Generalized Item Response Tree Models for Psychological Assessments

• In Event: 41.027 - Recent Advances in Modeling Response Style Bias in Rating Scale Data

Abstract

Item response theory (IRT) models are useful tools for analyzing categorical item responses in psychological, educational, and behavioral assessments. IRT models typically focus on understanding item responses, which are the outcomes of a respondent's choice made among other (discrete) options. A new type of IRT model has been proposed which focuses on explaining the internal cognitive or psychological decision processes, not only the results of those processes. These models often utilize a tree structure to describe a postulated internal decision process and hence, are referred to as an item response tree or branching model (e.g., Boeckenholt, 2012; De Boeck and Partchev, 2012).

Item response tree models can also be applied to capture features of item response categories. For instance, a tree can be used to illustrate a unipolar scale that includes choices such as ‘not at all sad’, ‘slightly sad’, ‘mostly sad’, and ‘completely sad’, while another tree can be utilized to describe a bipolar scale with options such as ‘completely sad’, ‘somewhat sad’, ‘somewhat joyful’, and ‘completely joyful’. A tree structure can also be specified to illustrate Likert scales with a middle response category (e.g., ‘neither sad nor joyful’, ‘perhaps’, ‘not sure’, ‘undecided’, ‘?’, etc.). In this case, an outcome is set to indicate the undecided (middle) category, while other outcomes are set to represent regular response categories. It should be noted that in the current practice of item response analysis, the differences between various response formats are usually ignored and the responses are analyzed as if they were interval scale data or ordinal scale data (e.g., using ordinal factor/IRT models).

Item response tree models have been utilized in a variety of applications, for instance, for differentiating slow and fast intelligence (Partchev and De Boeck, 2012, Ditrapani et al., 2016), modeling motivated misreports to sensitive survey questions (Boeckenholt, 2013), examining content and response styles in multiple-choice items (Plieninger and Meiser, 2014), and modeling skipped and non-reached item responses (Debeer et al., submitted). These prior approaches, however, seem to have some limitations. For instance, the models proposed by De Boeck and Partchev (2012) do not allow for trees with more than two branches and rely on one-parameter logistic models.

In this presentation, I will discuss a generalized item response tree modeling framework that has been proposed by Jeon & De Boeck (2016). By utilizing nodes as building blocks, the framework allows for specifying a possibly different IRT model at each node. Latent variables (or abilities) can be uni- or multi- dimensional and can be node-specific or shared between nodes. Similarly, item parameters can also be node-specific or shared across nodes. Explanatory variables for persons and items can be incorporated to investigate sources of hetoregeneity in a response scale. I will show that item response tree models constructed within the proposed framework can be readily estimated with a standard IRT software package. I will also discuss some useful extensions of item response tree models, for instance, which allow for hetoregeneity in respondents’ differential response styles.

Authors

• Minjeong Jeon, University of California at Los Angeles

• Paul De Boeck, The Ohio State University