A topic in the Open Knowledge Graph — a free, open map of 15,290 topics and the order to learn them in.
Multidimensional IRT (MIRT) extends standard unidimensional IRT to simultaneously assess multiple latent traits or dimensions. MIRT is essential for complex constructs like cognitive ability (which includes verbal, spatial, and mathematical abilities) or personality (which may measure multiple trait dimensions). MIRT provides more accurate ability estimates and separate scores for each dimension.
Standard IRT models you've already studied assume a single latent trait θ that accounts for all covariation among item responses. Every item's characteristic curve is a function of one number: where the person sits on one ability continuum. This works well when a test really is measuring one thing — but many real constructs are not unidimensional. A math test might require both quantitative reasoning and reading comprehension. A personality inventory might tap extraversion, conscientiousness, and neuroticism simultaneously. When items load on multiple dimensions, forcing a single θ onto the data produces biased parameter estimates and misleading scores. Multidimensional IRT (MIRT) addresses this by extending the latent space from a line to a vector space.
In MIRT, each examinee is characterized not by a scalar θ but by a vector of latent trait scores — for example, θ = (θ₁, θ₂) for a two-dimensional model. Your linear algebra and matrix operations prerequisites become directly applicable here: the key model parameters include a discrimination vector (analogous to factor loadings in CFA) for each item, which specifies how strongly and in what direction the item differentiates examinees in the multidimensional ability space. An item that requires both verbal skill and spatial reasoning will have positive discrimination on both dimensions; a purely verbal item will have near-zero discrimination on the spatial dimension. The probability of a correct response is then a function of the dot product between the examinee's ability vector and the item's discrimination vector — a multivariable generalization of the single-parameter logistic function.
Your prerequisite on confirmatory factor analysis is the bridge to understanding MIRT's structure. In CFA, you specify which observed variables load on which latent factors; in MIRT, you similarly specify which items are allowed to discriminate on which dimensions. A compensatory MIRT model (the most common) assumes that high ability on one dimension can offset low ability on another — a high-verbal examinee can compensate for moderate spatial skill when the item requires both. A non-compensatory model requires adequate ability on all relevant dimensions; strong verbal skill cannot compensate for very low spatial ability. The choice between these models depends on the theoretical structure of the construct being measured.
The connection to eigenvalues and eigenvectors (your soft prerequisite) appears in two places. First, exploratory MIRT (when you don't know the dimensional structure in advance) uses factor-analytic techniques on the item correlation matrix, with eigenvalues indicating how many dimensions explain meaningful variance. Second, the orientation of the multidimensional ability space is not uniquely identified — there are infinitely many rotations of the factor axes that fit the data equally well, just as in factor analysis. Choosing between oblique and orthogonal rotations, and interpreting what each dimension means, requires the same conceptual tools. MIRT thus sits at the intersection of IRT's probabilistic item modeling and factor analysis's dimensional decomposition — a synthesis that becomes essential for measuring complex psychological constructs with both precision and conceptual clarity.