How to read a statistic before you quote it

Most wrong numbers in circulation are not fabricated. They are real figures read without their footnotes: a provisional estimate quoted as final, a monthly change compared against an annual one, an index rebased halfway through the period. The statistics office published everything needed to get it right, in a methodology note nobody opened.

These are the checks worth running before a figure leaves your draft. None takes more than a minute.

1. Is it provisional?

Most economic series are published three times: a flash estimate, a provisional figure, and a final one, and the gap between the first and the last is routinely larger than the change everyone is reporting. Look for provisional, advance, flash, estimated or a revision policy in the metadata. If the series is revised, say which vintage you used and when you downloaded it — that single clause is what makes your number reproducible when the source has moved on.

2. Seasonally adjusted, or not?

Unemployment rises every January and tourism peaks every August in a way that says nothing about the trend. A seasonally adjusted series has that pattern removed so consecutive months can be compared; a raw series cannot be, and the only honest comparison for it is against the same month a year earlier.

Comparing an adjusted figure with a raw one — which happens constantly when two portals are mixed — produces a difference that is pure artefact. Adjusted series also get revised as the adjustment model re-estimates, which is check 1 again.

3. What is the index base?

An index of 112.4 means nothing without the year it is 100. Offices rebase every few years, and a rebased series is not comparable with the old one by subtraction: you need the linked series the office publishes for exactly this purpose, or the conversion factor in its note.

The same trap in a smaller form: a chart that starts at a suspiciously good year. Check what happens if you move the start by twelve months.

4. Nominal or real?

Any series in currency over time is meaningless in nominal terms once inflation is a few percent: “spending grew 5%” during 4% inflation is a 1% real increase, which is a different sentence. Look for constant prices, real, deflated, chain-linked — or deflate it yourself with the matching price index, not the headline one if the office publishes a specific deflator.

5. What is the denominator?

Counts almost always need to become rates before they compare. Per capita, per 100,000 inhabitants, per registered vehicle, per case — and the denominator must come from the same period and the same population as the numerator. Two errors recur: a numerator counting events and a denominator counting people, and a regional comparison where population is a year older than the count.

Watch for the change of unit that hides in a definition: “households” is not “people”, “cases” is not “patients”, “unemployed” is not “jobseekers”.

6. Is it a survey or a count?

A census or an administrative register counts; a survey estimates. An estimate has a margin of error and a confidence interval, and a change smaller than that margin is not a change at all — the correct sentence is “no statistically significant difference”. Breakdowns are where this bites hardest: a national survey is precise nationally and can be near-useless for one province, and the office usually flags exactly that with a footnote or a cell marked as unreliable.

Never silently drop those flags when you reshape a table.

7. Is there a break in the series?

Definitions change. A new classification (NACE, CNAE, ISCO, ICD), a new threshold, a new survey design, a change in what counts as a dwelling or a resident — each of these leaves a discontinuity, and the file rarely marks it in a column you would notice. The clue is usually in the methodology note or a footnote reading “not comparable with previous years”.

A two-decade chart that crosses three definitions is a nice-looking artefact.

8. What are you comparing against?

Base effects: a 20% rise measured against a collapsed month is a recovery, not a boom. Growth measured over a single quarter and annualised swings wildly. Cross-country comparisons of the same-named indicator often compare different definitions — even inside the EU, where Eurostat harmonises precisely because national figures are not comparable as published.

Where the answers live

All eight checks are answered in the same three places: the methodology or metadatos note attached to the dataset, the footnotes in the table itself, and the variable dictionary. In a portal API these sit in the dataset’s description and its documentation links — our guide on how to query a portal’s API covers how to fetch them alongside the data. If a dataset has none of the three, that is itself a finding worth reporting.

And if the figure you need is not published at all, ask for it: how to file a FOIA request. Before you republish what you downloaded, check what licence it carries.

The Spanish version of this guide is at Cómo leer una cifra antes de citarla.