How accurate is it?
Each file contains synthetic respondents. Their age, sex, education, state and employment come from real, anonymous US Census records (American Community Survey 2020–2024 public-use microdata, weighted to the population you choose). Their answers to your questions are generated: an AI model estimates how that population answers each question, and a statistical model turns those estimates into individual people whose answers hang together.
How we measure it
We state accuracy as what a file is worth in real respondents: the size of a real random sample that would be just as hard to tell apart from the true population. It is a number you can compare with a budget, and it does not hide behind a percentage. Most tests use national surveys of US adults held out from our own models' training. We write down our predictions before each test and record the misses.
| Test battery | Questions | A file from your questions is worth about |
|---|---|---|
| Social attitudes (CES 2022) | 14, 5-point agree | 50 or more real respondents |
| Civic engagement (CES 2024, items new that year) | 20, 5-point agree | 42 |
| Personality (IPIP Big Five)* | 50, 5-point | 34 |
| Finances and community (SHED 2018) | 32, mixed formats | under 25 overall; yes/no and select-all items within about 3 points |
* The personality comparison file is internet volunteers reweighted to US adults, not a probability sample. Figures are averages over three draws.
Where it is weaker
- Some topics. Accuracy depends mostly on what you ask, not on the answer scale. Opinions, personality and factual behaviors do best; mental health, social support, past experiences and how much people value specific things miss more. What works best.
- Politically charged items. Average answers are close, but fewer synthetic people sit on the fence than real people do.
- Links between different topics. Correlations between questions on unrelated topics come out about twice as strong as in real data.
- Narrow groups. Our tests are on US adults in general. Expect less for small or unusual groups, and say who they are as precisely as you can.
How to use it well
Use a file to test and improve a questionnaire, to plan sample sizes and write analysis code, and to teach. Do not run significance tests on a synthetic file as if its rows were real people, and do not present it as real data. For decisions that matter, run your survey with real people.