PsychologyOrbit
Sign in Join

What is Item Response Theory?

A family of models describing the probability of a given response to an item as a function of the respondent's standing on a latent trait and the properties of that item, rather than as a simple summed score.

From the PsychologyOrbit knowledge base, last updated 2 November 2026.

Knowledge base Updated 2 November 2026 1 views

Item Response Theory

A family of models describing the probability of a given response to an item as a function of the respondent's standing on a latent trait and the properties of that item, rather than as a simple summed score.

Definition

The common models for dichotomous items differ in how many item parameters they estimate:

  • 1PL / Rasch: difficulty only. All items assumed equally discriminating.
  • 2PL: difficulty and discrimination.
  • 3PL: adds a lower asymptote for guessing, used in multiple-choice testing.

For rating scales, the graded response and partial credit models extend the same logic to ordered categories.

The item characteristic curve is the central object: a function giving the probability of endorsing an item at each level of the trait.

History

Lord and Novick's 1968 book set out the theory; Rasch's model came from Danish educational measurement in 1960 and developed as a partly separate tradition with different commitments.

The Rasch tradition treats the model as a requirement that data must meet in order to support measurement. The IRT tradition treats models as descriptions to be selected on fit. The disagreement is philosophical rather than technical and it is why the two literatures often talk past each other.

Theory

What IRT buys over classical test theory:

  • Item and person parameters on the same scale, so difficulty and ability are directly comparable.
  • Invariant parameters in principle: item properties do not depend on the sample, and person estimates do not depend on which items were administered.
  • Information varies by trait level. Classical reliability assumes one precision for everyone. IRT shows a test can measure precisely in the middle of the distribution and poorly at the extremes, which is exactly where clinical decisions are made.

This last point is the strongest practical argument. It makes computerised adaptive testing possible: select the next item to be maximally informative at the current estimate, and reach the same precision with far fewer items.

Research

IRT underpins large-scale educational assessment, the PROMIS health measures, and the modern versions of many clinical instruments.

The assumptions are demanding and are often not checked: unidimensionality, local independence (responses independent given the trait), and correct functional form. Violations of local independence are common where items share wording or content and inflate apparent reliability.

Differential item functioning analysis is the IRT-based test of whether an item behaves differently across groups at equal trait level, and is the item-level counterpart of measurement invariance.

Sample requirements are substantial: 2PL models need several hundred respondents and 3PL considerably more.

Applications

Test construction and item banking, adaptive testing, scale shortening without losing precision, cross-form equating, and fairness analysis.

In clinical measurement, the practical gain is identifying that an instrument is imprecise exactly at the severity range where it is used to make decisions — something a single alpha coefficient conceals completely.

Criticism

  • Assumptions are strong and often unmet, particularly unidimensionality in clinical scales that are conceptually multidimensional.
  • Large samples are required, which puts it out of reach for much research.
  • The Rasch-versus-IRT dispute leaves practitioners without clear guidance on whether to fit the model to the data or the data to the model.
  • Complexity obscures. The output is harder to check, and errors are correspondingly harder to notice.
  • It does not establish validity. A well-fitting model of a construct that does not exist fits just as well.

References

  • Lord, F. M., & Novick, M. R. (1968). Statistical Theories of Mental Test Scores. Addison-Wesley.
  • Rasch, G. (1960). Probabilistic Models for Some Intelligence and Attainment Tests. Danish Institute for Educational Research.
  • Embretson, S. E., & Reise, S. P. (2000). Item Response Theory for Psychologists. Erlbaum.
  • Cella, D., et al. (2010). The Patient-Reported Outcomes Measurement Information System (PROMIS). Journal of Clinical Epidemiology, 63(11), 1179-1194.
Written for PsychologyOrbit and reviewed before publishing. It is a summary for professionals and students, not clinical guidance, and it is not a substitute for the primary sources listed above.