Latent trait models and dichotomization of graded responses (Q1091736)

The polychotomous Rasch model has been introduced in 1961 [\textit{G. Rasch}, Proc. 4th Berkeley Sympos. math. Statist. Probab. 4, 321-333 (1961; Zbl 0107.368)]. Since that time many researchers have contributed to the theory of analysis of data where the response consists of two or more alternatives or options. This paper discusses the compatibility of the Rasch model with dichotomization ''before the fact'' (suitable response format) or ''after the fact'' (collecting response data). The main result states that the unidimensional polychotomous Rasch model is not \(\xi\)-invariant under the joining assumption. However, \textit{F. Samejima}'s homogeneous logistic model [Estimation of latent ability using a response pattern of graded responses. Psychometrika 34 (1969) is \(\xi\)-invariant. At the end of the paper the \textit{R. Likert} scoring technique [Archives of psychology 140, 1-59 (1932)] is discussed.