Writing a long description requires understanding the data behind the graphic, not a post-hoc caption

What it is

A caption added just before launch is often written by someone who can no longer say whether a bar is a mean or a cumulative, whether a missing year is absent data or a zero. A long description has to be written by someone who understands the dataset, at the same time as the figure is made — not as the last appearance sentence on a ship list. Post-hoc writing produces looking-at-the-picture prose; concurrent writing produces a note that is accountable for the numbers.

“Understands the data” is not a demand that the author know accessibility jargon. It is a demand that they can answer what the encoding stands for, who is being compared, and which points cannot be read off the drawing.

Why it happens

A chart is a compression. What the compression drops — error bars, missing years, weighting, tax-inclusive or not — does not grow back inside the pixels. Only the person who did the compression still holds the dropped layer. A later writer sees a finished appearance and can describe slope and colour, but cannot decide whether the 2021 gap is missingness or zero, and cannot responsibly write “the difference is not significant” — that sentence comes from a test, not from ink.

Time works against the author too. When the data refresh, the figure redraws; a caption stuck beside it becomes a lie. Treating the description as a pre-launch copy patch means nobody owns “when the figure changes, the note changes.” Written into the charting flow — analysis notebook, chart config, a required field on the content model — the description shares a source with the data.

Studying it

Compare two authorship conditions: the analyst or content owner writes the note while producing the figure, versus someone with no data background captions it after visual lock. Score the same data questions, and have the original author flag factual errors in the post-hoc note (missingness treated as zero, cumulative treated as mean, a broken axis omitted).

Independent variables: whether the writer saw the source table, when the note was written (at charting / just before launch), whether anyone edited the note after a data refresh. Dependent variables: factual error count, trend-question accuracy, how often the note and figure diverge after a refresh.

A process audit is closer to production than a lab: sample shipped figures and ask who wrote the note, which table it rests on, and whether the last data change edited it. Silence is ownerless post-hoc text.

Where it stops holding

Stock icons and decorative infographics have no “data behind them”; they do not need an analyst author, and this claim does not apply. When a figure is republished and microdata cannot be obtained, the writer can only work from encodings already visible on the figure, and should mark “based on visible values, source table not checked” — an evidence bound, not a licence to invent a trend. Templated auto-descriptions of a point table can work for simple structures, but a generator cannot see the author’s reading (“this gap is the pandemic break”) and still needs a data-aware pass. Confidential data cannot be handed to outsourced copy for the sake of a description.

Applying it

• Make the long description a required product of the charting flow, first drafted by the person who analysed or supplied the numbers. Do not wait for visual lock and hand it to someone who has not seen the table.
• State encoding meaning and known gaps in the note (missing vs zero, broken axis, comparison grain). Those sentences can only come from data knowledge.
• On data refresh, regress the note with the table: a redrawn figure with an untouched note is a content defect.
• How to check: for a sample of figures, ask three questions — which table, who wrote it, did the last data change edit it. Any unanswered question, or a spot check that wrote missingness as zero, is a post-hoc failure.

Related

• Same group: J2.12.1 Long descriptions of charts should capture data trends, not only visual appearance · J2.12.2 A long description can sit in nearby visible text or on a linked page · J2.12.3 Interactive charts need an equivalent data table as an alternative presentation
• Nearby: J2.04 Alternative Text · J1.09 Cost and Timing of Accessibility · J5.08 Limits of Automated Checking
• Search terms: authoring long descriptions · post-hoc caption · data-informed description