Dataset–transformation–display schema
Identify the data object, not merely the portal
Do not cite only the landing page that contained a download button. Find title, creator or publisher, release date, version, repository and persistent identifier. For dynamic portals, record query date, selected filters and export format. Preserve the file actually used with a checksum or stable local identifier.
Read the documentation and codebook. A variable name does not define unit or target population. Determine whether values are weighted, seasonally adjusted, revised or suppressed. Such details may belong in methods or the figure caption.
Make calculations reproducible
The analytical claim is shaped before drawing begins. Filters, exclusions, joins, recoding and aggregation all matter. Keep a transformation log in actual order, naming input version, command or formula, output and reason. Never overwrite raw data. Separate source file, cleaned data and the exact plotting table.
- Check import and inferred data types.
- Handle missing and invalid values under a documented rule.
- Record case counts before and after filtering.
- Test calculations with small transparent examples.
- Confirm units and denominators before percentages.
- Export the precise chart data as a separate artefact.
When combining sources, compare definition, time period and population. A common axis can make substantively different measurements appear equivalent.
Explain consequential visual choices
Chart type and scale influence perception. Bars normally require a meaningful baseline; truncated axes can exaggerate differences. Colour must not imply an order absent from the data. Name sorting, smoothing, cumulative values and logarithmic scales. Display uncertainty where supplied and relevant.
A useful caption reads: “Figure 6. Annual median of measure Y, 2019–2025. Author's visualisation based on dataset X, version 2.1, variables a and b, DOI …; invalid values excluded and monthly values aggregated by year. Reporting for 2020 is incomplete.” This identifies source, author work and limitation without repeating methods.
- Label axes with unit and denominator.
- Do not rely on colour alone.
- Write a self-contained legend.
- Never remove outliers without a declared rule.
- Mark interpolation and smoothing.
- Align caption and prose claim.
Separate data licence from citation
An open licence may permit reuse while requiring attribution or particular downstream terms. Data without a visible licence are not automatically free. Record licence name, version, source and modifications. Small or sensitive groups may also create disclosure risk even where access was legitimate.
The media citations hub covers figures and tables. How to cite research data supplies the dataset reference. The citing sources overview explains short and full citations.
Final audit
Can another researcher find the exact file? Can they repeat selection and transformation? Do plotted values match the exported table? Are axes, units, denominators and uncertainty visible? Does the caption identify source and contribution? Is the licence decision recorded? Do prose and chart stay within the data's scope? Archive dataset state, code or formulas, chart and rights record together.
Test an alternative display. If a small axis decision reverses the impression, phrase the interpretation cautiously and explain the choice. The purpose is not maximum visual drama but a faithful view of structure and uncertainty. Retain rejected display drafts where they clarify why the final design was chosen.
Worked case: a changing denominator
A portal reports annual shares but changes the denominator. A rising percentage need not mean more cases. Export share and base, inspect yearly definitions and plot counts separately where useful. A smooth trend misleads when measurement or category rules changed.
Check minimum, maximum, missing years and totals. Compare selected observations with the portal and preserve query parameters. Limit prose to measure, period and population: “measured share rose” is narrower than “behaviour increased”. Mention reporting and sample changes. Treat later portal revisions as a new data and chart version rather than overwriting the basis of the published figure.
Official foundations
Run a blind interpretation check: give chart, caption and prose to someone who has not seen the transformation log. Can they identify unit, population, period, source and main limitation? Misreadings reveal missing labels. Then write alternative text that communicates the relevant pattern and uncertainty rather than listing every point.
Freeze the submission version of an automatically updated chart. A live embed may later show data different from those analysed. Export image and plotting table together, assign a version and connect them in the record. Visually distinguish estimates, interpolations and observations through line style, legend and wording. A forecast must never merge silently with measured values.