Takane-de Leeuw (1987) On the Relationship Between IRT and Factor Analysis of Discretized Variables

Summary

Takane and de Leeuw formally prove that the marginal likelihood of the two-parameter normal-ogive item-response model (IRT) is identical to that of factor analysis (FA) of dichotomized variables: the two are alternative formulations of the same latent-variable model, differing only in where the marginalization over the latent trait is performed. They extend the basic dichotomous result to multicategory data — both ordered and unordered — and treat pair-comparison data (from multiple-judgment sampling) as a special case of the unordered case, closing with a taxonomy of data types for the IRT/FA models.

Key Claims

• Formal equivalence. For dichotomous items, the marginal likelihood of the 2-parameter normal-ogive IRT model equals that of FA applied to the underlying dichotomized continuous responses. IRT discrimination ↔ factor loading, item threshold ↔ the cut-point of the latent response, and the subject "ability" ↔ the common factor score.
• Only marginalization differs. In the IRT formulation the latent response is dichotomized conditionally on the latent trait and then marginalized; in FA the marginalization order differs — but the resulting likelihood is the same. Hence Bock's "item factor analysis" name for full-information FA.
• Common-factor structure. With factor scores $f \sim N(0,I)$ and diagonal uniqueness $\Omega^2$ (local independence), the latent responses are $N(0, CC' + \Omega^2)$ — the standard FA covariance — with dichotomization defining a parallelepiped region of integration.
• Extensions. The equivalence extends to ordered categorical (graded-response) and unordered categorical data; pair-comparison data from multiple-judgment sampling is a special case of the unordered case. A taxonomy organizes the data types.

Concepts Introduced or Extended

• IRT–Factor Analysis Equivalence

Entities Mentioned

• Yoshio Takane
• Jan de Leeuw

Quotes

"It is clear from the discussion that IRT and FA are two alternative formulations of a same model … The only crucial difference is where the marginalization is performed."

My Take

A clean unification result: it dissolves the apparent boundary between the psychometric-testing tradition (IRT) and the multivariate-statistics tradition (FA of categorical data), showing they are one model seen from two angles. The payoff is practical — estimation machinery, identification conditions, and software from either literature transfer to the other, and it legitimizes "full-information item factor analysis" (Bock-Gibbons-Muraki) as literally factor analysis. On the wiki it ties the Factor Model and Rasch Model clusters together and underwrites the normal-ogive Gibbs samplers (Albert 1992; Béguin-Glas) that treat item parameters as factor loadings.