Why Data Quality Should Be Reviewed by Someone Else

The analyst who cleaned a dataset knows its history better than anyone. That expertise is essential—and it is also why the analyst should not be the only person who verifies data quality. Repeated exposure makes assumptions feel obvious, familiar workarounds become invisible, and checks designed by the same person who wrote the rules can reproduce the same misunderstanding.

Independent review adds a different line of sight. Its purpose is not to imply distrust or to repeat every task. It is to test whether the data and transformations support the study’s critical conclusions.

01Define independence proportionately

Independence is not a single standard. For a low-risk student project, it may mean a second analyst who did not write the cleaning code. For a multicenter trial, it may require separate data-management, statistical-programming, monitoring, adjudication, or quality functions with documented responsibilities.

The reviewer should be free to raise issues without having produced the result being defended. Conflicts and prior involvement should be visible. If complete organizational separation is impossible, use procedural safeguards: independent checklists, blinded review where appropriate, direct access to source specifications, and sign-off by someone accountable for quality.

02Review the highest-consequence data path

Do not begin with random checking across thousands of fields. Trace the variables most capable of changing the conclusion: eligibility, treatment assignment or exposure, primary and safety outcomes, follow-up time, major covariates, exclusions, and missingness indicators.

For each, compare the protocol, data dictionary, collection form or source, cleaning rules, derivation code, analysis dataset, and reported output. This end-to-end trace can reveal that a correct formula was applied to the wrong source field, a unit changed across sites, or an outcome definition drifted between documents.

ICH E8(R1) supports a proportionate focus on factors critical to quality. Independent review should follow the same logic: effort should concentrate where error would matter most.

03Use complementary verification methods

Useful methods include:

• independent code review for derivations and exclusions;
• duplicate programming of selected critical outputs;
• source-to-dataset trace and dataset-to-source verification where authorized;
• comparison of counts across raw, cleaned, and analysis datasets;
• checks for impossible sequences, duplicate identifiers, unit inconsistencies, and missingness patterns;
• recreation of table cells from a frozen dataset;
• review of manual overrides and cleaning-log entries; and
• reconciliation of protocol, registry, analysis plan, and manuscript definitions.

A reviewer should not merely run the original script and confirm that it completes. Re-execution tests reproducibility; independent logic tests whether the intended rule was correct and fully implemented.

04Preserve disagreement and resolution

Quality review needs a controlled issue log. Record the finding, evidence, consequence, owner, response, resolution, reviewer acceptance, and affected versions. Distinguish a cosmetic label from an error that changes participant inclusion or the estimated effect.

If a material issue remains unresolved, define who can accept the risk or require correction. The person whose work is under review should not unilaterally close the finding. Leadership should also protect reviewers from pressure to approve a deadline-driven result.

05Verify the correction

Fixing one value or function can affect many outputs. Link the approved change to a new dataset and code version, rerun impacted analyses, compare results, and document whether interpretation changed. MHRA data-integrity guidance emphasizes reconstructable processing and visible repeated modifications; this prevents the final “clean” output from erasing the route by which it was reached.

06Make review a release gate

Data-quality review is most useful before database lock and before manuscript submission, not after a journal asks a question. Define the minimum evidence required for release: completed critical-variable trace, resolved high-severity findings, reproducible dataset build, verified primary outputs, and named approvals.

Start with one independent exercise: give a second reviewer the primary-outcome specification, source extract, cleaning rules, and final table. Ask that reviewer to trace ten participants through the full pipeline and independently reproduce one key estimate. The discrepancies found will reveal where the institution needs stronger definitions, code, documentation, or decision rights.

Measure the review process itself. Track the severity and recurrence of findings, time to resolution, proportion discovered before lock, and whether corrections altered results. These indicators should improve the system, not rank or punish individuals. Recurring problems often identify weak forms, ambiguous definitions, or inadequate automation rather than careless staff.

Independence does not replace expert ownership. It gives that ownership a credible test.

This article is educational and intended for research purposes. It does not provide individual medical advice, diagnosis, or treatment. No patient data were used.