Multidimensional Rasch models for partial-credit scoring

Abstract

Rasch models for partial-credit scoring are discussed and a multidimensional version of the model is formulated. A model may be specified in which consecutive item responses depend on an underlying latent trait. In the multidimensional partial-credit model, different responses may be explained by different latent traits. Data from van Kuyk’s (1988) size concept test and the Raven Progressive Matrices test were analyzed. Maximum likelihood estimation and goodness-of-fit testing are discussed and applied to these datasets. Goodness-of-fit statistics show that for both tests, multidimensional partial-credit models were more appropriate than the unidimensional partial-credit model. Index terms: X2 testing, exponential family model, multidimensional item response theory, multidimensional Rasch model, partial-credit models, Progressive Matrices test, Rasch model.