Multidimensional Scaling of Sorting Data: A Comparison of Three Procedures

Dissimilarity measures (D) derived from sortings of stimuli can be submitted to multidimensional scaling (MDS) either directly, or after transforming them to profile distances (Δ) computed on the rows of the D matrix. The latter procedure was criticized by Drasgow and Jones (1979) who performed two simulation studies, which are criticized here in turn. In the present article two sets of real sorting data were used for comparing the results of MDS on D and Δ, both with each other and with the results of two other procedures: multiple correspondence analysis (by means of HOMALS) on the raw sorting data, and MDS on the pairwise similarity ratings of the same stimuli by the same subjects. The three procedures were compared both with respect to the final configurations and with regard to the fit of the corresponding distances to the data. These comparisons suggested that MDS on D is slightly superior to MDS on Δ. The latter analysis, however, yields results that are similar to those of the much more efficient HOMALS program. The differences, however, are on the average very small.