Hierarchical models for educational data: An overview

model checking The use of hierarchical models in statistical applications, and for educational data, is a promising but still underutilized approach. However, because these models are more complicated than many standard methods, it is important that we, as users and developers, not rush to use them before we understand them. We emphasize here, in support of the views on hierarchical models expressed in the 3 preceding papers by Draper, by Rogosa and Saner, and by de Leeuw and Kreft, the need to not diminish hard thinking about data and iterative model checking when fitting hierarchical models, the need for more and better software, the need to test methods to assure their proper calibration, and the need to produce supporting materials to aid analysts and users of hierarchical modeling methods. The RAND conference convened researchers knowledgeable of and inter-ested in the theory and application of hierarchical models, especially for education data. Their concerns involved the proper use and interpretation of such models, and the availability and quality of software. Hierarchical models are extremely promising tools for data analysis. As hierarchical modeling has become better understood and better supported, its applications have proliferated in such diverse fields, besides education, as health and medicine, quality assurance, demography, and remote sensing. Hierarchical models are the source of considerable excitement now, with the computing environment making their use widespread in this decade. Nevertheless, there is the danger that these fascinating but complex models will be oversold before their performance is adequately understood, and that their nominal operating characteristics may not be achieved in particular applications