Start With a Clear Review Question

Between two clinical data cuts, the hard question is rarely whether the file changed. It is whether the change matters. A lab value can move because of a data correction, a visit can appear because of late entry, and a dataset can look different because the underlying version is different. Before opening compare mode, write down the review question: which subjects, which domain or form, and which kind of change are you checking? Keep it narrow enough to finish in one review session, but broad enough to cover the records the study team will discuss. In a workspace built for clinical study data analytics, each study has its own space and each data cut is kept as its own version with a checksum, so the first step is to confirm that you and your reviewers are looking at the same two cuts. The checksum is not a substitute for review, but it helps prevent the common mistake of comparing a current cut against a stale download. Name the cuts in your notes, state the review question, and decide whether you are reviewing subject-level changes, dataset-level changes, or both.

Filter to the Subjects That Matter

Subject filters turn a cut comparison from a file-level audit into a focused review. Use them before compare mode, not after. If the question is about subjects with a specific event, a particular visit, or a set of identifiers supplied by the clinical team, filter to that subject set first and apply the same filter in both cuts. This avoids a diff that is dominated by unrelated records and makes it easier to see whether the same subject is represented consistently. Subject filters are useful throughout the workspace: forms, visit schedule, datasets, and documents can all be browsed with the same subject lens. That consistency matters when a change in an SDTM or ADaM dataset needs to be checked against the CRF entry, the visit schedule, or the protocol and SAP that describe the expected handling. When the filter is applied first, compare mode can show what changed within the population you actually intended to review. If the subject set itself changes between cuts, treat that as a finding to investigate rather than a filter problem.

Keep a record of the subject set you used. If the review expands, add subjects deliberately and note why. A stable subject list makes it possible to rerun the same comparison later, especially when the next cut arrives. It also separates two questions that are easy to blur: did the data for these subjects change, and did the group of subjects included in the review change?

Compare in Layers, Not in One Pass

Compare mode is most useful when you read the diff in layers. Start with the structure: which datasets, forms, documents, or outputs are present in each cut, and which are new, retired, or superseded. Then move to the subject level and check whether the same subjects appear with the same identifiers and the same visit structure. After that, review the fields and rows that matter to your question. A side-by-side view is helpful because it keeps the before and after in context, but the reviewer still needs a method.

  • Confirm the two cuts and their checksums.
  • Filter to the review subjects.
  • Compare the dataset or form at the structural level.
  • Inspect the changed fields and rows.
  • Cross-check the change against the source record or document.
  • Record the conclusion and the action.

Do not stop at a row count difference. A changed row may be a correction, a late entry, or a new derivation, and those are different review outcomes. If a document is involved, use page-by-page search to find the protocol, SAP, or annotated CRF passage that explains the expected value. The goal is not to prove that every diff is harmless; it is to produce a defensible account of what changed and why.

Tie the Diff Back to Documents and Outputs

Not every difference can be explained from the dataset alone. A changed value in an SDTM or ADaM dataset may be the result of a correction in CRF data, a lab data update, a change in a derivation, or a new version of a document that affects expected handling. This is where the rest of the study workspace matters. Protocols, SAPs, and annotated CRFs can be searched page by page, next to the data they describe, so a reviewer can move from a changed row to the passage that defines the expected value. Define.xml sits in the same study workspace as the datasets, so a reviewer can check it alongside the changed record, while TFL outputs can show whether a downstream result also changed. If a TFL output differs between cuts, check the ADaM dataset and SAP behind it rather than treating the output as an isolated file. If a lab value differs, compare the dataset record with the source lab data and the CRF entry. The point is not to read every document for every row. It is to use documents and outputs selectively to explain the differences that affect the review question, and to leave a clear link between the evidence and the decision.

Attach Findings to the Exact Record

A comparison is only useful if the team can act on it. Attach notes and issues to the exact form, field, row, or document page they concern, rather than collecting comments in a separate spreadsheet. That keeps the review trail close to the data and makes it clear whether a difference is a data issue, a documentation gap, a programming query, or an expected change between cuts. Be explicit about which cut contains the issue and which cut you compared it with. If the change involves unblinded content, check that the reviewer has the appropriate grant; blinding is respected in the workspace, and unblinded files, records, and search results remain hidden without that grant. Also remember that records are retired or superseded rather than deleted, and changes to membership, roles, unblinding, and records are written to an append-only audit trail. That does not replace a reviewer’s judgment, but it supports a clean account of the review. Before closing the review, confirm that every material difference has either an explanation, an open issue, or a documented decision to accept it. The final output should be a short, traceable record of what changed between the two cuts, which subjects were affected, and what the team decided to do next.

Categories: Blog

Sofia Andersson

A Gothenburg marine-ecology graduate turned Edinburgh-based science communicator, Sofia thrives on translating dense research into bite-sized, emoji-friendly explainers. One week she’s live-tweeting COP climate talks; the next she’s reviewing VR fitness apps. She unwinds by composing synthwave tracks and rescuing houseplants on Facebook Marketplace.