From an Event Target to a Calendar Date
A design may specify that an interim analysis occurs after a target number of events has accrued. That event target does not, by itself, identify a calendar date. The date depends on when participants enter, how long they are followed, and the event probabilities assumed by the projection. The practical task here is to evaluate the projected cumulative event count at candidate dates and find when it first reaches the target.
The date obtained from an expected event count is a planning forecast, not a guarantee about the observed trial. At a given date, the actual event count can be above or below its expectation. Keeping that distinction clear is essential when translating a statistical target into an operational timeline.
For a first-event endpoint, suppose the projection concerns a defined set of participants and participant \(i\) has probability \(p_i(t)\) of experiencing the event by date \(t\). If \(I_i(t)\) is 1 when participant \(i\) has had a first event by \(t\), and 0 otherwise, then the cumulative count is \(D(t)=\sum_i I_i(t)\). By linearity of expectation,
This expectation formula does not require participants' event indicators to be independent. It does require the count to be defined as stated: each participant contributes at most one first event. For a recurrent-event count, let \(N_i(t)\) be participant \(i\)'s number of events by \(t\); then \(D(t)=\sum_i N_i(t)\) and \(E[D(t)]=\sum_i E[N_i(t)]\). Summing first-event probabilities would count people with an event, not all recurrent events.
Evaluating Candidate Dates
A candidate-date calculation uses the same projection assumptions at every date. These typically include the enrollment schedule, participant follow-up, event probabilities or rates, and the handling of censoring and loss to follow-up. The projected population and event definition must also be held consistent. If these inputs change between candidate dates without a stated reason, the resulting date comparison is not meaningful.
For first-event outcomes, \(p_i(t)\) should represent the probability of an event by the candidate date given participant \(i\)'s enrollment time and relevant follow-up. A participant who enrolls later has less follow-up at the same calendar date. The introductory material on event projection covers constructing these probabilities; the key step here is to evaluate the resulting expected count at dates around the target.
In standard event projections, the cumulative count is nondecreasing over time: each participant's event-by-date indicator cannot change from 1 back to 0. Consequently, \(M(t)\) is nondecreasing. This gives a useful search principle: once candidate dates are found with \(M(t)<K\) and \(M(t)\geq K\), the target date lies between them, subject to the date resolution and model assumptions.
This justification follows directly from the event definition: having experienced a first event by an earlier date implies having experienced it by a later date as well. Taking expectations preserves the inequality, and summing the participant-level expectations preserves it again. The same reasoning applies to cumulative recurrent-event counts, which cannot decrease as time passes.
Worked Example: Finding a Date Between Two Monthly Projections
Suppose a hypothetical first-event projection concerns 100 participants enrolled by March 1. The model gives expected cumulative counts of 20 events on September 1 and 29 events on October 1. The design target is \(K=25\). Assume for this illustration that the projected expectation increases linearly between those dates, which are 30 days apart.
Bracket the target. Since \(20<25\leq29\), the target lies between September 1 and October 1. Under the stated linearity assumption, the fraction of the interval needed to gain five expected events is
The estimated elapsed time is \(0.5556(30)=16.67\) days after September 1. For a daily calendar, check the adjacent dates: September 17 is 16 days after September 1 and has projected expectation \(20+9(16/30)=24.8\); September 18 is 17 days after and has \(20+9(17/30)=25.1\). Thus September 18 is the first whole calendar date on which the projected expectation is at least 25. The fractional expected count is appropriate: it is a mean, not a claim that a fraction of an event will be observed.
Interpolation and Date Resolution
When projections are available only at a few dates, interpolation can give a convenient estimate between them. If \(M(t_1)<K\leq M(t_2)\), a linear interpolation estimates the fraction of the interval needed to reach the target as
This is an approximation, not a general property of event accrual. Event probabilities may change nonlinearly with follow-up, and enrollment patterns or event rates may change over calendar time. When the projection can be evaluated at daily dates, a direct daily check is preferable for selecting a calendar date. Interpolation is most useful for locating which dates to evaluate more closely or for communicating an approximate date from a coarse schedule.
Worked Example: Checking the Arithmetic Across Several Cohorts
Consider a hypothetical first-event projection with three enrollment cohorts containing 120, 80, and 60 participants. On July 15, their respective probabilities of an event by that date are 0.46, 0.20, and 0.05. The projected expectation is
Suppose the same model gives \(M(\text{June 15})=64.8\) and \(M(\text{August 15})=82.5\), and the target is 75 events. The target is not reached by July 15, but is reached by August 15. The bracket is therefore July 15 to August 15, a 31-day interval. Linear interpolation gives
The interpolated date is about three days after July 15. Check the daily dates under the linear-between-dates assumption: July 17 is two days after July 15, giving \(74.2+8.3(2/31)=74.7355\); July 18 is three days after, giving \(74.2+8.3(3/31)=75.0032\). Thus July 18 is the first date meeting the target on this interpolated projection. In an actual application, the candidate-date projections—not rounded intermediate values—should determine whether the target is met.
The essential arithmetic check is that the cohort-specific contributions sum to the stated total. A small discrepancy can change an interpolated date, especially when the expected count increases slowly around the target. Retain sufficient precision during calculations, and round the final date only to the resolution that the operational plan uses.
What the Projected Date Does and Does Not Mean
A projected date based on \(M(t)\) answers when the expected cumulative count reaches the target under the chosen assumptions. It does not say that the actual trial will have exactly \(K\) events on that date, nor does it give the probability that the actual count has reached \(K\). Those are questions about the distribution of the event count, not only its expectation.
For example, if \(M(t)=50\) when the target is 50, the expected count is 50. This does not imply \(P(D(t)\geq50)=0.5\), or any other particular probability. Variability, dependence, and the event-count distribution matter for that probability. Use a prediction distribution or simulation when the planning question is specifically about the chance of reaching the target by a given date.
A projected date also differs from an event-triggered operational date. An event-triggered look is scheduled in response to the observed event count reaching its threshold; a forecast based on \(M(t)\) estimates when that may happen. Operational processing, adjudication, and analysis preparation can add further delay, but those delays should not be silently folded into the event-reach date.
Worked Example: A Recurrent-Event Target
Suppose a hypothetical recurrent-event projection covers 20 participants. The mean number of events per participant is projected to be 1.1 by August 1 and 1.4 by September 1. The target is 25 total events. Assume the projected total increases linearly between these dates, which are 31 days apart.
At August 1 the expected total is \(20(1.1)=22\) events; at September 1 it is \(20(1.4)=28\) events. The target is between these expectations. Interpolation gives a fraction \((25-22)/(28-22)=3/6=0.5\), or 15.5 days after August 1. For whole calendar dates, August 16 is 15 days after August 1 and has expectation \(22+6(15/31)=24.903\) events; August 17 is 16 days after and has \(22+6(16/31)=25.097\) events. The projected first date meeting the target is August 17. Here the calculation uses expected recurrent-event counts, not the number of participants with at least one event.
Diagnostics, Alternatives, and Limitations
Before reporting a date, check that the projection behaves as the event definition requires and that the target is actually bracketed. If the projected cumulative count decreases between dates, investigate the calculation: a cumulative event count should not fall. Possible causes include inconsistent participant sets, an incorrect date or follow-up calculation, or a table that reports a different quantity at different dates. Also check that later enrollees are not inadvertently given more follow-up than earlier enrollees.
A target may not be reached within the stated horizon. In that case, report that the projection remains below target through the final candidate date; do not extrapolate a precise date without an explicit model and justification. If \(M(t)\) is flat near the target, small changes in assumptions can produce large changes in the estimated date. If candidate values are very coarse, evaluate additional dates around the crossing rather than reporting excessive precision from interpolation.
| Approach | What it answers | Main limitation |
|---|---|---|
| Candidate-date expectation \(M(t)\) | When expected events reach the target under specified assumptions | Does not quantify the probability of reaching the target by that date |
| Prediction distribution or simulation | How uncertain the event count or target date may be | Requires distributional assumptions and more computation |
| Observed event-triggered timing | When the actual event count reaches its threshold | Cannot be known in advance; operational delays may follow |
The candidate-date method is transparent and easy to audit, but its forecast depends on the enrollment and event assumptions. A date reported to the nearest day should not be interpreted as known to that precision when input assumptions are uncertain or the projection itself is only available monthly. Report the assumptions and date resolution alongside the forecast.
A common mistake is to treat the interpolated date as the date the trial will certainly reach its event target. The reasoning is wrong because interpolation concerns the projected mean, while the observed count varies around that mean. The correction is to describe the result as a model-based expected target date, and to use a prediction distribution if the probability or uncertainty of actual target attainment is needed.
Check Your Understanding
Answer each question, then compare with the brief solutions below.
- A projection has \(M(t_1)=36\) and \(M(t_2)=48\), with a target of 42. Under linear interpolation, what fraction of the interval follows \(t_1\)?
- For 50 participants, the mean recurrent-event count per participant is 0.8 by one date. What is the projected total event count, and why is it not necessarily the number of participants with an event?
- If \(M(t)\) is 41 at the final planning date and the target is 45, what conclusion is justified?
- Does \(M(t)=60\) for a target of 60 imply a 50% probability that the observed count is at least 60? Explain.
Solutions. (1) The fraction is \((42-36)/(48-36)=6/12=0.5\), so the interpolated date is halfway through the interval. (2) The expected total is \(50(0.8)=40\) events. Recurrent-event counts can include multiple events for one participant, so this total is not the count of participants with at least one event. (3) The target is not reached within the stated planning horizon under this projection; a later date requires an explicit extension of the projection. (4) No. \(M(t)=60\) states the expected count is 60; the probability of the observed count meeting the target depends on its distribution, not its mean alone.