Source record
Every data-driven page names the organization, dataset or report, measurement period, publication date, retrieval date, and a direct source link. When data are downloadable, we retain the release or version information needed to reproduce the result.
Transformations
Derived measures identify the formula, exclusions, normalization, assumptions, rounding, and treatment of missing data. A per-capita ranking must name the population estimate and year used as its denominator.
Units and universes
Firm, establishment, employer, nonemployer, application, and projected formation are different statistical units. We do not combine values from different programs into a new total unless the universes and methods are compatible.
Time comparisons
We do not mix seasonally adjusted and unadjusted values in the same comparison. Month-over-month and year-over-year changes are labeled separately, and historical comparisons use one revision vintage. One month can be newsworthy without being a durable trend.
Revisions
Government time series are revised. When a revision materially changes a headline number or conclusion, we update the page, change the updated date, and explain the correction or revision.
Administrative and survey data
Administrative counts can still contain classification and processing error. Surveys should disclose sample size, target population, field dates, recruitment, weighting, and whether the sample was probabilistic. A margin of error is not invented for a convenience sample.
Experiments
Original studies publish the sampling frame, exact questions or prompts, collection dates, tools or models, analysis rules, exclusions, missing-data treatment, and limitations. Observed association is not described as causation.