| The effectiveness of a set of procedures in assessing the degree of parallelism between alternate forms of a standardized test was investigated. A dimensionality assessment of large-scale databases was conducted using information obtained from multidimensional item response theory (MIRT) parameter estimates to identify item clusters, combined with a statistical assessment of the dimensional distinctness of the item clusters using a nonparametric conditional covariance comparison procedure. The set of procedures consists of an initial estimation of the optimal number of dimensions in the data (DETECT, Kim, 1994; Zhang & Stout, 1996), a parametric cluster analysis procedure (Miller & Hirsch, 1992) for identifying item cluster groupings, and a nonparametric statistical evaluation of the dimensional distinctness of the item clusters (DIMTEST, Stout, Douglas, Junker, & Roussos, 1993).;The dimensionality assessment results from two samples of examinees completing the same exam resulted in item clusters replicable at 88 percent. The item clusters were both statistically dimensionally distinct and substantively interpretable. The second comparison was between two samples of examinees completing two alternate forms of an exam. Unfortunately, use of the procedures proposed in the study, with the data for the second test form provided, resulted in an identification of item clusters that were neither dimensionally distinct nor substantively interpretable.;Because of the numerous composites of abilities that could account for the variability in item responses attributing to the item clustering, post hoc analyses were conducted using a subset of items hypothesized to have approximate simple structure. The MIRT clustering procedure used in evaluating the complete test form information was compared with the item cluster groupings identified using a conditional covariance clustering procedure (DETECT, Kim, 1994; Zhang & Stout, 1996). The results indicated that the conditional covariance clustering procedure, in combination with the genetic algorithm used for identifying the optimal number of statistically distinct dimensions, was more effective in replicating the hypothesized dimensional structure of the subset of items. |
