What an Interim Snapshot Represents
An interim review does not examine “the trial so far” in the abstract. It examines a particular set of records, selected under a particular timing rule and represented by a particular database state. A snapshot is therefore a defined data object: its contents depend on when outcomes are cut off, which participants have had enough time for the endpoint to be observed, which records have been received and processed, and which database version is used.
These distinctions matter even when reviewers see only blinded or pooled information. A snapshot can describe what was recorded by a date; it cannot automatically describe what will eventually be observed in the randomized population. Before interpreting interim information, reviewers need to know what records were eligible for inclusion and what follow-up those records represent.
The cut date is not necessarily the date on which a reviewer receives the data, nor the date a database is locked or exported. For example, a database extract generated on 10 July might include records received through 30 June. Those dates answer different questions: the first says when the extract was generated; the second says how far the data are intended to extend. A snapshot specification should make the distinction explicit.
Follow-Up Status Defines What Is Observable
A participant can be randomized but not yet have reached the endpoint’s assessment time by the cut date. Another participant may have reached that time, but the assessment may not yet have been collected, entered, or resolved. These are different states. “No event recorded by the cut date” is not equivalent to “the participant completed the relevant follow-up and had no event.”
A useful snapshot summary therefore separates at least the number randomized, the number who have reached the relevant follow-up time, the number with an ascertained outcome, and the number whose outcome is unresolved or missing. The appropriate categories depend on the endpoint. Time-to-event outcomes, repeated measures, and outcomes with visit windows require their own operational definitions; a single overall “followed” count can conceal important differences.
Worked Example: Separating Maturity From Outcome Availability
Suppose a hypothetical trial has randomized 100 participants. By the 30 June cut date, 72 have reached the protocol-defined 90-day assessment window. Of these 72, endpoint assessments are recorded for 66. Eight of those 66 recorded assessments meet the endpoint definition. The remaining 28 participants have not yet reached the 90-day window.
The assessment completion proportion among participants who have reached the window is \(66/72=0.9167\), or 91.67%. The unresolved number among mature participants is \(72-66=6\), and the number not yet mature is \(100-72=28\). These groups should not be combined as though they had the same follow-up status.
The recorded endpoint proportion among participants with an assessment is \(8/66=0.1212\), or 12.12%. This describes the 66 recorded assessments, not all 100 randomized participants and not even all 72 participants who reached the window. It does not establish the eventual endpoint proportion in the randomized cohort: six mature participants lack recorded outcomes, and 28 have not yet reached the assessment time.
The snapshot supports reporting those counts and the explicitly defined 12.12% recorded-assessment proportion. It does not, without additional assumptions and methods, justify treating every unrecorded outcome as a non-event or treating the observed proportion as the final trial result.
Why Cut Dates Can Change the Snapshot
A cut date determines which information can be present, but it does not give every participant equal opportunity to contribute an outcome. In an enrolling trial, participants randomized at different times have different follow-up durations at a common calendar cut. A calendar date is thus not itself a common participant-level follow-up time.
To see why a crude event proportion can shift with a later cut, divide the randomized participants into those with an ascertained outcome and those without one. Let \(N\) be the number randomized, \(R\) the number with an ascertained outcome by the cut date, \(Y\) the number of events among them, and \(U=N-R\) the number without an ascertained outcome. The recorded proportion is \(Y/R\). If the eventual event count among the \(U\) participants is \(Z\), the eventual proportion would be \((Y+Z)/N\). The relationship is
When both groups are nonempty, the eventual proportion is a weighted average of the recorded proportion and the eventual proportion in the currently unascertained group. The snapshot contains \(Y\) and \(R\), but generally not \(Z\). Therefore, the recorded proportion equals the eventual proportion only under special conditions, such as \(Y/R=Z/U\); that equality cannot be assumed merely because the data are blinded or because the observed denominator is large.
This identity is not a method for filling in missing outcomes. It clarifies why follow-up completeness and outcome availability affect interpretation. If ascertainment is related to prognosis, treatment, site, or calendar time, the currently observed participants may not represent those without recorded outcomes. Interim reviewers should examine the status and pattern of follow-up, not only the event count.
Worked Example: Comparing Two Cut Dates
In a hypothetical blinded trial, 80 participants are randomized, 40 in each coded group. The endpoint is assessed 30 days after randomization. At a 31 May cut, 50 participants have reached day 30: 32 have recorded assessments and 5 of those assessments report the endpoint. By a later 30 June cut, 68 participants have reached day 30: 60 have recorded assessments and 9 report the endpoint.
At the first cut, assessment completion among participants who reached day 30 is \(32/50=0.6400\), or 64.00%. The recorded-assessment endpoint proportion is \(5/32=0.1563\), or 15.63%.
At the later cut, assessment completion among participants who reached day 30 is \(60/68=0.8824\), or 88.24%. The recorded-assessment endpoint proportion is \(9/60=0.1500\), or 15.00%. The later snapshot has more recorded outcomes and better ascertainment among mature participants, but the difference between 15.63% and 15.00% is not by itself evidence that the underlying endpoint risk changed. The denominator, maturity, and participant composition also changed.
Because the data are pooled across coded groups, these calculations do not compare treatment groups. They summarize recorded outcomes under each snapshot’s rules. A reviewer should also verify that both cuts use the same endpoint definition, assessment window, and data-processing rules before treating the proportions as comparable.
Database Versions and Reproducible Interpretation
A database version is part of the meaning of an interim result. Between versions, records can be added, corrected, adjudicated, deduplicated, or reclassified. A revised count can therefore reflect data cleaning rather than new participant follow-up. Without version control, two reviewers may appear to disagree when they have actually analyzed different data states.
A snapshot should be traceable to a version identifier and a documented extract or freeze time. Its description should also state the cut date, the rules for including participants and endpoints, the follow-up status categories, and any relevant outstanding data queries or adjudications. If a later version is used, the change from the earlier version should be attributable to identifiable record-level or processing changes, rather than silently substituted.
Worked Example: Interpreting a Revised Database Count
Suppose a hypothetical snapshot from database version 1 contains 180 participants with recorded assessments and 24 endpoint events. After review, version 2 keeps the same 180 assessments but corrects two records from event to non-event and adds one late-confirmed event. The version 2 event count is \(24-2+1=23\).
The version 1 recorded-assessment proportion is \(24/180=0.1333\), or 13.33%. The version 2 proportion is \(23/180=0.1278\), or 12.78%. The numerical difference is \(12.78\%-13.33\%=-0.55\) percentage points (using unrounded proportions, \(-0.5556\) percentage points).
This difference is explained by documented changes to outcome records while the assessment denominator remains fixed. It is not evidence of a change in participant risk over time, nor does it represent a treatment contrast. A reviewer can interpret the version 2 count only if the correction and late confirmation are within the prespecified endpoint and cut-date rules; the late record must not be included solely because it is convenient.
Snapshot Checks Before Interim Review
A practical snapshot audit asks whether the data being reviewed answer the question the review is meant to address. It should reconcile the participant flow and the endpoint records, and it should make any limitations visible rather than burying them in an overall denominator.
State the data cut date and time zone, the extract or freeze time, and whether records received after the cut are excluded even if they refer to earlier events.
Define the participant set, endpoint, assessment window, and any requirements for maturity. Do not assume one denominator applies to every endpoint.
For the defined participant set, account for those not yet mature, mature with an ascertained outcome, and mature with an unresolved or missing outcome. Investigate unexplained discrepancies.
Record the version and relevant processing status, including important unresolved queries or adjudications. If versions differ, explain which records or rules changed.
Label counts and proportions by their actual denominator and follow-up status. Distinguish observed information from unobserved outcomes and avoid implying a treatment effect when the reviewed information does not support one.
Common Pitfalls and Alternatives
A common mistake is to treat “not recorded by the cut date” as “no endpoint event.” That reasoning silently converts incomplete follow-up or unresolved assessment into a negative outcome. The correction is to report maturity and ascertainment separately, and to describe a proportion using the denominator of participants whose outcomes are actually recorded. Even that observed proportion is descriptive of the recorded subset; whether it estimates a wider population quantity depends on assumptions and an appropriate analysis.
A second mistake is to compare summaries across cuts or database versions as if only the calendar date changed. A sound comparison first checks that the endpoint definition, eligible participant set, follow-up window, and processing rules are the same. If they are not, the summaries have different meanings. Similarly, a current snapshot should not be described as a final, complete outcome analysis unless the relevant follow-up and data processing are complete under the trial’s rules.
The main alternative to a simple snapshot proportion is not a more optimistic interpretation of the same incomplete records. Depending on the endpoint and review question, a prespecified analysis can account for differing follow-up or missing outcomes, or the review can wait for a later, more mature snapshot. Those approaches require explicit assumptions and methods. Snapshot documentation makes their inputs and limitations visible, but does not itself solve incomplete follow-up or define an inferential procedure.
Snapshot quality can be flagged by impossible date sequences, participants counted as mature before the applicable endpoint window, event records beyond the cut date, unexplained changes in denominators across versions, or counts that fail to reconcile to the participant flow. Such checks do not prove the data are error-free; they identify inconsistencies that should be resolved or disclosed before interpretation.
Check Your Understanding
Answer each question using the stated snapshot rules. Solutions follow the questions.
- A trial has 90 participants randomized. At the cut date, 54 have reached the endpoint assessment time, and 49 of those have recorded assessments. State the numbers mature with recorded outcomes, mature without recorded outcomes, and not yet mature.
- Among 49 recorded assessments, 7 meet the endpoint definition. Calculate the recorded-assessment proportion and state what it describes.
- An extract is generated on 12 August but includes records only through a 31 July cut date. Which date defines the information boundary, and what does the other date describe?
- Version 1 has 30 events among 200 recorded assessments. In version 2, one event is corrected to a non-event and two late-confirmed events are added, with the denominator unchanged. Calculate the version 2 proportion and explain whether the change demonstrates a change in risk.
Solutions. (1) Mature with recorded outcomes: 49. Mature without recorded outcomes: \(54-49=5\). Not yet mature: \(90-54=36\). (2) \(7/49=0.1429\), or 14.29%; this is the endpoint proportion among the 49 recorded assessments, not among all randomized participants. (3) The 31 July cut date defines the information boundary; 12 August is the extract-generation date. (4) Version 2 has \(30-1+2=31\) events, so \(31/200=0.1550\), or 15.50%. The change from \(30/200=15.00\%\) reflects revised records in the same denominator, not evidence by itself of a change in risk.