Why a Monitoring Schedule Needs More Than Dates
A monitoring schedule is part of the statistical procedure, not just a calendar reminder. The plan should make clear when the accumulating trial data will be analyzed, which hypothesis or hypotheses will be tested, and what decisions may follow. If any of these are left open, the analysis team may have to make choices after seeing accumulating results.
A schedule also needs operational detail. “Analyze at the halfway point” is incomplete unless the plan defines halfway by a measurable quantity, explains which participants’ outcomes are ready for analysis, and says whether the analysis can be delayed if the required data are incomplete. For a beginner’s plan, a clear sequence of two interim analyses followed by a final analysis is a useful starting point.
The schedule is not itself a rule for deciding statistical significance. It specifies opportunities to evaluate the evidence. The testing rule, including any thresholds used at those opportunities, must also be specified and justified. In particular, listing several planned tests does not make it valid to use the single-analysis significance level independently at every look without considering the resulting error rate.
Specify the Three Core Elements
A usable plan connects each analysis time to a defined hypothesis and a defined set of decisions. The same primary hypothesis may be tested at every look, but that should be stated rather than assumed.
| Element | What to specify | Example of clear wording |
|---|---|---|
| Planned analysis time | A measurable sample-size or calendar trigger, and the data that must be available | After 200 participants have been randomized and their primary endpoint outcomes are available |
| Tested hypothesis | The outcome, comparison, population, and direction or sidedness of the hypothesis | The two-sided primary comparison of the treatment and control event proportions |
| Decision opportunity | Which actions are permitted after reviewing the result | Continue the trial, or stop early only if the prespecified stopping criterion is met |
A decision opportunity is not a promise that a decision must change. For example, a planned interim analysis may allow the monitoring committee to recommend continuing as planned, while permitting early stopping only under a stated rule. The plan should distinguish a statistical finding from the action authorized by the protocol.
State the hypotheses in context. If \(p_T\) is the probability of the primary outcome in the treatment group and \(p_C\) is the probability in the control group, a two-sided comparison might be written as:
This statement identifies the comparison, but a complete plan should also identify the outcome definition and the population to which the comparison applies. If the primary question is directional, the alternative hypothesis must reflect that direction and be justified in advance. Do not switch between one-sided and two-sided testing after observing results.
Build a Two-Interim-Analysis Plan
Consider a hypothetical randomized trial planning to enroll 400 participants. The primary outcome is measured for every randomized participant. Suppose the design calls for two interim analyses before the final analysis. One straightforward schedule uses cumulative enrollment triggers of 200, 300, and 400 participants. The first two are interim looks; the third is the final analysis.
Enrollment count alone does not guarantee that outcome data are ready. The plan should therefore say whether each trigger means “randomized” or “randomized with primary outcome available,” and how missing or delayed outcomes will be handled. A trigger based on complete primary-outcome data is often easier to interpret than a trigger based only on enrollment if the outcome is not observed immediately.
Name the primary outcome and comparison, then write \(H_0\) and \(H_a\) precisely.
State cumulative enrollment or another operational trigger for each planned look, including the final analysis.
Say which hypotheses are tested and which decisions can be made at that look.
Clarify that analyses are not added merely because results or data summaries become available, and describe how delays or incomplete data affect the planned trigger.
Worked Example: Enrollment Triggers for Two Interim Looks
A hypothetical trial plans to randomize 400 adults equally to a treatment or control group. The primary outcome is assessed 30 days after randomization. The plan calls for two interim analyses and one final analysis. Specify the schedule and calculate the cumulative enrollment fractions at those analyses.
Analysis times. Plan the first interim analysis after 200 participants have been randomized and have reached the 30-day outcome assessment. Plan the second after 300 participants meet the same condition. Conduct the final analysis after all 400 randomized participants have reached the assessment, subject to the prespecified outcome-data rules.
The cumulative enrollment fractions are:
Thus, the interim analyses are planned at 50% and 75% of the target enrollment, and the final analysis is planned at 100%. These are enrollment fractions, not a claim that the information for the statistical test is exactly 50%, 75%, and 100%.
Hypothesis. If the primary outcome is binary, define \(p_T\) and \(p_C\) as the 30-day event probabilities in the treatment and control groups. At each of the three planned analyses, the plan may specify the same primary comparison, \(H_0:p_T-p_C=0\) versus \(H_a:p_T-p_C\ne0\).
Decision opportunities. State which interim findings may lead to an early stop and under what prespecified rule; otherwise, the default action is to continue toward the next planned analysis. State that the final analysis addresses the primary trial conclusion. The schedule is now specific about triggers, tested hypothesis, and decision opportunities without implying that an unadjusted threshold at each look controls the overall error rate.
Why Prespecification Matters
Let \(R_i\) denote the event that the null hypothesis is rejected at planned look \(i\), and suppose there are two interim looks and a final look. If the trial’s claim can be based on a rejection at any of these three analyses, the event of at least one rejection is:
This follows directly from the decision rule: a false-positive claim can occur if any permitted look rejects a true null hypothesis. Therefore, the set of planned decision opportunities is part of what defines the procedure whose error rate must be evaluated. Specifying the schedule in advance makes that set explicit; it does not, by itself, ensure that the overall probability of a false-positive claim is at the desired level.
This is the practical reason to write down the schedule before outcomes are reviewed. If investigators add a look after seeing an unfavorable trend, or move a planned analysis to a moment when the results seem especially promising, the actual set of opportunities no longer matches the stated plan. Earlier material on repeated significance testing establishes why repeatedly testing a true null can increase the chance of at least one false rejection. Here, the key planning task is to define the opportunities that a testing procedure must address.
Worked Example: Converting a Calendar Trigger Into an Operational Trigger
A hypothetical study begins recruiting on January 1 and expects to finish enrollment in 12 months. A draft says, “Perform interim analyses at six and nine months.” The primary outcome is assessed 30 days after randomization. Clarify what these instructions mean.
The draft does not say whether “six months” means six months after the first participant is randomized, after the first site opens, or after recruitment begins. It also does not say whether participants randomized near the trigger are included if their 30-day outcomes are not yet available.
A revised calendar plan could state: “The first interim analysis will occur on July 1, six months after the first participant is randomized, using primary-outcome data available for all eligible participants as of the data cutoff. The second will occur on October 1, nine months after the first participant is randomized, using the same outcome definition and data-cutoff rule.” If data are insufficient for the planned analysis, the protocol must state whether the analysis is delayed and what condition ends the delay.
Alternatively, the plan could use enrollment triggers. If 150 and 250 participants are expected to have completed the 30-day assessment by the corresponding dates, specify those triggers directly and say what happens if the expected counts are not reached. The dates and counts are hypothetical planning values; they are not interchangeable unless the protocol defines their relationship.
The revised wording removes ambiguity about the time origin, outcome availability, and data cutoff. The hypothesis and decisions still need to be stated separately; calendar details alone do not complete the monitoring plan.
Worked Example: Making the Tested Hypothesis and Decisions Explicit
A hypothetical two-group trial measures whether participants experience a defined adverse event by 90 days. Its draft monitoring statement reads, “Review safety and efficacy at 100, 200, and 300 participants.” Rewrite the statistical part for the primary efficacy question, assuming 300 participants is the planned maximum.
Let \(p_T\) and \(p_C\) be the probabilities of the defined 90-day outcome in the treatment and control groups. If the primary question is whether the probabilities differ, state:
Specify analyses after 100 and 200 participants have completed the 90-day assessment, followed by a final analysis after 300 have completed it. These are two interim looks and one final look. At each interim analysis, state whether the committee may recommend continuation or may recommend stopping for a prespecified efficacy or safety reason. At the final analysis, state that the primary efficacy hypothesis will be evaluated according to the protocol’s specified testing rule.
The rewritten plan is more precise because it names the outcome, comparison, timing condition, and permitted decisions. It does not invent a stopping boundary or assign a significance threshold; those require separate specification and justification.
Diagnostics, Alternatives, and Limits
A schedule is difficult to implement when its triggers cannot be determined consistently from trial records. Before enrollment, ask whether two independent analysts would identify the same date or participant count for each planned look. Ambiguous phrases such as “when enough data are available,” “around halfway,” or “if the trend warrants it” are warning signs unless the plan defines them with objective criteria.
Check the schedule against the way the outcome is collected. For a delayed outcome, a participant may be enrolled but not yet contribute an observed primary endpoint. Clarify whether the trigger uses randomized participants, participants with completed follow-up, or another defined quantity. Also state how missing outcome records, delayed sites, and data corrections are handled. Otherwise, the effective timing may drift according to informal choices.
The main alternative to a fixed schedule is a monitoring approach in which analysis timing is not restricted to those preplanned discrete triggers. That is a different monitoring design, not merely a less detailed calendar. Do not substitute one for the other without specifying the resulting opportunities and an appropriate testing procedure. The distinction between scheduled looks and continuous monitoring requires separate treatment.
A fixed schedule has limits. It may be operationally inconvenient if recruitment or follow-up is slower than expected, and a fixed calendar date can arrive before enough primary outcomes are available. A schedule also cannot answer by itself whether a trial should stop, how to adjust testing thresholds, or how to interpret treatment effects after stopping. Those questions require additional design and analysis choices.
Common Mistakes
- Listing dates but not data conditions. A date does not establish which participants’ outcomes are included. Define the data cutoff and outcome-availability rule.
- Calling enrollment fractions information fractions. The ratio of randomized participants to target enrollment is an enrollment fraction. Do not label it an information fraction unless information has been defined and calculated appropriately.
- Leaving the hypothesis implicit. State the endpoint, comparison, and direction or sidedness of the test at each planned look.
- Treating an opportunity as an automatic stop. A planned look permits an evaluation; the stopping action depends on the prespecified decision rule.
- Assuming a schedule controls error. Prespecification clarifies when tests occur, but repeated testing still requires an error-control strategy appropriate to the planned opportunities.
Check Your Understanding
Answer each question using the definitions and planning principles above. Short solutions follow the questions.
- A trial plans 240 participants, with interim analyses after 120 and 180 participants and a final analysis after 240. How many interim looks are planned, and what are the cumulative enrollment fractions?
- Why is “analyze when half the outcomes are ready” less precise than a prespecified participant-count trigger?
- For a two-sided comparison of treatment and control event probabilities, write the null and alternative hypotheses using \(p_T\) and \(p_C\).
- A plan permits an interim analysis but does not say whether the trial may stop early. What is missing?
- Does prespecifying three analysis times, by itself, guarantee control of the overall false-positive rate? Explain briefly.
Solutions. (1) There are two interim looks. The cumulative enrollment fractions are \(120/240=0.50\), \(180/240=0.75\), and \(240/240=1.00\) at the two interim analyses and final analysis. (2) “Half the outcomes” needs a defined denominator and a rule for which outcomes count as ready; a fixed trigger makes the analysis timing more reproducible. (3) \(H_0:p_T-p_C=0\) versus \(H_a:p_T-p_C\ne0\). (4) The plan must state the permitted decisions, including whether stopping is allowed and the prespecified condition for doing so. (5) No. The schedule identifies planned opportunities, but the testing procedure must still account for repeated opportunities to reject a true null hypothesis.