A Distribution Based Association Measure And Its Application In Dimensionality Assessment

The core of the procedure is a novel pairwise association measure based upon information theory and boundaries on bivariate distributio:Qs. Asymptotic results on the standard error of the measure allow to scan for anomalies in the pairwise item association matrix, allowing for the detection of serious local item  dependence issues in the test. To assess the more general underlying dimensionality of the test a divisive clustering procedure is used to search for structure among the test items. A criterion that balances the homogeneity within clusters and the heterogeneity between clusters is suggested to select an optimal partitioning within the set of cluster solutions. The method is illustrated using a range of simulated test data under both strict and essential uniand multidimensional conditions.