IRT — Item Response Theory

Dr. R. Düsing · Osnabrück University
IRT — Item Response Theory — Help
What Is IRT?

Item Response Theory (IRT) models how the probability of a correct answer depends on a person's ability θ (theta) and the properties of an item. Unlike Classical Test Theory (CTT), person ability and item difficulty lie on the same scale — this allows direct statements about which items are informative for which people.

IRT and Factor Analysis — Kinship and Difference

IRT and FA are more closely related than they appear at first glance: both model one or more latent variables that explain response behavior. Mathematically, the 2PL model is equivalent to a single-factor model for dichotomous data — the discrimination parameter a corresponds to a factor loading, θ corresponds to the person-specific ability parameter (the person's value on the latent scale). For ordinal items (Likert scales), the Graded Response Model and ordinal FA with polychoric correlations are closely related.

The decisive difference lies in the research question: factor analysis asks: how many latent dimensions are there, and which items load on which factor? The goal is the structure of the construct. IRT asks: how well does each individual item function? How precisely does the test measure at which point on the θ scale? The goal is the quality and usefulness of the items as well as the measurement precision of the test for specific people.

ICC — Item Characteristic Curve

For each item there is an S-shaped curve (ICC, also IRF — Item Response Function) that shows: how high is P(correct | θ)? The general model with four parameters (4PL):

4PL: P(θ) = c + (d − c) · 1 / (1 + exp(−a · (θ − b)))

a = discrimination (steepness) · b = difficulty (inflection point) · c = lower asymptote (guessing) · d = upper asymptote (carelessness)

The simpler models arise by successively fixing parameters:

3PL: P(θ) = c + (1 − c) / (1 + exp(−a · (θ − b))) [d = 1] 2PL: P(θ) = 1 / (1 + exp(−a · (θ − b))) [c = 0, d = 1] 1PL: P(θ) = 1 / (1 + exp(−a · (θ − b))) a equal for all items Rasch: P(θ) = 1 / (1 + exp(−1 · (θ − b))) a = 1 by axiom

Going from the 4PL to the Rasch model, flexibility decreases but interpretability and measurement-theoretic rigor increase. 2PL and 3PL are the most common models in applied assessment.

The Four Parameters

b — difficulty: θ value at the inflection point of the ICC — the point where the S-curve has its maximum slope (the transition from increasing to decreasing slope). For the 2PL, P(θ=b) = 0.5. For the 3PL/4PL, P(θ=b) = (c+d)/2. High b values → difficult item. b lies on the same scale as θ.

a — discrimination: steepness of the ICC at the inflection point. High a → steep curve → item discriminates well between ability groups. Typical: a ∈ [0.5, 2.5]. Values below 0.3 indicate a weak item.

c — pseudo-guessing probability: lower asymptote. Even at very low θ, P(correct) ≥ c. For multiple-choice with K options, c ≈ 1/K is plausible. Ignoring c (2PL) overestimates θ for low-ability persons.

d — upper asymptote (carelessness): maximum response probability. d < 1 models careless errors even among very able persons. Hard to estimate in practice — needs many high-ability persons in the dataset.

Rasch Model vs. 1PL — the Decisive Difference

This distinction is often blurred in textbooks. Both models have a identical across all items — but the philosophy is fundamentally different:

Rasch model: Georg Rasch (1960) defined his model not as a statistical data model, but as a measurement ideal. a = 1 is set by axiom — not because the data demand it, but because it is the precondition for specific objectivity: item difficulty b should hold independently of the sample tested, and person ability θ independently of the test used. If an item doesn't fit, it is removed or revised — not the model.

1PL model: a is estimated as a common parameter from the data. The model is fitted to the data. If a ≠ 1 comes out, that's not a problem. There is no axiomatic claim to measurement independence. In short: Rasch fits the data to the model — 1PL fits the model to the data.

Advantages of Rasch-conforming items: if items satisfy the Rasch model, strong measurement properties result: (1) Person-free item calibration — difficulty b is estimated independently of which persons took the test. (2) Item-free person measurement — θ is estimated independently of which item selection is used. (3) Item banking — items from different tests can be anchored on a common scale and interchanged (the basis for adaptive testing). (4) Fit statistics — items that don't fit are identified and can be targeted for improvement. These properties make the Rasch model the gold standard in psychometric quality assurance.

Test Information Function (TIF)

I(θ) shows how precisely the test measures at each ability level. It is the sum of the item informations:

I(θ) = Σᵢ Iᵢ(θ) with Iᵢ(θ) = [P'ᵢ(θ)]² / [Pᵢ(θ) · (1 − Pᵢ(θ))]

The standard error of the estimate is SE(θ) = 1/√I(θ). A test measures most precisely where many items with b ≈ θ are present and discriminations a are high. Extreme groups are often measured poorly.

Interpretation: the TIF shows relatively where a test measures precisely — not whether it reaches an absolute threshold. What matters is the shape of the curve: a narrow, high maximum means precise measurement in a narrow θ range (typical for selection tests); a broad, flat curve means uniform but less precise measurement across many ability levels (typical for screening). The standard error SE(θ) = 1/√I(θ) translates the information directly into measurement uncertainty on the θ scale — at SE = 0.5, the estimate typically lies within ±1 of the true value. More items shift the whole curve upward; harder or easier items shift the information maximum along the θ axis.

Wright Map — People and Items on One Scale

A unique feature of IRT: person ability θ and item difficulty b lie on the same scale. The Wright map (item-person map) shows both together. Items with b ≈ θ are most informative for a person. Items that are too easy (b ≪ θ) or too hard (b ≫ θ) contribute little to measurement. In this tool θ is vertical: on the left the person histogram (N(0,1) reference population), on the right the items at their difficulty b. Items with similar difficulty are fanned out horizontally side by side, and a colored tick on the θ axis marks the exact difficulty of each item. Dot size encodes discrimination a (steeper items = larger = more informative), and items within the teal band around the person's θ line (Max) are highlighted — that's exactly where the test measures precisely.

Response Pattern and θ̂ (MLE)

From an observed response vector (e.g. 1,1,0,0,0), θ̂ can be estimated via maximum likelihood: find the θ that maximizes the probability of the observed pattern. If someone answers everything correctly or everything incorrectly, the MLE is undefined (±∞) — in practice EAP or MAP estimators are then used.

References

Rasch, G. (1960). Probabilistic Models for Some Intelligence and Attainment Tests. Danish Institute for Educational Research.
de Ayala, R. J. (2022). The Theory and Practice of Item Response Theory (2nd ed.). Guilford Press.
Tutz, G. (2025). A Short Guide to Item Response Theory Models. Springer.
Mair, P. (2018). Modern Psychometrics with R. Springer.
Moosbrugger, H. & Kelava, A. (Eds.) (2020). Testtheorie und Fragebogenkonstruktion (3rd ed.). Springer.
Petersen, I. T. (2024). Principles of Psychological Assessment: With Applied Examples in R. Chapman & Hall/CRC.

📋 Example — Reading Comprehension Test
Max (grade 8) takes a standardized reading comprehension test with 5 items — from simple factual knowledge (item 1) to complex synthesis (item 5). How well does the test measure his true reading competence θ?
✓ Why IRT — in Addition to Classical Test Theory?
CTT/factor analysis and IRT are not competitors but complements: both assume that a latent ability generates the response behavior (for dichotomous data, the 2PL model is even formally equivalent to a single-factor model on tetrachoric correlations — a corresponds to a factor loading). The difference lies in the focus: in CTT/factor analysis the latent variable is the target of the analysis; reliability (α/ω) is then a single number for the whole test — it implicitly assumes equal measurement precision across the entire scale. IRT makes the individual item the object of study and shows explicitly that this assumption is often wrong: a test can be very precise in the middle and barely informative at the edges (see the test information function below).
Approach — From Item to Ability Estimate
①Item Characteristic Curve (ICC). The S-shaped curve shows P(correct | θ) — the probability of a correct response as a function of ability. → View ICC
②Choose a model — how many parameters does the item need? Rasch/1PL/2PL/3PL/4PL differ in whether discrimination, guessing, and carelessness are modeled. Rasch vs. 1PL is not a purely statistical distinction but a measurement-theoretic/philosophical one (details in Help).
③Test information I(θ) and SE(θ) — how precisely does the test measure where? This is the core advantage from the box above, now made visible: precision is not a constant but depends on θ. → View the test information function
④Wright Map — people and items on one scale. Do item difficulties match the ability distribution of the people (targeting)? → View the Wright Map
⑤From responses to θ̂. From a concrete response pattern, a person's ability can be estimated via maximum likelihood — try it below in the sidebar.
Items with more than two response categories? → Ordinal IRT Does an item work the same way for all groups? → DIF
ICC — Item Characteristic Curves
Test Information Function I(θ) and SE(θ)
Wright Map — Items and People on One Scale
How to Read the Wright Map

One shared scale θ (vertical): people on the left, items on the right — both on the same ability/difficulty axis. This is IRT's unique feature.

  • People — histogram of the reference population N(0,1).
  • Max — the current person; the teal band is his informative zone (|b−θ| ≤ 0.5).
  • Items — at their difficulty b; dot size = discrimination a (larger = discriminates more sharply). Items of equal difficulty are fanned out horizontally.
  • Tick (in item color) on the θ axis — marks each item's difficulty on the scale.
  • ◉Highlighted (halo + ring) = item near θ → the test measures precisely there; faded items are too easy or too hard for Max.

Do items and people overlap?

  • Good match (targeting): if items sit at Max's level (b ≈ θ), the test measures precisely there — small standard error. Ideally item difficulties exactly cover the θ range where the people lie.
  • Gap: in θ ranges without items, the test measures imprecisely even if people sit there — test information is low (see the TIF tab, where SE rises).
  • No overlap: if all items are too easy (b ≪ θ), Max solves practically everything → ceiling effect, θ is barely estimable (SE becomes very large). Mirror-image for consistently too-hard items → floor effect.

Try it: drag Max's θ slider out of the item cloud — once no item is highlighted anymore, there is no overlap, and the test is poorly matched to that ability. Item numbers correspond to the sidebar.

Model Formula & Interpretation
Item Parameter Table
ItemabcdP(θ)I(θ)
Flashcards