2021 Volume 48 Issue 2 Pages 53-68
Paired-comparison format questionnaires that measures two latent traits suffer from the problem of unidentifiability in the Thurstonian item response model. This study proposes a new method that enables us to identify this model. Specifically, we pro- pose to use every item twice in the questionnaire. In simulation studies, our proposed method recovered the respondent parameters at least as well as the control method, which required analysts to fix some of the parameters to their true values. In the real data analysis, the proposed method was applied to data obtained by two types of psychological scales, one of which contained only positively keyed items, whereas the other contained both positively and negatively keyed items. Item parameters and trait scores estimated using the former data resulted in large posterior standard deviations. In contrast, those estimated with the latter data exhibited better estimation proper- ties. Although it generally does not allow for precise measurement, we show that the proposed method is applicable to empirical analysis. It achieved progress, in that, it negates the need for the unrealistic assumption of fixing parameters to their true values.
