A researcher administers a math test where items require both numerical computation and reading comprehension. She fits a unidimensional IRT model. What is the most likely consequence?
AThe model will fail to converge because IRT cannot handle two-dimensional tests
BItem parameters will be biased and ability scores will conflate distinct skills, producing a single score that is difficult to interpret
CThe results will be equivalent to MIRT since IRT models adapt automatically to multidimensional data
DScores will be more precise because fewer parameters reduces estimation noise
Forcing a unidimensional model on multidimensional data produces biased parameter estimates — the single θ score will reflect a mixture of the underlying abilities in proportions that vary unpredictably across items. The scores cannot be cleanly interpreted as either verbal or quantitative ability. Option C is the key misconception: IRT does not adapt to multidimensional structure automatically; it simply tries to find the best-fitting line through a space that requires a plane.
Question 2 Multiple Choice
In a compensatory MIRT model, a student with very high verbal ability but only moderate spatial ability takes an item that requires both skills. What does the model predict?
AThe student will fail the item because both dimensions must exceed a threshold
BHigh verbal ability can offset moderate spatial ability, increasing the probability of a correct response
COnly spatial ability matters for spatially-loaded items, regardless of verbal ability
DThe two dimensions contribute independently with no possibility of offset
In a compensatory model, the probability of a correct response is determined by the dot product of the examinee's ability vector and the item's discrimination vector — this dot product allows strength on one dimension to compensate for weakness on another. A non-compensatory model would require adequate ability on all relevant dimensions. The distinction between compensatory and non-compensatory models is a theoretical choice that depends on whether the construct allows such trade-offs.
Question 3 True / False
In MIRT, each item has a discrimination vector that specifies how strongly and in what direction the item differentiates examinees across multiple ability dimensions.
TTrue
FFalse
Answer: True
This is the key extension from unidimensional IRT (where discrimination is a scalar) to MIRT. The discrimination vector indicates which dimensions the item loads on and how strongly. A purely verbal item has high discrimination on the verbal dimension and near-zero on the spatial dimension; an item requiring both has positive discrimination on both. The probability of a correct response is a function of the dot product between the examinee's ability vector and the item's discrimination vector — a direct application of linear algebra.
Question 4 True / False
A multidimensional IRT model has a unique correct orientation for its latent dimensions — there is mainly one valid rotation of the ability space that fits the data.
TTrue
FFalse
Answer: False
Just as in factor analysis, the orientation of the multidimensional latent space is not uniquely identified — there are infinitely many rotations of the factor axes that produce identical model fit. Choosing between oblique (correlated) and orthogonal (uncorrelated) rotations, and determining what each dimension means substantively, requires the same conceptual tools as factor analysis. This is why exploratory MIRT borrows heavily from factor analytic rotation methods and why dimension labels must be assigned by the researcher, not extracted mechanically.
Question 5 Short Answer
What is the fundamental difference between a unidimensional IRT model and a MIRT model in how they represent examinee ability, and why does this matter for complex psychological constructs?
Think about your answer, then reveal below.
Model answer: A unidimensional IRT model represents each examinee's ability as a single scalar θ on one latent continuum. A MIRT model represents ability as a vector of latent trait scores — one value per dimension. This matters because many real psychological constructs are not unidimensional: cognitive ability includes verbal, spatial, and mathematical components; personality includes extraversion, conscientiousness, and neuroticism. Forcing a single θ onto multidimensional data produces biased estimates and conflates distinct skills into an uninterpretable composite. MIRT yields separate, interpretable scores for each dimension.
The vector representation directly requires linear algebra: item parameters include discrimination vectors (like factor loadings), and the probability of a correct response is computed via dot products in the multidimensional ability space. This synthesis of IRT's probabilistic item modeling with factor analysis's dimensional decomposition allows MIRT to measure complex constructs with both precision and conceptual clarity — which is why it is essential when a test is designed to yield subscores or assess multiple distinct facets of a construct.