Nonparametric Multidimensional IRT (Item Response Theory) Approach with Applications to Ability Estimation and Test Bias

摘要

A determined case is made for the use of a nonparametric multidimensional monotonic IRT modeling framework with local independence replaced by the less restrictive assumption of essential independence. The concept of essential dimensionality is then introduced to count the number of dominant latent dimensions. Consequences of this more general approach include the consistent estimation of ability on a common scale using a natural class of estimators, uniqueness of the latent ability when essential unidimensionality holds, a theoretical treatment of test bias, an IRT based notion of validity, and a reassessment of the importance of the concept of item parameter invariance.