Comparison Canonical Correlation Interbattery Analysis ...291 Comparison of Canonical Correlation...

291

Comparison of Canonical Correlationand Interbattery Factor Analysison Sensation Seeking and Drug Use DomainsG. J. Huba, M. D. Newcomb, and P. M. Bentler

University of California, Los Angeles

The relationships between different types of sen-sation-seeking tendencies and the use of 26 sub-stances are studied for a group of 1,068 adoles-cents. The methods of canonical correlation analy-sis with dimension rotation and maximum likeli-hood interbattery factor analysis are contrasted inthe data set. Several major patterns are found, andit is concluded that the relationship between druguse and sensation-seeking tendencies is not a gen-eral one.

In recent years there has been a resurgence ofinterest in different ways to explore predictor-criterion relationships using factor analysis ofonly those dimensions that are common to pairsof domains or by improved interpretability of ca-nonical correlation solutions (Bentler & Huba,in press; Browne, 1979; Cliff & Krus, 1976;Skinner, 1977, 1978; Tucker, 1958). Two mathe-matically similar but philosophically differentmodels underlie the bulk of the methodologicalsuggestions. The manifest variable alternative,canonical correlation analysis, is a method for

determining linear combinations of the originalvariables under the constraint that compositesformed in the two domains be mutually or-

thogonal within domains and maximally cor-related with corresponding combinations in theother domains. Skinner (1977, 1978) has sug-

APPLIED PSYCHOLOGICAL MEASUREMENTVol. 5, No. 3, Summer 1981, pp. 291-306@ Copyright 1981 Applied Psychological Measurement Inc.0146-6216/81/030291-16$1.80

gested several ways of altering the componentsso as to rotate dimensions for maximal corre-

spondence to original variables, whereas Cliffand Krus (1976), as well as Bentler and Huba (inpress), considered ways to rotate the derived ca-nonical dimensions to positions where the pat-tern of loadings (or weights) meets some &dquo;simplestructure&dquo; or simplicity criterion. Thus, the firstapproach to the problem of structuring the co-variation in common between domains involvesthe use of canonical correlation analysis witheither a correspondence or simplicity rotation toimprove interpretability. Advantages of the ca-nonical correlation approach are the relative in-expensiveness of the procedure, the wide circula-tion of efficient computer programs, the de-

terminacy of scores, the transformation-freenessof the solution, and the large-sample test for thenumber of significant dimensions (Mardia,Kent, & Bibby, 1979).

Several decades ago, Tucker (1958) proposedthe interbattery factor analysis model as a

means of determining those factors in commonto two domains of variables. At the time, Tuc-ker’s least squares solution to the problem ofcommunality estimation for factor analysis wasingenious; an explicit solution was available

when there was focus on the factors influencingboth domains. Recently, Browne (1979) showedthat maximum likelihood estimates of the factor

loadings in the interbattery factor analysismodel can be obtained by a column rescaling of

Downloaded from the Digital Conservancy at the University of Minnesota, http://purl.umn.edu/93227. May be reproduced with no cost by students and faculty for academic use. Non-academic reproduction

requires payment of royalties through the Copyright Clearance Center, http://www.copyright.com/

292

the canonical correlation loadings matrices.

Browne’s comprehensive study of the interbat-tery method showed that many advantages of ca-nonical correlation analysis could be transferredto the interbattery factor analysis model. First,since maximum likelihood interbattery factoranalysis loadings are a column rescaling of thecanonical correlation loadings, the wide avail-ability of computer programs for canonical cor-relation analysis (and loadings matrices) makesinterbattery factor analysis generally available.Second, the relationship to the scale-free canoni-cal solution by a column rescaling also estab-lishes the scale-freeness of the interbattery solu-tion. Third, as Browne has noted, the fact thatmaximum likelihood estimates have been ob-tained makes it possible to apply the result ofRao (1973)-that the standard tests for the num-ber of significant canonical correlations (and di-mensions) are also tests for the number of inter-battery factors. Tucker (1958) and Browne

(1979) have commented, however, on the rela-tively high indeterminacy of the interbattery fac-tor analysis model, and Browne’s estimates werederived under the usually tenable assumptionthat domains of variables are equally predictiveof the underlying interbattery factors.The major difference between canonical cor-

relation and interbattery factor analysis lies inthe focus upon correlations among sets of de-rived linear combinations versus loadings of theoriginal variables on inferred, &dquo;error-free,&dquo; la-tent constructs or factors. The canonical correla-tion model assesses the way in which weightedsums of the original variables interrelate. Thecorrelations across domains will not, in general,be perfect. The interbattery factor model, in

contrast, assumes that the same factors (i.e.,cross-set perfectly correlated factors) influencethe variables in the two domains. The interbat-

tery model determines how many such factorsthere are, as well as the pattern of correlationbetween variables in the two domains implied bythe interbattery factors.This paper compares the methods in two do-

mains that have recently generated theoretical

interest. Repeated research efforts have consis-tently shown that drug-taking behaviors are sig-nificantly related to sensation-seeking behaviors(Khavari, Mabry, & Humes, 1977; Segal, Huba,& Singer, 1980; Zuckerman, 1978, 1979; Zuc-kerman, Buchsbaum, & Murphy, 1980). In-

terestingly, the current theoretical debate aboutthe function of drug use in satisfying sensation-seeking needs (Segal et al., 1980), as opposed toan interpretation that sensation-seeking ten-

dencies and drug-taking proclivities are general-ly &dquo;caused&dquo; by the same biological processes(Zuckerman et al., 1980), tends to parallel thedifferent philûsûphit:~ underlying the canonicalcorrelation and interbattery factor analysis ap-proaches. Zuckerman et al. (1980) favors the&dquo;common cause&dquo; interpretation, whereas Segalet al. (1980) take a more dynamic approach toexplaining the relationships. The dynamictheoretical approach is more fully explicated inthe DOMAIN model of Huba and Bentler (inpress).

Method

Participants

The participants in this study were a group of1,068 individuals in an ongoing longitudinalstudy of adolescent growth (see Huba, Wingard,& Bentler, 1979, 1980). The numbers of adoles-cents of each gender by grade were: 153 tenth-grade boys, 130 eleventh-grade boys, 81 twelfth-grade boys, 275 tenth-grade girls, 248 eleventh-grade girls, and 181 twelfth-grade girls.

Drug Use Indicators

The 26 drug use measures were frequencies ofuse in the past 6 months. The responses for thedifferent drugs were rated on a 7-point anchoredscale (1 = never, 2 = once, 3 = a few times,4 = once a month, 5 = once a week, 6 = once a

day, 7 = more than once a day). The full list ofdrug use categories is presented in Table 1.



293

Sensation-Seeking Items

Participants also completed a brief question-naire about their sensation-seeking tendencies.Individuals were asked, &dquo;How often do you feelthe following way?&dquo; Abbreviated versions of the16 items are presented in Table 3. Participantsresponded to the items on an anchored 5-pointscale ( = never, 2 = rarely, 3 = sometimes,4 = usually, 5 = always).The 16 items were chosen as follows. Zucker-

man, Eysenck, and Eysenck (1978) present thederivation of the Sensation-Seeking Scales,Form V, which are 10-item, forced-choice in-

dices for Experience Seeking, Thrill and Adven-ture Seeking, Disinhibition, and Boredom Sus-ceptibility. In order to provide brief indices, fourpositive choice statements from each constructwere selected based on a high loading in Ameri-can and British samples and on judged rele-vance for adolescents. In the current question-naire, Items 1, 5, 9, and 13 were designed to bemeasures of Experience Seeking, Items 2, 6, 10,and 14 were measures of Thrill and Adventure

Seeking, Items 3, 7, 11, and 15 were hypothe-sized to be items assessing Disinhibition, andItems 4, 8, 12, and 16 were presumed to be mea-sures of Boredom Susceptibility. The value ofcoefficient alpha for Experience Seeking was.43, for Thrill and Adventure Seeking it was .70,for Disinhibition it was .68, and for BoredomSusceptibility it was .43. Since the internal con-sistency coefficients for the scales were only mar-ginally acceptable, analyses using both originalitems and derived scales are reported.

Data AnalysisFor the first two analyses, the four sensation-

seeking scale scores and the 26 drug use itemswere used in a canonical correlation analysiswith the sensation-seeking scales forming oneset and the drug use items forming the remain-ing set. The loadings (correlations of variableswith canonical variates) were input into an or-thogonal asymmetric rotation procedure called

two-matrix orthosim in the program developedby Huba, Palisoc, and Bentler (in press). The ro-tated canonical correlation loadings constitutethe first analysis. For the second analysis, theunrotated canonical correlation loadings wererescaled using Browne’s (1979) results so thatmaximum likelihood interbattery factor load-ings were obtained. The interbattery factor load-ings were also orthogonally rotated using theorthosim method in Huba, Palisoc, and Bent-ler’s (in press) program. For the third and fourthanalyses, the 16 original sensation-seeking itemsand the 26 drug use items were used in a canoni-cal correlation analysis with the loadings againrotated. An interbattery factor analysis solutionfor these data was also derived and orthogonallyrotated.

Canonical Correlation Rotation

In asymmetric orthogonal canonical correla-tion rotation, the two matrices of loadings yield-ed by canonical correlation analysis are trans-formed using a different orthonormal matrix foreach set of loadings. That is, let A be ap x r ma-trix of the loadings of p Set 1 variables on r ca-nonical variates. Let B be a q x r matrix of load-

ings of the q Set 2 variables on the r canonicalvariates. These two matrices of loadings are thecorrelations of the original variables with the ca-nonical variates or, alternately, the standardizedregression coefficients of the original variates onthe canonical variates. Let U be the r x r diago-nal matrix of the canonical correlations, i.e., thecorrelations between the derived variates for Set1 and the derived variates for Set 2. As shown byBentler and Huba (in press), the problem inasymmetric orthogonal canonical correlation ro-tation is to find the r x r orthonormal matricesTi and T2 so that the matrices C and D, as de-fined in Equations 1 and 2, satisfy some overallsimplicity criterion that is a joint function of thetwo matrices. Bentler and Huba (in press) applyBentler’s (1977) orthosim criterion and call theprocedure that finds the matrices Tl and Tz theasymmetric two-matrix orthosim method.



294

As a point of clarification, it should be notedthat in this paper the canonical correlation load-

ings, rather than the weights, were rotated. Cor-responding to the loadings matrices A and B forthe two sets of variables are the matrices G and

H, which are p x r and q x r matrices, respec-tively, of the weights to be applied to the vari-ables in each set to obtain the standardized ca-nonical variates. The loadings matrices A and Bare reiated to the weights matrices and the ma-trices RXx and R,,Y of the correlations among thevariables in the two sets through the relation-ships depicted in Equations 3 and 4.

It is important to realize that simplicity is beingsought in the rotated loadings matrices AT, andBT~ rather than in the rotated weights matricesGT, and HT2. Reasons for maximizing the sim-plicity criterion as a function of the loadingsrather than of the weights have been consideredin detail by Bentler and Huba (in press). Con-ceptually, it should be remembered that the ca-nonical loadings matrices A and B are anal-ogous to factor loadings matrices, whereas theweights matrices G and H are analogous to fac-tor score coefficient matrices. (See Timm, 1975,for a discussion of weights and loadings ma-trices.)After rotating the matrices of canonical vari-

ates by the transformation matrices, it is neces-sary to determine the correlations among the ro-tated variates in the two sets. It can be easilyverified that the canonical correlations after ro-

tation are defined by Equation 5.

An orthogonal rotation that guarantees that V issymmetric can be developed by constraining Ti

to be equal to T,. Such solutions were con-sidered by Cliff and Krus (1976) and are calledsymmetric rotations by Bentler and Huba (inpress). Different weighting systems for de-

termining the common T matrix under this con-straint have been considered by Bentler and Hu-ba. This paper used the more general case thatdoes not constrain the two transformation ma-trices to be equal.Some examples of canonical correlation rota-

tion that additionally utilize &dquo;split samples&dquo; toassess robustness are given by Huba, Wingard,and Bentler (1979, 1980), Segal et al. (1980,chap. 8). and Wingard, Huba, and Bentler

(1980). The rotations can be made simply andinexpensively (about 1 second for those reportedhere) and in the Fortran program of Huba, Pali-soc, and Bentler (in press).

Maximum Likelihood Interbattery FactorAnalysis

Browne (1979) showed that canonical correla-tion loadings and maximum likelihood esti-mates of the factor loadings in the interbatteryfactor analysis model developed by Tucker

(1958) are related by a rescaling when an identi-fication condition is imposed on the interbatterysolution. Let E be the (p + q) x r supermatrix ofunrotated canonical variate loadings formed byconcatenating the two matrices (A, B) of canoni-cal loadings for the variable sets, i.e., as in

Equation 6.

Browne (1979, p. 82) then showed that the

(p + q) x r matrix F of maximum likelihoodloadings of the total set of variables on r inter-battery factors is derivable from E and the ca-nonical correlations U by the relationship shownin Equation 7.

That is, the columns of E, the canonical variateloadings, are rescaled by the square root of thecanonical correlations to yield interbattery fac-tor loadings.



295

In rotating the loadings, Browne (1979) re-

stricted his attention to the case of a common

transformation (T) matrix for F1 = AU ~, andF~ = BU 5. By introducing the column rescalingfirst and only then rotating by a common trans-formation matrix, Browne noted that &dquo;we have

restricted our attention to those sets of inter-bat-

tery factors which are equally related to the twosets of observed variables.&dquo; (1979, p. 83)

Comparison of Canonical Rotation and Inter-battery Solutions

As sets of mathematical equations, the canoni-cal correlation and interbattery solutions are re-lated to one another before rotation through thecolumn rescaling. Since the same constant is

used to rescale corresponding columns of the ca-nonical loadings matrices A and B, canonicalloadings that are large in absolute value will berelatively large in absolute value in the interbat-tery solution. Thus, dimensions will generally be&dquo;labeled&dquo; or &dquo;interpreted&dquo; in the same manner.One point of difference for the unrotated solu-tions, which is more apparent than real, is thatthe smaller interbattery factors will have theirloadings most reduced in magnitude from thecanonical loading matrices because the rescalingwill be by the square root of a small canonicalcorrelation. It should be noted, however, that a

large loading and a small canonical correlationare being &dquo;traded&dquo; for a small loading and a&dquo;canonical&dquo; correlation of unity. In the rotatedsolutions, the canonical correlation and inter-battery solutions will generally be most similar ifthe two transformation matrices (Ti, T2) em-ployed in the canonical rotation are approxi-mately the same and the canonical correlationshave similar values.In summary, there are two major differences

between canonical correlation analysis and in-terbattery factor analysis. The canonical correla-tion analysis forms linear composites of ob-

served or manifest variables, since the underly-ing model does not hypothesize measurement er-ror in the variables. In contrast, interbattery fac-tor analysis is a true score or latent variable

method that explicitly defines variables with er-ror. Second, and perhaps most critically, canoni-cal correlation analysis forms composites thatare only imperfectly related across sets; the ca-nonical correlations are a measure of the magni-tude of association between variates formed inthe two domains under study. Thus, compositesin one set may be highly correlated with com-posites in the other set but will only be perfectlycorrelated when there are one or more variablesin the second domain that are exact linear com-

binations of one or more variables in the first do-main. This situation almost never occurs in

practice. On the other hand, in interbattery fac-tor analysis, dimensions are generated that arepresumed to be common &dquo;causal&dquo; elements forvariables in the two different domains (in thefactor analytic sense). The corresponding di-

mensions in the two domains are thus perfectlycorrelated, since they are generated under theconstraint that the same factors operate in thetwo sets of measures. Thus, a major part of thechoice between canonical correlation analysisand interbattery factor analysis lies in the inves-tigator’s intention. If the analyst wishes to de-termine the common latent variable sources ofvariance for a set of variables that is grouped in-to two domains, the interbattery model is prob-ably the more appropriate one. If, on the otherhand, the investigator wishes to choose a smallnumber of linear combinations of the originalvariables in each of the two sets in such a way asto maximize the correlations between the do-

mains, then the canonical correlation model is

probably more appropriate.

Results

In the canonical correlation analysis of thefour Sensation-Seeking Scale scores against the26 drug use items, it was found that there was asignificant relationship between the domains:x 2(104) = 490.86, p < .001. The size of the ca-

nonical correlations (.52, .27, .23, and .16 for thefour dimensions) seems to indicate that there isone major set of relationships spanning domainswith the possibility of several minor across-do-



296

main relationships. Application of the Bartlettstepdown method indicates that with two dimen-sions partialled, residual covariation among thedomains remained: X2 (48) = 84.04, p < .001;with three dimensions partialled, there was

no statistically reliable remaining covariation:X2(23) = 28.5, p = .20. Consequently, the hy-pothesis that there are three dimensions span-ning the domains could not be rejected while hy-pothesis of smaller dimensionalities were rejec-ted.

In the canonical correlation solution, it is pos-sible to determine (1) the multivariate associa-tion (or correlation) among the domains and (2)how well the variables in one domain predict thevariables in the other domain. To quantify theassociation, Cramer and Nicewander (1979) sug-gested six coefficients, of which two, ys and y,,seem to have reasonably good properties as in-dices of squared multivariate association. Forthese data, the square root of ys was .33, whilethe square root of y6 was .32. Either coefficientindicates moderate levels of association betweendomains. Stewart and Love (1968) have sug-

gested a directional coefficient of multivariatepredictability, the redundancy index. Assumingthree significant dimensions, the value of the re-dundancy index when predicting drug use fromthe Sensation-Seeking Scales was .0567, indicat-ing that about 6% of the variance in the drug usevariables is explained by that domain. Whenpredicting sensation seeking from drug use, thevalue of the redundancy index was .1285. Eitherway of quantifying multivariate association orpredictability allows the conclusion that the re-lationships among domains are moderate in

magnitude.Although three dimensions were statistically

necessary, it is possible that less than three maybe required for meaningful interpretation.Therefore, the one-dimensional solution was

examined, as well as solutions with two and

three rotated dimensions in each domain. Ca-nonical rotations were accomplished using theasymmetric two-matrix orthosim procedure.The loadings for the alternate solutions are

shown in Table 1, which also presents the rota-tion-adjusted canonical correlations.The three-dimensional solution seems inter-

pretable. It is interesting to note that the first di-mension in each of the three different solutionsis also about the same: This observation would

argue that the orthogonal rotation utilized doesnot destroy a large general relationship betweencertain aspects of drug use and the Sensation-Seeking Scales. The authors’ interpretation ofthe three-dimensional solution is as follows. Thefirst dimension in the sensation-seeking domainis almost synonymous with the Disinhibition

scale and is also slightly correlated with theBoredom Susceptibility and Thrill and Adven-ture-Seeking scales. The second dimension is al-most perfectly correlated with the ExperienceSeeking scale and is also quite highly correlatedwith the Boredom Susceptibility scale. The thirddimension is quite highly correlated with Thrilland Adventure Seeking.The first rotated dimension in the drug use do-

main has large loadings for patterns of drug use,which are perhaps indicative of early stages ofexperimentation with substances (see Huba,Wingard, & Bentler, 1981). The frequency of us-ing beer, wine, liquor, marijuana, and hashishis correlated highly with the dimension. The sec-ond dimension is correlated highly with a varietyof drugs, many which might be classified as&dquo;hard&dquo; substances. Variables that correlate atleast moderately with this dimension include

marijuana, hashish, coffee, barbiturates, seda-tives, amphetamines, other stimulants, LSD andother psychedelics, amyl nitrite, cocaine, andPCP. Finally, the third dimension has high cor-relations with the usage frequency of such sub-stances as cigarettes and marijuana as wellas barbiturates, sedatives, amphetamines, in-

halants, nonprescription cough medicines, andPCP.The pattern of adjusted canonical correlations

shown in Table 1 indicates that the first sensa-

tion-seeking dimension (Disinhibition) is cor-

related .41 with the first drug use dimension (al-cohol and marijuana use), .15 with the second



297



298

use dimension (polydrug use), and .21 with thethird drug use dimension (cigarettes and harddrugs). The second sensation-seeking dimension(Experience Seeking) is correlated at a nontriviallevel (.27) only with the second dimension in thedrug use domain. Finally, the third dimension ofsensation seeking (Thrill and Adventure Seek-ing) is positively correlated (.23) with the use ofalcohol and marijuana and negatively correlated(-.17) with a dimension indicative of the use ofcigarettes and hard drugs. It should be empha-sized that it is the nonzero, off-diagonal canoni-cal correlation that might be argued to providetheoretically important &dquo;dynamic&dquo; information.7 .... k,-..L Y ..- .... Ij ivtl

about how types of sensation seeking are relatedto more than one type of drug use.Corresponding to the different canonical solu-

tions is a maximum likelihood interbattery fac-tor analysis solution which, before rotation, isrelated to the canonical solution by a simplecolumn rescaling. Browne (1979) has stated thatRao (1973) showed the correspondence betweenthe Bartlett (or Lawley approximation) stepdowntest for the number of canonical correlations ex-

ceeding zero and a test for the number of inter-battery factors. Consequently, it may also beconcluded that three maximum likelihood inter-

battery factors are necessary. One, as well as twoand three, rotated interbattery factor solutionsare shown in Table 2.To maintain comparability with the previously

discussed canonical solution, attention will befocused on the three-factor solution loadings.The first factor has loadings for Disinhibitionand Thrill and Adventure Seeking in the sensa-tion-seeking domains as well as the usage fre-quency of beer, wine, liquor, marijuana, andhashish. The second factor has loadings for Ex-perience Seeking and Boredom Susceptibility inthe sensation-seeking domain as well as loadingsfor the usage frequency of marijuana, coffee,LSD, and cocaine. The third factor has a nega-tive loading for Thrill and Adventure Seeking aswell as positive loadings for cigarettes and mari-juana and a much smaller loading for inhalants.

In general, it would seem that many of thesame interpretations can be drawn from either

the canonical correlation or interbattery factoranalysis results. First, the large loadings in cor-responding places are generally about the same.The interbattery loadings are &dquo;smaller,&dquo; but thisis due primarily to the rescaling by the squareroot of the canonical correlations and not an in-teraction of the rescaling and transformationmethod. In general, the columns in the rotatedinterbattery and canonical solutions are propor-tional, with relatively large loadings remaininglarge. A seeming point of discrepancy betweenthe two analyses is that the first interbattery fac-tor has loadings for both Thrill and AdventureSeeking and Disinhibition, while these scales aresplit over two dimensions in the canonical re-sults. The difference can be reconciled by notingthat the interpretation of the first and third vari-ates in the canonical solution, coupled with theasymmetric nature of the adjusted canonical

correlations, is pointing toward the same rela-tionship.Information about the degree of association

between the drug use and sensation-seeking do-mains is given in different places in the two typesof analyses. In the canonical solution, the rathersmall size of the canonical correlations points tothe fact that the two domains, while significantlyrelated, overlap to only a modest degree. In theinterbattery analysis, the rather modest loadingsand small communalities for the variable in thetwo domains point out the same phenomenon.From both types of solution, then, it may be

inferred that Disinhibition tendencies are re-

lated to a pattern of alcohol and marijuana use:The canonical correlation results merely showthat the two sorts of tendencies are related, whilethe interbattery model further postulates thatDisinhibition sensation-seeking tendencies andfrequencies of alcohol and marijuana use sharea common cause. Choosing between the twomodels, of course, must be done on grounds thattranscend mere obtained empirical relation-

ships. It is reassuring, however, to find that thedifferent models will allow for approximately thesame statement of the degree and type of rela-tionship. Similarly, both the interbattery and ca-nonical correlation models permit the argument



299

Table 2

Rotated Loadings for One through ThreeInterbattery Factor Solutions

of Sensation Seeking Scales and Drug Use

that Experience Seeking and Boredom Suscepti-bility tendencies are related to the use of a vari-ety of &dquo;hard drugs,&dquo; while part of Thrill andAdventure Seeking is modestly related in a nega-tive direction to cigarette and marijuana use.Since there were relatively low levels of in-

ternal consistency for the individual sensation-seeking scales, it seemed desirable to rerun the

canonical correlation and interbattery analysesusing the original items in order to determinewhether the general pattern of specific resultsand the relative findings for the different modelschange in any major way.In the canonical correlation analysis of the 16

sensation-seeking items versus the 26 drug usevariables, there was a significant relationship



300

-continued on the next page-



301

Table 3, continued

between domains (416) = 952.84, p < .001.The size of the 15 canonical correlations (.60,.30, .28, .21, .21, .20, .19, .17, .16, .13, .12, .11,.10, .09, .09, .05) suggests that either two orthree dimensions are necessary to span the do-mains. As in the analysis with the Sensation-Seeking Scale scores, two dimensions were

found to be inadequate to fully explain the co-variation among domains: x2 (336) = 393.39,p < .02, but three dimensions were adequate: X2(299) = 307.11, p = .36.When different measures of multivariate cor-

relation are examined, it is again found that theconclusion is accepted that the domains are onlymodestly interrelated. The square root of Y5 is

.24, while the square root of y, is .23. Examiningthe redundancy coefficients, 6.93% of the vari-ance in the drug use items is explained by thesensation-seeking items, whereas 6.34% of thevariance in the sensation-seeking items is ex-

plained by the drug use variables.As in the analysis with the scale scores, al-

though three dimensions are statistically neces-sary to explain the observed pattern of across-domain covariation, it is possible that a lower di-mensionality solution will be judged to be more

satisfactory. Again, the one-dimensional solu-tion as well as the two- and three-dimensionalrotated solutions were examined. The loadingson the orthogonally transformed variates are

shown in Table 3, which also shows the adjustedcanonical correlations.From Table 3 it appears that the general inter-

pretations made for the scale scores are repli-cated at the item level. Indeed, even with theslight change in battery composition, the pa-rameter estimates for the drug use variables arealmost identical in numerical value when com-

pared to the results for the scale scores given inTable 1. (Note that the third drug use dimensionin Table 1 is the second drug use dimension inTable 3 and vice versa). The results are so con-ceptually similar that no additional discussionneeds to be made.

Corresponding to the canonical correlation

solutions of the sensation-seeking items versusthe drug use variables, one through three inter-battery factor solutions were developed by re-scaling the canonical loadings. The interbatteryfactors were orthogonally rotated using or-

thosim. The rotated loadings are shown in Table4. Again, the pattern of loadings is predictable



302

-continued on the next page-



303

Table 4, continued

from the analysis of scale scores previously con-sidered. Similarly, the interbattery solution ap-pears to support the same conclusions as the ca-nonical correlation solution.

Discussion

It is apparent that despite a few minor dis-crepancies, and bearing in mind the scaling dif-ferences between the canonical correlation and

interbattery factor analysis methods, the alter-nate solutions presented very similar results. Inall cases, sensation seeking and drug use weremoderately associated in the sample of adoles-cents.

When comparing the canonical correlationand interbattery factor analysis solutions thatused the Sensation-Seeking Scales with the anal-ogous two solutions using the sensation-seekingitems, scale integrity was not always retained.For example, the entire Experience Seekingscale was preserved in Dimension II (always con-sidering the three-factor solutions) of the canon-ical correlation analysis (Table 3), while three ofthe four items were retained in Dimension III ofthe interbattery factor analysis solution (Table4). In general, the canonical correlation analysespreserved the scale integrity more effectivelythan did the interbattery solution. However, thiswould be expected in light of the conceptual andmethodological differences in the two methods.

Using the full Sensation-Seeking Scale (FormV-10 items per scale), Ridgeway and Russell(1980) found a pattern of internal consistencycoefficients similar to that of the authors forthese scales in their sample of college students.They obtained alpha coefficients ranging from.48 to .69 that are identical in terms of rank or-der of magnitude with those obtained in thisstudy. The results of this study and those ofRidgeway and Russell (1980) clearly question thehigh scale reliabilities reported by Zuckerman etal. (1978). On the other hand, an average coeffi-cient alpha of .56 was obtained in this study,whereas Ridgeway and Russell (1980) obtainedonly a slightly higher average of .59 on all fourscales using 10 items per scale. This seems to in-dicate that the four-item scales have greatereconomy and parsimony than the 10-item scales,while sacrificing a negligible amount of internalconsistency. In general, the different sensation-seeking constructs are somewhat abstract, andthe indicators are only moderately correlatedwith one another.There are two conceptual perspectives on how

sensation seeking and drug use interrelate. Oneperspective, advocated by Zuckerman (e.g., Zuc-kerman et al., 1980), is that each person has anindividual optimal level of physiological arousaland that drug use and sensation-seeking tenden-cies are motivated by this physiological need.For example, a person who has a high innate



304

physiological need for stimulation will seek outthose experiences that will increase their arousallevel, such as exciting and stimulating activities(and thus would score high on sensation-seekingscales) and mood-altering drugs (and would like-wise report high drug use). This perspective re-flects a self-medicative or compensatory use ofdrugs and exciting activities to achieve an opti-mal level of physiological arousal. The secondperspective (see Segal et al., 1980) emphasizesthe psychological need (personality trait) for

stimulation, sensory variety and novelty, andtheorizes that drug use is caused by this person-ality need for novel stimulation. This perspectiveemphasizes the role of drug use to augmentemotional state rather than to compensate for aphysiological arousal imbalance. The two statis-tical methods contrasted in the analyses in thispaper parallel these two perspectives, such thatthe canonical correlation approach stresses thelatter argument, implying that drug use is

caused by sensation seeking, while the interbat-tery factor analysis approach assumes that someunderlying factor (presumably a physiologicalneed for arousal and stimulation in this case)causes both drug use and sensation-seeking ac-tivities. Since the results from both methods are

quite similar, it is not possible to say which ofthe two perspectives the data favor.Although the analyses leave little doubt that

there is a moderate relationship between sensa-tion seeking and drug use, similar to previous re-search on noninstitutionalized populations (e.g.,Zuckerman, 1979), the association is not a gen-eral one as would perhaps appear to be true forindices equivalent to the number of drugs usedby college students (Segal et al., 1980) or addicts(Kaestner, Rosen, & Apel, 1977; Kohn, Barnes,Fishlinsky, Segal, & Hoffman, 1979). In order togain a substantive understanding of the results,it is necessary to examine several patterns that

emerged from the analyses.Disinhibition, which is characterized by a de-

sire for active social affiliation and participa-tion, is related to a pattern of alcohol and can-nabis use. A similar finding regarding alcoholwas reported by Schwartz, Burkhart, and Green

(1978) for a sample of college students. Alcoholand cannabis are well noted for facilitating so-cial interaction, at least in small to moderate

quantities.Thrill and Adventure Seeking is related to two

patterns of drug use. Similar to Disinhibition, itis positively related to alcohol and cannabis usebut negatively related to the (more commited?)use of cigarettes, sedatives, amphetamines, in-halants, PCP, and marijuana. The content of theThrill and Adventure Seeking scale is related torisky and frightening physical endeavors. Indi-viduals drawn to these types of activities mayhave a strong awareness of and concern for theirphysical well being and thus may avoid the regu-lar use of those drugs that are likely to damagetheir bodies or to diminish physical responsivityand agility. Concomitantly, for adolescents,marijuana and alcohol may be used either tomake mundane activities more exciting or be-cause it is thrilling to engage in illegal activitieswith modest but not overwhelming penalties.Experience Seeking and Boredom Susceptibil-

ity are related to the use of stimulants (caffeine,marijuana, amphetamines, cocaine), depres-sants (sedatives, barbiturates), and major mood-altering drugs (LSD, other psychedelics, PCP).Both of these sensation-seeking scales are con-cerned with novelty, uniqueness, curiosity, andnonrepetitiveness. These drug effects apparentlyfacilitate Experience Seeking and combat Bore-dom Susceptibility.The results in this paper support the connec-

tion between various types of sensation-seekingbehaviors and several drug use patterns. Theconnection seems specific rather than global.Whether the results reflect underlying factorsinfluencing both domains or whether one do-main has a causal relationship with the other re-mains equivocal. Interpretations of the data

supporting either argument could be made. Thecontrast of the two methodological approaches(canonical correlation and interbattery factoranalysis) clearly substantiates the effectivenessand utility of either. The appropriate choice ofone model needs to be based on conceptual andtheoretical concerns, since statistical adequacy



305

has been demonstrated for both, at least in thisexample. Longitudinal data and analyses mighthelp partition the covariation into commonsources and &dquo;reciprocal&dquo; effects.

References

Bentler, P. M. Factor simplicity index and transfor-mations. Psychometrika, 1977, 42, 277-295.

Bentler, P. M., & Huba, G. J. Symmetric and asym-metric rotations in canonical correlation analy-sis : New methods with drug variable examples. InN. Hirschberg (Ed.), Multivariate methods in thesocial sciences: Applications. Hillsdale NJ: Erl-baum, in press.

Browne, M. W. The maximum-likelihood solution ininterbattery factor analysis. British Journal ofMathematical and Statistical Psychology, 1979,32, 75-86.

Cliff, N., & Krus, D. J. Interpretation of canonicalanalysis: Rotated vs. unrotated solutions. Psycho-metrika, 1976, 41, 35-42.

Cramer, E. M., & Nicewander. W. A. Some sym-metric, invariant measures of multivariate asso-ciation. Psychometrika, 1979, 44, 43-54.

Huba, G. J., & Bentler, P. M. A developmental theoryof drug use: Derivation and assessment of a cau-sal modeling approach. In P. B. Baltes & O. G.

Brim, Jr. (Eds.), Life-span development and be-havior (Vol. 4). New York: Academic Press, in

press.Huba, G. J., Palisoc, A. L., & Bentler, P. M.

ORSIM2: A FORTRAN program for symmetricand asymmetric orthogonal rotation of canonicalvariates and interbattery factors. American Statis-tician, in press.

Huba, G. J., Wingard, J. A., & Bentler, P. M. Begin-ning adolescent drug use and peer and adult inter-action patterns. Journal of Consulting and Clini-cal Psychology, 1979, 47, 265-276.

Huba, G. J., Wingard, J. A., & Bentler, P. M. Longi-tudinal analysis of the role of peer support, adultmodels, and peer subcultures in beginning adoles-cent substance use: An application of setwise ca-nonical correlation methods. Multivariate Be-havioral Research, 1980, 15, 259-280.

Huba, G. J., Wingard, J. A., & Bentler, P. M. A com-parison of two latent variable causal models foradolescent drug use. Journal of Personality andSocial Psychology, 1981, 40, 180-193.

Kaestner, E., Rosen, L., & Appel, P. Patterns of drugabuse: Relationships with ethnicity, sensation

seeking, and anxiety. Journal of Consulting andClinical Psychology, 1977, 45, 462-468.

Khavari, K. A. Mabry, E., & Humes, M. Personalitycorrelates of hallucinogen use. Journal of Ab-normal Psychology, 1977, 86, 172-178.

Kohn, P. M., Barnes, G. E., Fishlinsky, M., Segal,R., & Hoffman, F. M. Experience-seeking char-acteristics of methadone clients. Journal of Con-sulting and Clinical Psychology, 1979, 47,980-981.

Mardia, K. V., Kent, J. T., & Bibby, J. M. Multivari-ate analysis. New York: Academic Press, 1979.

Rao, C. R. Linear statistical interference and its ap-plications. New York: Wiley, 1973.

Ridgeway, D., & Russell, J. A. Reliability and validityof the Sensation-Seeking Scale: Psychometricproblems in Form V. Journal of Consulting andClinical Psychology, 1980, 48, 662-664.

Schwartz, R. M., Burkhart, B. R., & Green, B. Turn-ing on or turning off: Sensation seeking or tensionreduction as motivational determinants of alcoholuse. Journal of Consulting and Clinical Psy-chology, 1978, 46, 1144-1145.

Segal, B., Huba, G. J., & Singer, J. L. Drugs, day-dreaming, and personality: A study of collegeyouth. Hillsdale NJ: Erlbaum, 1980.

Skinner, H. A. Exploring relationships among mul-tiple data sets. Multivariate Behavioral Research,1977, 12, 199-220.

Skinner, H. A. The art of exploring predictor-criteri-on relationships. Psychological Bulletin, 1978, 85,327-337.

Stewart, D., & Love, D. A general canonical correla-tion index. Psychological Bulletin, 1968, 70,160-163.

Timm, N. H. Multivariate analysis with applicationsin education and psychology. Monterey CA:

Brooks-Cole, 1975.Tucker, L. R. An inter-battery method of factor

analysis. Psychometrika, 1958, 23, 111-136.Wingard, J. A., Huba, G. J., & Bentler, P. M. A lon-

gitudinal analysis of personality structure andadolescent substance use. Personality and Individ-ual Differences, 1980, 1, 259-272.

Zuckerman, M. Sensation seeking. In H. London & J.Exner (Eds.), Dimensions of personality. NewYork: Wiley, 1978.

Zuckerman, M. Sensation seeking: Beyond the opti-mal level of stimulation. Hillsdale NJ: Erlbaum,1979.

Zuckerman, M., Buchsbaum, M. S., & Murphy, D.L. Sensation seeking and it biological correlates.Psychological Bulletin, 1980, 88, 187-214.

Zuckerman, M., Eysenck, S., & Eysenck, H. J. Sensa-tion seeking in England and America: Cross-cul-tural, age, and sex comparisons. Journal of Con-sulting and Clinical Psychology, 1978, 46,139-149.



306

Acknowledgments

This research was partially supported by Grant Num-ber DA01070 from the National Institute on DrugAbuse. The authors thank Julie Honig and SuongLuong for production assistance and Arthur L. Pali-soc for programming aid.

Author’s Address

Send requests for reprints or further information toG. J. Huba, Department of Psychology, University ofCalifornia, 405 Hilgard Avenue, Los Angeles CA90024.



Comparison Canonical Correlation Interbattery Analysis ...291 Comparison of Canonical Correlation...

Documents

Transcript of Comparison Canonical Correlation Interbattery Analysis ...291 Comparison of Canonical Correlation...