UNIVERSITY OF ILLINOIS LIBRARY AT URBANA-CHAMPAIGN BOOKSTACKS Digitized by the Internet Archive in 2012 with funding from University of Illinois Urbana-Champaign http://www.archive.org/details/bimultivariatere138joha Faculty Working Papers BIMULTIVARIATE REDUNDANCY MAXIMIZATION Johny K. Johansson and R. Narayan 0138 College of Commerce and Business Administration University of Illinois at Urbana-Champaign r FACULTY WORKING PAPERS College of Commerce and Business Administration University of Illinois at Urbana-Champaign December 19, 1973 BIMULTIVARIATE REDUNDANCY MAXIMIZATION Johny K. Johansson and R. Narayan #138 . BIMULTIVARIATE REDUNDANCY MAXIMIZATION 1 by Johny K. Johansson and R. Narayan University of Illinois Introduction The relationship between two sets of variables is often analyzed with the help of canonical correlation techniques. The interpreta- tion problems are often severe, however, as soon as more than one pair of variates are significant at the pre-selected level. Stewart and Love (1968) have suggested a method, called "redundancy analysis", to deal with these problems. Miller and Farr (1971) pointed out that the redundancy measure would remain invariant with respect to any ortho- gonal rotation of the complete set of canonical variates, and that, consequently, canonical correlation was only one special case of a general redundancy analysis. It can be argued that since the redundancy measure provides a straightforward interpretation of the degree to which two sets of variables covary, one focus of bimultivariate analysis ought to be the maximization of the redundancy contribution from as small a set of variates as possible. As a start in that direction, this paper presents an approach to the maximization of the partial redundancy attributable to the first pair of variates. 'Several persons contributed to this paper. Jagdish Sheth encouraged us to focus upon the problem and gave valuable feedback throughout; Charles Lewis gave exceedingly helpful assistance on the theory part; and Joseph Kolman and Maurice Tatsuoka contributed many valuable ideas to the testing of the optimizing approach. Funds were made available by the University of Illinois Computer Services Office and the Bureau of Economic and Business Research. The authors want to thank the people involved but also absolve them of the responsibility for any remaining errors. The Theory Miller and Farr (1971) show that the redundancy attributable to the first linear combination of the Y variables Gi i? equal to 2 RED, = (L 6 /tr(R YY ))*(5 r GF.) where L« stands for the sum o1 of :he Y variables upon 6j > tr (Ryy) stands for the trace of the correlation matrix of the Y's (equal to the number of criterion variables), and the F. , i=l,..., I 2 stand for the successive orthogonal factors of the X variables. Because these orthogonal factors together span the space of the X's completely, we have I 2 * 2 so that the redundancy becomes a product of the loadings and the squared multiple correlation: RED Gi . (L Gi /tr(R YY ))*R 2 SiX The canonical correlation technique maximizes the latter component of the product, whereas a principal component analysis of the Y's would maximize the loading part. In general, then, neither of these two approaches would maximize the redundancy measure. 2 In what follows, we will treat the Y's as "dependent" and the X's as "independent" -- an inverse relationship is dealt with similarly but yields no new insights so is ignored here. Also, in what follows, the redundancy measure will always refer to the first linear combination of Y's unless otherwise stated. Finally, both the Y's and the X's are assumed standardized. To derive an expression of the redundancy measure — our objec- tive function -- in the original variables Y and X we proceed as follows. Let W r denote the m by 1 vector. of variable weights for 1 the first linear combination G-,. Then we. have Y W^ = B, , with the dimension of the Y matrix equal to n by m, n denoting the number of observations, m the number of Y's. Then for the loadings we have R YG 1 : R YY W G ] R again denoting the correlation matrix. Since we want the squared loadings we need *il } \ = w g{ r yy r yy w g 1 - the T superscript indicating transpose. We also note for future use that tr(Ryy) - m . As for the squared multiple correlation, we have first XB * G 1 , as the predicted value of G , with B denoting the I by 1 vector, of '.parar meter weights. Using a least squares fit, we compute B as B = R XX R XG 1 4 To get a measure of the squared simple correlation between the «. actual and the predicted G^s -- which is the squared multiple correla- tion we are looking for — we compute first V G 1 G ] = B R XX B K BR x Gl • where V stands for the variance. Then we get VG 1 G 1 (R" 1 R W ) T R IL - -xx xy g/ xy g 1 " w gJ r yy w Gi B "g^xy r xx r xy w g 1 W G YY G 1 1 for the correlation between the predicted and actual G's. The complete objective function cag then be written RED G ■ ( W 4 (R V V )2W S1 >< W gJ R XY R XX R XY % > "s,"rY \ ' - which is to be maximized under a normalizing constraint such as W r W r = 1 , G l G l • or W as • o IN 4 IN • rt IN • ft IN o • rt IN rt • .» IN >0 • rt IN o • rt o rt • IN e " O ■ •8 n • n •« lA ■ rt • ft IN • rt CN m «N • rt M CO • CN IN e • IN IN f e .s e •© • rt e .8 e • to o n rt in • IA IN M rt • IN M rt IN rt • e M 10 M rt • IN rt M • in pd m • © e o • " o •z .5 e NT e rt •8 O • M M • IN M M • NT •N ■0 • in rt rt • n M B • <~ • IN ia • IN r* rt • rt e m *8 art *s in • •! e • rt o •> •8 M rt • •* M .5! d> • w-i N • ia in 4 • ■» IN • o IN 1 e .3 o o • IN o do •8 *> • o o .3 IA •IN e IN N • rt M ■a • rt in • rt id IS • ia •s % in . rt IN IN • O IN 1 0* • OS c 0- • rt o N> • r* e •9 • O o '8 • —* e rt • m in • rt M i* • rt IN lA • rt in •a • •» N *p • « IN • — 1 • ft rt • o • ft o ■» • IN o rt ' S 00 •8 '8 ■0 • ia M • rt r* <• • rt •1 in M IN • IN IN IA -* 1 m «A ' 3 •a *3 N • IN o IN •9 '8 References Cooley, W. W. and P. R. Lohnes, Multivariate Data Analysis , New York: Wiley, 1971. Hiramelblau, D. M., Applied Non-Linear Programming , New York: McGraw- Hill, 1972. Miller, J. K. and S. D. Farr, "Bimultivariate Redundancy: A Comprehen- sive Measure of Interbattery Relationship", Multivariate Behavioral Research , July 1971. Stewart, D. and W. Love, "A General Canonical Correlation Index", Psycho- logical Bulletin , September, 1968.