A Scaled Threshold Model for Measuring Extreme Response Style. (February 2020)

• Record Type: Journal Article
• Authors: Lubbe, Dirk; Schuster, Christof
• Is Part Of: Journal of educational and behavioral statistics. Volume 45:Number 1(2020)
• Page Start: 86
• Page End: 107
• Publication Date: 2020-02
• Subjects: extreme response style -- item response -- latent trait -- graded responses -- scaled thresholds; Educational statistics -- Periodicals; Social sciences -- Statistical methods -- Periodicals; 370.2
• DOI: 10.3102/1076998619859541
• Languages: English
• ISSNs: 1076-9986

Abstract

Extreme response style is the tendency of individuals to prefer the extreme categories of a rating scale irrespective of item content. It has been shown repeatedly that individual response style differences affect the reliability and validity of item responses and should, therefore, be considered carefully. To account for extreme response style (ERS) in ordered categorical item responses, it has been proposed to model responder-specific sets of category thresholds in connection with established polytomous item response models. An elegant approach to achieve this is to introduce a responder-specific scaling factor that modifies intervals between thresholds. By individually expanding or contracting intervals between thresholds, preferences for selecting either the outer or inner response categories can be modeled. However, for a responder-specific scaling factor to appropriately account for ERS, there are two important aspects that have not been considered previously and which, if ignored, will lead to questionable model properties. Specifically, the centering of threshold parameters and the type of category probability logit need to be considered carefully. In the present article, a scaled threshold model is proposed, which accounts for these considerations. Instructions on model fitting are given together with SAS PROC NLMIXED program code, and the model's application and interpretation is demonstrated using simulation studies and two empirical examples.