Multidimensional scaling was developed by psychometricians, namely R. N. Shepard (1962) and J. B. Kruskal (1964). Its purpose is to deduce indirectly the dimensions a respondent uses to evaluate alternatives. The reason for using the indirect approach is that, in many cases, the attributes may be unknown and respondents may be unable or unwilling to represent their reasons accurately. As already mentioned, multidimensional scaling requires an object-by-object similarity matrix as an input.
Initially popularized, however, multidimensional scaling relies on judged similarity. That is, respondents indicate how similar pairs of objects are directly rated (e.g. on a 1–10 scale). This can be a burdensome task since for p objects p(p-1)/2judgments are needed. Still, the use of similarity judgments is relatively easy for respondents, especially when they cannot or do not want to reveal the basis for their opinion.
The results of multidimensional scaling depend on (a) the sample chosen to judge similarity and (b) the objects whose similarity is judged and the quality of input data. Multidimensional scaling derives dimensions that appear to be used by those rating a particular set of objects.
The basic type of multidimensional scaling involves deducing graphical models of alternatives (e.g. brands) alone (simple space) from similarity data.
Some early applications of multidimensional scaling accepted apparent dimensions as “truth” without question or validation, which often proved to be disastrous. It is advisable to use multidimensional scaling as a generator of hypotheses rather than as a final model of the market. Any important result should be confirmed on a separate sample with a separate method, such as direct questioning, before the results are given too much credence.
Multidimensional scaling generates a configuration in which the relative positions of the brands are unique. The picture can be changed by several operations without changing the relationship among the interpoint distance in some of the algorithms (assuming the Euclidean distance is used, which it almost always is).
A major problem in data collection is the burden on respondents as the number of alternatives increases (e.g. 20 alternatives require 190 pairs). However, if respondents are “homogeneous”, it is possible to have different subjects rate a different pair.