Measuring Response Styles in Likert Items Using Item Response Tree Models

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

Abstract

It is well-known that labels of items and their response categories can elicit response processes that may distort the measurement of interest. Textbook examples include such response-style (RS) effects as acquiescence, middle-scale, or extremity responding. For example, some respondents may avoid using extreme response categories even if they feel strongly about a topic, whereas other respondents may prefer to select extreme response categories. Although these reactions are driven by the question format and not by the question content, it is important to account for them because they complicate the interpretation and affect the commensurability of ratings across respondents. For example, does a ``strongly agree'' response signal an extreme attitude or a person's RS? In this discussion round, I will argue that the recently proposed class of item-response-tree models provides a flexible framework for understanding how response styles may affect answers to attitudinal questions. Facilitating the disassociation of response styles and attitudinal traits, item-response-tree models can provide powerful process tests of how different response formats may affect the measurement of substantive traits. I will review the results on a recently conducted empirical study which utilized three distinct item formats to measure the same personality construct. An analysis of the data by the Graded Response model yielded similar factorial solutions of the personality construct across formats. However, when applying IR-tree models tailored to the different response formats, I find that two of the three response conditions give rise to substantial RS effects. The results of this study demonstrate that item-response-tree models provide a valuable and easy-to-implement addition to the toolbox of attitude and personality researchers since these models offer new ways to test and investigate experimentally whether response formats affect the measurement of latent traits.

Author

• Ulf Bockenholt, Northwestern University