Frequently asked questions
The basics
What is synthetic survey data?
A data file of respondents who do not exist, whose answers behave like a real survey's: realistic averages, realistic spread, and questions that hang together the way they do for real people. Each respondent is grounded in an anonymous US Census record, and the answers are generated.
Who is it for?
Researchers, analysts, market researchers, teachers and students who want to test a questionnaire, plan an analysis or practice with realistic data before fielding a survey.
What can I use it for?
Trying out a questionnaire before you pay for fieldwork; spotting questions that don't hang together or say the same thing twice; writing and testing analysis code; planning sample sizes; and teaching survey methods and statistics with data that looks like the real thing.
Can it replace a real survey?
No. Use it to design, test and teach. Do not publish it as real data, use it for decisions about individual people, or run significance tests on it as if its rows were real people.
Accuracy
How accurate is it?
We measure accuracy as what a file is worth in real respondents. In our tests on national surveys of US adults, a file made from the questions alone was about as informative as a real survey of 30-65 people. That holds however many rows you order: more rows give you more people to analyze, not more accuracy.
Which questions work best?
Accuracy depends mostly on what you ask, not on the answer scale. Opinions on broad social and political issues, personality and self-description, and everyday behaviors asked as yes/no or select-all do best. Mental health, social support, past experiences and how much people value specific things need more caution.
How do you test it?
We compare synthetic files with real surveys that our models never saw: election and civic surveys, a Federal Reserve household survey, a national health information survey, Census records, and large personality and interest datasets. We write down our predictions before each test and publish the misses.
Does it work for small or specific groups?
Our tests are on US adults in general. You can narrow by age, sex, education, employment and state, or describe the group in plain English, but expect less accuracy for small or unusual groups. Describe them as precisely as you can.
How the answers are made
Are the respondents real people?
No. Each one starts from a real, anonymous American Community Survey record (2020-2024), which supplies their age, sex, education, state and employment. Everything else, including every answer to your questions, is generated.
Does an AI role-play each respondent?
No. Role-played answers come out too tidy and too consistent. An AI model reads each question and predicts how the population answers it; a statistical model then writes individual answers, with the person-to-person variation and inconsistency real respondents show.
How do you know which questions go together?
We built our own model that predicts, from the wording of two questions, how strongly real people's answers to them are related, including when one question is worded in reverse. It was trained on how hundreds of thousands of people answered real surveys. On a benchmark of 87,153 question pairs that no model trained on, it was the most accurate of ten openly available models we tested.
Where is it weaker?
Fewer synthetic people sit on the fence on politically charged questions than real people do, and questions on unrelated topics come out somewhat more related than they really are. The read-me in every download says what to watch for.
Using the service
What question types can I use?
Agree-disagree (5 and 7 points), describes-me (5), satisfaction (5), importance (4), how-much-bothered (4), never-always (5), yes/no, select-all-that-apply, number answers, and demographics filled from the Census record. Not yet supported: unordered single choice (such as 'what kind of car do you drive') and open text.
How many questions can I include?
Up to 120 per data file. Each rating statement, yes/no question or number question counts as one, and each option in a select-all question counts as one. Demographic columns (age, sex, education, state, employment) are free. For a longer survey, split it into two files, or write to info@likelypeople.com.
What do I get?
A CSV with one row per respondent and one column per question, ready for SPSS, R, Stata or Excel; a codebook; and a read-me that says what the file is worth.
How long does it take, and what does it cost?
Most files are ready in 5-20 minutes. The price is per data file, shown before you pay, with no subscription.
Can I get more than one sample?
Yes. Each purchase includes up to 5 independent draws of the same study: a different set of people from the same population, answering the same questions. Teachers can give each student their own sample, and you can see how much results vary from sample to sample. Each draw is worth the same as the first; together they do not add up to more accuracy. Changing your questions, population or number of respondents counts as a new file.
Can I use my own questionnaire?
Yes, if you have the right to use it. Some published questionnaires are copyrighted; that is between you and the copyright holder.
Do I need IRB approval to use synthetic data?
Making a synthetic file involves no human subjects: no one is recruited, surveyed or identified, and the Census records we start from are anonymous public-use data. So generating a file does not by itself call for institutional review board (IRB) review, and you can design and test a questionnaire before you apply. Your institution makes the final call, and any later survey of real people will need its usual review.
What happens to my questions?
We use them to make your file, and, as the Terms of Use explain, to improve the service. We never publish them in a form that identifies you or give them to other customers. See the Privacy Policy.
Something not covered? Write to info@likelypeople.com.