Item response theory: how IRT turns your answers into a score

Short answer: Item response theory (IRT) scores a test from the pattern of your answers. Columbia University's Mailman School of Public Health calls it a "family of mathematical models" that link a trait you cannot observe directly to the answers people give. An IRT test compares your right and wrong answers with each item's calibrated curve, finds the ability level, called θ, that best explains them, and reports it with a standard error. So your score comes with a range.

How IRT turns your answers into a score

1. The items are calibrated first. When an adaptive test is built, its items are pretested and calibrated "using actual candidates", then linked onto one scale (Seo, 2017).
2. Your answers form a pattern. θ is estimated "based on the string of correct and incorrect responses provided by the examinee" (Klein and Kovacs, 2024).
3. The model finds your most likely ability. In their words, "θ is the point in the ability scale where the given string of responses is most likely to have occurred."
4. θ is placed on a reported scale. On the adaptive test in Klein and Kovacs's study, θ compares you with the norming sample in standard deviation units. Each test converts θ to its own reporting scale.
5. The error is reported as a range. The fewer items a test has near your ability, the larger the standard error, so the report gives a band around your score.

What item response theory is, and how it differs from counting right answers

Storme and colleagues (2019) put the aim plainly: IRT models "the probability of a response to an item as a function of the latent trait or ability of the test taker". Which trait an IQ test targets is covered in our guide to what an IQ test measures.

The older approach is classical test theory (CTT). Schmukle (2025) gives its core: your observed score is your true score plus an error score, written X = T + E.

How the two compare:

• What gets modeled. CTT models your total score. IRT models your chance of getting each item right, given your ability and that item's properties (Storme and colleagues, 2019).
• Measurement error. CTT treats the error as the same size for every test taker, and IRT lets it vary, as Wikipedia's article on item response theory puts it. (Wikipedia is a secondary source.)
• Dependence on the sample. Yang and Kao (2014) call CTT "test- and sample-dependent". When an IRT model's assumptions hold, they add, ability estimates do not depend on the test and item parameters do not depend on the sample.
• Simplicity. CTT scoring is easy to compute and to explain, the same Wikipedia article says, and IRT's gains "only hold when the assumptions of the IRT models are actually met".

The item characteristic curve and the a, b and c parameters

Every item has its own curve, the item characteristic curve (ICC). Yang and Kao describe it as a picture of "the true relationship between the trait and the responses to the item". Columbia's page says "the curve is S-shaped (Sigmoid/Ogive)".

Three parameters shape that curve.

• Difficulty, b. Seo (2017) calls b "an index of item difficulty". It is the point on the θ scale where you have a 50% chance of a correct answer, when there is no guessing. Harder items sit further to the right.
• Discrimination, a. The slope of the curve at its middle. A high a gives a steep curve and a low a a flat one, in Seo's description. Sources differ on its usual range. Seo gives a practical upper limit of about 3.0. Columbia's page (read November 3, 2026) says about 2.
• Guessing, c. The floor of the curve, the chance that someone of very low ability still gets the item right. For a four-option item, c "would be approximately 0.25", per Wikipedia's article (read November 3, 2026).

θ has no built-in unit: it "lacks a natural metric, thus needing the model to establish one", and people and items share one scale, "called a logit scale" (Schulze and colleagues, 2025). The name comes from the log-odds link between θ and your answers.

The main models: 1PL, 2PL and 3PL

The models differ in which of a, b and c each item gets. Antoniou and colleagues (2022) set out the formulas. You do not need them to read a score.

• 1PL and the Rasch model. Only difficulty varies: P = e^(θ − b) ÷ (1 + e^(θ − b)), and "the item slopes are fixed to unity". Georg Rasch published Probabilistic models for some intelligence and attainment tests in 1960, through the Danish Institute for Educational Research, per Wikipedia. Every item has the same slope, so the curves run parallel.
• 2PL. Discrimination varies too, so curves can cross: P = e^(a(θ − b)) ÷ (1 + e^(a(θ − b))), "with steeper slopes associated with greater discrimination and vice versa".
• 3PL. A guessing floor is added: P = c + (1 − c) × e^(a(θ − b)) ÷ (1 + e^(a(θ − b))).

Seo and also Yang and Kao credit Allan Birnbaum with the logistic models behind the 2PL and 3PL. His chapters appear in Statistical Theories of Mental Test Scores (1968), the book by Lord and Novick.

A worked 1PL example

Take one item with difficulty b = 0. By the 1PL formula (a calculation, not a figure from any real test):

• At θ = −1, one unit below the item, your chance of a correct answer is 0.269.
• At θ = 0, where ability matches difficulty, the chance is 0.500. Even odds.
• At θ = +1, one unit above, it is 0.731.

So an item tells the model most about people near its difficulty.

Same number right, same score?

Under the Rasch model, on the same set of items, yes. Antoniou and colleagues (2022) write that the total score is "a sufficient statistic" for this model. That means the number right carries everything the model needs to estimate θ, even though different patterns can give the same total.

A calculation shows it. Take three items at b = −1, 0 and +1. Invented person A gets the easy and middle items right. Invented person B gets the middle and hard ones. Under the Rasch model, by maximum likelihood, both get θ = 0.80.

Under the 2PL, the pattern matters. Give the items discriminations of a = 0.5, 1 and 2. By the same calculation, A gets θ = 0.50 and B gets θ = 1.91. Each answer's weight comes from its discrimination, not its difficulty. Schulze and colleagues (2025) note that then "the sum score of the scaled items is no longer an equivalent of the IRT person score".

Difficulty still matters, in a different way. It sets where each item's curve sits on the scale. So when people answer different items, as on an adaptive test, the same number right on harder items can mean a higher θ.

The College Board says on its SAT scoring page (read November 3, 2026) that on the digital SAT, two students with the same number correct in a section "may earn differing section scores based on the characteristics, including difficulty level, of the particular questions they answered correctly."

How θ is estimated

The common route is maximum likelihood (ML). The model multiplies the item probabilities into one likelihood for your pattern, then calculates "the maximum likelihood estimate of ability", as Columbia's page explains. The peak is your θ.

ML has a gap. It works "only when there is a mixed response pattern", Seo (2017) writes, meaning some answers right and some wrong. By the 1PL formula, if you get every item right, the likelihood keeps rising with θ, so the estimate never settles.

The Bayesian fix is expected a posteriori (EAP) scoring. It starts from a prior, an assumption about how ability is spread before any answers are seen, which Seo says "usually assumes a standard normal distribution". Your answers update the prior into a posterior, and EAP takes "the mean of the posterior distribution of ability" as your θ.

Item information: why your score is a range

Each item adds information about θ. Columbia's page says item information is the chance of a right answer times the chance of a wrong one, largest "when the probability of answering correctly or wrongly are equal, i.e. 50%." Test information is the sum over items, and the standard error is 1 ÷ its square root (Xiong and colleagues, 2026).

An idealized calculation, not WhatIQ's or any real test's figures:

• A 24-item 1PL test with every item at your ability gives information of 24 × 0.25 = 6, a standard error of about 0.41.
• On the IQ scale, with its standard deviation of 15 (Voncken and colleagues, 2019; see the θ-to-IQ section), that is about 6.1 points, a 90% interval of about ±10 points.
• The same 24 items, all at b = 0, taken by someone at θ = +2 give information of about 2.52 and a 90% interval of about ±15.5 points.

That is why a score is wider at the extremes. Helland-Riise and colleagues (2024) saw the effect in 113,671 adolescents' responses to a Norwegian figural matrices test used in military selection. They report "precision that is generally poor on the upper part of the scale", since the test "was deliberately designed to screen at the lower part of the ability scale". Our guide to how accurate IQ tests are covers other sources of error.

Adaptive testing: choosing the next item

A computerized adaptive test (CAT) picks each item from the item information function: given your current estimate, it chooses "the most informative item among the remaining items" (Seo, 2017). Summing up earlier studies, Seo reports that "the number of administered items in CAT is on average 50% shorter than when a paper-pencil test is used, with equal precision".

The digital SAT is multistage adaptive. The College Board says: "Based on how you perform on the first module, the second module of questions will either be more difficult or less difficult." It credits "use of Item Response Theory (IRT) methodology" for precise measurement with fewer questions.

An adaptive test need not score your pattern. ETS says the GRE General Test (read November 3, 2026) is "section-level adaptive", yet "Within each section, all questions contribute equally to the final score."

Where IRT shows up in IQ and ability tests

Riverside Insights says the Woodcock-Johnson W score is "a mathematical transformation of the Rasch model, based on item response theory" (its blog, from a webinar first aired May 6, 2021). For the Stanford-Binet Fifth Edition, publisher PRO-ED says its optional Change-Sensitive Scores "use modern item response theory scaling to convert the raw scores into criterion-referenced scores". Our Stanford-Binet guide covers the test itself.

Matrix puzzles have been studied with IRT too. Antoniou and colleagues (2022) fit models "from the Rasch through to the four-parameter models" to Raven's Coloured Progressive Matrices in 1,127 children aged 5 to 11. For one form, which model fit best gave "a very mixed picture": different fit criteria favored different models. Our guide to Raven's Progressive Matrices explains the puzzle format.

From θ to the IQ scale

Klein and Kovacs (2024) describe θ on the adaptive test in their study as "a standardized z-score", where θ = 0 is the norming sample's average.

The step to IQ is the deviation formula. Voncken, Albers and Timmerman (2019) give the rule: multiply the z score by the standard deviation of the IQ scale, 15, and add its mean, 100. So if θ is standardized on a norming sample, IQ = 100 + 15 × θ. By that rule, θ = +1 gives 115. Our IQ bell curve guide shows how scores spread around 100.

That mapping is our inference from those two sources. No source we read states it for a named IQ test, and each test sets its own conversion. It also needs a norm group, the people the test is scored against. Without one, θ only places you relative to the group the items were calibrated on. Our IQ formula guide covers deviation IQ and z, and our IQ percentile guide turns a score into a rank.

What that means for your score: Sam is invented for this example. Sam takes a test that sets θ against its norm group and gets θ = +1, so by the rule above the score reads 115. The range printed around it is the standard error, converted to IQ points. A different test, with its own items, model and norm group, or another sitting of the same test, can place that score somewhere else. A score can differ between tests and between sittings, so read the range as well as the number.

What IRT does not do

IRT makes better use of each answer. It cannot add answers that were never given.

• A short test stays imprecise. In the idealized calculation above, 24 perfectly targeted items still leave a standard error of about 6 points. Storme and colleagues (2019), studying an online reasoning test taken by 2,949 job applicants, wrote: "Extracting reliable information from short tests remains a real challenge from a measurement perspective". Our guide to how long an IQ test takes compares test lengths.
• Model fit has to be checked. Schulze and colleagues (2025) found one data set where "the Rasch model does not fit the data very well".
• The scoring method is not the assessment. A clinical assessment, such as the WAIS given by a psychologist, is more than a scoring method, so IRT does not make an online test clinical.

Taking a test that uses IRT

To try one, WhatIQ's homepage (read November 3, 2026) describes its test as 24 picture puzzles, about 15 minutes, with IRT scoring. It reports your score on the standard IQ scale with the range it most likely sits in, and calls the result "not a clinical assessment".

Frequently asked questions

Can you explain item response theory in a simple way?

Each question on an IRT test has a curve showing how likely people at each ability level are to get it right. The test takes your right and wrong answers and finds the ability level, θ, that makes that pattern most likely (Klein and Kovacs, 2024). Then it reports that level with a range.

How does IRT scoring work?

Items are calibrated in advance on real test takers. Your answers are compared with each item's curve, and the model estimates your ability, θ, by maximum likelihood or by expected a posteriori scoring (Seo, 2017). θ is then put on the test's reporting scale and reported with a range.

What is the difference between IRT and classical test theory?

Schmukle (2025) describes classical test theory as observed score equals true score plus error, X = T + E. IRT models each item separately and lets the error vary by ability. Yang and Kao (2014) say CTT "is test- and sample-dependent", while IRT puts items and people on one scale (Seo, 2017).

What is classical test theory in psychology?

In classical test theory (CTT), your true score is your score free of measurement error. It can never be known, only estimated by your observed score, Wikipedia's item response theory article notes. Schmukle (2025) writes observed score as true score plus error, X = T + E, and the standard error of measurement as SD × √(1 − reliability), where √ means square root. That Wikipedia article says CTT treats this error as equal for every test taker.

Can two people with the same number right get different IRT scores?

It depends on the model. Under the Rasch model the total score is "a sufficient statistic" (Antoniou and colleagues, 2022), so the same count on the same items gives the same θ. Under the 2PL and 3PL, the pattern matters. The College Board says two digital SAT students with the same number correct "may earn differing section scores".

Does IRT make an online IQ test accurate?

Not on its own. IRT draws more information from each answer, but a short test still has a limit on precision, and the model has to fit the data. A score can also differ between tests and sittings. An online test that uses IRT is still not a clinical assessment by a psychologist.