Why Interim Completeness Matters
An interim package is a dated view of a trial, not a guarantee that every relevant observation has arrived. Some participants may not yet have reached an assessment time, some assessments may be overdue or unobtainable, and adverse-event reports may still be moving through review and reconciliation. These situations have different meanings. If they are summarized together as “missing,” reviewers can mistake immature data for negative findings, or treat incomplete ascertainment as evidence of safety.
The earlier material on setting the interim data cut and validating a snapshot establishes why the analysis cut and package must be tied to a defined data state. Here the question is how to judge what that state can support. The goal is not to repair an incomplete snapshot by assumption. It is to describe what is observed, what is unresolved, how the unresolved part could affect interpretation, and what checks or follow-up are needed.
Describe the Data Before Interpreting It
For each interim endpoint or safety summary, reviewers need a denominator that matches the question being asked. The number randomized, the number treated, the number due for assessment, and the number with an observed assessment are not interchangeable. A package should make these counts visible by treatment group and, when useful, by site, enrollment period, or time since randomization. It should state the data cutoff, the assessment window, and the rule used to classify a record as observed, pending, or missing.
A missingness percentage is informative only with its denominator and reason. For a scheduled binary endpoint, for example, report the number randomized, the number who have reached the assessment window, the number assessed, and the number overdue or otherwise unavailable. Reasons such as withdrawal, missed visit, technical failure, and ongoing follow-up may point to different risks. A participant who has not yet reached the visit is not an observed nonresponder; counting that participant as one would silently impose an outcome value.
A concise supplemental check can ask whether data availability differs across randomized groups or across clinically important participant groups. Such a difference does not by itself prove bias. It does mean that an observed-case summary may describe different subsets of participants in the groups being compared. Reviewers should also check that the snapshot’s counts reconcile with the trial’s tracking systems and that late-arriving records are identified rather than silently mixed into a later data state.
Worked Example: Bounding an Interim Binary Outcome
Hypothetical data. At a cut, 50 participants have been randomized to each of two groups. The scheduled response outcome is available for 40 participants in Group A, of whom 24 responded, and 45 participants in Group B, of whom 27 responded. The remaining outcomes are unresolved: 10 in Group A and 5 in Group B. Assume only that each unresolved outcome is either response or no response; do not assume that it resembles the observed outcomes.
Among participants with an observed outcome, the response proportions are \(24/40=0.600\) in Group A and \(27/45=0.600\) in Group B. These are observed-case summaries, not response proportions for all randomized participants.
A simple worst-to-best bound for each group assigns all unresolved outcomes first to no response and then to response. For Group A, the lower and upper bounds are \(24/50=0.480\) and \((24+10)/50=0.680\). For Group B, they are \(27/50=0.540\) and \((27+5)/50=0.640\). Thus the Group A minus Group B difference could range, under these assignments, from \(0.480-0.640=-0.160\) to \(0.680-0.540=0.140\), or from −16 to 14 percentage points.
These bounds are deliberately broad and do not estimate the likely treatment effect. They show that the observed equality of 60% response does not settle the all-randomized comparison: the unresolved outcomes could change its direction. Before using the calculation, verify that the stated randomized denominators and unresolved counts are correct, and distinguish participants whose assessment is not yet due from those whose outcome is overdue.
Delayed Safety Reports Need Their Own Completeness Check
A safety count at an interim cut is a count of reports available under specified definitions and data flows by that date. It is not necessarily a count of every event that has occurred. The package should clarify the event ascertainment period, the reporting sources, the date through which records were received, and whether reports remain under follow-up, coding, or reconciliation. For serious or otherwise important events, useful checks include comparing the safety database with relevant source or tracking listings, reviewing report and receipt dates, and describing unresolved cases by group and status.
Rates can account for differing observed exposure time, but they do not fix incomplete reporting. If a group has \(E\) reported events and \(T\) observed person-time, the crude rate is \(E/T\). Interpreting it as a rate of occurrence requires, among other things, that the event definition and ascertainment be comparable and that the exposure-time denominator be appropriate. A rate based on events reported by the cut may still be biased downward if reports arrive late, especially if reporting delays differ by group or site. Nor does an event rate alone describe event severity, recurrence, or the number of participants affected.
Worked Example: Checking Delayed Adverse-Event Reports
Hypothetical data. By the interim cut, Group A has 3 received serious adverse-event reports and 150 person-months of observed exposure; Group B has 2 received reports and 160 person-months. A reconciliation listing identifies 2 additional potential reports for Group A and 5 for Group B that are not yet in the package. For this illustration, suppose each listed potential report would qualify under the same event definition if confirmed. The listing is a check on reporting completeness, not an assertion that every potential report is a confirmed event.
The rates using only received reports are \(3/150=0.0200\) and \(2/160=0.0125\) events per person-month. Multiplying by 100 gives 2.00 and 1.25 reported events per 100 person-months, respectively. These are descriptive rates from the records available at the cut.
If all potential reports were confirmed, the corresponding illustrative upper counts would be 5 for Group A and 7 for Group B. The rates would then be \(5/150\times100=3.33\) and \(7/160\times100=4.38\) events per 100 person-months. This range is not a confidence interval and the upper calculation is not a substitute for adjudication. It demonstrates that the apparent ordering of the rates could change if pending reports are confirmed. The DMC needs the report status, timing, and clinical details before drawing a safety conclusion.
Uneven Follow-Up Changes What a Proportion Means
An interim comparison can have uneven follow-up because participants enrolled at different times or because visits occur on different schedules. For a fixed-time endpoint, the proportion among participants who have reached that time answers a different question from the proportion among everyone randomized. Even within the assessed subset, the groups can differ in calendar time, sites, or participant mix. A package should therefore show how much follow-up has accrued and how many participants are eligible for each time-specific summary.
For time-to-event outcomes, a crude event proportion ignores how long each participant was observed and how censoring occurred. A Kaplan–Meier estimate can accommodate varying observation times under its assumptions, including independent censoring with respect to the event process in the relevant analysis population. At an interim review those assumptions cannot simply be presumed from the display. The number at risk over time, censoring patterns, follow-up distribution, and data maturity should be examined. If follow-up is very short or highly uneven, the estimated curve may be unstable or not yet informative for the time horizon of interest.
Worked Example: Comparing a Fixed-Time Outcome With Unequal Maturity
Hypothetical data. There are 30 randomized participants in each group. At the cut, 18 in Group A and 27 in Group B have reached the scheduled 90-day assessment. Four of the 18 assessed in Group A and 6 of the 27 assessed in Group B had the event by day 90.
Among participants who have reached the assessment, the observed event proportions are \(4/18=0.222\) in Group A and \(6/27=0.222\) in Group B. But maturity differs: 12 participants in Group A and 3 in Group B have not yet reached day 90. The equal observed proportions do not mean that the groups have equally complete 90-day data.
If unresolved outcomes are simply left out, the comparison uses 18 people in one group and 27 in the other. If they are instead treated as no event, the all-randomized proportions become \(4/30=0.133\) and \(6/30=0.200\), but this incorrectly assigns outcomes to 15 participants who have not reached the assessment. The defensible interim presentation keeps the observed, not-yet-due, and overdue counts separate and explains the difference in maturity. It does not infer a group difference from these crude calculations.
Supplemental Checks and Limits
A practical review can proceed in a fixed sequence. First establish the data state and definitions; next quantify what is available; then check whether delays or missingness are patterned; finally show how uncertainty about unresolved records limits interpretation. The purpose is to make the boundary between observed evidence and assumption explicit.
Verify the data cutoff, endpoint windows, randomized and treated counts, and the numbers due, observed, pending, overdue, or unavailable.
Compare assessment availability and report status across groups, sites, and enrollment periods. Investigate discrepancies against the relevant tracking records.
Give counts and reasons for missing or delayed records, and show follow-up or exposure time in a form that matches the endpoint.
Use transparent bounds or sensitivity summaries where appropriate, state their assumptions, and identify what additional data or reconciliation would change the review.
Simple bounds are useful because they require few assumptions, but they may be wide. Imputation or model-based adjustments can be more precise only by adding assumptions about the unobserved outcomes and the mechanism generating missingness. For example, an analysis assuming outcomes are missing at random conditional on recorded information is not justified merely because recorded variables are available. It requires a defensible model and adequate measured predictors. A DMC should be able to distinguish observed results from results that depend on such assumptions.
Likewise, exposure-adjusted rates are an alternative descriptive summary to raw counts when exposure time differs, not a cure for differential reporting or a replacement for clinical review. For fixed-time outcomes, restricting to those who have reached the time point is clearer than counting not-yet-due participants as failures, but it can still select a subset whose composition differs. No single summary resolves every limitation. The package should make the limitation visible and avoid suggesting that a data-quality check establishes the absence of events or a treatment effect.
Common Mistakes
- Calling every blank “missing.” A visit not yet due is pending, not an overdue assessment. Use schedule and status information to classify it.
- Using randomized participants as an outcome denominator without stating the rule. A proportion over all randomized participants is meaningful only when outcomes are defined for all or unresolved values are handled transparently.
- Reading delayed reports as no events. “Not received by the cut” does not mean “did not occur.” Reconcile report status and timing before interpreting counts.
- Assuming a rate corrects incomplete ascertainment. Person-time adjusts the denominator for exposure duration; it cannot recover events that were not reported or recorded.
- Treating a sensitivity bound as a likely estimate. Worst-to-best assignments show possible impact under stated limits, not the most plausible outcome or a probability statement.
Check Your Understanding
Answer each question, then compare with the solutions below.
- In a group of 40 randomized participants, 28 have reached a scheduled outcome visit, 24 have an observed outcome, and 4 are overdue. How many are pending, and how many are missing or overdue at the cut?
- In a hypothetical group, 5 events are reported over 250 person-months. Calculate the reported event rate per 100 person-months, and name one limitation of that rate.
- Why is it incorrect to classify participants who have not reached day 90 as having no event by day 90?
- What does a worst-to-best bound for unresolved binary outcomes communicate, and what does it not communicate?
Solutions. (1) Twelve have not yet reached the visit, so 12 are pending; 4 are overdue and unavailable, so 4 are missing at the cut under the stated classification. The 24 observed plus 4 overdue account for the 28 who have reached the visit. (2) \(5/250\times100=2.00\) reported events per 100 person-months. It may understate occurrence if reports are delayed or incomplete, and it does not describe severity or unique participants. (3) Their 90-day outcome is not yet observed; assigning no event gives them an outcome value without evidence and can distort the comparison. (4) The bound shows the range of group summaries possible under extreme assignments of unresolved outcomes. It is not a likely-effect estimate, confidence interval, or probability that the true value lies in the range.