Deep-Rasch as an Alternative to Rasch Modeling under Assumption Violations and Small Sample Sizes

In certain situations, it may be challenging to fully exploit the advantages of modern test theory, including Rasch modeling and item response theory (IRT), when applied to real data. Although Rasch modeling tends to be more robust than IRT for small sample sizes, it still requires that the assumptions of unidimensionality and local independence be satisfied. In practice, these assumptions are often violated, which can lead to less accurate analyses and reduced validity of the results. Deep-Rasch, which integrates deep learning with Rasch modeling, has been proposed as an alternative measurement framework to overcome these limitations. This study examines the potential of Deep-Rasch as an alternative to Rasch modeling using student response data from 17 final semester examinations at Universitas Terbuka (UT), with sample sizes ranging from 33 to 11,504 students. Most examinations consisted of 30 multiple-choice items. The analyses showed that several datasets violated one or both assumptions of Rasch modeling. Nevertheless, Deep-Rasch performed comparably to conventional Rasch modeling in estimating item difficulty and student ability parameters, as well as in predicting student responses. Remarkably, for the smallest sample size (\emph{n} = 33), Deep-Rasch exhibited slightly better performance than Rasch modeling.