Every poll is skewed.
Most of them hide it.
DataRadio is built on the assumption that its own numbers are wrong in knowable ways — and that saying so is more useful than pretending otherwise.
The question that started it
When someone says "the world thinks X" — who exactly was asked?
Almost always: a few thousand people, in a handful of wealthy countries, reachable by phone or panel, willing to spend twenty minutes on someone else's questionnaire. That is a real finding about a real group of people. It is not the world. The gap between those two things is where most confident claims about global opinion actually live.
The honest response is not to stop measuring. It is to measure, then publish the shape of the error next to the result — every time, without being asked, including when it is embarrassing.
Three things that are true about all polling
Who answers is never who exists
People who take online political surveys skew young, connected, male, educated and unusually interested in politics. No amount of sample size fixes that — a million self-selected answers is a million answers from the same skewed pool.
What we do: rake every nation's respondents to that country's real population on age, gender and education, using UN and UNESCO figures. Where a country has no benchmark, results are published labelled unweighted rather than quietly presented as representative.
Sample sizes are smaller than they look
Weighting fixes composition and costs precision. Correcting a badly skewed sample means a few respondents carry a lot of the estimate, and the effective sample size falls — sometimes far below the raw count. A poll reporting "n = 2,000" may be carrying the statistical weight of 600.
What we do: publish the Kish effective sample size and derive every margin of error from it, not from the raw count. We also cap how much any one respondent can count, which means extreme skew stays visibly uncorrected instead of being papered over.
The question shapes the answer
Ask people to agree or disagree with a one-sided statement and they agree more than they should. Write one side beautifully and the other as a strawman and you have authored the result. Pitch it at university reading level and you have measured education while calling it politics.
What we do: every item states both positions as a real person would hold them, in randomised order. Reading level is capped and the test suite fails the build if any question breaches it — the English bank averages grade 4.5.
Translation is not a formality
A translated attitude question is a different question. If "should" becomes "must" in one language, the gap between two countries is part opinion and part translation — and nothing downstream can separate them again. That gap is exactly what a global poll is selling.
What we do: poll items follow translate → review → adjudicate, and only adjudicated wording counts toward cross-national comparison. Every answer stores which wording the respondent actually saw.
What we refuse to do
- Publish a national figure before it means anything. A country needs 100 respondents before it is broken out. Below that, a "national result" is a handful of people with a flag next to them.
- Fill gaps with models. Where there is no data, the site says there is no data. It does not impute a plausible number and present it in the same typeface as a measured one.
- Fingerprint anyone. It would improve our detection of ballot-stuffing. It is also the capability that makes a database of political opinion dangerous. We decline it.
- Let money buy the answer. Members' responses are weighted exactly like everyone else's, and no published figure is computed differently for people who pay. If that changed, the site would be worthless — including to the people paying.
The part most sites leave out
Here is what DataRadio cannot tell you, stated plainly because you will not find it in the footnotes elsewhere:
- Nothing about countries where the internet is restricted, expensive or watched. Silence there is not moderation, and we do not infer it.
- Nothing causal. This is a cross-section of opinion, not an explanation of it.
- Nothing reliable about people who will not answer political questions online — a group that is certainly not politically random.
- Weighting corrects observed composition on three variables. It cannot correct for the politically disengaged being scarce on a site about politics.
Seed data — read this first
5,211 of the 5,213 respondents in this corpus (99.9%) were generated, not collected. They were written into the database so the site could launch populated rather than empty, and — this is the part that matters — they are counted in every aggregate exactly like real respondents: the nation atlas, the world reading, the weighting, the significance tests, the blocs and the phenotype shares.
So while this notice is here, a national figure on this site is not a measurement of that nation. Nothing on this site should be cited as evidence about a population until the seeded share is small and the figure's own n is large enough to stand without it.
How they were built: each seeded respondent draws demographics from that country's real population marginals (UN World Population Prospects age and sex structure, UNESCO educational attainment), then takes a position from a hand-written country archetype plus individual variation and a demographic tilt. Lived-condition answers come from the country's published World Bank indicators. That makes the seed internally consistent and demographically plausible — it does not make it true. A country prior is an assumption about a nation, and the entire argument of this site is that you have to ask people instead of assuming.
Seeded rows are flagged in the database, and every real respondent retires one of them from their own country — so a nation's sample size holds steady while its composition becomes real, and this section disappears by itself once the last seeded row is gone. 2 real respondents so far.
How it stays checkable
Claims about rigour are worth nothing unless someone can check them, so:
- Every calibration run records the live state of the model — each item's discrimination, which items are currently failing — and the run that produced any figure can be recovered.
- The languages page publishes translation grade per language, including when the honest answer is "none yet".
- Every nation page states whether its figures are weighted or unweighted, with the benchmark source and year.
- Every calibration run is recorded, so any position the site has ever shown can be explained by the state of the model on the day it appeared.
Join the world.
Six minutes, anonymous, free. The more people answer, the less skewed the picture gets — and we will keep telling you exactly how skewed it still is.
Take the poll How it works