Selecting Testlet Features With Predictive Value for the Testlet Effect

High-stakes tests often consist of sets of questions (i.e., items) grouped around a common stimulus. Such groupings of items are often called testlets. A basic assumption of item response theory (IRT), the mathematical model commonly used in the analysis of test data, is that individual items are independent of one another. The potential dependency among items within a testlet is often ignored in practice. In this study, a technique called tree-based regression (TBR) was applied to identify key features of stimuli that could properly predict the dependence structure of testlet data for the Analytical Reasoning section of a high-stakes test. Relevant features identified included Percentage of “If” Clauses, Number of Entities, Theme/Topic, and Predicate Propositional Density; the testlet effect was smallest for stimuli that contained 31% or fewer “if” clauses, contained 9.8% or fewer verbs, and had Media or Animals as the main theme. This study illustrates the merits of TBR in the analysis of test data