An SDTM dataset can pass structural checks and still misrepresent a study observation. The difficult work is often deciding what a record means before deciding where it belongs. A date, a category, and a value are insufficient if the collection context has been lost.
CDISC SDTM supplies the tabulation model. This article proposes a review approach for applying that model thoughtfully; it is not a substitute for the applicable implementation guide, therapeutic-area guidance, or submission requirements.
1. Start with collection intent
For every unusual field, retain the case report form question, permissible responses, units, timing, and related instructions. A measurement taken before dosing and one taken after an intervention may share a label while answering different clinical questions. Mapping by label alone can erase that distinction.
A hypothetical infusion form might capture a planned dose, prepared amount, and administered amount. These should not be treated as interchangeable simply because each is numeric. Write a short semantic decision describing which observation each field represents, then establish the mapping with the appropriate standards expertise.
2. Resolve relationships before flattening data
Repeated observations, parent-child assessments, and supplemental information need deliberate handling. Reviewers should be able to understand which records are related and why. A convenient join key in the operational database does not automatically establish the correct clinical relationship in tabulation data.
Ask a reviewer unfamiliar with the source system to inspect a small set of participant records. If they cannot reconstruct the collection sequence or distinguish an observation from its qualifier, the mapping specification needs additional explanation. This record-level exercise complements conformance checking.
3. Separate vocabulary from meaning
A controlled term can be valid while its assignment is wrong. CDISC controlled terminology supports consistent representation; the semantic decision still requires study context. Maintain the terminology version used by the study, document extensions where applicable, and distinguish a coding correction from a change in mapping logic.
For a recurring issue, keep both the rejected mapping and the rationale for the accepted one. That decision history is especially valuable when new programmers join or when a later transfer resembles an earlier exception without being identical.
4. Review the package, not only the datasets
Dataset contents, metadata, annotated collection materials, and reviewer explanations should tell the same story. A discrepancy between them is a review problem even when the dataset itself is syntactically valid. FDA study data standards resources are the appropriate starting point for identifying submission expectations and supported standards.
We recommend a semantic review log organized by clinical concept rather than file. Give each decision an owner, supporting source, affected domains, and resolution status. This makes the consequence of a mapping change visible before it propagates into analysis.
Reliable tabulation does more than standardize columns. It preserves enough meaning for another person to interpret an observation without relying on the memory of the person who collected or mapped it.
