Brian Habing

Department of Statistics

University of South Carolina


Locally Dependent, Monotone, Unidimensional Item Response Theory

The most commonly used item response theory models are those assuming a unidimensional latent trait (d=1), monotone item response functions (M), and local independence (LI). If these assumptions are violated the data is typically either modeled incorrectly or by assuming a multidimensional latent trait (d>1), M, and LI. The use of a d>1 trait is not always desirable however. In many cases there is a single dimension of interest, with the remainder being nuisance dimensions. In this case the primary goal is often to estimate the loss of information when a d=1 model is used incorrectly. Another reason that d>1 may be undesirable is that many of the statistical tests of d=1, M, LI are based on conditioning on a unidimensional latent trait and measuring the departure from LI. The d>1 models thus do not directly relate to the commonly used measures of departure from d=1, M, LI. Recently, these issues have been addressed by Ip (1998) and Pashley and Reese (in press). We examine their procedures and discuss two weaknesses in their approaches: inability to model guessing, and failure to examine which of the underlying correlation matrices are possible. A proposal for dealing with these difficulties is made, and some of the remaining issues in doing so are discussed. Joint work with Louis Roussos, Law School Admission Council.


Back to Colloquium Series