The purpose
Make every consequential number reproducible and every interpretation proportionate to the data. Start from the reporting question, original dataset or study, its documentation, and the intended claim. A chart is part of the evidence argument.
How to apply this skill
- Obtain the original table, release, or study where accessible. Record source, retrieval time, release vintage, observation period, geography, population, units, denominator, and definitions. Distinguish preliminary estimates, revisions, forecasts, and observed outcomes.
- Inspect coverage and quality: missingness, exclusions, sampling, measurement changes, duplicates, selection effects, and uncertainty. For studies, read methods and limitations rather than relying on the abstract or press release. Identify whether the result is peer reviewed, preprinted, corrected, or retracted when relevant.
- Preserve a reproducible transformation from recorded inputs to each published number. State filters, joins, exclusions, rounding, and formulas. Compare totals and sample rows with the original. For small arithmetic, show the calculation; for repeated transformations, save executable code and inputs or authorized retrieval instructions.
- Check interpretation: percentage versus percentage points, count versus rate, nominal versus adjusted values, mean versus median, and relative versus absolute change. Correlation, temporal sequence, or a before/after chart alone does not establish causation.
- Test whether reasonable alternative denominators, windows, definitions, or missing-data treatment change the conclusion. Report sensitivity when material. Compare uncertainty intervals and practical magnitude; a single significance threshold cannot carry the story.
- Label charts with units, dates, source, and necessary caveats. Use encodings and axis choices that preserve meaningful comparisons. Include an accessible text takeaway and the underlying values or a useful table. A graphic should answer a question the prose raises.
What this produces
Return an analysisRecord containing input provenance, definitions, transformation steps/code, result table, validation checks, uncertainty, sensitivity findings, claim-ready wording, chart specification, and limitations. Tie each derived number to a claim ID and a reproducible calculation.
When publication must pause
Hold numerical claims when results cannot be reproduced or definitions are incompatible. Repair the calculation or narrow the comparison. Keep consequential uncertainty in the article, not only a methodology appendix. If original data is inaccessible, clearly distinguish a quoted published figure from an independently reproduced finding. Do not invent precision, missing rows, or causal certainty.
Sources behind this skill
Quality, transparency, and uncertainty principles are informed by the Office for Statistics Regulation Code of Practice. These workflow artifacts are Follow.fyi's design.