Data provenance and reproducibility: trace a result back to its inputs

Define the claim before collecting files

Begin with the sentence the final statistic is intended to support. A mean for eligible calibrated measurements is not the same target as a mean for every imported row. The W3C provenance overview distinguishes the data objects, activities and responsible agents involved in a result. For a small project, this can be recorded in a plain table rather than a complex system. The essential point is to preserve relationships: which source was used, which operation transformed it and which output was produced.

Separate reproduction from new evidence

The National Academies report distinguishes computational reproducibility from replication with newly collected data. Re-running the same calculation on the same input checks one kind of consistency. Collecting a new sample to address the same scientific question checks another. A result can be reproducible and still depend on biased sampling, a wrong formula or an unsuitable model. State what the check establishes. Avoid using the word verified as though reproducing a spreadsheet total proved every substantive conclusion attached to that total.

Follow a hypothetical calculation chain

The imported values are 6, 8, 10, 12, 14, 22, 10 and 14. The last two rows duplicate earlier record identifiers, so their removal is based on identity, not merely equal values. The eligibility exclusion is separately recorded. Different rows can legitimately share a value; deleting all repeated numbers would be an unjustified transformation.

Preserve raw inputs and explicit transformations

Keep the original received file unchanged and create a derived file for cleaning. Record the operation, parameters and reason for each change. For the example, a duplicate rule and an eligibility rule are two different activities with different implications. The first removes repeated records; the second changes the population included in the summary. A reader needs both explanations to interpret the final mean. Manual edits should be logged with the same care as scripted ones, because an undocumented correction can be impossible to reconstruct even when the final formula is visible.

Record metadata that changes interpretation

The W3C data-publication recommendations address descriptive information and version history. Useful metadata includes variable definitions, units, observation dates, missing-value codes, geographic scope and the source’s own revision identifier. A column named value is insufficient if one file uses centimetres and another metres. A retrieval date is also different from the date the observations were collected. Preserve both when available. Metadata should explain the meaning of the data rather than merely repeat the file name in a more decorative format.

Keep source authority separate from data suitability

An official source can be authoritative for its own reporting scope while still being unsuitable for a different question. A current table may overwrite earlier values, or an administrative dataset may cover only registered participants. Record those boundaries before combining sources. A link proves where a claim can be checked; it does not eliminate the need to read definitions and limitations. If an original source is unavailable, label the accessible substitute and avoid suggesting that a secondary copy has been independently confirmed against the original record.

Design a useful reproduction package

A compact package can contain the input snapshot, transformation instructions, a data dictionary and a result table with calculation details. Include the software or environment information needed to reproduce operations that depend on it, without treating a long list of tools as a substitute for clear steps. For sensitive data, access restrictions may prevent public distribution; document the restriction and the permitted verification route. Reproducibility should not be pursued by publishing personal information that the research question does not require.

Provenance checklist

• State the target population and statistic.
• Preserve source identifiers and dates.
• Keep original inputs separate from derived files.
• Log duplicate, eligibility and missing-data rules.
• Record units and variable definitions.
• Retain calculation steps and environment details.
• Explain what reproduction does and does not establish.

Ask a second reader to trace one reported number backward through the chain. If they can identify its source rows, exclusions and formula, the documentation is doing useful work. Any gap they encounter is an actionable improvement: add the missing rule or version reference rather than relying on the original author’s memory.

Sources and verification

• Overview of the PROV family — World Wide Web Consortium — 2026-10-23
• Reproducibility and Replicability in Science: summary — National Academies of Sciences, Engineering, and Medicine — 2026-10-23
• Data on the Web Best Practices — World Wide Web Consortium — 2026-10-23