IRT–Factor Analysis Equivalence

Definition

The IRT–factor-analysis equivalence is the result (Takane-de Leeuw 1987) that the two-parameter normal-ogive item-response model and factor analysis of dichotomized (more generally, discretized) variables are the same latent-variable model, with identical marginal likelihood. Item-response theory (IRT) and factor analysis (FA) are two formulations of one model; they differ only in where the marginalization over the latent trait is carried out.

Key Ideas

• Parameter dictionary. IRT discrimination $a_i$ ↔ factor loading; IRT difficulty/threshold ↔ the cut-point that dichotomizes a latent continuous response; subject ability $\theta$ ↔ the common factor score $f$.
• Shared covariance. With $f \sim N(0,I)$ and diagonal uniqueness $\Omega^2$ (local/conditional independence), the latent responses have the classic FA covariance $CC' + \Omega^2$; observed categorical responses arise by thresholding, so the likelihood integrates a multivariate normal over a parallelepiped.
• Marginalization is the only difference. IRT dichotomizes conditionally on the latent trait and then marginalizes it out; FA marginalizes in a different order — but yields the same marginal likelihood. Estimation is marginal maximum likelihood in both.
• "Item factor analysis." The equivalence is why Bock renamed his IRT approach item factor analysis — full-information FA of categorical items is IRT.

How It Works

Let a subject's latent responses be $y = Cf + \varepsilon$ with $\varepsilon \sim N(0,\Omega^2)$ diagonal, so $y \sim N(0, CC'+\Omega^2)$. Each observed item $x_i$ is the sign of $y_i$ relative to a threshold $r_i$ ($x_i=1$ if $y_i>r_i$). Conditional on $f$ the items are independent (local independence), giving the IRT conditional likelihood $\prod_i \Phi(\cdot)$; marginalizing over $f\sim N(0,I)$ recovers exactly the FA marginal likelihood of the dichotomized variables. The argument extends to ordered categories (multiple thresholds → graded response) and unordered categories (and pair-comparison data as a special case).

Why It Matters

• Method transfer. Identification conditions, rotation, estimation algorithms, and software cross over between the psychometric (IRT) and multivariate-statistics (FA) traditions.
• Bayesian estimation. It underwrites the normal-ogive Gibbs samplers (Albert 1992; Béguin-Glas) that treat item parameters as factor loadings with Gaussian latent responses.
• Conceptual clarity. "Latent trait model," "item factor analysis," and "factor analysis of tetrachoric/polychoric correlations" name one object, not three.

Open Questions

• Where to marginalize in practice. The formulations are likelihood-equivalent but computationally different; the better estimation route depends on dimension of $f$ and number of items/categories.
• Beyond the normal ogive. The exact equivalence uses the probit (normal-ogive) link; the logistic (2PL) link is only approximately a factor model.

Related

• Factor Model
• Rasch Model
• Latent Class Model
• Item Response Theory
• Takane-de Leeuw (1987) On the Relationship Between IRT and Factor Analysis of Discretized Variables
• Yoshio Takane
• Jan de Leeuw