MCA is a multivariate factor analytical method that was designed as an extension of (simple) correspondence analysis (Lebart et al ., 2006). Its use in the framework of free sorting task dates back to the early 1990s (Van der Kloot and Van Herk, 1991). MCA can be regarded as a principal components analysis on categorical data (Greenacre, 1993). With free sorting data, we consider that there are as many categorical variables as subjects where the categories for each variable (subject) are the groups formed by the subject under consideration. Thereafter, the data from each subject are expressed as a matrix of indicator variables (see Fig. 7.1d). These individual matrices are horizontally merged to form a super-matrix formed of dummy variables.As stated above, column j (say) is associated with a category, that is, a group of products associated with a given subject. The average p of this column reflects the proportion of products contained in the group under consideration. Subsequently, column j is standardized by dividing it by p p j j . This kind of standardization is very typical with correspondence analysis, and is backed up by several considerations, the discussion of which is beyondthe scope of this paper.As a final step, principal components analysis is performed on the standardized super-matrix ofdummy variables. It is worth noting that Takane (1981) proposed a method of analysis called MDSORT for analyzing sorting data. The rationale behind this method is very appealing. Suppose that we seek an axis to depict the products. Naturally, we expect that, for a given subject, products that are in the same group should be very close to each other, whereas products in different groups should be far removed from each other. In other words, the representation axis should discriminate as much as possible the groups givenby the subject under consideration. At the panel level, we may seek a representation axis that, on average, discriminates as much as possible the groups given by the subjects. Subsequent axesmay be sought following the same strategy of analysis after imposing orthogonality constraints between the successive axes. From the derivation of the solution to this discrimination problem, it turns out that, as a matter of fact, we are led to the same solution as MCA. This remark wasalso stressed by Van der Kloot and Van Herk (1991), who stated that their program for running MCA gave outcomes that are identical to those of MDSORT. Takane (1982) designed yet anotherprocedure, called IDSORT, for analyzing sorting data. IDSORT can be seen as a refinement over MDSORT since it makes it possible to take account of individual differences among the subjects. This is done by adopting a strategy of analysis that combines MCA and INDSCAL methods. Thebottom line is that IDSORT, similarly to INDSCAL, yields a representation space for the products and a set of weights associated with the subjects that reflect the importance they attach to the various dimensions of this representation space. In the same vein, Qannari et al.(2009) proposed a method of analysis called SORT-CC.
đang được dịch, vui lòng đợi..
