.jpg){kind=link}
- Published
- Updated
Rerunning the same code and reproducing a table does not by itself establish correct analysis. Independent checking should examine scientific definitions, derivations and calculations, rather than treating repetition of a shared incorrect assumption as successful validation of the result.
Define the question and assumptions
Define the checking scope first: primary endpoints, populations, models and critical derivations carry different risks. Let a second programmer work from the approved specification rather than merely copy the first implementation. Prespecify numerical tolerances and discrepancy resolution. Matching programs cannot establish correctness if both implement an inappropriate interpretation of the scientific question.
Hypothetical study example
In a hypothetical study, two programs select baseline from the same incorrect field and produce identical tables. A reviewer traces an example participant to source information and finds the defect. The team corrects the definition and all affected derivations instead of accepting numerical agreement as sufficient evidence for approval.
Design review checks
- Provide independent interpretation of critical definitions and traceability to underlying records.
- Record logical causes of discrepancies as well as their numerical differences.
- Link checked code and data versions to the final analysis package.
Record the decision and its limits
Accepted evidence shows scope, method and resolved issues. Controls may be proportionate to risk rather than identical for every formatting change. Do not hide unresolved discrepancies. Final approval covers both computational agreement and consistency with the authorised scientific rules, while any claim of independence accurately describes how checking was actually performed.
Frequently asked questions
- Do two matching outputs prove correctness?
- No. A shared source or definition error can appear in both.
- Does independence always require different software?
- Not necessarily. Independent reasoning and appropriate verification coverage are central.