A test built on method, with its limits stated

The questions are our own. The figures are generated by rule, in code, and each one is checked to have exactly one correct answer. The wrong options are built so that you can't find the right one by elimination without understanding the pattern.

In each session, the test draws the questions from a larger bank, one from each difficulty level before repeating a level, and orders them from easiest to hardest. The forms vary from person to person, which makes it hard for answer keys to circulate.

The result is calculated with Item Response Theory, using the three-parameter model. Each question has three characteristics: its difficulty, how well it separates people with more ability from people with less, and the chance of getting it right by guessing.

Your ability is the value that best explains your pattern of right and wrong answers. That's why getting a hard question right is worth more than getting an easy one right, and guessing doesn't produce a high number. The same kind of model is used in ENEM, Brazil's national exam. Read more in Item Response Theory, without the formulas.

The structure follows ICAR (International Cognitive Ability Resource), a public-domain battery administered online to nearly 97,000 people (Condon and Revelle, 2014). We use its published figures, not its questions. Our questions are original.

Every test has measurement error, and a short test taken online without supervision has more error than a full test given by a psychologist. That's why, alongside the number, we show the range your result most likely falls in, with 95% confidence.

With 37 questions, the standard error is around 5 to 6 points in the middle of the scale, which gives a range of about 10 points on either side.

The value is converted to the usual IQ scale (mean 100, standard deviation 15) by comparing you with people in your age group. Reasoning about new problems changes over adult life, and comparing by age takes that into account.

Today's norms are provisional. They use published parameters, mainly those from ICAR, adjusted because people who take part in online research score higher than the general population; the age curve follows the literature (Salthouse, 2009). They will be recalculated with the responses of people who take the test, separately for each language, and the report always says which norms were used.

• The test has a ceiling and a floor: for adults aged 20 to 29, it doesn't distinguish above about 138 or below about 57. These limits shift at other ages, and the report shows yours.
• Response time doesn't change the number: it is only used to catch answers given too fast. Education and self-rating don't change the number either.
• Answers given too fast, or a very unusual pattern of right and wrong answers, get the result flagged as unreliable, and retaking it is free.
• We don't give an IQ per area: with only a few questions per area, the error would be too large.
• We don't use a "Brazil average" or rankings by country. The percentile in the report is theoretical: your position on the normal curve.

With your permission, your answers go into a research dataset without your email, IP address or payment data. Once the sample is large enough, the model re-estimates the difficulty of each question and the age norms, reweights the sample by education and checks whether any question favors one age group.

Adopting new norms is a human decision, not an automatic one. The 16 to 19 age group stays provisional even with a large sample, because people who are still in school can't be reweighted by education. You can remove your answers from the research in the Account tab.