Computerized classification testing with the Rasch model

Abstract

If classification in a limited number of categories is the purpose of testing, computerized adaptive tests (CATs) with algorithms based on sequential statistical testing perform better than estimation-based CATs (e.g., Eggen & Straetmans, 2000). In these computerized classification tests (CCTs), the Sequential Probability Ratio Test (SPRT) (Wald, 1947) is applied to determine when and which classification decision is to be taken. In practice, the procedure is always truncated at a maximum test length (TSPRT). Stochastically Curtailed SPRT (SCSPRT) (Finkelman, 2008) uses additional stopping rules. If the Rasch model (1960) is used as the item response theory (IRT) model, applying the TSPRT and SCSPRT is elegant and simple. The performance of the TSPRT- and SCSPRT-based CATs are compared using different item selection methods. It is shown that the TSPRT and SCSPRT procedures are much better than optimal traditional linear tests. Results with the Rasch model are compared to results with other IRT models.