How stories are made
Methodology
How the data behind each visual story is sourced, versioned, calculated, and reviewed, and where AI assists.
This page describes the practices behind every visual story. Each story also has its own methodology section with the specific source, calculations, exclusions, and definitions it uses.
Sources
We use data from the organizations that produce it, usually government statistical agencies or established research institutions, and link to the exact source and release. Each story credits its sources as those organizations request and states the basis on which the data can be reused.
Four different dates
Data pages can be misleading when dates are blurred together, so we keep four apart:
- Observation period: the time the data describes, for example prices in 2024.
- Source release: when the publisher released those figures.
- Retrieval: when we downloaded them. Downloading old observations again does not make them new data.
- Page update: when we last made a meaningful change to the story.
Versioned data
Each download is validated (structure, units, duplicates, coverage, missing values, and unexpected changes) and saved as a version with a checksum. A story is pinned to one approved version. A new release is first recorded as a proposal with a summary of what changed, and only replaces the approved version after review. Earlier versions are kept, so a story can be restored to them.
Calculations
Charts, tables, downloads, and the numbers quoted in the text are all produced by the same code from the same data version. Values are calculated at full precision and rounded only for display. Missing values are shown as unavailable and are never treated as zero.
Interpretation
We try to separate what the data shows from what might explain it, and we label explanations as inferences when the data does not test them directly. We avoid presenting correlation as causation, ranking groups that are not comparable, or showing more precision than an estimate supports.
AI assistance
Claude, an AI model made by Anthropic, assists with research, data-processing and chart code, and drafting. Its output is reviewed, and no number is published because a model wrote it; it must be computed from the cited data.
Corrections and updates
Report errors through the contact page. We correct mistakes promptly and note substantive corrections and data updates on the story page.