Sie sind auf Seite 1von 17

Behavioral Intention Formation: The Interdependency of Attitudinal and Social Influence

Variables
Author(s): Michael J. Ryan
Source: Journal of Consumer Research, Vol. 9, No. 3 (Dec., 1982), pp. 263-278
Published by: Oxford University Press
Stable URL: http://www.jstor.org/stable/2488622
Accessed: 12-01-2016 10:16 UTC

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at http://www.jstor.org/page/
info/about/policies/terms.jsp

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content
in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship.
For more information about JSTOR, please contact support@jstor.org.

Oxford University Press is collaborating with JSTOR to digitize, preserve and extend access to Journal of Consumer Research
.

http://www.jstor.org

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
Behavioral Intention Formation:
The of Attitudinal
Interdependency
and Social Influence Variables

MICHAEL
J. RYAN*

Fishbeinand Ajzenhave proposeda theoryinwhichbehavioralintentionformation


is a functionof the separable effects of attitudeand the social norm.A variable
networkis deduced from their writingsthat explicitlymodels complex variable
interdependenciesnot previouslysubjectedto empiricaltesting.The findingsfrom
an experimentaltest supporta model in which normativevariables mediate the
effects of both cognitiveand normativeinformationon behavioralintentions.Im-
plicationsforthe theory'scentralequationsand for informationprocessingin gen-
eral are discussed.

Fishbein's (1967) model of behavioral intentions has FISHBEIN AND AJZEN'S THEORY
spawned extensive researchinvestigatingboth the the-
In relatingattitudesto behavior, Fishbein copes with the
ory and its applications in a number of disciplines (see
traditional attitude-behavior discrepancy by arguing that
reviews by Ajzen and Fishbein 1973; Farley, Lehmann,and
this gap is due to inadequateconceptualizationand mea-
Ryan 1981). A good deal of this researchhas appearedin
surementand to the need to consider "other variables" in
the consumerbehaviorliterature,as reviewed by Ryan and
Bonfield (1975), and more recent work continuesto appear additionto attitudes.He addressesthe issue of "other vari-
(Ahtola 1976; Carnegie-MellonSeminar1978; Dickson and ables" by combining attitudewith a variable described as
Miniard 1978; Fishbein 1976; Glassman and Fitzhenry the "subjective norm" (Fishbein and Ajzen 1975, Chapter
1976; Lutz 1977, 1978a, 1978b; Miniardand Cohen 1979; 7), which is designed to capture the social influences of
Miniardand Dickson 1979; Ryan 1978; Ryan and Bonfield relevantothers. The basic Fishbeinparadigmis that behav-
1980; Ryan and Etzel 1976; Ryan and Holbrook in press; ior is affected by behavioral intention which, in turn, is
Ryan and Peter 1976). A majorconcernof recent work has affected by attitudeand the subjective norm.
been that earlier regression testing did not capturethe the- The central equations in the theory appearas follows:'
ory's richness. In this spirit, the present researchexpands B - BI= (Aact)w1 + (SN)w2 (1)
the theory's behavioral intention paradigmin order to ex-
n
plicitly model variable interdependencies not previously
examined. A structuralequation methodology (Jdreskog Aact= E Biai (2)
and Sorbom 1978) is employedthat, in additionto modeling
k
complex interdependencies,explicitly considers measure-
ment error. The following sections outline the basic theory SN= NBjMCj (3)
j=1
and discuss its empirical supportand conceptualunderpin-
nings, propose a more complex general model, and report where
on a behavioral intention formation experiment in which
informationabout attitudinaland normativebeliefs is ma- B = overt behavior
nipulated.
'Fishbein and Ajzen (1975) have recently changed their algebraicsym-
*Michael J. Ryan is Associate Professorof Marketing,GraduateSchool bols. Althoughsome of theirnew notationappearsin recentresearch(e.g.,
of Business Administration,The University of Michigan, Ann Arbor, MI Miniardand Cohen 1979), it is inconsistentwith the notationused in the
48109. Research was funded by the University of Alabama Faculty Re- majority of published empirical studies. The new notation also equates
search Grant Committee and the Graduate School of Business Faculty attitudinal(Be)and normative(NBj)beliefs by using the same notation(Be)
Research Fund, Columbia University. The author is indebted to Claes for both. Since a conceptualdistinctionis maintainedbetweenthese beliefs
Fomell, Morris B. Holbrook, Jerry C. Olson, Srinivas Reddy, and two and theirrespectiveoutcomes and referents,consistency suggests thatthey
anonymous JCR reviewers for their comments on earlier drafts of this also be distinguished symbolically. For these reasons the new notation
paper. appears confusing, and the original model's symbols were used in this
paper.
263
? JOURNAL OF CONSUMERRESEARCH0 Vol. 9 0 December 1982

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
264 THEJOURNALOF CONSUMERRESEARCH

BJ = behavioralintention FIGUREA

Aact = attitudetoward behavioralact FISHBEINANDAJZEN'SINTENTION PARADIGM


FORMATION

Bi= the expectation(i.e., the probabilityor improb- ~BLa Aact


ability) that the performanceof a specific be-
havior will lead to an ith outcome
ai = the positive or negative evaluation of the ith Stimulus BI
outcome Conditions

n = the numberof salient outcomes


N8j SN 3
SN = the subjectivenorm (i.e., overall perceptionsof
what relevant reference groups or individuals NOTE:AdaptedfromFishbeinand Ajzen(1975, p. 334)
think the actor should do)
NBj = the expectation(i.e., the probabilityor improb- of the relationshipsamong attitudinaland normativevari-
ability) that the performanceof a specific be- ables, a literal interpretationof Equation 1 and Figure A
havior is expected by a jth group or individual has led othersto assume independence(Miniardand Cohen
1979, 1981; Ryan 1978; Ryan and Bonfield 1975). How-
MCi = the motivationto comply or not to comply with ever, a close scrutinyof Fishbein and Ajzen's earlierwrit-
the expectation of the jth group or individual
ings (1975; Fishbein 1976), togetherwith their more recent
k = the numberof salient groups or individuals positions (Fishbein and Ajzen 1981), 'suggests that the
schema may have served as an oversimplificationfor the
wowI-= empiricallydeterminedstandardizedregression sake of elegance, or as a point of departure.
coefficients
The predictive ability of Equation 1, incorporatingnor- THE INTERDEPENDENCY OF
mative structure(>NBjMC1)instead of SN as the second ATTITUDINAL AND NORMATIVE
predictorvariable, has received empiricalsupportin a num-
ber of studies reviewed by Ajzen and Fishbein (1973) and
VARIABLES
Ryan and Bonfield (1975). The parameter estimates for A numberof statementsAjzen and Fishbein make about
Equation 1 excluding B were also found to be consistent specific portionsof the model imply more complex variable
across 37 studies conductedin a varietyof situations(Farley relationshipsthan those shown in Figure A. They consider
et al. 1981). The relationshipbetween Aact and >Biai has belief formationand change processes to be the main force
been empiricallysupported(Ajzen and Fishbein 1972; Jac- driving the model (1975, Chapter5). They call these an-
card and Davidson 1972), as has been the need to include tecedents, which are described as prior subjective proba-
Aact as a moderatorof the >Biai and B! relationship(Lutz bilities that determineattitudes, primarybeliefs. The gen-
1973; Ryan 1974, 1978). One experimental study (Lutz eral definition they assign to "belief," as incorporatedin
1977) manipulatedBi and ai and demonstratedsubsequent Equations2 and 3, involves the link an individualperceives
changes in Aact and B1. In contrast to the amount of re- between any two concepts or objects. There are three types
search investigatingthese issues, the variable networkhas of beliefs (Fishbein and Ajzen 1975, pp. 131-135):
not been tested in its entirety, and only two studies (Glass-
man and Fitzhenry 1976; Miniardand Cohen 1979, 1981) * Descriptive beliefs, derived from direct experience
have investigated SN.2 Ryan (1978) also attemptedto de- * Informationbeliefs, formed by accepting informa-
velop a social influence variable different from SN. The tion from some source
results from these studies are reportedin the following dis-
cussion of the theoretical network within which Aact and * Inferentialbeliefs, derived througha process of in-
SN occur. ference from descriptive, informational,or otherin-
In terms of both change and formationprocesses, Fish- ferentialbeliefs
bein and Ajzen (1975) consistently argue for a chain of The notion of inferentialbeliefs creates the possibility that
effects that proceeds from stimulusto 2Biaiand INBjMCj, attitudinalbeliefs (B1)may be. formed from normativebe-
which influence Aact and SN, respectively. Aact and SN, liefs (NBa),and vice versa. Fishbeinand Ajzen consistently
in turn, affect BI. A schematic representationof their for- acknowledge this possibility in terms of the formationof
mation paradigmis shown in Figure A. WhereasFishbein normative(1975, pp. 304, 306, 314) and attitudinalbeliefs
and Ajzen do not furnish an explicit conceptualdiscussion (1975, p. 304):

20thers (Bagozzi 1981a, 1981b; Bentlerand Speckart1979, 1981) have Not only may an item of informationto which a person is
reportedtesting networks containing SN. However, they have employed exposed during an influence attemptaffect one of the deter-
measuresthat deviate from Fishbein's operationalprocedures.These stud- minantsof the intention-say, the attitudetowardthe behav-
ies will be considered in the discussion section of this paper. ior-but it may also have an impact on the second determi-

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
BEHAVIORAL
INTENTION
FORMATION 265
nant of intention, the subjective norm. Consider, for Fishbeinand Ajzen (1981) questionMiniardand Cohen's
example, a person who observes that his best friend receives findingsby raising a numberof technical issues, which the
$5 for tutoringa student.Formationof this descriptivebelief latter dispute (Miniard and Cohen 1981); they also claim
may lead him to inferthattutoringa studentis financially that the findings actually supportthe theory because it ac-
rewarding, and this belief may in turn increase his attitude
counts for attitudinal-normativeoverlap.3While the argu-
towardtutoringa student. At the same time, the descriptive
belief may also lead the person to infer that his best friend ments of each pair of authorsare quite complex, the main
thinkshe should tutorthe student.This inferentialbelief may dispute-involvesthe degree of separationbetween attitudes
increasethe subjectivenorm thatmost importantothersthink and social norms, with Miniard and Cohen claiming that
he should tutor a student. Alternatively,once the person has the currentconceptualizationrendersthem inseparable,and
changed his attitude in a favorable direction, he may also Fishbein and Ajzen holding that attitudesand social norms
infer that most importantothersalso hold a favorableattitude have different effects in the model, depending on the sit-
towardtutoringa studentand then make the furtherinference uation under study. Although Fishbein and Ajzen cite em-
that these referents think he should perform this behavior. pirical support for their position, the cited studies do not
An influence attemptcan thus have an impact effect even if contain SN (as Miniard and Cohen point out). However,
it provides informationthat is directly relevantfor only one
determinantof intentions. The strengthand directionof this
Gur-Arie, Durand, and Beardon(1979) found that, as pre-
kind of impact effect will depend on the extent to which the dicted, Aact was a strongerpredictorof intentionsfor opin-
two componentsare relatedand the directionof the relation- ion leaders than was SN, while the weighting was reversed
ship (p. 402). for nonopinionleaders. Thus the study by Gur-Arieet al.,
which contains SN and accuratelypredictsrelative weight-
Yet Fishbein and Ajzen consistently maintainthat there is ing, supportsFishbein and Ajzen.
utility in separatingattitudinaland normativevariables, de- Fishbein (1976) has acknowledgedthat the social influ-
spite the possibility that they may be highly correlated ence variablesare underdeveloped.The theory's conceptual
(Fishbein 1976; Fishbein and Ajzen 1981). This utility frameworkis sufficiently vague to serve its heuristicfunc-
would be more apparentif these variablescould be shown tion.4 However, the acknowledgmentof complex attitudinal
to have differentialeffects within the context of the entire and normative variable interdependencies suggests that
model. Equations 1, 2, and 3 may have outlived their usefulness.
Fishbein (1976) has maintainedthat regression is the best
Evidence and Criticisms of Attitudinal- estimation procedurecurrentlyavailable, despite its well-
NormativeIndependence known limitations. Fortunately,recent structuralequation
methodologies allow more complex modeling than do the
On the basis of the high correlationsreportedbetween threecentralequations.Furthersteps in this direction,along
attitude and social influences and Aact changes that oc- with the research setting used in this study, are discussed
curredfollowing either Bi or NBj manipulations,Ryan and in the following section.
Bonfield (1975) addressthe problemof developing a social
influence variable independentof attitude.BorrowingKel- A REFORMULATED OPERATIONAL
man's (1961) three processes of social influence-compli- MODEL
ance, identification, and internalization-they reason that
an actor, under the influence of anotherperson or group, Marketingresearchershave long recognized social influ-
may play a role that is not congruentwith his own attitudes ences on behavior (Bourne 1957) and on expected attitu-
toward a behavior. The motivation for playing the role dinal outcomes (Haley 1968). Attitudeis a centralconcept
would be to attainrewardsunder the other's control (com- in buyer behavior models (cf. Howard and Sheth 1969); it
pliance) or to meet the other's own role expectations(iden- is common to segment individuals on the basis of attitude
tification), regardlessof their compatibilitywith the actor's similarity within and dissimilarity between groups (Wind
value system (internalization).Ryan and Bonfield identify 1978). Based on a large body of communicationresearch
this constructas social compliance(SC); however, attempts (McGuire 1973; Sternthal, Phillips, and Dholakia 1978),
to operationalizeit and show its independencefrom Aact attempts to influence attitude formation are often made
are equivocal (Ryan 1978). throughthe use of "expert" informants.For example, an
Miniardand Cohen (1979) reportthat both SN and MC endorsementof flouridefrom a dental associationhas been
were affected by manipulationsof normativeinfluence and used to enhance the belief that decay preventionwill result
attitude, and criticize the theory for failing to separateat- from the use of Crest toothpaste (Shuchman and Riesz
titudinaland normativevariables. They speculatethat their 1975). Ryan (1974, 1978) found that attitude and social
findingsmay have been due to respondents'use of expertise influenceare quite good at predictingintentionsto purchase
as a source of informationalsocial influence which, taken toothpaste brands. Although attitudinal outcomes varied
as evidence about reality, would be incorporatedin atti-
tudes. Interestingly, their position is consistent with Fish-
3The last three pages of Miniardand Cohen's (1981) article is a post-
bein and Ajzen's notion of belief formation-namely, that script answeringFishbein and Ajzen's (1981) criticism.
a primaryattitudinalbelief may be inferredfrom other be- 'See Kaplan (1969) for a discussion of the role of vagueness and am-
liefs arising from a variety of external sources. biguity in theory construction.

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
266 THE JOURNAL OF CONSUMER RESEARCH

FIGUREB
PROPOSEDINTENTION
FORMATION MODEL
STRUCTURAL

N4

Xi 2

NOTE: Where ryij,,Bii= standardized path coefficients and (j=error terms (residual variances)

across brandsand situations(Ryan and Etzel 1976), "den- information (NI) directed toward normative beliefs (NBj)
tist" consistently emerged as a referent. Together with would affect both normativeand attitudinalbelief structure
these findings, the aforementionedargumentssuggest that formation (y11, Y21' Y12 Y22), The Y12 crossover effect is
when intentionstowarda previouslyunknownphenomenon likely to be strongwhere a normativereferent-say, a den-
are formed, and there is a key referent, variable relation- tist-may serve as a source of informationin forming an
ships will appearas shown in Figure B.5 attitudinalbelief about, say, decay prevention.
The model is true to the Fishbein and Ajzen intention Perhapsnormativebeliefs should be included merely as
formationparadigmshown in Figure A, with the exception additionalbeliefs (Bi) in attitudinalstructure.In this study,
of the crossover relationships among attitudinaland nor- however, they are expected to be related (P12' 21) but
mative variables (12 P21 and 342).The crossover rela- separable. Attitudes may be formed on the basis of infor-
tionships are, however, suggested in their writings. To be mation including that provided by an expert referent;they
specific, informationalimpact on behavioralintentionsoc- may also be influenced by referentexpectations (McGuire
curs only throughbelief formationandthe mediatingeffects 1973). Thus, normative beliefs (XNBjMCj)should influ-
of attitudeand the social norm (,y11,P41'P54, and Y22 32 ence Aact formation. Previous researchon toothpastepur-
353). Furthermore,through the process of secondary and chase intentionsalso found strongerAact than social influ-
inferentialbeliefs, informationmay affect beliefs otherthan ence beta weights (wo > w, in Equation 1; Ryan 1978),
those toward which it is directed. Hence, cognitive infor- plus a joint attitudinal social-influence effect (Ryan and
mation (CI) aimed at attitudinalbeliefs (Bi) and normative Peter 1976). Perhapstoothpaste,being privatelyconsumed
(cf. Bourne 1957), is more susceptible to internalthan to
'To be consistent with LISREL (Joreskogand Sorbom 1978) notation, external influences. Consequently, it is hypothesized that
the firstsubscriptreflectsthe variableto which and the second the variables Aact has a strongerdirect link than SN to BI (P54 > 53),
from which the path is drawn. but that social influence variables also have an impact

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
BEHAVIORAL
INTENTION
FORMATION 267

throughAact mediation,which is reflectedin both the direct EXHIBIT


(P53) and the indirect (r42) links. In addition, there is no OUTCOMESAND REFERENTS
ELICITED
direct link between Aact and SN. The relationshipsprevi-
ously found between attitudinaland normativevariablesare Bj Outcomes
expected to be caused by belief interdependencies.If these B1 Price or economy
expected beliefs are the primarydeterminantsof Aact and B2 Taste or flavor
B3 Cavityor decay prevention
SN, once they are formed there is no reason to expect that B4 Fresh, pleasant breath
an internalizedpredisposition(Aact) would be relatedto an
external expectation (SN). That is, consistent with Kel- NB. Referents
NB1 Dentist
man's processes discussed in the preceding section, nor- NB2 Children
mative beliefs (XNBjMCj)may be internalizedthroughthe NB3 Husband
process of inferentialbelief formation,thus influencingat-
titudinalbeliefs (XBiai) and overall affect (Aact). However,
SN falls more in the domain of an actor playing a role,
motivated primarily by externally mediated rewards-a toothpaste.A fictitious new brand,designated "Brand0,"
type of social compliance. Thus SN is relatively more in- was used so that attitudesand normscould be formedsolely
dependentof internalpredispositionsthanthe variousbelief from informationprovided in the experiment. Operation-
formationprocesses. alization of the model involved the identificationof salient
The pragmaticadvantagesof the theoreticalmodel shown outcomes and referents, the constructionof measuringin-
in Figure B over the Fishbein and Ajzen approachare now struments, and the design of written communications to
more apparent.The latterrelies on beta weights from Equa- formulatemore positive attitudinaland,normativevariables
tion 1 to determine the relative influence of Aact and SN for experimentalversus control groups.
on behavioral intentions. For example, the research cited
regarding relative normative-attitudinalprominence used Salient Outcomes and Referents
weights derived from OLS regression (Fishbein and Ajzen
1981; Gur-Arieet al. 1979; Miniardand Cohen 1979; Ryan Salient outcomes and referentswere determinedwith an
and Bonfield 1975). Using beta weights in this fashion as- elicitationtechniquecommon to this type of research(Ryan
sumes that an interactioneffect is not present, despite the and Etzel 1976). The technique employed nondirective
fact that it has been demonstratedto exist (Ryan and Bon- questions to obtain free responses, which were then ana-
field 1980; Ryan and Peter 1976), due either to poor mea- lyzed as to content and separatedinto groups on the basis
surement discriminantvalidity or to inherent variable re- of common meanings. Questioning referred to outcomes
lationships. In addition, the beta weight analysis does not and referentsrelevant to the purchase and usage of tooth-
provide very rich insights. For example, a small SN beta paste in general;Brand 0 was not mentionedto these sub-
weight may be interpretedas indicatingweak social influ- jects. Naturalbreaks in the frequency of mentioned items
ence effects on BI, when in fact it may have a strong in- were used to separatesalient items, shown in the Exhibit,
fluence throughthe mediatingeffects of Aact. from nonsalientitems. Consistentwith Haley's (1968) ben-
The empiricalportionof this paperutilizes an experiment efit segmentationresearch, these housewives sought decay
to test the model shown in Figure B. The hypothesized prevention. The dentist referent found by Ryan and Etzel
model is comparedboth to simpler forms that more closely (1976) also recurred.
captureFishbein and Ajzen's conceptualization(Figure A)
and to more complex alternatives. Measures
The elicited outcomes and referents were used to con-
METHOD structBrand0 measures.A set of measureswas constructed
to consider belief (Bi and NBJ), evaluation (ai), attitude
Sample and IntentionObject toward the act (Aact), and behavioral intention (BI). The
Data were collected at two points in time, the first to single-item scales commonly used in this type of research
determine salient outcomes from which measures and an were modified to create multiple-item scales, in order to
experimentcould be designed, the second to administerthe obtain reliability estimates and avoid bias from adjective
experiment.The panel constructedfor use in this study was specificity. The bipolar adjectives used in the Bi, NBj, and
composed of 80 membersof various churchgroups located BI scales included the single set of adjectives commonly
in the SoutheasternUnited States. The panel memberswere used in previous research(likely-unlikely; cf. Jaccardand
white, Anglo-Saxon, Protestanthousewives who were mar- Davidson1972) andtwo additionalsets (probable-improbable
ried, had children living at home, and were predominantly and possible-impossible). For example:
middle-agedmembers of the lower-middle social stratum.
All subjectsvolunteeredto serve in the experimentin return (B3) Brand0 toothpasteusagewouldlead to decaypreven-
for monetarydonationsto their respective churches. tion:
The productchosen to representthe intentionobject was possible impossible

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
268 THEJOURNALOF CONSUMERRESEARCH

(NB1) Withrespectto Brand0 toothpaste,my dentistwould employed, and attitudinaland normative structuralscores
expectme to purchaseanduse it: (:Biai and ENBJMCJ)were divided by the numberof sa-
probable -: .: - - : - : - : :- improbable lient items (four and three, respectively). Thus, BI, Aact,
and SN scores ranged from -3 to +3, and >Bi ai and
(BI) Whenit is introduced,I intendto buy Brand0 tooth-
paste: INBjMCj scores ranged from -9 to +9. Similarly con-
structedBi, ai, Aact, and BI measurespreviously had been
likely - - : - : unlikely shown to possess internal consistency and concurrentva-
The adjectivesused in the a, scales includedgood-bad and lidity (Ryan 1978).
two pairs taken from the Fishbein and Raven (1962) AB It should be noted that the measuresconfound purchase
scale-i.e., wise-foolish and beneficial-harmful. For ex- and usage. In this operational setting, the purchase act
ample: would be performed by the housewife, whereas the ex-
pected outcomes would accrue from family usage. While
(ad) Forme, low-pricedtoothpasteis: the distinction between purchase and usage intentions is
good _: .:- - :- :- : :- bad interestingin its own right, purchaseis necessaryto obtain
Aact was measuredby the same adjectivesas those used in usage in the present situation, and this confounding does
the ai scales, plus a fourthpair, rewarding-punishing,e.g.: not mitigate testing of the relationshipsshown in Figure B.
A similar criticism can be made concerningthe specificity
(Aact) My purchaseanduse of Brand0 toothpaste is: level of the outcomes and referentsthat were derivedat the
rewarding -: - _: - :- :- : :- punishing productlevel. In fact, Ryan and Etzel (1976) have shown
that outcomes change across existing toothpaste brands
Use of the same adjectives to measure different con- aimed at different market segments. However, previous
structs may result in common method variance. However, experience(Ryan 1974) revealedthatlaboratoryprocedures
a previous study (Schwartz and Tessler 1972) compared were not able to produce strong belief changes for existing
different measurementprocedures and provided evidence brandsfor which belief structureswere strongly in place,
indicatingthatthis possible artifacthas not favorablybiased as a result of usage and advertising. The proceduresused
the evidence supportingthis model. In addition, Ryan and in the present study attemptto overcome this methodolog-
Bonfield(1980) employedpersonalinterviewsduringwhich ical limitation by perceptually constructing an artificial
questionnaireitems were verballyadministeredand respon- brandthat varies in its ability to deliver expected outcomes
dents gave verbal replies throughthe use of rulers identi- and in its referentexpectations. Consequently,the research
fying scale points. In this case, although majordeviations investigates attitudeformation, not change (CarnegieMel-
from accepted practice were avoided, statementsusing the lon Seminar1978). When used to evaluatenew brandsprior
same adjectiveswere separatedfrom one another,and some to manufacture,such proceduresare referredto as concept
scale directions were reversed to lower the possibility of testing (Tauber 1977).
response-setbias.
Motivation to comply (MC.) and the subjective norm
(SN) were operationalizedwith single-item measures, fol- ExperimentalProcedures
lowing Fishbein's procedures(Fishbein and Ajzen 1975), Booklets were designed to produce two levels each of
as follows: cognitive and normativeinformationas the result of a per-
suasive communication, after which the various measures
(MCI) Withrespectto toothpastepurchaseandusage: were presented. Subjects were told that the elicitation ses-
I wantvery I wantvery
much to - - :- : : much not to sion, which had taken place one month earlier, had deter-
mined that they were typical in terms of what they sought
do as my dentistexpects. in a toothpaste.In additionto standardinstructions,the first
(SN) to me wouldthink:
Mostpeoplewho areimportant page of the booklet contained the following statementof
I should _: : : : : :- I should not purpose:
purchaseBrand0 toothpastewhenit is available. Many of today's shoppers have called for more objective
product informationin advertisements.The purpose of this
Bipolar ( + 3 to -3) summativescoring was used for all study is to examine how shoppers like yourself use infor-
scales.6 To maintain scale perspective, each of the sum- mation from an impartialsource.
mative scores was divided by the number of scale items This booklet contains information from the testing of a
new brand of toothpaste. The information is accurate and
unbiased, having been derived solely for the company's use.
6Fishbein(1976) has recently suggested thatthe MCjstatementmay not In grantingpermission to use this information,the company
be viewed as bipolarby respondents.He maintainsthat bipolarscoring is has requested that it and the brand name be anonymous.
theoreticallyappropriateand suggests a modificationusing a negative an-
chor, such as "I want to do the opposite of ...." Miniardand Cohen Consequently, the brandwill be referredto as "Brand 0."
(1981) used this modificationand were criticized by Fishbein and Ajzen The booklet containeda page of text followed by a sum-
(1981) for not using unipolarscoring. The readeris referredto the sources
for details, but it seems safe to conclude that MCj's operationalizationis mary of information about Brand 0. Four forms of the
not a settled issue. booklet, identical in appearancebut each containing a dif-

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
BEHAVIORAL
INTENTION
FORMATION 269

ferent descriptionof Brand 0, were randomly distributed TABLE 1


to subjects. The purpose was to producea 2 x 2 factorial
CORRELATIONALRESULTS
design with 20 subjects in each cell, by varying the infor-
mation pertainingto price and dentist expectations. Based Pooled within-cell correlation
on the previous discussion of inferentialbeliefs and on ear-
lier findings(Lutz 1975), it was expected that manipulating BI IB,a, Aact YNBjMCJ
the most frequently elicited attitudinaloutcome and nor-
mative expectationwould produce changes in the same di- 1B,A, .42
Aact .57 .66
rection in otherbeliefs, thus influencingthe entire structure INB,MCJ .51 .52 .69
(EB ai or INBJMCJ).The control group received the fol- SN .68 .40 .60 .55
lowing version:
In summary,"Brand0" is Pooled within-cell regression
1. The samepriceas competitivebrands. BI Aact 1B,a, SN INB,MCJ
R2 b~~1
2. Averagein tasteandflavorappeal. b2 3 b4

3. Averagein cavityprevention. .51 b .25b .53b


.29b .22b ,40b
4. Averagein breathfresheningandtoothbrightening. .72b .51 b .06c .51 b .Q9d

5. Endorsedby about50%of dentists. aStandardized regression coefficients


bp < 0.001
The following change was made for the cognitive infor- Cp < 0.46
dp < 0.44
mation experimentalgroup:
1. Pricedmuchlowerthancompetitivebrands,
plicit incorporationof such error in the structuralmodel.
and for the normativeinformationexperimentalgroup: The measurementmodel for the endogenous variables is
shown in Figure C.7 Since a one-item measurewas used to
5. Endorsedby 90%of dentists.
indicate SN, X8 and E8 must be constrainedto be equal to
No informationwas given about childrenand husbandref- one and zero, respectively. Like the structuralmodel, the
erents. measurementmodel was specified a priorifrom the theory.8
The next partof the booklet containedthe questionnaire. Fishbein and Ajzen (1975, pp. 229-235), on the other
Measureswere groupedin the following order:ai, MCj, Bi, hand, specified an unweighted additive model. Whetherit
NBj, Aact, SN, andBI. While the orderingfollows the con- is weighted or unweighted is an importantempiricalques-
ventional wisdom of proceeding from specific to general tion; consequently, the model using individuallyestimated
construct measures, there is evidence that reversing this Xivalues will be comparedto one with Xiconstrainedto be
proceduredoes not influence goodness-of-fittests (Miniard equal for each construct.
and Dickson 1979).
Data Check
Method of Analysis The data were examined to see whether they produced
The path coefficients shown in Figure B were estimated results consistent with those from the extensive body of
with LISRELIV (J6reskogand Sorbom 1978). The primary correlational evidence. Correlationalresults adjusted for
advantageof the structuralequation methodology over the experimentaleffects are shown in Table 1. The within-cell
traditionalcentral equations (1, 2, and 3) is that it simul-
taneously estimates all path coefficients, including the 7Since the exogenous variables (CI and NI) are merely dichotomous
crossover paths, which are ignored in the regression ap- levels of information, each produced by an experimentalmanipulation,
proach. they have no measurementmodel.
8A more faithful representationof the operationalizationfor 2Biaiand
The proposed model, as derived from the theory, is un- Y2NBjMCj is precludedby technical limitations.This would entail treating
deridentified. The identifying conditions for the LISREL each outcome and referentas a latentvariableindicatedby theirrespective
model usually differ from the classical rankand ordercon- multiple belief measures. Thus, TBjaj and 2NBjMCjwould become sec-
ditions, but in this case they are identical. To remove this ond-orderlatents. Unfortunately,a procedurefor estimatingsecond-order
endogenouslatents is not available. Thus in the presentoperationalization,
problem, I12 and I21 were constrainedto be equal. This the measurementerrors (E.) refer to respective outcomes and referents.
solution was chosen on the grounds that it would yield an Fishbein and Ajzen view belief structureas a composite and allow for
identified model while losing the least amount of useful inconsistent beliefs. The present operationalization,which posits unidi-
theoretical information;only directionality is lost, rather mensional composites, presumesbelief consistency, which was borne out.
than an entire path. in the dataanalysis. Had the compositesbeen multidimensional,this would
have necessitated changing the model specification (for an example, see
LISREL also has an advantageover the more traditional Bagozzi 1981b). In the measurementof SN, the empiricalspecificationof
proceduresin that it estimates measurementerrorand, be- the constructwith one measure is a shortcoming.Unfortunately,the lack
cause of its simultaneousestimationprocedures,allows ex- of conceptualdevelopment precludesthe inclusion of other indicants.

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
270 THE JOURNAL OF CONSUMER RESEARCH

FIGUREC
FORMATION
INTENTION MODEL
MEASUREMENT
X1 ~~~Price I
Taste2
I I Decay E3
~~~~~~~~~~4
__ 6~~~~~~~4
6
-6

:NBjMC 6__ Children 6E


Husband E7

SN should-should not E8

Xg foolish-wise E9
Aact F good-bad E 0
x,,
X12 harm-benef icial E
reward-punish E 12

BI X1 improbable-probable E 14
XiN
possible-impossible E 5

NOTE:Whereel endogenousvariablemeasurementerrorsand XI= standardizedpathcoefficientbetweenan observedindicationand respectiveendogenousvariable

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
BEHAVIORALINTENTIONFORMATION 271

TABLE2
PARAMETER ESTIMATES, CRITICALRATIOS, RELIABILITIES,AND VARIANCE PROPORTIONS

Measurement model
Standardized Proportions of
factor Critical variance
loadings ratios Reliabilities extracted

Y.B,ai .87 .63


Price (X1).57 .32
Taste (X2) .89 5.49 .79
Decay (X3) .75 5.00 .56
Breath (A4).91 5.55 .83
INB,MCJ .83 .62
Dentist (A5) .86 .74
Children (X6) .67 6.69 .68
Husband (A7).82 8.92 .66
Aact .97 .88
Foolish-wise (Ag) .93 .83
Good-bad (X10).98 19.00 .96
Harm-beneficial (X11).94 15.79 .88
Reward-punishing (X12).91 14.30 .83
BI .93 .81
Likely-unlikely (X13).83 .69
Improbable-probable (X14).94 11.38 .91
Possible-impossible (X15).90 10.56 .83

Causal model Proportions of shared variance

Standardized Critical Structural Total


weights ratios model model

Exogenous paths
'Yii .10 1.03 1. IB,a. .60 .36
'Y21 .33 3.45 2. ENBJ MC, .47 .29
'Y12 .29 2.62 3. SN .67
'Y22 .29 2.92 4. Aact .78 .69
5. BI .71 .50
Endogenous paths
I12 .50 5.10
I21 .22 5.10
I32 .75 7.81
I41 .31 2.47
P42 .63 4.98
PM .45 4.74
P54 .48 4.88

correlationmatrixshows considerablepairwisecovariation. ious fit. These findings suggest that the data are typical-
The within-cell regression using Equation 1 (Aact and SN e.g., that nonsalient beliefs were not included and salient
to predictBI) producedresults (R2 = 0.51) consistent with beliefs were not excluded-and that furthertests are appro-
previous correlationalstudies (see reviews by Ajzen and priate.
Fishbein 1973; Farley, Lehmann,and Ryan 1981; Ryan and
Bonfield 1975). The prediction of BI from YBiai and
ENBjMCjproduced an R2 of 0.29, which suggests that, RESULTS
consistent with the theory, Aact and SN are better predic-
tors. Regardless of the set of predictors(Aact and SN or
MeasurementModel Fit
EBiai and ENBjMCJ)used, the beta weights suggest that A correlationratherthan a covariance matrix was ana-
the social influence variable predominates.When all four lyzed because the matterof interestis the relative strengths
predictorswere used, R2 increasedto 0.72 and, consistent of the paths, not the predictedvalue of BI. Model testing
with the theory, b2 andb4 in Table 1 become nonsignificant, began by comparingthe overall goodness-of-fit statisticfor
thus indicatingthat Equation 1 produces a more parsimon- the model when Kjwas estimated (X2l4 = 198.16, p <

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
272 THEJOURNALOF CONSUMERRESEARCH

0.001) to the fit obtained when Xi was constrainedto be may account for the lower predictive ability of Equations
equal for each construct(X22o= 216.28, p < 0.001). Fol- 1 and 2 when perfect measurementis assumed. The differ-
lowing the procedure suggested by Bentler and Bonett ences in shared variance between the structuraland total
(1980), the difference between these two x2 values (X2 = model are a direct reflection of the differing amounts of
18.12) was statisticallysignificant(p < 0.01). This finding measurementerror.9For example, when varioussources of
suggests thatthe model utilizing unequalXivalues provides measurementerrorare removed, the model explains 71 per-
a better fit. Consequently, the detailed results, shown in cent of BI variance.That this rathersimple model including
Table 2, are for the model using unequally weighted con- measurementerror("total model") explains 50 percentof
structindicators. BI variance is a point in its favor.
Since the measurementerrors (E) and structuraldistur- With the exception of the path from "cognitive infor-
bance (C) estimateshave little absolutemeaning, they were mation" to YBiai (yl 1; see Figure B), all path coefficients
used to calculatemore familiarreliabilityestimates(Werts, were statisticallysignificant.An overall goodness-of-fitwas
Linn, and Joreskog 1974) and sharedand extractedvariance obtained for a model with y,, constrainedto be equal to
proportions(Fornell and Larcker1981), which are conser- zero (Xl25 = 199.25, p < 0.001), which was not statisti-
vative analogs to explained variance. These estimates are cally different (XI = 1.09) from the full model. This in-
shown in Table 2. To serve as reference indicators, mea- dicates that deviating from the theoreticalmodel by drop-
sures fixed to unit variancedo not have criticalratiovalues. ping y,, does not alter the fit. Since neither model yields
The critical ratios are distributednormally;hence any val- a superiorfit, parsimonydictates dropping yul solely on an
ues greaterthan 1.65 (one-tailed)are statisticallysignificant empirical basis. However, given the strong prior theory,
at the 0.05 level. weak empirical disconfirmation, and additional analysis
The measurement model results indicate that the con- concerning overall fit reportedbelow, -y1lwas retained in
structsare well-specified. The reliabilityestimatesfor each the model. The position taken throughoutthis analysis is
endogenous construct, which are identical to Cronbach's that the proposed theoretical model serves as a reference
(1951) alpha except for the relaxationof the assumptionof against which evidence is needed to produce acceptable
equally weighted items, exceed acceptable values. More alternatives.The suspicion of measurementproblems, dis-
revealing are the more conservative "proportionsof vari- cussed in the next section, also has a bearing on the atti-
ance extracted," which indicatethe percentageof variance tudinalvariablerelationshipsand thus suggests that discon-
in the constructsaccounted for by the measures. Whereas firmationof the EBiai-Aact link may be premature.
the majorityof measurementvariance is accountedfor by Having accepted unequally weighted measures and all
each of the four variablesfor which multiplemeasureswere estimatedpath coefficients as the best representationof the
available, the multiplicative variables (YBiai = 0.63, model proposed in Figure B, attentionturns to how well
ENBjMCj = 0.62) explain a smaller amount of measure the total model fits the data. The overall goodness-of-fit
variance than do the others (Aact = 0.88, Bi = 0.81). statistic, shown in Table 3, suggests that the model does
While the individualmeasurereliabilitiesand loadings (Ki) not sufficiently explain all observed sample covariances
are uniformly high for the Aact and BI indicators,there is (Xl24 = 198.16, p < 0.001). However, use of the X2 as an
more variationamong the multipliedvariable indicatorre- absolute index of fit is open to question, and there is some
liabilities and loadings. This is to be expected because the agreementthat it should be used as a guide ratherthan as
bipolar adjectives composing the Aact and BI instruments a rule (see review by Bentler 1980). In the present study
are meant to be parallelform items. On the other hand, the the amount of explained variance supportsthe theoretical
weighted beliefs that compose EBiai and ENBjMCj,while model; however, the failure to reject the residual matrix
meeting the consistency presumptionof the latent variable null hypothesis suggests that more is needed in terms of
techniquein this instance, are composites based on distinct variables, paths, and so on. Two considerationsneverthe-
outcomes and referents. It is interesting to note that the less indicate that the proposed model is adequate.
price belief, which was directly manipulated,was the least First, consonant with the work of Maruyamaand Mc-
reliable and contained the smallest measure variance ac- Garvey (1980), the mean absolute value of the differences
countedfor by cognitive structure(EBiai). Partof the price between the data and the model-reproduced correlation
belief error may be attributableto common method vari- matrix (excluding diagonal elements) was found to be
ance, which is discussed later. 0.051, whereas the mean correlationwas 0.541. Thus, the
discrepanciesbetween observed and predictedrelationsare
Causal Model Fit small. However, the x2 value is sensitive to lack of fit
anywherein the matrix, and an examinationof the residual
Under "Causal model" in Table 2, the proportionsof matrix revealed that larger discrepanciesoccur in attempt-
shared variance are separatedinto those for the structural ing to fit the data to the standardpsychometricassumption
model, which assume perfect measurement,and those for that correlations among the measurement errors (Eij) are
the total model, which reflect the inclusion of measurement zero. In fact, relaxing this assumptionfor 15 of the error
error. These proportions, which are goodness-of-fit indi- covariances selected on an empirical basis (for procedures
cators, show the amount of variance in the endogenous
variables accounted for by their respective predictors.The 9For a discussion of the relationshipsamong reliability, extractedvar-
dichotomous scoring of the exogenous variables (CI, NI) iance, and goodness-of-fit, see Fornell and Larcker(1981).

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
FORMATION
INTENTION
BEHAVIORAL 273

TABLE 3 error, the overall interpretation of the structuralmodel,


CHI-SQUARE GOODNESS-OF-FIT TESTS FOR VARIANTS OF which is of primaryinterestin this paper, does not change.
THE PROPOSED MODEL Hence the evidence presentedthus far is supportiveof the
model shown in FigureB. The analysesthatfollow consider
Degrees of discrepancies between correlated and uncorrelated mea-
freedom 2
Model surement-errormodels.
The second considerationinvolves nested formulations,
1. Proposed model (FigureB) 114 198.16a
which were compared to the proposed model following
More restricted: Bentlerand Bonett's (1980) suggestions and the precedents
1a. Removing P42 P12 116 259.91a set by Bentler and Speckart (1979, 1981). Data fit com-
1b. Removing P42 115 219.68a parisons (shown in Table 3), were made for three more
1c. Removing P12 115 226.77b restrictedmodels (la, lb, and 1c), which were obtainedby
Less restricted: constraining the crossover paths in Figure B to be zero,
id. Adding ,,3, p31 112 194.22a thereby producing models more in line with the one rep-
1e. Adding p,.3 113 195.20a resented in Figure A. In all three cases, the x2 values in-
1f. Addingp31 113 197.71b creased and the differences were statistically significant,
1g. Addingp., P31 112 194.22a
1h. Adding P. 113 194.23a suggestingpoorerfits thanthose obtainedwith the proposed
model. Five less restrictedmodels (ld-lh) were obtained
Alternative path: by replacing zero constraints with estimated crossover
1i. Replacing P42 with P, 1 114 205.50a paths; in all five cases, the small decreases in x2 values
were not statisticallysignificant. Thus the proposedmodel
Parameter significance tests is simpler and as viable as more complex alternatives.Re-
gardlessof statisticalsignificance, the fit decreasesobtained
1 vs. 1a. 2 61.75a
with the more restrictedmodels are of considerablygreater
1 vs. 1b. 1 21.52a
1 vs. 1c. 1 28.61a magnitudethan the fit increases obtained with the less re-
1 vs. 1d. 2 3.94 strictedmodels. An alternativemodel (li) replacingP42 with
1 vs. 1e. 1 2.96 343 produced a slightly poorer fit (Xl24 = 205.50, p <
1 vs. if. 1 .45 0.001) than the proposed model. The proposed model
1 vs. 1g. 1 3.94
1 vs. 1h. 1 3.93 standsup well when comparedto competing specifications.
a< 0.001
b < 0.01 Findings ConcerningVariableRelationships
Given supportfor the proposed model, the standardized
see Sorbom 1979) resulted in an acceptable fit (x99 = path coefficients for the causal model (Table 2) are of in-
105.63, p > 0.31). Since 10 of the 15 measurementerror terest. While the low weight estimated for Yii (0.10) in-
correlationsinvolved outcomes (El - E4) and attitudeitems dicates that the exogenous variable cognitive information
(E9 - E12) which, consistentwith acceptedprocedures,used had a small effect on cognitive structure(XBiai) relative to
the same set of evaluative adjectives, this finding suggests the effects of normative information(Y12 = 0.29), when
the presenceof common methodvariance. (Whenrestricted measurementerrorswere correlatedthese weghts were ap-
to these 10 E, x2 = 120.35, p > 0.13.) While this post proximatelyequal. Consequently,weight interpretationsare
hoc explanationis speculative, it is strongly reinforcedby tenuous. Nevertheless, the manipulationof normative in-
the finding that the structuralmodel interpretationdid not formation and the formation of normative structure(NS)
change significantly. The relative path coefficient values both had effects on cognitive structure,as did the manip-
remained largely unchanged, with the exception of in- ulation of its direct antecedent,cognitive information(CI).
creases in -Ylland P41and the statisticalsignificanceof the Normative structure(ENBjMCj)had approximatelyequal
-Yllpaths (p < 0.01) for both the 15 and 10 measurement- relationshipswith its hypothesizedcauses (Y21 = 0.33, Y22
error correlation models. These paths involve the latent = 0.29, and I21 = 0.22), includingcognitive information
variables (XBiai and Aact) with measurementerror inde- (Y21). Thus the expected crossover effects were present.
pendenciesthat have been called into question. These find- These findings are tempered by the unreliablemeasure of
ings, which empirically pinpoint the lack of overall fit, the price belief, which was the directly manipulatedbelief
suggest possible measurement problems and support the in attitudinalstructure, and by the possibility of method
overall good fit for the structuralmodel. In fact, removing variance.The introductionof correlatedmeasurementerrors
the measurementmodel completely to obtain a full infor- did increase the value of Y IIrelative to otherpaths, as well
mation maximum-likelihoodfit (i.e., assuming the latent as the reliabilityof price. While the process underlyingthis
variables to be perfectly measured, which is essentially a result is unknown, the result itself does not alter the con-
path analytic model) produces a X=2 = 10.14, p > 0.52. clusion regarding manipulation crossover effects even
Thus the lack of fit appearsattributableto method variance thoughrelative weight interpretationsinvolving y1I and 141
which, when introducedinto the model, causes a slight Yii are tentative.
and 041 inflation. With or without correlatedmeasurement The standardizedweights representingexogenous vari-

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
274 THE JOURNAL OF CONSUMER RESEARCH

able relationships(7Yii= 0.25) are generally smaller than Implicationsfor BehavioralIntentionTheory


those representingendogenous variable relationships (,ij
= 0.48). These differences may be due to the attenuating Test Methods. Previousmethodsused to test aspects of
effects of the CI and NI binary scoring. The weights in- this theory, based on independentOLS tests of models de-
volving the exogenous variables (yi) also serve as an ex- rived from Equations 1 through3, do not providethe depth
perimentalmanipulationcheck (Bagozzi 1977) indicating of analysis necessary to capturethe complexity of the the-
thatmain and interactiveexperimentaleffects were present. oretical network. Given the availabilityof techniques such
For the endogenous variables, the relationshipbetween as the one used in this paper, the centralequationsseem to
the attitudinalvariables (P41 = 0.31 is weaker than that have outlived their purpose. This paper clearly demon-
between the normativevariables(P32 = 0.75). In addition, stratesthe advantagesof path analytic procedures(see also
normative structuremore strongly affects attitude (P42 = Dickson and Miniard 1978).
0.63) than does cognitive structure(P41 = 0.31), although
their relative influences were equal with correlatedmea- MeasurementError. The proceduresused in this paper
surementerror. Aact nevertheless appearsto be related to explicitly accounted for measurementerrorwhich, if pres-
both normativeand attitudinalvariables.On the otherhand, ent and not acknowledged, can bias the structuralmodel.
SN is not related to attitudinal variables; the addition of Few of the previous studies investigatingthis theory have
paths indicating such relationships (the less restricted reported reliability estimates or used multiple-item mea-
models in Table 3), did not significantly improve the fit. sures, much less isolated the effects of measurementerror.
While the normativevariables were expected to predomi- When reliabilitieshave been reported,they have been high;
nate in the predictionof behavioralintention(BI), viewing despite the high reliabilitiesreportedby Ryan and Bonfield
the Aact and SN weights (P54 = 0.48, p53 = 0.45) in (1980), for example, measurementerrordid influence the
isolation is misleading. Takentogether,the resultsfrom the structuralmodel results. The existence of method variance
entire model suggest that CI and NI influence BI, and that would contradictSchwartz and Tessler's (1972) findings,
the effects of both CI and NI are mediated by normative which were obtained with less powerful analyses. It seems
and attitudinalvariables. In addition, the effect of CI on reasonable that the similarity of items, together with the
Aact is mediatedby both EB ai and XNBJMCJ, whereasthe simple paper-and-pencilformat, would produce common
effect of NI on SN is mediated only by XNBjMCJ.These method variance, thereby biasing latent variable relation-
findings supportthe contention that attitudinaland norma- ships upward, as suggested by the findings in this study.
tive variables have separable roles in the model, despite Yet Bagozzi (1981b) and Ryan and Bonfield (1980) were
their interdependency. able to predict actual behavior that was observed indepen-
dently of the proceduresused to measurethe predictorvari-
ables. In fact, to avoid method variance, Bagozzi deviated
DISCUSSION markedlyfrom Fishbein's notion of ai in his measurement
It seems clear from the findings thatbehavioralintention procedures;however, in subsequent analyses of his data,
is not a function of parallel, independentsets or of insep- he claimed convergent, discriminant,concurrent,and pre-
arable attitudinaland normativevariables, but of a rather dictive validity (1981a) and also utilized a model in which
complex set of interdependencies.The complex influences various expectancy-value and overall-affect measurement
that were found are consistent with Ajzen and Fishbein's errors were correlated (1982). Experimental approaches
writings. The specific finding-that normative variables comparingequivalentsamples, each utilizingdifferentmea-
more stronglymediatethe experimentaleffects of cognitive sures that would be considered valid representations of
and normativeinformationon intentionsthan do attitudinal evaluation, would seem to be an appropriateway to begin
variables-is of minor importanceby itself; it follows the investigatingthe existence, nature,and impactof correlated
expectation that an expert informant'sendorsementwould measurementerror.
influence subjects' organizationof new informationabout
an unfamiliarbrand. More to the point, this finding is apt Past Studies. The third implication for behavioral in-
to be situation-specific,and will take on importantmeaning tentiontheory follows directly from the firsttwo: past stud-
only when it can be viewed with the results from studies ies incorporatingsimpler analyses that supportthe theory
conducted in different situations (cf. Bentler and Speckart should be questioned, whether they supportit by manipu-
1981). This first test of the complete Fishbein and Ajzen latingthe situation(Ajzen and Fishbein 1970), or by a priori
intention formationparadigmreveals complex interdepen- prediction of attitudinal or normative variable predomi-
dencies among attitudinaland normative variables. Six nance (e.g., Gur-Arieet al. 1979; Wilson, Mathews, and
issues that emerge from this study's findings are discussed Harvey 1975). Ryan and Bonfield (1980) found that fewer
in the following two sections. Four involve past findings respondentswho scored low on normativeand high on at-
concerningaspects of the tested network;the fifth considers titudinalbeliefs, or vice versa, performedan overt behavior
those findings within the context of recent studies that have than did those who scored either high or low on both mea-
examinedexpandedstructuralaspects of the model. Lastly, sures. While this result is unusual enough to demand a
implicationsfor the role of normativevariablesin consumer replication, it does suggest that there is much we do not
information-processingresearchare discussed. know about the interdependencyof attitudinaland norma-

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
BEHAVIORAL
INTENTION
FORMATION 275

tive variables. Previous studies, using simplermethodsthat studies that employ it difficult. (The previously cited Ryan
do not account for simultaneousrelationships, should be and Bonfield (1980) study did not employ simultaneous
used as guidance for furtherelaborationand research. estimationprocedures, nor did it include SN.) These three
structuralequation studies, togetherwith the presentwork,
Interdependency. The interdependency of attitudinal provide guidance for elaboratingthe basic theory. Research
and normative belief variables that largely motivated this along these lines is importantto the field of consumerbe-
researchraises the questionof whetherthese variablesought havior with regardboth to direct applicationsof the theory
to have been considered to be separate in the first place. (e.g., Engel, Blackwell, and Kollat 1978; Sternthal and
The holistic approachtaken in this study revealed different Craig 1982) and to the framework it provides for other
mediating effects for these variables, thus supportingthe researchareas (e.g., Crosby and Taylor 1981; Mitchell and
utility of the "distinct but related" position. This evidence Olson 1981).
is not without its limitations, however. The test for dis-
criminantvalidity-i.e., that the proportionof measure- ImplicationsBeyond Behavioral
ment variance accounted for by the constructexceeds the IntentionTheory
sharedvariancein the structuralmodel (Fornelland Larcker
1981)-was barely met (for >Biai, 0.63 > 0.60; for The strong role of normativevariables found in this re-
2NBJMCJ, 0.62 > 0.47; for Aact, 0.88 > 0.78; and for BI, searchhas implicationsbeyond behavioralintentiontheory;
0.81 > 0.71). However, the variablesthemselves were op- indeed, it is congruentwith hypotheses and evidence out-
erationalizedin a situationin which they were not expected side the theory in question. For example, others have ar-
to be independent; thus, additional evidence should be gued that attitudinaland normativeinfluences are interde-
sought in situations where independence is theoretically pendent (Karlins and Abelson 1970; Krech, Crutchfield,
appropriate(see discussion by Fishbein and Ajzen 1981). and Ballachey 1962; Sherif 1958), and empirical studies
Elaborations of the present methodology should also be have found that (1) normative informationwas more fa-
pursued. Holbrook (1981), for example, has used tech- vorably evaluated if it supported the listeners' views
niques such as conjoint analysis (to which the presentdata (Manis, et al. 1974); (2) respondentsmore readily accepted
are not amenable) in conjunctionwith path analysis to ex- an attitude in line with reference group majority opinion
plore the mediatingeffects of beliefs. When combinedwith (Cvetkovich and Baumgardner1973); and (3) respondents
structuralmethodsthatincorporatemeasurementerror,such were more likely to act in accord with stated attitudes if
techniquesmay shed additionallight on the natureof atti- such attitudes were consistent with the actions of others
tudinaland normativebelief interdependency. (Norman 1975). Although these studies are not always di-
rectly comparable, they do support the general notion of
Behavior. The present study did not include behavior, attitudeand social variable interdependency.
although three studies that employed structuralmodeling Inferential belief formation is hypothesized to have
have includedit (Bentlerand Speckart1979, 1981; Bagozzi caused the results of the present study. This process may
1981b). Bentler and Speckartfound that behavior was di- account for the widely observed finding that highly expert
rectly influencedby attitudesand previous behavior in ad- sources producemore positive attitudestowardthe position
dition to being mediatedby intentions, and that the effects advocatedin a message than do less expert communicators
of attitude and intention on behavior may depend on the (Sternthal,Phillips, and Dholakia 1978). Despite previous
substantive domain under consideration. Unfortunately, failure (Ryan 1978), it may be possible to develop a nor-
their results are severely limited by the use of retrospective mative construct that reflects only Kelman's compliance
self-reports of behavior, commonly thought prone to de- process and is thus independentof attitude (which would
mandartifacts,andby Aact, SN, andB1 measuresthatdiffer then reflect the processes of identificationand internaliza-
markedlyfrom Fishbein's procedures.While they did foot- tion); nevertheless, it seems more promising to accept in-
note Fishbein and Ajzen (1975, pp. 226-228) for examples terdependency.This view suggests that normativevariables
of acceptable measure alternatives,Bentler and Speckart's may prove useful in extending informationacquisitionand
measures do not resemble those examples. Their studies retrieval paradigms (cf. Bettman 1979). Consequently, it
also do not include attitudinalor normativebeliefs. Bagozzi seems useful to speculate about situationswhere attitudinal
(1981b) found that intentions did mediate the and normativeinterplaywould be cued.
attitude-behaviorrelationshipand that past behavior atten- In general, cuing seems likely to occur when concept
uated the attitude-intentionrelationship. He reportedthat learningis takingplace-e.g., in eitherlimited or extensive
SN did not have a separable effect (1981a, 1981b). Ba- problem solving (Howard 1977). There are a number of
gozzi's SN measure does reflect the general theoretical accepted situations (cf. McGuire 1973) that increase this
meaning, but it also introducesthe notion of a discussion likelihood, such as unavailabilityof product or brand in-
with valued others and uses "ought to-ought not to" as formation,the consumer'sdesire to abstract(Wright1975),
bipolar anchors. Fishbein (1976) has claimed that subtle and the availabilityand credibilityof an informantsuch as
deviations may change conceptual meaning and empirical an opinion leader (Gur-Arieet al. 1979). In the lattercase,
results. Unfortunately,there is no commonly accepted way beliefs may be formed on both an inferentialand an infor-
of measuringSN, which makes comparisonsamong the few mational basis (cf. Minard and Cohen 1981). There are a

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
276 THEJOURNALOF CONSUMERRESEARCH

numberof accepted methods for increasingthe saliency of REFERENCES


relevant others (see Sternthal and Craig 1982) that may
Ahtola, Olli T. (1976), "Toward a Vector Model of Intentions,"
enhance attitudinal-normative interplay (e.g., use of a
in Advances in ConsumerResearch, Vol. 3, ed. Beverlee B.
spokespersonwho is identifiedwith the product,brand, or Anderson, Cincinnati:Association for ConsumerResearch,
its usage situation). 481-484.
In concept utilizations-for example, in some cases of Ajzen, Icek and MartinFishbein (1970), "The Predictionof Be-
routine problem solving, or when informationretrieval is havior from Attitudinaland Normative Variables," Journal
internal (Bettman 1979)-there is general agreementthat of ExperimentalSocial Psychology, 6, 466-487.
informationis selectively cued and that cuing may not be and Martin Fishbein (1972), "Attitudes and Normative
entirely consistent over time. While both attitudinalout- Beliefs as FactorsInfluencingBehavioralIntentions," Jour-
comes and normativereferentsare easily cued for familiar nal of Personality and Social Psychology, 21, 1-9.
brands in research settings (e.g., Ryan and Etzel 1976), and MartinFishbein (1973), "Attitudinaland Normative
Variables as Predictorsof Specific Behaviors," Journal of
there is no researchthat suggests situationsin which such
Personality and Social Psychology, 27, 41-57.
cuing would occur in actual decision environments.How- Bagozzi, RichardP. (1977), "StructuralEquationModels in Ex-
ever, Sternthalet al. (1978) provide a detailed review of perimentalResearch," Journal of MarketingResearch, 14,
the conditions under which informantexpertise and trust- 209-226.
worthiness have been found to affect social influence and 1a), "An Examinationof the Validityof Two Models
(198
attitude, and this could serve as guidance for hypothesis of Attitude," Multivariate Behavioral Research, 16,
development. 323-359.
(1981b), "Attitudes, Intentions, and Behavior:A Test of
Some Key Hypotheses," Journal of Personality and Social
Psychology, 41, 607-627.
CONCLUSION (1982), "A Field Investigation of Causal Relations
Among Cognitions, Affect, Intentions, and Behavior," un-
This researchhas demonstratedthe usefulness of consid- publishedworking paper, Institutefor Consumerand Behav-
ering intentions as though they were formed from interde- ioral Research, University of the Saarland,West Germany.
pendent yet separable attitudinaland normativevariables. Bentler, Paul M. (1980), "MultivariateAnalysis with LatentVari-
These findings are subject to the usual limitationsconcern- ables: Causal Modeling," AnnualReview of Psychology, 31,
ing generalizabilityand the possibility of demand artifacts 419-456.
that is inherent in the tightly controlled artificial setting and Douglas G. Bonett (1980), "Significance Tests and
employed. Some speculationhas also been offered to guide Goodness of Fit in the Analysis of CovarianceStructures,"
future investigation into the different conditions and pro- Psychological Bulletin, 88, 588-606.
cesses that may give rise to differentmodel configurations. and George Speckart (1979), "Models of Attitude-Be-
A secondary objective involved the usefulness of havior Relations," Psychological Review, 86, 452-464.
and George Speckart(1981), "Attitudes 'Cause' Behav-
Joreskog's methodology for simultaneously examining
iors: A StructuralEquationAnalysis," Journal of Personality
complex theoretical networks and the impact of measure- and Social Psychology, 40, 226-238.
ment error.With regardto measurementerror,the issue of Bettman, James R. (1979), An InformationProcessing Theoryof
common method bias was reopened concerning the attitu- ConsumerChoice, Reading, MA: Addison-Wesley.
dinal items. This is a fruitful area for future research, but Bourne, Francis S. (1957), "Group Influence in Marketingand
the present evidence is not sufficiently damaging to cause Public Relations," in Some Applications of Behavioral Re-
us to question the theory's basic validity. Much less can be search, Chapter 6, eds. R. Likert and S. P. Hayes, Paris:
said aboutthe normativevariables. In contrastto attitudinal UNESCO.
investigations, only a few studies have incorporatedSN, Carnegie-MellonUniversityMarketingSeminar(1978), "Attitude
and they appearto have used somewhatdifferentmeasures; Change or Attitude Formation?An UnansweredQuestion,"
multiple SN indicators have yet to be developed. When Journal of ConsumerResearch, 4, 271-276.
Cronbach, Lee J. (1951), "Coefficient Alpha and the Internal
investigatingthese issues, researchersmust be careful not Structureof Tests," Psychometrika, 16, 297-334.
to repeatthe errorsof the past (cf. Ryan andBonfield 1975), Crosby, Lawrence A. and James R. Taylor (1981), "Effects of
which have largely involved inadequateattentionto rules Consumer Information and Education on Cognition and
of correspondence. Fortunately, conceptual guidance is Choice," Journal of ConsumerResearch, 8, 43-56.
abundant(Ahtola 1976; Fishbein 1976; Fishbein and Ajzen Cvetkovich, George, and Steve R. Baumgardner(1973), "Atti-
1975, 1981; Lutz 1976; Miniard and Cohen 1981; Ryan tude Polarization: The Relative Influence of Discussion
1978); there is a need to focus it primarilyon improved Group Structureand Reference Group Norms," Journal of
conceptualizationand operationalizationof SN and its an- Personality and Social Psychology, 2, 159-165.
tecedents. Dickson, Peter R. and Paul W. Miniard(1978), "A FurtherEx-
aminationof Two LaboratoryTests of the ExtendedFishbein
Attitude Model," Journal of Consumer Research, 4,
[ReceivedNovember 1981. Revised July 1982.] 261-266.

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
BEHAVIORAL
INTENTION
FORMATION 277

Engel, James F., Roger D. Blackwell, and David T. Kollat A Validation Study," unpublishedPh.D. Dissertation, Uni-
(1978), ConsumerBehavior, Hinsdale, IL: Dryden. versity of Illinois, Urbana, IL.
Farley, John U., Donald R. Lehmann, and Michael J. Ryan (1975), "First-Orderand Second-OrderCognitive Effects
(1981), "Generalizingfrom 'Imperfect'Replication," Jour- in AttitudeChange," CommunicationResearch, 2, 289-299.
nal of Business, 54, 597-610. (1976), "Conceptual and OperationalIssues in the Ex-
Fishbein, Martin (1967), "Attitude and the Predictionof Behav- tended Fishbein Model," in Advances in Consumer Re-
ior," in Readings in AttitudeTheoryand Measurement,ed. search, Vol. 3, ed. Beverlee B. Anderson, Cincinnati:As-
MartinFishbein, New York: John Wiley, 477-492. sociation for ConsumerResearch, 469-476.
(1976), "Extending the Extended Model: Some Com- (1977), "An ExperimentalInvestigationof Causal Rela-
ments," in Advances in Consumer Research, Vol. 3, ed. tions Among Cognitions, Affect, and BehavioralIntention,"
Beverlee B. Anderson, Cincinnati:The Association for Con- Journal of ConsumerResearch, 3, 197-208.
sumer Research, 491-497. (1978a), "Rejoinder," Journal of ConsumerResearch,
and Icek Ajzen (1975), Belief, Attitude, Intention and 4, 266-270.
Behavior:An Introductionto Theoryand Research, Reading, (1978b), "Rejoinder," Journal of ConsumerResearch,
MA: Addison-Wesley. 4, 276-278.
and Icek Ajzen (1981), "On ConstructValidity: A Cri- Manis, Melvin, S. Douglas Cornell, and JeffreyC. Moore (1974),
tique of Miniard and Cohen's Paper," Journal of Experi- "Transmissionof Attitude-Relevant InformationThrough
mental Social Psychology, 17, 340-350. a CommunicationChain," Journal of Personalityand Social
and BertramH. Raven (1962), "The AB Scales: An Op- Psychology, 1, 81-94.
erational Definition of Belief and Attitude," Human Rela- Maruyama, Geoffrey and Bill McGarvey (1980), "Evaluating
tions, 15, 35-44. Causal Models: An Application of Maximum-Likelihood
Fornell, Claes and David F. Larcker(1981), "Evaluating Struc- Analysis of StructuralEquations," Psychological Bulletin,
tural Equation Models with Unobservable Variables and 87, 502-512.
MeasurementError," Journal of MarketingResearch, 18, McGuire, William (1973), "Persuasion, Resistance, and Attitude
39-50. Change," in Handbookof Communication,ed. I. S. Pool et
Glassman,MyronandNancyFitzhenry(1976), "Fishbein
's Sub- al., Chicago: Rand McNally, 216-252.
jective Norm: TheoreticalConsiderationsand EmpiricalEvi- Miniard, Paul W. and Joel B. Cohen (1979), "Isolating Attitu-
dence," in Advances in Consumer Research, Vol. 3, ed. dinal and Normative Influences in Behavioral Intentions
Beverlee B. Anderson,Cincinnati:Associationfor Consumer Models," Journal of MarketingResearch, 16, 102-1 10.
Research, 477-480. and Joel B. Cohen (1981), "An Examinationof the Fish-
Gur-Arie, Oded, Richard M. Durand, and William 0. Beardon bein-Ajzen Behavioral Intentions Model's Concepts and
(1979), "Attitudinaland Normative Dimensions of Opinion Measures," Journal of ExperimentalSocial Psychology, 17,
Leaders and Nonleaders," Journal of Psychology, 101, 309-339.
305-312. and Peter R. Dickson (1979), "Item Order Effects in
Haley, Russell (1968), "Benefit Segmentation:A Decision Ori- Expectancy-ValueAttitudeInstruments,"in EducatorsCon-
ented ResearchTool," Journal of Marketing, 32, 30-35. ference Proceedings, eds. N. Beckwith et al., Chicago:
Holbrook, Morris B. (1981), "Integrating Compositional and AmericanMarketingAssociation, 44, 4-8.
DecompositionalAnalyses to Representthe InterveningRole Mitchell, Andrew A. and Jerry C. Olson (1981),"Are Product
of Perceptions in Evaluative Judgments," Journal of Mar- AttributeBeliefs the Only Mediator of Advertising Effects
keting Research, 18, 13-28. on Brand Attitude?" Journal of Marketing Research, 18,
HowardJohn A. (1977), ConsumerBehavior:Applicationof The- 312-332.
ory, New York: McGraw-Hill. Norman, Ross (1975), "Affective-Cognitive Consistency, Atti-
and Jagdish N. Sheth (1969), The Theory of Buyer Be- tudes, Conformity, and Behavior," Journal of Personality
havior, New York: John Wiley. and Social Psychology, 1, 83-91.
Jaccard,James H. and Andrew R. Davidson (1972), "Towardan Ryan Michael J. (1974), "An Empirical Test of a Predictive
Understandingof Family Planning Behaviors:An Initial In- Model andCausalChainDerivedFromFishbein'sBehavioral
vestigation," Journal of Applied Social Psychology, 2, IntentionModel and Applied to a PurchaseIntentionSitua-
228-235. tion," unpublishedD.B.A. dissertation,University of Ken-
Joreskog, Karl G. and Dag Sorbom (1978), LISREL:Analysis of tucky, Lexington, KY.
Linear StructuralRelationships by the Method of Maximum (1978), "An Examinationof an AlternativeForm of the
Likelihood, Version IV, Release 2, Chicago: National Edu- Behavioral Intention Model's Normative Component," in
cational Resources, Inc. Advancesin ConsumerResearch, Vol. 5, ed. H. Keith Hunt,
Kaplan,Abraham(1964), The Conductof Inquiry, San Francisco: Ann Arbor:Association for ConsumerResearch, 283-289.
Chandler. and E.H. Bonfield (1975), "The Fishbein Extended
Karlins,Marvinand HerbertI. Abelson (1970), Persuasion, New Model and ConsumerBehavior," Journal of ConsumerRe-
York: Springer. search, 2, 118-136.
Kelman, Herbert C. (1961), "Processes of Opinion Change," and E. H. Bonfield (1980), "Fishbein's IntentionsModel:
Public Opinion Quarterly, 25, 57-78. A Test of Externaland PragmaticValidity," Journal of Mar-
Krech, David, RichardS. Crutchfield,and Egerton L. Ballachey keting, 44, 82-95.
(1962), Individual in Society, New York: McGraw-Hill. and Michael J. Etzel (1976), "The Natureof Salient Out-
Lutz, RichardJ. (1973), "Cognitive Changeand AttitudeChange: comes and Referents in the Extended Model," in Advances

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions
278 THEJOURNALOF CONSUMERRESEARCH

in ConsumerResearch, Vol. 3, ed. Beverlee B. Anderson, tudinalData," in Advancesin Factor Analysisand Structural
Cincinnati:Association for ConsumerResearch, 485-490. EquationModels, eds. K. G. J6reskogand D. Sorbom,Cam-
and Morris B. Holbrook (forthcoming), "Importance, bridge, MA: Abt Books, 171-184.
Elicitation Order, and Expectancy X Value," Journal of Sternthal,Brian and C. Samuel Craig (1982), ConsumerBehav-
Business Research, in press. ior: An Information Processing Perspective, Englewood
and J. Paul Peter(1976), "Two OperationalModifications Cliffs, NJ: Prentice-Hall.
for Delineatingthe Relative Strengthof Attitudinaland Social Lynn W. Phillips, and Ruby Dholakia (1978), "The Per-
Influenceson PurchaseIntentions," in EducatorsConference suasive Effect of Source Credibility: A Situational Analy-
Proceedings, ed. K. L. Bernhardt,Chicago: AmericanMar- sis," Public Opinion Quarterly, 286-314.
keting Association, 39, 147-150. Tauber, Edward M. (1977), "Forecasting Sales Prior to Test
Schwartz, Shalom H. and RichardC. Tessler (1972), "A Test of Marketing," Journal of Marketing, 41, 80-84.
a Model for Reducing MeasuredAttitude-BehaviorDiscrep- Werts, C. E., R. L. Linn, and K. G. J6reskog, "Interclass Re-
ancies," Journal of Personality and Social Psychology, 24, liability Estimates: Testing StructuralAssumptions," Edu-
225-236. cational and Psychological Measurement,34, 25-33.
Sherif, Muzafer (1958), "Group InfluencesUpon the Formation Wilson, David T., H. Lee Mathews, and James W. Harvey
of Norms and Attitudes," in Readings in Social Psychology, (1975), "An Empirical Test of the Fishbein Intention
eds. E. E. Macoby, et al., New York: Holt, Rinehart & Model," Journal of ConsumerResearch, 1, 39-48.
Winston. Wind, Yoram (1978), "Issues and Advantages in Segmentation
Shuchman, Abrahamand Peter C. Riesz (1975), "Correlatesof Research," Journal of MarketingResearch, 15, 317-337.
Persuasibility:The Crest Case," Journal of MarketingRe- Wright, Peter (1975), "ConsumerChoice Strategies:Simplifying
search, 12, 7-11. vs. Optimizing," Journal of MarketingResearch, 12, 60-67.
Sorbom, Dag (1979), "Detection of CorrelatedErrorsin Longi-

This content downloaded from 14.139.242.60 on Tue, 12 Jan 2016 10:16:39 UTC
All use subject to JSTOR Terms and Conditions

Das könnte Ihnen auch gefallen