Why the Schedule Needs More Than a Date
An interim analysis is not fully specified by saying that it will occur “about halfway through the trial.” The protocol needs to say what statistical information will trigger each look, which data count toward that information, and what happens between reaching the trigger and making the interim decision. Calendar dates are useful for planning staff and committee meetings, but they are forecasts: recruitment, follow-up, event occurrence, and data processing do not proceed at perfectly predictable rates.
A clear schedule separates the statistical trigger from the operational sequence. The trigger defines when the planned information target has been reached. Operational rules define the data cut-off, checks, analysis preparation, and decision timing. This distinction helps prevent a meeting date from being mistaken for a statistical look or a convenient calendar date from silently changing the design.
State the Information Scale and the Trigger
Information is tied to the estimator or test statistic used in the design. In a simple model, it can be expressed as the reciprocal of the variance of an estimator. For example, suppose independent observations in two treatment groups have a common, known standard deviation \(\sigma\), with \(n_T\) and \(n_C\) evaluable observations. For the estimated difference in group means, \(\bar{x}_T-\bar{x}_C\), the variance and corresponding information are
This formula assumes independent observations, a fixed common variance, and the stated group sample sizes. In other designs, information must be defined on the scale used by that design; a raw participant count is not automatically an information measure. The protocol should identify the endpoint, analysis population, information calculation, and any estimator or nuisance-parameter assumptions that determine the target.
For each look, state the planned information target or fraction and the operational rule for reaching it. If the design uses an event-based trigger, define exactly which events count and how their information is assessed. If the design uses evaluable participants, define evaluability and specify whether the count is required in total or by treatment group. The trigger should not depend on a vague phrase such as “sufficient data” when a measurable rule can be written.
Worked Example: Converting Information Fractions to Sample-Size Targets
Suppose a hypothetical two-group study plans a continuous outcome, equal allocation, a common standard deviation of 10 points, and a maximum of 200 evaluable participants per group. Assume the independent-observation model above. The planned looks are at information fractions 0.50, 0.75, and 1.00.
At equal sample sizes \(n_T=n_C=n\), the variance of the difference is \(100(1/n+1/n)=200/n\), so \(I=n/200\). At the maximum, \(n=200\) per group and \(I_{\max}=200/200=1.00\). The three information targets are therefore \(I_1^*=0.50\), \(I_2^*=0.75\), and \(I_3^*=1.00\).
Solving \(n/200=I_k^*\) gives \(n=200I_k^*\). Thus the targets correspond to 100, 150, and 200 evaluable participants per group, respectively, or 200, 300, and 400 in total. A protocol can use these group-specific evaluable counts as operational trigger measures under the stated model. These counts are targets, not guaranteed calendar dates; losses to follow-up or unusable outcomes may delay their attainment.
For instance, at the first look the rule could require at least 100 evaluable outcomes in each group, with the information calculation verified using the prespecified method. The protocol should also state how it handles a small imbalance or an information value that does not match the nominal fraction exactly, rather than assuming that a rounded participant count always gives the intended information.
Deriving and Checking an Information Trigger
The reciprocal-variance definition provides a direct check that a sample-size rule really corresponds to its stated information target. In the same two-group model, increasing either group size reduces the variance and increases information. If group sizes differ, the information is determined by both sizes, not by total enrollment alone.
The second equality follows by combining the two fractions inside the parentheses. It shows why equal total enrollment can yield different information when the allocation between groups differs. A protocol that sets a total randomized-patient count as its trigger should therefore justify that this count adequately represents information under its allocation and evaluability assumptions.
Worked Example: Checking Whether an Information Target Has Been Reached
Consider a hypothetical snapshot in the same continuous-outcome setting, with 104 evaluable participants in the treatment group and 96 in the control group. Let the common standard deviation be 10 points. Suppose the interim trigger requires information of at least 0.50.
The variance of the estimated mean difference is
Thus the information is
Under an exact rule requiring \(I\geq0.50\), this snapshot does not meet the trigger: its information is about 0.0008 below the target. The small difference might be operationally minor, but it cannot be silently rounded upward if the protocol defines an exact threshold. The schedule should prespecify whether a tolerance, a recalculated actual information value, or a further accrual step determines when the look proceeds.
Define Data Availability and Decision Timing
Reaching an information trigger and having a decision-ready analysis are related but distinct conditions. The trigger is determined by the trial’s prespecified information measure. Readiness also depends on whether the data needed for the analysis have been collected, checked, and made available according to the protocol. For example, endpoint events may require adjudication, and a continuous endpoint may require completion of a specified follow-up window.
The protocol should make the handoff from trigger to decision reproducible. It should identify the data cut-off rule, the checks required before analysis, the responsible roles, the expected interval for preparing and reviewing the analysis, and the planned decision meeting. It should state whether the analysis uses data available by a fixed cut-off or requires all qualifying outcomes through the trigger to be finalized. These details help ensure that different groups do not apply different meanings to “the interim data.”
Worked Example: An Event-Driven Look and Its Data Cut-Off
Suppose a hypothetical time-to-event trial plans looks at 80, 160, and 240 confirmed primary events. For illustration, under equal randomization the approximate information for a log hazard ratio may be \(I\approx d/4\), where \(d\) is the number of qualifying events. This approximation assumes the specified equal-allocation event-based analysis; the protocol should use the information definition established for its actual design.
The planned targets are approximately 20, 40, and 60 information units, giving fractions \(20/60=1/3\), \(40/60=2/3\), and \(60/60=1\). Now suppose the trial has 154 confirmed events and six additional potential events awaiting adjudication. Under a rule requiring 160 confirmed events, the trigger has not been reached. The six pending cases do not count merely because they may later qualify.
Suppose the protocol also forecasts eight additional confirmed events per month. From 112 confirmed events, the expected time to 160 is \((160-112)/8=6\) months. If the forecast date is June 30, the event target is forecast around December 31. If the protocol allows one month for adjudication and two weeks for analysis preparation and the decision meeting, the meeting is forecast about six weeks after that target date, around mid-February. These are planning estimates, not substitutes for the confirmed-event trigger. The protocol should say how a delay in adjudication or event accrual changes the calendar forecast without changing the event definition.
Permitted Schedule Changes and Failure Modes
A prespecified schedule can still encounter delays, overshoot, or changes in trial operations. The protocol should distinguish routine operational adjustments from changes that could alter the statistical design. For example, rescheduling a meeting because a data extract is delayed does not by itself redefine the information target. By contrast, changing the trigger after access to emerging treatment comparisons may affect the operating characteristics of the planned sequential procedure. Such a change should not be made informally.
Specify who may propose and approve schedule changes, what documentation is required, and whether the change must be assessed by the statistician responsible for the design before any relevant unblinded results are available to decision-makers. State how actual information at the analysis will be used under the planned method, including what happens if accrual passes a target while data are being prepared. Do not assume that the boundary intended for a target fraction can automatically be applied to any overshot information value; follow the design’s prespecified rule.
Common warning signs include a calendar date being treated as proof of information attainment, randomized counts being substituted for evaluable counts without justification, pending outcomes being inconsistently counted, or a meeting being moved in response to an observed treatment difference. The correction is to keep the statistical trigger measurable, apply the same data rules at every look, and route schedule changes through the stated review and approval process.
Information-based triggers are often preferable when accrual and follow-up rates vary, because they tie the look to the amount of statistical information rather than an uncertain date. Fixed calendar looks can be easier to coordinate, but may occur at materially different information levels than planned. Event-count triggers are transparent for event-driven outcomes, but event counts approximate information only under the assumptions of the design. A participant-count trigger may be practical for some endpoints, but requires a clear link between the count and actual information. Whichever approach is selected, it must be consistent with the planned analysis and its sequential error control.
Check Your Understanding
Answer each question using the definitions and rules in this tutorial.
- In the equal-allocation continuous-outcome example, what evaluable sample size per group corresponds to an information fraction of 0.75?
- With 104 and 96 evaluable participants and a common standard deviation of 10, does an exact information threshold of 0.50 trigger the look?
- A protocol requires 160 adjudicated events. There are 158 confirmed events and four events awaiting adjudication. Has the trigger been reached?
- Why is a forecast date not equivalent to an information-based trigger?
- What should happen if the planned information target is overshot while the interim analysis is being prepared?
Solutions. 1. Since \(I=n/200\), set \(n/200=0.75\), giving \(n=150\) evaluable participants per group. 2. No. The variance is \(100(1/104+1/96)\approx2.0032\), so information is approximately \(1/2.0032=0.4992\), below 0.50. 3. No. The rule requires 160 adjudicated events; the four pending events do not count until confirmed. 4. A date forecasts when the trigger may be attained, but accrual, follow-up, or data processing can differ from the forecast; the prespecified information measure determines attainment. 5. Apply the overshoot rule specified for the design, using the actual information and approved analysis procedure. Do not automatically use a boundary or rule intended for the nominal target without checking that the design permits it.