnostics, Alternatives, and Limitations

Before relying on a pooled-rate projection, reviewers should confirm that the numerator and denominator refer to the same event definition, analysis population, and snapshot. They should check that the denominator is actual time at risk rather than time since randomization for participants no longer at risk. Useful aggregate diagnostics include event accrual over successive calendar periods, person-time by enrollment cohort, follow-up duration, censoring patterns, and major changes in eligibility or event ascertainment. These summaries can reveal departures from a stable-rate projection without requiring arm-specific results.

A rate may appear to change because the underlying hazard changes with time since randomization, because recent enrollees have had less opportunity to experience events, or because censoring and follow-up differ from the planning assumptions. If the endpoint is a first event and competing events prevent that endpoint from occurring, the observed cause-specific event rate is not the same as the cumulative probability of experiencing that event. Recurrent events also require a clearly specified counting process and denominator; a first-event calculation cannot simply be applied to repeated events without checking the model.

The main alternatives answer different questions. An external planning rate avoids using interim trial outcomes, but may not match the enrolled population or current follow-up. A pooled interim rate reflects the trial’s aggregate experience, while retaining treatment masking, but may combine different arm rates and be imprecise. An arm-specific rate can describe each group more directly, but requires access to treatment-separated information and therefore is not a blinded pooled update. The choice is not merely computational: it determines what information is exposed and what the estimate can support.

A second frequent error is to treat the interim rate as a precise forecast and increase follow-up until a point estimate reaches the event target. The correction is to distinguish an estimated rate from a future event count: both rate uncertainty and future variation matter, and actual person-time may differ from the plan. The adaptation rule should specify in advance how the rate is calculated, what data reviewers may see, how projections are used, and what operational limits constrain any response.