MULTIDIMENSIONAL SCALING ACCORDING TO ... - Science Direct

4 downloads 0 Views 318KB Size Report
BO EKEHAMMAR. Psychological Laboratories, University of Stockholm, Sweden. ABSTRACT. The purpose of this study was to compare the outcomes of some ...
36 (1972) 79-84;

Acta Psychologica

Not to be reproduced

MULTIDIMENSIONAL DIFFERENT FOR

0

in any form without

North-Holland written permission

SCALING

Company

from the publisher

ACCORDING

VECTOR

SUBJECTIVE

Publishing

TO

MODELS

SIMlLARITY

*

BO EKEHAMMAR Psychological

Laboratories,

University

of Stockholm,

Sweden

ABSTRACT The purpose of this study was to compare the outcomes of some multidimensional scaling methods, which could be derived from four different vector models for subjective similarity. The analysis was made on both theoretical and empirical material. The differences in the results between the various methods could be regarded as negligible in practical contexts.

Multidimensional scaling may be viewed as a summarizing term for different techniques aiming at revealing the underlying dimensionality of a perceptual, emotional or other subjective nature (EKMAN, 1970). Dimension analyses are often based on measures of dissimilarity, proximity or distance, which directly or indirectly, are obtained from an individual’s responses to stimuli. Perhaps the most usual technique is to allow the individual to perform some kind of similarity judgment of stimuli. The different types of such scaling methods may be regarded as being based on different kinds of models for subjective similarity or proximity. One group of such models has been called ‘content models’ by EKMAN and SJ~BERC (1965), since similarity in these cases is regarded as the degree of common content (‘communality’) in relation to total content (‘totality’) for the percepts compared. By describing such a model in vector terms (see e.g. EKMAN, 1963), at least four different geometrical models can be obtained by defining communality and totality differently (EKEHAMMAR, 1972). Of these four theoretical cases, in the following called cases I-IV, case I was reported first by EKMAN and LINDMAN (1961), and case II by EKMAN et al. (1964). If only qualitative variation is assumed to be present between stimuli, i.e. the stimuli are experienced * The study was supported by a research grant to Professor D. Magnusson from the Swedish Council for Social Science Research. 79

80

B. EKEHAMMAR

as being equally intensive, the similarity equations for the four theoretical models mentioned may be expressed mathematically as: Sij=COS

cos

sijx $!lij

$qj

(case 11)

($Qj/2) ’

COS

Sij=COS

(case I)

y?ij,

COS

Sfj=COS

($Oij/2),

(case ru) (case Iv)

fjlij,

where sij is the similarity estimate and pii is the angle between stimulus vectors i and j. It will be observed that case I and case IV give mathematically identical expressions, which, in order of size, are between the values obtained according to case II and case III respectively. To obtain methods as a basis for multidimensional scaling, the above equations may be solved for cos cpij. By these formulae, the estimates of similarity can then be transformed into cosines, which may be treated with conventional factor analytical methods. The purpose of this study was to compare the outcomes of the different scaling methods that can be derived from the theoretical similarity equations mentioned. Since everyone of these equations had some empirical support for at least some stimulus material (EKMAN and LINDMAN, 1961; EKMAN et al., 1964; EKEHAMMAR, 1972) it may be of practical interest to study differences in results obtained by the different methods. This analysis was made theoretically by comparing the different similarity functions, and empirically from the different

methods

by comparing the factor structures

for the same data.

THEORETICAL ANALYSIS Of the four theoretical

similarity equations for cases I-IV, there are

two (cases I and III) in which cos vij # Sij. Solving the similarity equation in case II for cos qii gives, according COS fJJfj=

By solving equation

the similarity

is obtained

to EKMAN (1970):

‘2 (Sij +

equation

j/S$-

si.j’).

(1)

in case III for cos qai a third-degree

: COS3

$?ij+

COS’

paj=2Sij2,

(2)

81

MULTIDIMENSIONAL SCALING

which was solved iteratively for different values of sii by the NewtonRaphson method (see e.g. HILDEBRAND, 1956, p. 447). To demonstrate the differences between the methods, cos qq =.f(s~) is reported graphically for each case in fig. 1.

0.6 -

0.6 -

0.0 Fig. 1.

0.2

0.4

0.6

0.6

1.0

sij

Cosine values (cos q 1.0 were rotated to simple structure according to the Varimax principle. The resulting factor matrices according to case II and case III are reported in table 2. An inspection of the factor matrices shows that the same structure is present, but with higher factor loadings for case III, since the analysis there is based on systematically higher cosines. The maximum difference between corresponding factor loadings for this material was only 0.08. To obtain a quantitative expression for similarity in factor structure between the various methods, coefficients of congruence were calculated according to Tucker (HARMAN, 1960, p. 257) for corresponding factors. The results are given in table 3. The values in table 3 indicate almost complete factorial agreement.

MULTIDIMENSIONAL

SCALING

83

TABLE2 Rotated factor matrix based on cos q,tj according to formula (l), case II, and formula (2), case III (within parentheses) for the same empirical material.

Stimulus

Factor B

A

I II III IV V VI VII VIII IX X

0.65 0.24 0.05 0.82 0.77 0.70 0.08 0.24 0.14 0.05

(0.73) (0.30) (0.10) (0.84) (0.82) (0.73) (0.14) (0.30) (0.21) (0.10)

0.34 0.68 0.85 0.15 0.06 0.00 0.72 0.17 0.14 0.16

(0.39) (0.74) (0.87) (0.21) (0.10) (0.05) (0.74) (0.24) (0.20) (0.21)

C - 0.06 0.08 0.16 0.24 0.10 0.37 0.29 0.72 0.75 0.64

(0.00) (0.14) (0.23) (0.30) (0.17) (0.46) (0.36) (0.73) (0.78) (0.72)

TABLE3 Similarity in factor structure, calculated as coefficients of congruence for corresponding factors, obtained according to cases I-IV for the same empirical material.

Factor

Case I (IV) vs. case II

Case I (IV) vs. case III

Case II vs. case III

I II III

0.9993 0.9989 0.9986

0.9991 0.9990 0.9989

0.9969 0.9960 0.9950

COMMENTS

Both the theoretical and empirical analyses show that the multidimensional scaling methods, based on different theoretical vector models for subjective similarity, give only negligible differences in outcomes for the same data. Provided that one or another of the similarity models described (cases I-IV) gives a satisfactory approximation of the similarity mechanism for a certain stimulus material, this implies that it is not very important to know which of the models that best describes the similarity mechanism. This is of interest for those who wish to apply the method reported. Without any great inconvenience, the mathematically simplest principle, cos 976,= sij, can be used generally, i.e. the dimension analysis can be made directly on similarity data (possibly after linear trans-

B. EKEHAMMAR

84 formation

to values between 0 and 1). The last-named mode of procedure

has also been

used in some

later investigations

EKMAN, 1969 ; EKEHAMMAR, 1971;

(e.g.

BRATFISCH and

MAGNUSSON, 197 I ; MAGNUSSON

and EKEHAMMAR, 1972; MAGNUSSON and EKMAN, 1970), in which also intuitively

meaningful

results were obtained. (Accepted

December

14, I 97 1.)

REFERENCES BRATFISCH,0. and G. EKMAN, 1969. Subjective and objective Psychological

Reports

EKEHAMMAR,B., 1971. A psychophysical ceptions of Rorschach

intelligence factors.

25, 607-620. approach

cards. Perceptaal

to the study of individuals’perand Motor

Skills 33, 951-965.

EKEHAMMAR,B., 1972. A comparative study of some multidimensional for subjective

similarity.

Scandinavian

EKMAN, G., 1963. A direct method for multidimensional 28, 33-41. EKMAN, G., 1970. Comparative ques.

Reports

Stockholm,

from

vector models (in press).

Joournal of Psychology

studies OH multidimensional the Psychological

ratio scaling. Psychometrika scaling and related

Laboratories,

The

techni-

University

of

Supplement 3.

EKMAN,G., T. ENGEN,T. K~~NNAPAS and R. LINDMAN,1964. A quantitative principle of qualitative similarity. Journal of Experimental Psychology 68, 530-536. EKMAN, G. and R. LINDMAN, 1961. Multidimensional ratio scaling and multidittlerzsiorral simikarity. Stockholm,

Reports from the Psychological

Laboratories,

The University of

No. 103.

EKMAN, G. and L. SJBBERG,1965. Scaling. Atmual Review of Psychology HARMAN, H. H., 1961. Modern factor

analysis.

Chicago:

HILDEBRAND,F. B., 1956. Introduction

to numericalanalysis.

16, 45 l-474.

University of Chicago Press. New York: McGraw-Hill.

MAGNUSSON,D., 1971. An analysis of situational dimensions. Perceptual SkiNs 32, 851-867. MAGNUSSON,D. and B. EKEHAMMAR,1972. An analysis replication.

Reports from the Psychological

of Stockholm,

of situational

Laboratories,

and Motor

dimensions.

A

The University

no. 344.

MAGNUSSON,D. and G. EKMAN, 1970. A psychophysical approach to the study of personality traits. Multivariate Behavioral Research 5, 255-274.

Suggest Documents