Béguin-Glas (2001) MCMC Estimation and Some Model-Fit Analysis of Multidimensional IRT Models

Summary

Béguin and Glas give a fully Bayesian Gibbs-sampler estimation procedure for normal-ogive item-response models that generalizes Albert's (1992) two-parameter data-augmentation scheme in three directions at once: adding a guessing parameter (three-parameter model, 3PNO), moving to multidimensional ability (a $Q$ -dimensional normal ogive = full-information factor analysis), and supporting multiple populations and incomplete designs. Factor-structure restrictions ("subscale factor analysis") let one test hypotheses about the ability structure, and posterior-predictive checks provide model-fit analysis. It is essentially the Bayesian/MCMC counterpart to Bock-Gibbons-Muraki full-information factor analysis (TESTFACT).

Key Claims

Generalizes Albert (1992). The engine is the same latent-Gaussian data augmentation ( $Z_{ij}$ underlying each binary response, so the conditionals are Gaussian/truncated-Gaussian), extended to a guessing parameter and to a vector ability $\boldsymbol\theta_i\in\mathbb R^Q$ .
Three-parameter normal ogive (3PNO). With a guessing probability $c_j$ , a second layer of augmentation is introduced — a latent indicator for whether a correct response came from "knowing" vs. "guessing" — restoring conditional conjugacy so Gibbs still applies.
Multidimensional model = factor analysis. The $Q$ -dimensional normal ogive is the normal-ogive/full-information factor model; the paper labels the with-guessing and without-guessing versions $(Q+2)$ PNO and $(Q+1)$ PNO. Item discriminations are factor loadings, and the relationship to factor analysis (McDonald 1982; Mellenbergh 1994) is exploited directly.
Multiple populations and incomplete designs. The sampler handles data from several groups and matrix-sampled / incomplete item administration — the unified framework (Bock-Zimowski 1997) for differential item functioning, item-parameter drift, non-equivalent-groups and vertical equating, and computerized-adaptive-testing calibration.
Identifiability via factor-structure restrictions. Because factor models are rotationally indeterminate, restrictions on the loading (factor) matrix pin down the solution and let one test specific hypotheses about which items load on which latent dimension ("subscale factor analysis").
Model fit. Posterior-predictive model-fit analysis is used to assess dimensionality and fit; illustrated on simulated (parameter-recovery) and real data.

Concepts Introduced or Extended

Entities Mentioned

Quotes

"The procedure is a generalization of a procedure by Albert (1992) for estimating the two-parameter normal ogive model."

"It is shown that restrictions can be imposed on the factor matrix for testing specific hypotheses about the ability structure."

My Take

This is the natural sequel to Albert (1992) and a good illustration of how far the probit data-augmentation idea stretches: once you can impute the latent normal behind a binary item, you can bolt on a guessing layer (another latent indicator) and swap the scalar ability for a vector without leaving the Gibbs framework — so the "hard" psychometric models (3PL, multidimensional/full-information factor analysis) become routine MCMC. The genuinely useful engineering is the double augmentation for the 3PNO (knowing-vs-guessing) and the recognition that the multidimensional normal ogive simply is a factor model, so identifiability is handled with the same loading-matrix restrictions factor analysts already use. For the wiki it extends the IRT and data-augmentation threads from Albert's unidimensional 2PL to the modern multidimensional/multi-population setting, and connects the psychometric factor-analysis machinery to Bayesian estimation.