台大心理系

 In the field of social science, accurately measuring individuals' preferences among alternatives is essential, with two prominent measurement formats—Likert Scale (LS) and Comparative Judgment (CJ)—often employed for this purpose. While LS is widely used because it is well-known and easy to apply, CJ offers a more direct measurement of preferences (Cheng et al., 2021). However, CJ poses challenges due to its ipsative nature, which can hinder the interpretability and comparability of estimated participant scores. Böckenholt (2004) and subsequent researchers (e.g., Xiao et al., 2017) proposed models to combine data acquired via LS and CJ, leveraging the strengths of both measurement methods. Despite these significant advances, the “joint Thurstonian models” prior researchers applied neither could adequately model the data acquired through ordinal scales nor account for the heterogeneity of item parameters and thresholds, and were insufficient for detecting latent differences between LS and CJ data. To address these limitations, we proposed the Joint Thurstonian Generalizability Theory (JT-GT) and examined the new JT-GT models in terms of their applicability and validity in analyzing a data set of academic attribution collected in both LS and CJ formats.
College students reported on their multiple causal ascriptions after a real-life experience of a major academic success/failure (i.e., grades received from a midterm: nsucceeded = 336; nfailed = 282). In addition to the commonly assessed academic attributions of effort and ability, this study also included other attributions, namely, task difficulty, luck, and mood. We then used different mathematical models to generate indicators from LS data alone, CJ data alone, or combined LS and CJ data. Subsequently, we compared the predictability of these different indicators on students' academic adjustment, such as their goal orientations, procrastination, and implicit theory of intelligence.
In the success condition, the Böckenholt model achieved the lowest Bayesian Information Criterion (BIC), suggesting that integrating LS and CJ data within a joint Thurstonian framework is appropriate. Moreover, the predictive validities varied across indicators based on the criteria of academic adjustment. For example, students with higher latent ability attribution scores in the Böckenholt scale, β = 0.24, and those identifing mood on the Likert scale as the primary attribution of their academic success, B = 0.86, were more inclined to endorse an ability goal orientation before the exam, R² = 0.084, adjusted R² = 0.079. Taken together, this study suggests that, at least in the context of academic success, compared with single-format scores, indicators revealed by combining LS and CJ formats within the joint Thurstonian framework yielded incremental validity showing unique predictive power for students’ academic adjustment.