Extending Locally Dependent Item Response Models for Analyzing Psychological and Social Surveys

Wake Forest University School Of Medicine, Winston Salem NC

Investigators

Abstract

This research concerns the development of flexible statistical and psychometric methods for analyzing item-response data from educational tests and social surveys. The project contains two main components. The first component is the methodological development of item response models and extensions, specifically the locally dependent hybrid kernel models for dichotomous and polytomous responses. These methods are applicable to items that do not function independently after conditioning on subject effect. Examples of locally dependent items include items that have a common reading stem (as in a reading comprehension test), or items that survey related quantities such as the frequency and the intensity of a feeling (as in a psychological test). The project is expected to advance standard item response models in several ways: (1) the handling of dependency within item clusters with separable submodels, (2) the incorporation of multiple scales, (3) the accommodation of item- and person-specific covariates, and (4) the exploitation of rating scales of items. The second component of this project addresses applications. Three data sets from three different areas - education, psychology, and health-related social study - have been identified, and each will be analyzed using the locally dependent models. Item response models are increasingly used in calibrating scientific instruments used for measuring human traits and behavior. Recent examples include large-scale educational assessment and health-related quality-of-life research. This research supplements the most commonly used item response model - the unidimensional model - with flexible ways of dealing with potential minor deviations from the unidimensionality assumption. The benefits of the project include providing more flexible analytic methods that are compatible with the standard item-response models, which means that interpretability is retained, and adding novel features such as the handling of a wide range of response data (e.g., data with covariates or data that can be formulated as a rating scale). Given the increasing interest in the applications of refined item-response models to newer and often more complex tests and surveys, this project is expected to have an impact on improving both the quality of instruments and the subsequent analysis of gathered responses. This research is supported by the Methodology, Measurement, and Statistics Program and a consortium of federal statistical agencies as part of a joint activity to support research on survey and statistical methodology.