Building the DMC Data Package

Purpose and Scope of the Package

A DMC data package is the organized set of materials prepared to support a committee’s review at a defined interim point. Its job is to make the information relevant to that review complete, traceable, and consistent across displays. It is not a substitute for the DMC’s judgment, and assembling it does not authorize changing an endpoint, analysis population, or review question after seeing interim results.

The starting point is the validated interim snapshot: the fixed, documented data cut from which the package is produced. Package builders should preserve its identity and use the approved review plan and charter to determine which materials belong in the review. The package should make clear what data are included, what time period they cover, and which audience may receive each item. A polished layout cannot correct an unclear data cut, an inconsistent denominator, or an unintended disclosure.

The package’s exact contents depend on the trial and the DMC’s remit. The inclusion of a display does not, by itself, imply that a finding is clinically important or that a recommendation should follow. Those are matters of review, not document assembly. The focus here is how to build and quality-check the collection so its contents can be reviewed as intended.

Build From a Defined Package Plan

Before producing tables or figures, create a package plan that maps each review question to a display, its source, its population, its time window, and its intended audience. This step prevents a common failure: producing a large collection of available outputs without confirming that it answers the questions the DMC was asked to review.

A package plan is useful even when the final set is short. It distinguishes required materials from optional context, and allows reviewers to see whether an expected display is absent rather than assuming it was overlooked. If a planned display cannot be produced, state that fact and its reason; do not leave a silent gap.

Core Displays and Their Reconciliation

Trial-status material gives the committee a compact operational context: for example, the enrollment target, participants enrolled to the cut, sites open or closed, and relevant follow-up status. Each count needs a definition and a reference date. “Enrolled” should not be used interchangeably with “randomized,” “treated,” or “completed.” Where the trial uses multiple arms or stages, the display should make the grouping clear while respecting the audience rules.

Participant disposition accounts for what happened to participants from a stated starting population through the relevant milestones. The categories should be mutually exclusive when presented as a partition, and counts should reconcile. For a simple flow from randomized participants to current status, the basic check is:

This identity applies only when the three status categories are exhaustive, mutually exclusive, and defined at the same cut. If participants can have more than one reason for discontinuation, the number of reasons is not expected to sum to the number of discontinued participants; label those reason counts as potentially overlapping. Similarly, a disposition table based on treated participants must not be presented as if its denominator were all randomized participants.

Protocol-deviation material should identify the classification rules and reporting window used. Useful fields can include deviation category, date or period, affected participant count, and whether the deviation was considered important under the trial’s prespecified definitions. Counts of deviations and counts of participants with at least one deviation are different quantities. If one participant has two deviations, a deviation-event count can exceed the participant count; the table must name which quantity it reports.

Safety and Efficacy Materials

Safety summaries should use the definitions, time windows, groupings, and denominators specified for the review. At a package-assembly level, confirm that labels distinguish participants from events, that the observation period is explicit, and that the source population is clear. A participant may experience multiple events, so a total event count is not generally a count of unique participants. The detailed interpretation of safety listings and summaries belongs to the committee’s review, not to the act of packaging.

Efficacy displays should likewise follow prespecified endpoints, analysis populations, visit or follow-up definitions, and presentation conventions. The package plan should identify each display and its provenance; the package builder should not select a more favorable endpoint, add an unplanned subgroup, or change a time point because a different view appears more informative after looking at the data. If a planned output is unavailable or a prespecified analysis cannot be completed, disclose the limitation clearly.

Every display benefits from a concise caption or footnote stating the population, data cut or observation window, units, and any non-obvious denominator or derivation. Include a contents list and a package-level note that identifies the snapshot and version. For any audience-specific set, check not only the file name but also the contents, metadata, and rendered output; a correct label does not prevent an incorrect version from being included.

Quality Checks, Alternatives, and Limitations

A useful package-level quality check is to test both internal arithmetic and consistency between materials. For example, the randomized total in the status overview should agree with the disposition denominator when both refer to the same population and cut. Differences may be legitimate if definitions differ, but then those definitions should be explicit. A disagreement without an explanation is a quality issue, not a matter for readers to guess through.

The reconciliation identity above follows directly from the partition of the population. If every randomized participant belongs to exactly one of completed, discontinued, or ongoing status at the cut, the sets are disjoint and their union is the randomized population. Therefore, the size of the whole equals the sum of the sizes of its parts. If the sum fails, at least one count, category definition, or population alignment is wrong or incomplete. This is why reconciliation is a substantive check, not merely a formatting preference.

Common diagnostics include:

• Verify that each count has a named population, denominator, date, and unit where relevant.
• Recalculate sums and percentages from the displayed counts, allowing only stated rounding differences.
• Compare shared totals across status, disposition, deviation, and other sections when definitions match.
• Check that event counts are not mislabeled as participant counts and that overlapping categories are identified.
• Confirm that captions and footnotes match the prespecified analysis and the actual output.
• Review the rendered package for omitted pages, stale dates, inconsistent version identifiers, and unintended audience content.

A package assembled manually from separate analyst files may be quick for a single review, but it increases the risk of inconsistent labels, stale outputs, and undocumented edits. A controlled programmatic workflow improves repeatability and makes it easier to regenerate the same package from the same snapshot. It still requires human review: automation cannot determine whether a display answers the intended question, whether a caveat is understandable, or whether the right audience receives the right set.

A further alternative is to provide every available output “for completeness.” This can obscure the planned review, increase the chance of inconsistent versions, and expose material beyond the intended scope. A focused package is preferable, with clearly identified supplemental material when justified and authorized. The limitation of any package remains important: it organizes evidence but does not guarantee data completeness, remove uncertainty, or replace DMC deliberation.

Common Mistake