An evaluation of the validity of multidimensional scaling methods : for the representation of cognitive processes (1980)
AuthorsFraser, Christopher Oshow all
This study is an evaluation of the issues involved in providing a meaningful psychological interpretation of multidimensional scaling solutions, for example to regard them as valid representations of the cognitive processes involved in generating the data.
The various metatheoretic approaches that have been developed to define appropriate procedures for the quantification of psychological attributes are discussed and evaluated. It is argued that much current psychological research is based on an inappropriate paradigm. In particular it is argued that the emphasis on magnitude estimation to generate psychological data is misplaced and scales derived from weak-ordered judgements are much to be preferred.
Extending these arguments to the multidimensional case, it is argued that most applications of multivariate methods in psychology have shown insufficient recognition of the theoretical implications of using a particular technique. The application of any method of data analysis such as multidimensional scaling is only appropriate if it can be shown that the assumptions implicit in the scaling model are satisfied for that set of empirical data. In addition some variations in the scaling model, such as subjective metrics scaling, involve additional assumptions which need to be explicitly formulated and tested.
These metatheoretic limitations, as well as evidence on the frequent occurrence of violations of its basic assumptions suggest that multidimensional scaling configurations can at best be attributed with only a limited degree of psychological significance. It is suggested that such value as it does possess is limited to the evaluation of non-dimensional structural hypotheses derived from some prior substantive theory.
An empirical example is presented demonstrating that even when there appears an obvious and intuitive interpretation of the dimensions of a MDS configuration, the solution may be completely inappropriate as a model of the underlying cognitive processes. A second example however describes a more appropriate and successful application of multidimensional scaling methodology. A theoretical interpretation of emotion labelling based on Guttmans (1957 ) facet theory, was shown to be substantially confirmed in the structure of a MDS configuration.