Why the Timing Rule Matters
An interim analysis needs both a statistical rule and an operational schedule. The schedule determines when the data are examined; it also affects how much event information is likely to be available at that time. In a time-to-event trial, a planned calendar date does not guarantee a planned event count, and a planned event count does not guarantee a predictable calendar date.
The choice is therefore not simply between an “early” and a “late” analysis. A fixed-date schedule makes the analysis date easier to communicate and coordinate, but the information available on that date can vary. An event-triggered schedule targets a specified amount of event information, but the date when the target is reached is uncertain. Either schedule can be appropriate if it is operationally feasible and the statistical design is calibrated to the actual stopping rule.
What Each Schedule Predicts Well
A fixed-date look has a relatively predictable calendar date. That can help with staffing, committee meetings, vendor arrangements, and communication with sites. Its limitation is that event accrual may differ from forecast. If fewer events have occurred than expected, the analysis may contain less information than planned. If more have occurred, it may contain more. The date alone does not establish that the planned information level has been attained.
An event-triggered look has a more predictable event count, assuming the target is defined in terms of events that are actually available for analysis. The calendar date is less predictable because event accrual varies with recruitment, follow-up, event rates, and other operational factors. If the target is reached later than expected, the trial may need to continue longer than planned, with consequences for costs, staffing, and the overall timeline.
These are operational distinctions, not substitutes for a statistical design. The number and timing of analyses, the information available at each analysis, and the stopping rules must be specified and evaluated together. A team should not treat a missed event target as permission to analyze early under a boundary calibrated for a different information level. Likewise, changing an analysis date or event target can change the procedure’s operating characteristics.
Quantifying the Risk of a Late Event Target
A simple model helps make the timing trade-off concrete. Suppose, for planning purposes, that events accrue according to a homogeneous Poisson process with rate \(\lambda\) events per month. If \(D(t)\) is the number of events by month \(t\), then \(D(t)\sim\operatorname{Poisson}(\lambda t)\). This approximation assumes a constant event rate over the period and independent increments. It is a planning model, not a guarantee that actual trial data follow a Poisson process.
For an event target of \(K\), the target has been reached by month \(t\) exactly when \(D(t)\ge K\). Thus the probability of reaching the target by that date is the upper tail of the event-count distribution:
where \(T_K\) is the time at which the \(K\)th event occurs. Under the constant-rate model, the waiting time to each event has mean \(1/\lambda\) and variance \(1/\lambda^2\). The time to the \(K\)th event is the sum of \(K\) such waiting times, so
The expected date is not a deadline that the target is likely to meet with certainty. For example, at the date \(K/\lambda\), the expected event count equals \(K\), but there remains substantial probability that fewer than \(K\) events have occurred. This distinction is central when judging whether an event-driven schedule is operationally realistic.
Worked Example: Chance of Reaching a Target by the Planned Date
Suppose a hypothetical trial plans its interim look after 120 events. The planning assumption is an average of 10 events per month. What is the approximate probability that the target is reached by month 12?
Under the planning model, \(D(12)\sim\operatorname{Poisson}(10\times12)=\operatorname{Poisson}(120)\). The target is reached by month 12 if \(D(12)\ge120\), so the probability is \(P(D(12)\ge120)\). Using a normal approximation with continuity correction, \(D(12)\) is approximated by a normal variable with mean 120 and standard deviation \(\sqrt{120}\approx10.95\).
Under these assumptions, the chance of having at least 120 events by month 12 is about 52%; the chance of being short is about 48%. The expected date, \(120/10=12\) months, is therefore not a reliable promise that the target will be met by month 12. The approximation is for planning, and its accuracy should be checked against a more suitable accrual model if the rate is not reasonably constant.
Comparing the Schedules Under Different Accrual Scenarios
The comparison becomes clearer when the same trial is considered under more than one plausible accrual scenario. A fixed date keeps the calendar commitment unchanged across scenarios, while the event count at that date changes. An event-triggered schedule keeps the event target unchanged, while the anticipated date changes. Scenario analysis is useful because a single average forecast can conceal the operational consequences of slower accrual.
Worked Example: A Fixed Date Versus a 120-Event Trigger
A hypothetical trial is considering an interim analysis at month 15 or when 120 events have accrued, whichever schedule is selected in advance. Under a faster scenario, the average accrual rate is 8 events per month; under a slower scenario, it is 6 events per month. Compare the expected timing and the information available at month 15.
Under the faster scenario, the expected time to 120 events is \(120/8=15\) months. Under the slower scenario, it is \(120/6=20\) months. Thus a 120-event trigger has an expected date of month 15 in the faster scenario and month 20 in the slower scenario; these are expectations, not guaranteed dates.
At a fixed month-15 look, the expected event counts are \(8\times15=120\) in the faster scenario and \(6\times15=90\) in the slower scenario. Under the Poisson approximation, the slower-scenario count is \(D(15)\sim\operatorname{Poisson}(90)\). The probability of already having 120 events is approximately
So, under the slower scenario, a fixed month-15 look would almost certainly occur before the 120-event target. An event-triggered look would instead wait until the target is reached, with an expected date of month 20 under the constant-rate assumption. The fixed date offers calendar predictability; the event trigger better targets the intended event count. Neither choice removes the need to plan for the slower scenario.
Information, Event Availability, and Operational Lag
The event count is often used as a practical planning measure of information in time-to-event studies, but the relationship is not universally exact. The information relevant to a particular test can also depend on the allocation, censoring, follow-up, and analysis model. In particular, event counts that have occurred are not always the same as events that have been verified, cleaned, and included in the analysis data.
An event-trigger rule should state which events count and what operational steps are required before the analysis. If the trigger is based on adjudicated events, the target is reached later than it would be if based on preliminary reports. If the trigger is based on reported events, the analysis may still have to wait for verification and data processing. A forecast that ignores this distinction can predict an event-trigger date accurately but still predict the analysis date poorly.
Worked Example: Separating the Trigger Date From the Analysis Date
In a hypothetical trial, 100 adjudicated events are required for an interim look. Based on the planning model, the 100th adjudicated event is expected at month 16. The protocol and operational plan estimate that verification and data preparation will take 1.5 months after the threshold is reached. What dates should the team distinguish in its planning?
The expected event-trigger date is month 16. The expected analysis-ready date is \(16+1.5=17.5\) months. If the plan instead described the look simply as “at month 16,” it would confuse reaching the event target with completing the work needed to analyze the data.
These are expected dates, not guarantees. The actual target date can be later or earlier than month 16, and the processing interval can also vary. The schedule should therefore identify the target, the event definition, and the operational steps between reaching the target and conducting the analysis. The statistical boundary must correspond to the planned analysis and stopping rule, not merely to an informal forecast date.
Choosing and Checking a Schedule
A useful choice starts by asking which quantity must be most predictable. If the analysis needs to be coordinated with a fixed external date, a calendar look may be preferred, but the design should assess the likely event count and consequences of arriving with less information than planned. If the analysis is intended to occur at a specified event-information level, an event trigger may be preferred, but the team should assess the distribution of its calendar date and the cost of a delayed target.
| Planning question | Fixed-date look | Event-triggered look |
|---|---|---|
| What is fixed? | The calendar date | The event-count target |
| What can vary? | Events and information available at the date | The calendar date when the target is reached |
| Main operational strength | Easier to coordinate around a known date | Targets a specified event count |
| Main timing risk | Fewer events than planned by the fixed date | The event target is reached later than expected |
Before finalizing either schedule, check that the forecast assumptions are plausible, that the event definition matches the analysis data, and that the operational plan allows for verification and preparation time. Compare more than one plausible accrual scenario. Then confirm that the statistical analysis plan and group sequential design use the chosen timing rule and any relevant information levels consistently.
A practical warning sign is a forecast that gives a single target date without showing what happens if accrual is slower. Other warnings include treating expected event counts as guaranteed counts, counting events differently in the forecast and analysis, or assuming that a calendar-date look automatically has the information level planned for it. Observed accrual should be monitored and projections updated using current data, but updates to a forecast do not by themselves authorize changing the prespecified analysis rule.
Common Pitfall and Limitations
Wrong reasoning: “The expected date for 120 events is month 12, so we can schedule the event-triggered look for month 12 and expect to have 120 events.” The correction is that an expected date is an average over possible accrual paths. At month 12, the event count can be above or below 120; the probability calculation, not the expected date alone, describes the chance of meeting the target by then.
The Poisson calculations here rely on a constant-rate approximation. Real event accrual may change as enrollment changes, participants accumulate follow-up, or event rates differ across participants. Operational delays may also vary. More detailed scenario models may therefore be needed for planning. Regardless of the forecasting method, neither a fixed date nor an event trigger guarantees a particular amount of statistical information, and neither should be altered informally after looking at the emerging treatment results.
Check Your Understanding
Answer each question using the stated assumptions. Solutions are included after each question.
- A target is 80 events, and the planning rate is 5 events per month. What is the expected time to the target under a constant-rate model?
Solution: \(E(T_K)=K/\lambda=80/5=16\) months. This is the expected time, not a promise that the target will be reached by month 16.
- At a fixed date, a trial expects 70 events under its slower scenario but planned an analysis around 100 events. What is the central scheduling concern?
Solution: The calendar date remains predictable, but the analysis may have substantially fewer events, and therefore less event information, than planned. The team should assess the statistical and operational consequences rather than assume that the planned information level has been attained.
- Under a Poisson planning model, \(D(10)\sim\operatorname{Poisson}(50)\). Write the probability that a 55-event target has been reached by month 10.
Solution: The target is reached when \(D(10)\ge55\), so the probability is \(P(D(10)\ge55)=1-\sum_{j=0}^{54}e^{-50}50^j/j!\). A numerical value would require evaluating this Poisson tail.
- A protocol triggers a look at 90 verified events, but the forecast counts preliminary event reports. What should be corrected in the planning process?
Solution: The forecast should match the trigger definition by accounting for verification, or the operational plan should explicitly model the delay from preliminary reports to 90 verified events. The analysis date should also allow for any data preparation required after the threshold is reached.