Four Multi Attribute Model

A Comparison of Four Multi-Attribute Models in the Prediction of Consumer Attitudes Author(s): Michael B. Mazis, Olli T. Ahtola, R. Eugene Klippel Source: The Journal of Consumer Research, Vol. 2, No. 1 (Jun., 1975), pp. 38-52 Published by: The University of Chicago Press Stable URL: http://www.jstor.org/stable/2489045 Accessed: 18/04/2009 05:47 Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at http://www.jstor.org/action/showPublisher?publisherCode=ucpress. Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit organization founded in 1995 to build trusted digital archives for scholarship. We work with the scholarly community to preserve their work and the materials they rely upon, and to build a common research platform that promotes the discovery and use of these resources. For more information about JSTOR, please contact [email protected].

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

http://www.jstor.org

A

Corn o

arison es

in

of

Four

Multi-Attribute

Prediction Attitudes

the

Consumer

of

MICHAEL B. MAZIS

OLLIT. AHTOLA R. EUGENE KLIPPEL* Three experiments were conducted on the predictability of four multiattribute models. Results support using the "adequacy" formulation over the Fishbein and Rosenberg models if the investigator's goal is maximization of explained variance. The utility of eliciting salient beliefs and the nonequivalence of value and prominence components is shown also.

Although instrumentality or expectancy theory has appeared in the psychological literature for at least four decades (Lewin, 1935; Tolman, 1932), only in the last few years have marketinig researchers begun to predict and understand consumers' attitudes and behavior directly from cognitive structure. Using salient beliefs which underlie consumer attitudes has considerable appeal to marketers because they often seek to change consumer preferences, and attitude change strategies can be most effective only if the true determinants of attitudes are known. Unfortunately, the consumer behavior literature has produced considerable confusion since a number of substantially different multi-attribute models have been developed to predict consumer attitudes. The purpose of this study is to clarify the situation by empirically testing the predictability of four approaches which have been used or proposed in consumer attitude studies. The three experiments which follow attempt to specify the circumstances under which particular models may be most appropriate and to indicate methodological problems which affect predictability of each model.

expectancy value theories is the notion that the strength of a tendency to act depends upon (1 ) the strength of the expectancy that the act will be followed by a consequence and (2) the value of that consequence to the individual. Various investigators have used this formulation to predict the strength of a tendency to act in such diverse situations as economic decisions, achievement-oriented behavior, maze behavior, and attitudes. The first component of the model, expectancy, focuses on the fact that when an individual chooses among alternatives with uncertain consequences, he anticipates that his choice will be followed by particular outcomes. Expectancy is the measurement of the likelihood that positive or negative outcomes will be associated with or follow from a particular act. Lewin (1944) and Edwards (1954) have referred to this component as the subjective probability of the consequences, given the act. Tolman (1955) and Rotter (1954) have used the more common term, expectancy, in their approaches to examining the determinants of behavior. In addition to the probability of an action-outcome association, the individual has some affective orientation (value) toward particular outcomes. To express attraction for a given outcome, Edwards (1954) employs the economist's description, utility; Lewin (1944) and Tolman (1955) use the Gestalt psychologist's designation, valence; and Rotter (1954) uses the learning theorist's term, reinforcement value. These two components determined what has been called performance vector, force, behavioral potential, aroused motivation, subjective expected utility, or attitude by various researchers.

THEORY Expectancy-Value Approach The expectancy-value approach has been used by numerous psychologists over the years to isolate the determinants of motivated behavior. Common to all * Michael Mazis is Associate Professor of Marketing, University of Florida; Olli T. Ahtola, Assistant Professor of Marketing, University of Florida; R. Eugene Klippel, Associate Professor of Marketing, The Grand Valley State College. 38

JOURNAL OF CONSUMER RESEARCH e Vol. 2 * JuLne1975

COMPARISONOF FOUR MULTI-ATTRIBUTE MODELS Expectancy-value theory also specifies how expectancy and value combine in determining choices. Although there are many possible ways of combining the two components, the force or motivation to perform a given act is usually determined by multiplying the value and expectancy components for each outcome and calculating the algebraic sum across outcomes. We would predict the strength of this impulse action to be a monotonically increasing function of the sum of the product of the expectancy and value. Rosenberg's Model Strongly influenced by the means-ends orientation, a number of attitude researchers at the University of Michigan began to focus on attitude in an expectancyvalue framework. Helen Peak stated that the affect attached to an attitude object is a function of: "(1) the judged probability that the object leads to good or bad consequences, and (2) the intensity of the affect expected from those consequences" (Peak, 1955, p. 154). In his doctoral dissertation at the University of Michigan, Milton Rosenberg (1953; 1956) developed a model to quantify the idea that attitude toward an object is "related to the ends which the object serves" (Peak, 1955, p. 153). Rosenberg collected data from 112 subjects dealing with attitudes toward free speech for Communists and attitudes toward removing Negro housing segregation. TI'ooperationalize his model, measurements were obtained on three specific variables: (1) judgments about the satisfaction derived from 35 goals, such as gaining power over people, everyone being assured a good standard of living, America having prestige in other countries, and all people having equal rights (value); (2) judgments about the probability that free speech for Communists or segregation would lead to or block the attainment of each of the 35 listed goals (expectancy); and (3) measurement of overall affect toward segregation and free speech for Communists. The Rosenberg model, which has been widely referred to in subsequent attitude research, was quantified as follows: i 16

where AO- attitude toward an object I, perceived instrumentality, the belief about the potential of the object for attaining or blocking the realization of valued state i value importance, valued state i's imporVitance to a person as a "source of satisfaction" n number of valued states.

39 Direct application of the Rosenberg model, using abstract values (not product characteristics), has been reported in several consumer attitude studies. The value model was used by Bither and Milier (1969), Hansen (1970), and Klippel and Bither (1972) to predict automobile, restaurant, book and mouthwash preferences. While values might have limited usefulness in product categories where clear differences on characteristics exist, they may be useful for product categories where brand differences are largely illusory. While Rosenberg's formulation was conceptualized around values, the model would appear equally valid using other cognitive elements, including product characteristics. Fishbein's Model Fishbein's model, based on behavior theory principles of mediated (secondary or conditioned) generalization, has hypothesized that attitude toward an act (or object) is a function of (a) the strength of beliefs about an act (or object) and (b) the evaluative aspects of these beliefs. Algebraically, attitude toward an act is expressed as follows: it

B, ai

Lat 1,=~1

where A,(.t attitude toward performing an act B, - an individual's belief about the likelihood that the behavior in question will result in outcome i = the person's evaluation of outcome i n - number of beliefs. Computationally, Fishbein's model is similar to Rosenberg's formulation, although they were derived from behavior theory and cognitive consistency theory, respectively. Although there are several theoretical and methodological differencesbetween the various theories that have dealt with the belief-attitude relationiship,the importantpoint is that all of them lead to the hypothesis that an individual'sattitude toward any object is a function of his beliefs about the object and the evaluative aspects of those beliefs (Fishbein, 1967, p. 395). However, no empirical evidence has been presented to show that the Rosenberg and Fishbein models, which use different wording to measure components, will produce equivalent findings. Differences Between Rosenberg's and Fishbein's Model The similarity of these two theories is striking but several differences should be noted. First, while Fish-

THE JOURNAL OF CONSUMER RESEARCH

40 bein's formula refers to the attitude toward an act of behavior Rosenberg is talking about the attitude toward an object. This difference is relevant here because Fishbein makes a distinction between these two kinds of attitudes whereas Rosenberg evidently does not. However, the study reported by Rosenberg in 1956 dealt with the attitude toward "allowing members of the Communist Party to address the public" which is clearly an attitude toward behavior even though no specific situation was defined temporally and spatially as has been done in Fishbein's studies. Second, it should be noted that while Fishbein's B, measures how likely it is (or to what extent) the behavior in question would result in a salient outcome, Rosenberg's Ii measures to what extent (or how likely it is) the behavior in question would result in an outcome or to what extent (or how likely it is) the behavior will result in blocking the outcome. That is, Fishbein measures only the expectancy of attaining the outcome while Rosenberg's "perceived instrumentality" measures also the expectancy of the opposite outcome. In other words, Rosenberg specifies two opposite consequences with respect to a value achievement, while Fishbein specifies only one. A given outcome may be quite improbable (low B^) through the behavior in question but it also may not create an obstacle or a hindrance to the achievement of that outcome later or by other means (which is what blocking means). That is, blocking goes beyond the mere improbability of outcome through a given behavior. For this reason, it could be argued that Rosenberg's Ik component has more discriminatory power than does Fishbein's Bi. The third difference arises from Fishbein's and Rosenberg's different theoretical background. Rosenberg bases his model on the functional approach to attitudes (Katz, 1960; Smith, 1949) whereas Fishbein explains the underlying theory of his model in terms of mediated (secondary or conditioned) generalization within a behavior theory framework. This leads Fishbein to accept any kinds of salient belief statements to form the elements in his model while Rosenberg tries to find consequences and goals which reflect what he calls important values. "Means to end" type of thinking may also have led Rosenberg to define his "perceived instrumentality" component in terms of "blocking" versus "'attaining"the valued state. Adaptation, of Rosenberg's and Fishbein's Models Substituting Importance for Evaluation in Value Component. One of the first adaptations of the Rosenberg and Fishbein expectancy-value models by applied behavioral scientists was the replacement of the evaluation term by a measure of importance. This substitution is quite understandable since Rosenberg referred to

this dimension as "value importance," thereby confusing many investigators. Industrial psychologists, attempting to predict the determinants of worker productivity or attitude toward employment, have often used this revised definition of value. For example, Vroom (1966) attempted to predict the attractiveness of organizations to students in industrial management for beginning their career. Subjects were asked: (1) to rate the importance of 15 goals, e.g., opportunity for advancement, variety in work assignments, and friendly and congenial associates; and (2) to select the three most "promising" organizations for them and to rate the 15 goals in terms of the extent to which they believed attainment was likely within each of the three organizations. While instrumentality is measured in a similar fashion as in Rosenberg's model, value is equated with importance. In the field of consumer behavior, the model developed by Cohen (Cohen and Ahtola, 1971; Cohen and Houston, 1971) follows the same procedure as Vroom's model except the objective is to identify salient product benefits, and then to measure how important these benefits are to the individual and to what extent he perceives each brand to possess these want satisfying properties. Cohen's measurement of expectancy differs slightly from Fishbein's and Rosenberg's measure since "possession" measures, indicating how much or how little of the product benefit is possessed by each brand, are obtained. However, the major difference is the substitution of importance for measurement of value. As Cohen (1972) points out: The advantageof this procedure is that if every one has the same direction of preferencealong the dimension (e.g., more is better-as in the case of cavity prevention for toothpaste), we are not simply asking how importantthe dimension is, but rather how important the attainment of a valued state is. This is similar to Rosenberg's approach, and under the assumption of common direction of preference, the greater the importance the greater the value. (p. 6) In other words, Cohen restricts the values in the model to those which are universally evaluated positive (or negative). Providing that the polarity of evaluation is highly correlated with the importance, Cohen's method should give results similar to Rosenberg's method. "Adequacy-Importance" Model. The most widely used multi-attribute model appearing in the consumer behavior literature has been labeled the "adequacy-importance" model (Cohen, Fishbein and Ahtola, 1972) and can be expressed as follows: h Pr De where

COMPARISONOF FOUR MULTI-ATTRIBUTE MODELS A, - an individual's attitude toward an object Pi importance of attribute (dimension) i for the person his evaluation of object with respect to the Di attribute dimension i n -number of attribute dimensions -

There are several variations in how the above components are measured such as the semantic differential versus constant sum scales, normalized versus nonnormalized scale values, positive-negative versus positive only scale values, direct measure Di versus o's distance from the ideal point, use of different Minkowski metrics, and so forth. These variations are discussed in detail by Wilkie and Pessemier (1972) anci will not be repeated here. This model clearly differs from the expectancy-value formulations developed by Rosenberg and Fishbein in two respects. First, it substitutes importance for evaluation in measurement of the value component. Second, product dimensions, rather than specific characteristics are used. Careful examination of the ''adequacy-importance"' model reveals that asking respondents to evaluate how satisfactory an object is with respect to an attribute dimension does not parallel Rosenberg's "perceived instrumentality" or Fishbein's "belief strength" measures. Rosenberg's "perceived instrumentality" as well as Fishbein's "belief strength" would be high whenever the individual strongly believes that the value is achieved (e.g., the brand possesses the attribute) regardless of how satisfactory the value itself is perceived to be. It is the purpose of the value component (Vi or a1) to measure the satisfaction with the value. As Cohen, Fishbein and Ahtola (1972) have pointed out, there is little question that the above procedure confounds both components of either Rosenberg's or Fishbein's model. This measure may be a relatively direct, though ambiguous, measure of Ii Vi or Bi ai, but in any case it is not either single component of any expectancy-value model. Cohen, Fishbein and Ahtola (1972) suggested that this measure be called "adequacy." It should be noted that use of the "adequacy" type of model is not confined to consumer behavior; industrial psychologists have used a somewhat similar approach. For example, Lawler (1968) and Lawler and Porter (1967) asked managers to judge how helpful a number of factors-pay, promotion, prestige, security, autonomy, and opportunities to use skills and abilities-were in obtaining rewards for their organization.1 The importance of the six factors to each man1 There is a slight conceptual difference between the Lawler and Porter technique and the approach used by most marketers. While marketers tend to ask how the satisfactory brands are with respect to certain attributes, Lawler and Porter asked respondents to judge how helpful certain factors (attributes)

41 ager was measured also. This procedure closely follows the two component "adequacy" model used by Bass, Hansen, Hughes, Lehmann, Sheth, Talarzyk and others in consumer behavior studies. While the "adequacy-importance" model is not an expectancy-value model, it does fall into the more general classification of linear compensatory multi-attribute models. Under some circumstances (to be explained later in this paper), the adequacy model may predict consumer attitudes more accurately and be more useful in developing attitude change strategies. Since the "adequacy" model is widely used and may offer some significant advantages (e.g., ease of measurement) over expectancy-value models, its relative predictability should be studied further. With a wide variety of multi-attribute models available, it was felt that some empirical comparison of models would be useful in providing guidance to consumer researchers. Three experiments were developed to contrast the relative predictability of four important approaches: Fishbein's model, "adequacy" model and Rosenberg's model using both values and product characteristics. The first experiment compared the four models across four product classes-mouthwash, cigarettes, toothpaste and automobiles. Since only two brands per product class were evaluated, analysis was restricted to examination of inter-individual correlation coefficients. Experiment #2 focused solely on soft drinks and compared the predictability of the "adequacy" and Fishbein models. The purpose of this experiment was to show that when complex product attributes are considered, the Fishbein model in its current formulation may offer very poor prediction. The third experiment was developed to clarify some methodological issues raised as a result of experiment #1 and to undertake additional data analysis. Three brands per product class were evaluated so that intraindividual as well as inter-individual correlation coefficients could be examined. EXPERIMENT ONE Method Four hundred and sixty-four undergraduate marketing students at the University of Florida were administered one of four questionnaires, each of which employed a different methodology, either Fishbein's model, Rosenberg's model with values, Rosenberg's model with product characteristics or the "adequacyimportance" model. Questionnaires were randomized within classes to avoid biases due to group differences. Each responjient provided responses for four product were in obtaining organizational rewards. In both cases emphasis is placed upon dimensions rather than the level or direction of a dimension.

THE JOURNAL OF CONSUMER RESEARCH

42 categories, two brands per product category. Only two brands per category were included to avoid subject fatigue. Instructions indicated that attitudinal responses were to be provided only when both brands in a product category had been used by a respondent. automobiles, The product categories-toothpaste, mouthwash and cigarettes-were selected for three reasons. First, it was felt that students could supply informed judgments about these products. Second, toothpaste, automobiles, and mouthwash have been used in previous consumer attitude studies (Bass and Talarzyk, 1972; Bither and Miller, 1969; Klippel and Bither, 1972; Mazis, 1972; Mazis, 1973) and information concerning salient attributes and values was available to the investigators. Third, cigarettes were included because several writers (Levy, 1959; Martineau, 1957) have hypothesized that consumers select cigarette brands based on images rather than product characteristics; therefore, the Rosenberg value model might be a useful predictive tool. Five attributes were used for

all product classes except automobiles where seven attributes were included. Ten values were included for all four product classes. (See Table 1 for a listing of attributes and values used.) Each questionnaire consisted of three parts. On the first page, the respondent was asked to provide his overall attitude toward the eight brands on seven-point bipolar scales labeled "extremely high appeal" and "extremely low appeal" at the end points. The second portion of the instrument contained evaluative or importance questions for all four product categories. For example, the Fishbein questionnaire asked, "When you are considering purchasing a car, how would you evaluate the following: An automobile which . . . gets poor gas mileage." Seven-point scales labeled "good" and "bad" at the endpoints were used. The Rosenberg attribute instrument asked customers to respond to "Buying a car which gets poor gas mileage provides me with . . ." on a seven-point scale labeled "maximum satisfaction" and "maximum dissatisfaction" at the

TABLE 1 BELIEFS USED IN FIRST EXPERIMENT EXPERIMENT #1 Product Dimensions ("Adequacy"Model) 1. 2. 3. 4. 5.

Effectiveness Color Price Taste/Flavor Kills germs

Product Characteristics (Fishbein &RosenbergModels) Mouthwash 1. 2. 3. 4. 5.

Lacks effectiveness Dark color Inexpensive Pleasant taste or flavor Potent germ-killingqualities

Values (R.osenbergValue Model) 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.

Keeping in good health Associating with the opposite sex Recognized as a leader Socializing with others Projectinga young dynamic image Protection against physical harm Thrifty Meeting new people Careful in personal cleanliness Complimentedby others

Cigarettes 1. 2. 3. 4. 5.

Flavor Tar & Nicotine Content Taste Strength Ease of draw

1. 2. 3. 4. 5.

Distinctive flavor Low tar & nicotine content Lacks taste Strong Draws easily

1. 2. 3. 4.

Decay prevention Taste/Flavor Freshens mouth Whitens teeth

1. 2. 3. 4.

High in decay preventioningredients Does not have.a pleasant taste Has strong mouth-freshening ingredients High in whitening characteristics

1-6. See values #1-6 for mouthwash 7. Avoiding worry &anxiety 8. Participatingin risky situations 9. Projectinga mature image 10. Living a sensible life

Toothpaste

5. Price

1-9. See values #1-2 and 4-10 for mouthwash 10. Living a sensible life

5. Low in price

Automobiles 1. 2. 3. 4. 5. 6.

Safety Gas mileage Styling Repair record Acceleration Quality of workmanship Price

1. 2. 3. 4. 5. 6. 7.

Safe Poor gas mileage Lacks currenthigh styling Rarely needs repairs Rapid acceleration Quality of workmanship Low in price

1-6. See values #2, 3, 4, 5, 7, 10 for mouthwash 7. Projectinga mature image 8. Viewed as intelligent 9. Living a sensible life 10. Competitive

43

MODELS COMPARISONOF FOUR MULTI-ATTRIBUTE endpoints. The Rosenberg value questionnaire posed questions in a similar format except that the evaluative statements were worded in the following form: "Associating with the opposite sex provides me with . . Subjects recorded their answers on a "maximum satisfaction-maximum dissatisfaction" scale. The adequacy instrument asked subjects to "indicate the degree of importance you attach to each of these attributes when buying a particular product" on sevenpoint bipolar scales. Endpoints were labeled "of more importance" and "of less importance" (Bennett and Scott, 1971), wording which encourages respondents to use a greater proportion of the scale than "important-unimportant," since all dimensions used are somewhat salient. Comparisons against an ideal brand used in several studies (Bass and Talarzyk, 1972; Talarzyk, 1969) and criticized by Cohen et al. (1972) were omitted. The third part of all four instruments consisted of instrumentality questions, using seven-point bipolar adjective scales. Fishbein's model asked, "What is the probability that . . . Pontiac GTO gets poor gas mileage" on a "probable-improbable" continuum. Rosenberg's attribute instrument stated, "Owning a car which gives poor gas mileage is attained by the purchase of Pontiac GTO," with responses recorded as to level of agreement. The adequacy model asked respondents to "Please indicate your judgment as to whether the particular brand gives you the desired satisfaction . . . Pontiac GTO: gas mileage." Responses were recorded on a scale labeled "very satisfactory-very unsatisfactory," coded from 7 to 1 to indicate the degree of satisfaction. The Fishbein and Rosenberg models used response categories from +3 to -3.

Results While several procedures for dealing with expectancy-value data have been advocated by different investigators, this study followed the practice of first multiplying the two components for each attribute or value then summing across beliefs to arrive at a total attitude score for each individual by product. Thus, each individual had eight attitude scores which could be compared against overall affect toward eight brands. The data were analyzed by computing correlation coefficients for each brand across individuals. Correlation coefficients for each brand and expectancy-value model are shown in Table 2. Correlations varied considerably: the average correlation for the adequacy model as .514 while r's were .410 and .414 for the Fishbein and Rosenberg attribute models, and .314 for the Rosenberg value model. The poor predictability of the value model is evident for all product categories except mouthwash. Surprisingly, there were only minor differences in average correlations for the four product classes. The highest average correlation was attained for automobiles (r - .47), while cigarettes received the lowest correlation (r - .35). The correlation coefficients for mouthwash, cigarettes and toothpaste were nearly identical. To compare the overall predictability of the four models (in a repeated measures design), analysis of variance using correlation coefficients as dependent variables (Jones, 1970) was employed. This procedure, using Z-values, is an extension of the techniques outlined by Fisher (1954, pp. 197-204) and Snedecor (1962, pp. 178-180) The chi-square value for groups (models) is compared with the chi-square value for

TABLE 2 CORRELATION COEFFICIENTS FOR FOUR EXPECTANCY-VALUE FORMULATIONS EXPERIMENT #1 "AdequacyImportance" Formulation

Fishbein Formulation (Product Characteristics)

Rosenberg Formulation (Product Characteristics)

Rosenberg Formulation (Values)

Mouthwash Scope Listerine Cigarettes Kent Marlboro Toothpaste Crest Colgate

(n = 100) .546 .182 (n-=42) .468 .673 (n = 116) .396 .621

(n = 96) .270 .406 (n = 31) .338 .551 (n = 119) .407 .272

(n = 94) .298 .395 (n = 39) .148 .353 (n = 117) .417 .320

(n = 98) .490 .341 (n = 35) .074 .236 (n = 112) .260 .237

.366

Automobiles

(n = 142)

(n = 148)

.473

Pontiac GTO Volkswagon Average r F3,28 = 3.71, p < .05.

(n = 136)

(n = 139)

.561 .588

.408 .574

.578 .483

.348 .305

.514

.410

.414

.314

Average r

.355

.366

44

THE JOURNAL OF CONSUMER RESEARCH

replications (brands), which is used as an error term to form an F-ratio. The significant (p < .05) F-value shown in Table 2 indicates that the average correlations obtained by the four models were not all equal. Posteriori analysis (Table 3) conducted on all six pairs of r's (Fisher, 1954, pp. 203-204), using multiple t-tests and the more conservative Tukey H. S. D. procedure (Kirk, 1969, pp. 88-90) revealed that the "adequacy-importance" formulation offers better predictability than the other three models. Statistically significant differences (p < .05) in predictability were found between all forinulations with the exception of the Fishbein and Rosenberg product characteristic models. While it had been assumed that the Fishbein and Rosenberg instruments will yield equivalent results, no empirical test of alternative measurement strategies had been conducted. This study indicates that no significant difference in predictability results in using the two formulations. Finally, this study shows that the model using abstract values offered the poorest predictability. Discussion While these results support the proposition that the "adequacy" and value models offer the best and the worst prediction of brand affect, respectively, among the four formulations tested, there are several issues which should cause the reader to view these findings as equivocal. As a result, two additional experiments are reported to contrast more clearly differences between approaches. First, it should be noted that all the values and product characteristics used may not have been salient for the subject population under investigation. Beliefs and TABLE 3 POSTERIORIANALYSIS OF CORRELATION COEFFICIENTSFOR FOUR EXPECTANCY-VALUE FORMULATIONS USING t-TESTS EXPERIMENT #1 Fishbein Rosenberg Formulation Formulation (Product (Product "AdequacyImportance" CharacterCharacterFormulation istics) istics) Fishbein Formulation (Product Characteristics) Rosenberg Formulation (Product Characteristics) Rosenberg Formulation (Values) * p < .05.

values selected for use in this study were in most cases identical to those used by other researchers; however, often these researchers have failed to indicate their exact procedure for eliciting salient beliefs and values. In addition, due to different subject populations and temporal factors, salient beliefs and values used in othei samples may not have been relevant for this investigation. Second, there was a change in the procedure used by Rosenberg which may have affected the results obtained from using his model. While Rosenberg's measure of instrumentality was obtained by having subjects sort values into 11 categories ranging from "completely attained" to "completely blocked" by the object, this study followed Bither and Miller's (1969) adaptation of Rosenberg's methodology. Respondents were asked to indicate their level of agreement with a statement, such as "Keeping in good health is enhanced by the purchase of Scope Mouthwash," rather than specifying the amount that purchase of a brand tended to block or enhance the attainment of a particular value state. Third, only one measure of affect was used, a sevenpoint bipolar scale labeled "extremely high appeal" and "extremely low appeal," was used (Bither and Miller, 1969; Klippel and Bither, 1972). Since this measure has not been used very often in attitude research studies and since it was the only dependent measure used, the reliability of the measure may be questioned. Fourth, since only two brands per product category were used, only inter-individual analysis was undertaken. Several researchers (Bass, 1972; Nakanishi and Bettman, 1973; Talarzyk, 1969) have argued that intra-individual analysis of multi-attribute data, focusing on correlations for each individual, is a more appropriate method of analyzing the data since the models are designed to predict choice behavior of a single person. Finally, prediction of brand affect was the sole criterion for choosing the "best" model. However, it has been argued that understanding or representation of cognitive structure is a more appropriate measure of the "goodness" of a model (Bettmian, Capon and Lutz, 1974; Wilkie, McCann and Reibstein, 1974). EXPERIMENT TWO

2.68*

Method 2.61*

.67

4.96*

2.33*

2.36*

Two hundred and eleven students in eight freshman classes at Western Illinois University were used in this study. Twenty-nine subjects were discarded from the analysis because of missing data or noncompliance with instructions (i.e., failure to complete accurately other parts of the instrument which was used for another research investigation), leaving 182 usable responses. All subjects provided judgments of three soft drink

45

MODELS COMPARISONOF FOUR MULTI-ATTRIBUTE TABLE 4

Results

BELIEFSUSED IN SECOND EXPERIMENT EXPERIMENT #2 ProductCharacteristics (Fishbein Model)

ProductDimensions ("Adequacy"Model) Soft Drinks 1. 2. 3. 4.

Flavor Sweetness Carbonation Calorie Content

1. 2. 3. 4.

Lemon-lime Flavor Sweet Carbonated Calories

brands, Fresca, Squirt and Regular 7-Up. Identification of salient beliefs associated with the choice between brands of a soft drink was accomplished by using a free association test to identify beliefs in subjects' habit-family-hierarchies (Fishbein, 1967). Fifty-two undergraduate students were asked, "what kinds of things (e.g., product attributes, consequences of consuming the product, etc.) do you look for or take into consideration when you select a brand of soft drink for your own consumption as a refreshment?" The mean, median and modal number of concepts supplied was four. The most frequently mentioned concepts were taste/flavor, price, carbonation, sweetness and calorie content/diet drink.2 Since the dependent variable focused on having subjects evaluate several brands which would be distributed free, price was excluded, and the remaining four concepts were included in the experiment (see Table 4). Each subject was administered the measures in the following order: (1) Fishbein's ai; (2) Fishbein's Bj; (3) importance measures (Pi); (4) "adequacy" measures (D); (5) attitudes toward selecting brands: Scales for Fishbein's ai and Bi were selected from Fishbein and Raven's (1962) A- and B-scales, respectively. Each concept, e.g., carbonation in a soft drink, was evaluated on three seven-point bipolar scales, goodbad, unpleasant-pleasant and satisfying-rewarding. These scales were summed to produce an a.1for each concept. Similarly, subjects responded to each brandconcept association (Bi), e.g., Fresca is carbonated, on three seven-point scales, probable-improbable, unlikely-likely and true-false. Importance measures were collected using a ninepoint scale labeled, "not at all important to me" and "very important to me" at the endpoints. Measures of satisfaction were obtained using the procedure described in experiment #1. Finally, attitudes toward selecting brands were obtained by asking subjects to to drink respond to the statement, "selecting after this study is" on four bipolar scales labeled, pleasant-unpleasant, foolish-wise, good-bad and punishingrewarding. 2 The manner in which these concepts were operationalized to represent the Fishbein model was approved by M. Fishbein (personal communication).

The data in Table 5 reveals that the "adequacy" model offered better prediction of attitude toward selecting a soft drink brand to consume after the study than did the Fishbein model. Comparison of interindividual correlation coefficients for complete models (Pi Di vs. Bi AJ) shows that correlation coefficients for 7-Up, Fresca and Squirt were .47, .61, .62 and .13, .35, and .32 for the "adequacy" and Fishbein model, respectively. The t-tests conducted on the transformed correlation coefficients (through r to Z transformations) show that there were statistically significant differences in predictability between the "adequacy" and Fishbein models across all three brands. Intra-individual analysis was accomplished by computing correlation coefficients for each individual, using his judgments on the three soft drink brands as observations. Since the distribution of correlation coefficients approximates the normal distribution, t-tests were performed to test the differences between the distribution of the Z transforms of correlation coefficients for the "adequacy" and Fishbein models. The results are consistent with the inter-individual analysis; there were statistically significant differences between the Di only vs. Bi only models (t = 8.07, 181 df, p < .001) and between the P, Di and Bi Ai models (t_ 11.01, 181 df, p < .001). As indicated previously, there has been some controversy concerning the similarity between "adequacy" and expectancy-value models. While it has been maintained that the importance or prominence component in the "adequacy-importance" model is similar to the value importance (Vi) or evaluation (ai) component in the Fishbein model, Table 5 shows substantial differences between expectancy-value and "adequacy" models when the at or Pi components are removed or retained in the models. Eliminating the prominence component enhances the predictability of the "adequacy" model. Inter-individual analysis shows correlaTABLE 5 INTER-INDIVIDUAL CORRELATION ANALYSIS EXPERIMENT #2 "Adequacy"Model

7-Up Fresca Squirt

Fishbein Model

Di Onlya

Pi D,b

B,iOnlya

B, a,b

.55 .69 .74

.47 .61 .62

.06 .13 .03

.13 .35 .32

a t-values for differences between correlation coefficients for Di only vs. Bi only = 4.66, 6.83 and 8.77 for 7-Up, Fresca and Squirt, respectively. All differences are significant at the .001 level. b t-values for differences between correlation coefficients for Pi Di vs. Bi ai = 3.61, 4.22 and 3.76 for 7-Up, Fresca and Squirt, respectively. All differences are significant at the .001 level.

46

THE JOURNAL OF CONSUMER RESEARCH

tion coefficientsincreasingfrom .47, .61 and .62 to .55, .69 and .74 when importanceis deleted from the model. On the other hand, removingthe evaluationdimension substantiallyreduces the predictabilityin the case of the Fishbeinmodel, with correlationcoefficients declining from .13, .35 and .32 to .06, .13 and .03 when the a, component is removed. This finding is obtainedin the case of Fishbein'smodel because there is not homogeneousperceptionof the concepts relating to selection of a soft drink. For example, while some people prefersweet soft drinks,otherspreferbeverages which lack sweetness or possess some intermediate level of sweetness.More is not necessarilybetter. EXPERIMENT THREE Method

Experiment#3 was conducted in a similar fashion to the first experimentwith several significantchanges. Three differentproduct classes: deodorant, shampoo and fast food restaurants,were selected. Each subject evaluatedthree brands or outlets per product class to facilitate intra-individualanalysis and responded to only one product class in which he was familiarwith all brands. In addition, once knowledgeof all brands in a product category was confirmed, subjects were randomlyassignedto completeone of four conditions: the Rosenbergvalue model, the "adequacy"model, or eitherthe Fishbeinor Rosenbergmodels using product characteristics.Thus, each individualcompletedone of the 12 questionnairesdesignedfor the study based on

his appearance in one of the cells of the 3 X 4 (products by models) design. Four hundred and ninety-nine undergraduate marketing students were administered questionnaires, with subjects randomly assigned to the four model conditions. Fifty-seven subjects were eliminated from the study due to their failure to complete all items. The subsequent analysis was based on 442 responses, with 158, 123 and 161 responses for deodorant, shampoo and restaurants, respectively. Sample sizes ranged from n ---26 to n -44 for the 12 cells in the design. The procedure used for obtaining salient product characteristics was similar to the free association test used in experiment #2. A group of 130 subjects were asked, "what kinds of things (e.g., product attributes, consequences of using the product, etc.) do you look for or take into consideration when you select a brand of for your own use?" In addition, a second step in the procedure was added for shampoo since many respondents supplied nondescriptive concepts, such as "makes hair feel good" or "makes hair look good." A group of 41 subjects were asked to specify salient choice concepts more specifically as follows: Please describe the following brandsof shampoo with respect to the following 3 attributes (cleaning ability, how hair feels after shampoo and how hair looks after shampoo) and with respect to any other characteristics you look for when you purchase a brand of shamnoo.

TABLE 6 BELIEFS USED IN THIRD EXPERIMENT EXPERIMENT #3 ProductDimensions ("Adequacy"Model)

Values (Rosenberg Value Models)

ProductCharacteristics (Fishbein & Rosenberg Models) Restaurants

1. Cleanliness 2. Price 3. Service 4. Taste of Food 5. Location

1. 2. 3. 4. 5.

Clean Low in Price Quick Service Not Greasy Nearby Location

1. 2. 3. 4. 5. 6. 7.

1. 2. 3. 4. 5. 6. 7.

Clean Hair Thoroughly Low in Price FragrantScent Leaves Hair Soft Leaves Hair Manageable Gentle on Hair Leaves Hair Shiny

1. Experiencing Change and Variety 2. Being Thrifty 3. Socializing with Other People

Shampoo Cleaning Hair Price Shampoo Scent Making Hair Soft Making Hair Manageable Gentleness Making Hair Shiny

1. 2. 3. 4. 5. 6. 7.

Good Looking Admired by Opposite Sex Careful in Matters of Personal Cleanliness Complimentedby Others Socializing with Other People Keeping in Good Health Associating with Opposite Sex

1. 2. 3. 4.

Careful in Matters of Personal Cleanliness Associating with Opposite Sex Avoiding Worry and Anxiety Socializing with Other People

Deodorant 1. 2. 3. 4. 5.

PreventingOdor Price Deodorant Scent PreventingOdor Skin Irritation

1. 2. 3. 4. 5.

Prevents Odor Low in Price FragrantScent Prevents Wetness Does Not IrritateSkin

COMPARISONOF FOUR MULTI-ATTRIBUTE MODELS Give your description in terms of specific aspects of the listed attribute. Do not give overall evaluative judgmentssuch as "cleanspoorly" or "makeshair feel good after a shampoo."It might help if you imagine that you are describing the brands of shampoo to a person whose tastes you don't know and who is totally unfamiliarwith each shampoo brand. Through this procedure, five salient attributes were identified for deodorant and fast food restaurants and seven were specified for shampoo. Since it was felt that obtaining salient values through a free association test would be extremely difficult, respondents were asked to indicate important values through a direct questioning procedure. A group of 155 subjects were supplied a list of 47 values obtained from Rosenberg (1956) and Bither and Miller (1969) and were asked (for fast food restaurants): Which of the following do you feel might be responsible for or underliethe selection of a fast food restaurant (e.g., McDonald's or Burger King) by college studentshere at the Universityof Florida. Please place a check mark beside all statementswhich you feel are relevant. Values mentioned by two-thirds of the students responding were considered salient for this study. Using this method, 3, 4 and 7 values were included for restaurants, deodorant and shampoo, respectively. To operationalize the instrumentality component of Rosenberg's model, subjects were asked to respond to statements in the following format:

47 Results Inter-individual Analysis. Based on inspection of the inter-individual correlation coefficients in Table 7, several interesting findings emerge. As in experiment #2, adding the importance element to the "adequacy" model consistently reduces the amount of correlation with the dependent variable; however, no clear pattern appears for the three expectancy-value models. Overall, the evaluative component has little effect on the predictability of the Fishbein and Rosenberg models. Also, while the Rosenberg value model offers the greatest predictability of attitude toward the act, with average correlation coefficients for deodorant, shampoo and restaurants of .84, .71 and .66 for the complete model and .78, .70 and .68 for the reduced model without the value component, it predicts no better than the "adequacy" model when behavioral intention and behavior are dependent variables. Since values are rather abstract cognitive elements, they might be expected to be less useful than product-specific beliefs in predicting behavior or action. Fishbein's (1967) finding that the expectancy-value score will be more predictive of attitude than behavioral intention or behavior is supported in the data as well. For example, the average correlation for the complete Fishbein and Rosenberg models dropped from .59 to .44 and from .75 to .58, respectively, when the dependent variable was changed from A,l,t to BI; the overall reduction in average correlation was .12. As expected, there were less dramatic differences in pre-

Buying a deodorant which does not irritate skin is completely attained

:

:

:

:

:

completely blocked

by buying SURE DEODORANT SPRAY (regular scent) It was felt this form more closely paralleled Rosenberg's original procedure than the method used in experiment #1. Finally, measures were obtained on three dependent variables compared with the one measure of affect used in the proceeding two experiments. First, attitude toward the act of buying the three brands was obtained by four seven-point bipolar scales labeled foolish-wise, good-bad, harmful-beneficial and rewarding-punishing at the endpoints. Second, behavioral intention measures were collected. The questions were in the following form: "Buying (brand) the next time I purchase a (product) is . ." Responses were recorded on sevenpoint scales labeled extremely probable-extremely improbable. Attitude toward the act and behavioral intention measures followed the format used by Ajzen and Fishbein (1972). Third, respondents specified their behavior indicating how often they purchase each of the brands on seven-point scales with endpoints of always buy and never buy.

diction when behavior replaced BI as the criterion variable, i.e., the average r dropped by .05. The inter-individual correlation coefficients were analyzed by using r to Z transformations and through x2 analysis to examine overall differences among models, products and interactions (Jones, 1968). This procedure was similar to the method in experiment #1, except that an interaction term could be calculated since independent rather than repeated measures were obtained across products. Overall, there were no statistically significant differences among the four models, except when Aact was used as a dependent variable in comparing two component models; the differences were due to the low predictability of the "adequacy" model when attribute importance was included. There was no significant main effect for models and for interaction between models and products. Intra-individual Analysis. The intra-individual analysis was performed in a similar manner as in experiment #2 with correlation coefficients computed for

THE JOURNAL OF CONSUMER RESEARCH

48 TABLE 7

AVERAGE INTER-INDIVIDUAL CORRELATION COEFFICIENTS BY PRODUCT CLASS FOR THREE DEPENDENT VARIABLES EXPERIMENT #3 "AdequacyImportance" Formulation Di only Pi Di

Restaurants

(n = 31) .59 .74 (n = 31) .50 .63 (n = 41) .52 .61

Average

.67

Deodorant Shampoo

Fishbein Formulation (Product Characteristics)

Rosenberg Formulation (Product Characteristics)

Rosenberg Formulation (Values)

B. only Bi ai Attitude Toward Act

li

i only 1i V;

.68

(n =44)

.70

.46

.32

.68

.66

.59

.58

.58

.74

.75

.68 (n =41)

.60

.54

Ij V;.

(n = 44) .59 .61 (n = 34) .64 .65 (n = 40) .62 .47

.81

(n =26) .69

only

(n = 39) .78

.84

(n = 32) .71 (n = 39)

Behavioral Intention Deodorant Shampoo Restaurants

.74 .54 .63

.56 .48 .58

.53 .49 .28

.55 .53 .22

.57 .41 .47

.42 .34 .38

.64 .54 .57

.62 .56 .56

Average

.64

.53

.44

.44

.49

.38

.59

.58

Behavior Deodorant Shampoo Restaurants

.70 .51 .56

.49 .45 .51

.52 .45 .51

.52 .45 .31

.55 .33 .20

.42 .45 .41

.64 .53 .39

.55 .54 .40

Average

.59

.49

.49

.43

.37

.43

.53

.50

each individual across three brands. The correlation coefficients were transformed to Z-scores3 and multivariate and univariate analysis of variance was conducted on the 3 X 4 factorial design for the three dependent variables, treating the set of dependent variables as a vector in the case of multivariate ANOVA (Cramer, 1967). When analysis was undertaken across all three dependent variables simultaneously through MANOVA, F-tests using the Wilks' lambda criterion revealed sta3 Correlation coefficients above .999995 were converted to .999995 to avoid problems of perfect correlations which have an infinite Z-score.

tistically significant F-values for both one component (F 9,1042) and two component models 2.62, df df (F 9,1042) for the models variable. Ex2.35, amination of test statistics for univariate F-tests shows that there were significant differences for each dependent variable for the one component models, but a statistically significant F-value was obtained only when Behavioral Intention was used as the dependent variable for the two component models. Adding the value importance component in the Rosenberg product characteristic and value formulations tended to increase predictability substantially. Overall, the results appear to support the proposition that there is no substantial -

-

TABLE 8 SOURCES OF VARIATION FOR Z TRANSFORMS OF INTER-INDIVIDUAL CORRELATION COEFFICIENTS BY PRODUCT CLASS FOR THREE DEPENDENT VARIABLES EXPERIMENT #3 One Component Models BI

Anct Source Models (col.) Products (row) Interaction * p < .05.

Two Component Models Behavior

Aact

BI

Behavior

x2

df

x2

df

x2

df

x2

df

x2

df

x2

6.03 1.67 1.94

3 2 6

6.48 2.31 1.65

3 2 6

7.18 2.18 3.05

3 2 6

10.09* 4.69 3.28

3 2 6

5.15 .82 2.40

3 2 6

4.03 1.59 2.09

df 3 2 6

49

MODELS COMPARISONOF FOUR MULTI-ATTRIBUTE TABLE 9 AVERAGE INTRA-INDIVIDUAL CORRELATION COEFFICIENTS BY PRODUCT CLASS FOR THREE DEPENDENT VARIABLES EXPERIMENT #3 "AdequacyImportance" Formulation Di

only

Pi Di

Fishbein Formulation (Product Characteristics)

Rosenberg Formulation (Product Characteristics)

B% only

1,

Bi a,

only

I Vi

Rosenberg Formulation (Values) hI, only

Ii Vi

Attitude Toward Act Deodorant Shampoo Restaurants

.994 .916 .872

.992 .893 .896

.950 .957 .824

.956 .945 .880

.904 .926 .597

.893 .945 .709

.982 .995 .545

.979 .995 .558

Average

.959

.955

.928

.933

.854

.878

.833

.961

Deodorant Shampoo Restaurants

.990 .954 .930

.989 .961 .952

.934 .935 .714

.928 .953 .637

.610 .808 .846

.756 .862 .815

.982 .988 .840

.985 .986 .701

Average

.972

.975

.897

.885

.779

.815

.966

.958

Behavioral Intention

Behavior Deodorant Shampoo Restaurants

.991 .851 .838

.989 .858 .912

.919 .934 .653

.916 .912 .608

.802 .798 .696

.914 .844 .747

.977 .993 .537

.983 .991 .247

Average

.962

.947

.874

.854

.769

.849

.954

.945

and consistent difference in predictability across the three products when the four two-component models are compared. DISCUSSION AND SUMMARY Careful examination of the correlation coefficients in experiments #1 and #3 reveals some useful findings. It is obvious that the correlations in experiment #3 are considerably higher than those in experiment #1. Two reasons may be advanced for this discrepancy: different operational measures and different product classes were used. However, the method of eliciting salient beliefs and values was likely responsible for much of the differential. Whereas beliefs and values used by

other investigators were included in the first experiment, a more careful pretest procedure which involved questioning subjects similar to those completing the final instrument was employed in the final two studies. The increase in predictability for the Rosenberg value model (r -_ .31 to r - .74) was particularly dramatic. One reason for the relatively poor prediction of multiattribute models in consumer behavior studies compared with psychological research may be due to the prevalence of non-salient beliefs in marketing studies. Based on the results of the third experiment, which used a careful pretest procedure to elicit values, values might be as appropriate as product characteristics in discovering the determinants of attitude.

TABLE 10 MULTIVARIATE AND UNIVARIATE ANOVA FOR ONE COMPONENT AND TWO COMPONENT MODELS F-TESTS EXPERIMENT #3 One Component Models

Two Component Models

ANOVA Source

Models Products Interaction p <.05. p < .01. **P

< .001.

ANOVA

MANOVA

Aact

BI

Behavior

MANOVA

Aact

2.62** 7.68*** 1.92*

2.90* 17.48*** 2.95**

7.00*** 4.06 * 1.84

3.88** 11.83*** 2.95**

2.35* 6.657*** 2.19**

1.92 12.55*** 3.42**

BI

6.05*** 7.28*** 2.03

Behavior

2.37 13.78*** 3.59**

50

Also, it should be noted that the conversion of r's to Z transformsand the subsequentparametricanalysis was not without problems.Average correlationcoefficients were materially affected by the many high correlationsat or near 1.0; the median r was .91 for Aact. One problem with this form of analysis is that the goodness of Z transformsdecreasesfor extremely high r's and for small sample sizes (Hays, 1963, pp. 530-531). For example, the difference between an r - .996 and .998 is .3470 Z-scores, while the difference between r - .500 and .502 is only .0027. Small differences in r's on the extreme have substantially more impactthan differencesbetween moderater's. In addition, if a subject supplies identical scores for the three dependentvariablesor JB,Ja>,an r - 0.0 is obtained, which further complicates analysis. When r's equal to 0.0 were removed from the analysis, median Fs increasedto .93 for Aact. It may be argued that a nonparametricprocedure, such as used by Talarzyk (1969), might be more appropriatefor this data. However,Talarzyk'smethod of using each individual'slPi Di score to predictthe ranking of brands based on the dependent variable score suffersfrom similarproblems.Due to the large number of ties in ratingbrands, any rankingmethod has very little power in detecting small differences between models since ties must be randomlyassigned.Removing ties further obscures analysis and is questionable statistically. From a theoreticalperspective,the empiricalresults provide some important distinctions between "adequacy" and expectancy-valuemodels. Consistentwith most previousstudies,nonnormalizedprominence(importance) weights reduced the predictabilityof the "adequacy-importance" model. However, eliminating the evaluativecomponentfrom expectancyvalue models either reduced predictability,as in the case of experiment #2, or had little overall impact. The results of this study illuminatethe differencesbetween "adequacy" and expectancy-valuemodels, which were discussed by Cohen et al. (1972) but not shown empiricallybefore this research.Clearly,the Pi component in the "adequacy"model is not equivalentto the a, and Vi componentin the Fishbein and Rosenbergmodels. The fact that the correlationsbetween the Pi and a, components in experiment #2 (a repeated measures design) were -.05, .31, .30, -.38 for flavor,carbonation, sweetness and calorie content, respectively,indicates that there is little agreementbetween the two measureseven with respectto direction. However, while the "adequacy"model is not an expectancy-valuemodel, it provides better prediction than the Fishbein and Rosenbergformulations.In experiment #1, average correlationswere .51, .41 and .41 for the "adequacy,"Fishbein and Rosenbergmodels, respectively,across the eight brands evaluated.In experiment#2 inter-individualanalysisyielded r .66

THE JOURNAL OF CONSUMER RESEARCH for the one component "adequacy" model (r

-

.57 for

two-componentmodel) versus -r .27 for the Fishbein model. Intra-individualanalysis produced transformed mean Z values of r -_ .97 for the "adequacy" model and r - .33 for Fishbein model. The third experiment produced smaller but consistent differences between the three formulationswhen product characteristicswere used as beliefs. Across the three dependent variables-Aact, BI and Behavior,the "adequacy" model yielded higher correlationsfor both inter- and intra-individualanalysis. For inter-individualanalysis, r

-.67, .64 and .59 for the one component "ade-

quacy" model, while mean correlationswere r = .59, .44, and .43 for the Fishbein model and f = .58, .38 and .43 for the Rosenbergformulation.The intra-individual analysissupplieda similarportrait:f - .96, .97 and .96 for the Di only "adequacy"model; rf- .93, .89 and .85 for the Fishbein model; r - .88, .82 and .85 for the Rosenbergmodel. If the investigator'sgoal is prediction of attitudes and behavior,the "adequacy"model appearsto be superior to expectancy-valueformulations.However, if understandingthe determinantsof attitudeis the major goal of research,the choice of an appropriatemodel is more difficult.Three situationsshould be distinguished in order to assess the relative utility of competing multi-attributemodels. First, for many product categories the relationship between the amount of an attributecontained by a brand (Bi) and affect toward that brand is a linearly increasingfunction.In general,the three productcategoriesused in experiment#3, deodorant,shampooand restaurants,are illustrativeof this "moreis better"situation. That is, when ai remainsconstant irrespective of the level of Bi and there is homogeneityof the goodness or badness across salient attributes,the evaluative componentadds little to both understandingand intraindividualprediction.For inter-individualpredictions, homogeneityof goodness or badness of each attribute across consumersin additionto inter-attributeevaluative homogeneityis requiredin orderfor the evaluative component to be unnecessary.Therefore,while there probably would be only small differencesin predictability among alternativemodels, the "adequacy"formulation is likely to be the model of choice since it may offer a slight increasein predictiondue to diminished measurementerror while sacrificinglittle in understandingconsumers'cognitive structuressince the ai or V, componentsin expectancy-valuemodels would be constant. Second, for product categoriesin which at remains constant irrespectiveof the level of Bi but there are heterogeneousperceptions about the preferreddirection and magnitude (slope) of attribute evaluations across subjectsand attributes,the Fishbeinand Rosenberg models are likely to yield more insight into the determinantsof motivatedbehavior.In this case, some

MODELS COMPARISONOF FOUR MULTI-ATTRIBUTE consumers believe "more is better," while others feel "less is better." For example, many consumers prefer cigarettes which are "weak" because this connotes low tar and nicotine content, while others stress the "strength" of a cigarette since this characteristic is associated with flavor. Also, many consumers have a strong preference for an automobile which is nonpolluting; however, a substantial number of car buyers would look unfavorably on a vehicle with low amounts of emissions because they believe such an automobile would lack acceleration and produce poor gas mileage. Third, for situations where ai is dependent on the level of Bi, neither Fishbein's nor Rosenberg's model is likely to yield good prediction or understanding of cognitive structure; while satisfactory prediction may be obtained through the "adequacy" model, little understanding of behavioral determinants is likely. Soft drinks, the product category used in experiment #2, is a good example of a class of stimuli in which a more complicated expectancy-value approach (Ahtola, 1973) is needed to adequately assess the structure of beliefs which underlie attitude toward an act. It should be noted that Fishbein uses belief statements which specify only the direction of belief. When only the direction of a belief is specified, e.g., carbonation or sweetness, subjects may not be homogeneous in their use of these concepts. For example, if a person uses the scales purely as probability scales and he knows that Fresca is "carbonated," he will give maximum scores on the scales whether he knows that it is only "slightly carbonated" or whether he knows that it is "very carbonated," as long as he is sure that it is "carbonated." Another person who also is sure that Fresca is "slightly carbonated" may give relatively low scores because Fresca is only "slightly carbonated." It is argued here that Fishbein's Bi does not discriminate between perceived possession and probability. This same problem will arise in the measure of value. If a person likes, say, carbonated soft drinks, but only if they are "slightly carbonated," he will find it confusing or even impossible to respond to the question whether or not he likes carbonated soft drinks because his liking depends on the amount of carbonation. Since the Fishbein and Rosenberg models rely on response to a directional belief, such as sweetness or carbonation, they are unable to accurately account for nonmonotonicity. For example, consumers may desire a brand which is somewhat sweet or carbonated but may respond negatively to brands which are unsweet or very sweet and lack or are very carbonated. 1i and Bi scores do not discriminate between the amount of a characteristic a given brand possesses and the probability that it contains a given level of a characteristic. Therefore, in the case of soft drinks, the Fishbein model produced very poor predictability of Aa(.t and provided little understanding as well. It is quite clear that before a multi-attribute model is applied to a par-

51

ticular marketingproblem,careful attentionshould be devoted to eliciting salient beliefs and values and to understandingthe nature of these cognitive elements. REFERENCES Ahtola, Olli T. "An Investigation of Cognitive Structure Within Expectancy-Value Response Models," unpublished doctoral dissertation, University of Illinois, 1973. Bass, Frank M. "Fishbein and Brand Preference: A Reply," Journal of Marketing Research, 9 (November 1972), 461. Bass, Frank A. and W. Wayne Talarzyk. "Attitude Model for the Study of Brand Preference," Journal of Marketing Research, 9 (February 1972), 93-6. Bennett, Peter D. and Jerome E. Scott. "Cognitive Models of Attitude Structure: 'Value Importance' is Important," Proceedings. Combined Spring and Fall Conferences, American Marketing Association, 1971, 34650. Bettman, James R., Noel Capon, and Richard J. Lutz. "Cognitive Algebra in Multi-Attribute Models," Journal of Marketing Research, 12 (May 1975), forthcoming. Bither, Stuart W. and Stephen J. Miller. "A Cognitive Theory View of Brand Preference," Proceedings. Fall Conference, American Marketing Association, 1969, 280-6. Cohen, Joel B. and Olli T. Ahtola. "An Expectancy X Value Analysis of the Relationship between Consumer Attitudes and Behavior," Proceedings. Second Annual Conference, Association for Consumer Research, 1971, 344-64. Cohen, Joel B. and M. J. Houston. "The Structure of Consumer Attitudes: The Use of Attribute Possession and Importance Scores," Working Paper #2, College of Commerce and Business Administration, University of Illinois, 1971. Cohen, Joel B., Martin Fishbein, and Olli T. Ahtola. "The Nature and Uses of Expectancy-Value Models in Consumer Attitude Research," Journal of Marketing ReseaF ch, 9 (November 1972), 456-60. Cramer, Elliot M. "Revised MANOVA Progiram," The University of North Carolina at Chapel Hill, December, 1967. Edwards, Ward. "The Theory of Decision Making," Psy chological Bulletin, 51 (1954), 380-417. Fishbein, Martin. "A Behavioral Theory Approach to the Relations between Beliefs about an Object and the Attitude Toward the Object," in M. Fishbein (ed.), Readings in Attitude Theory and Measurement. New York: Wiley, 1967, pp. 389-400. Fishbein, Martin and Bertram H. Raven. "The AB Scales: An Operational Definiition of Belief and Attitude," Human Relations, 15 (1962), 35-44. Fisher, R. A. Statistical Methods for Researchers. New York: Hafner, 1954. Hansen, Flemming. "Consumer Choice Behavior: An Experimental Approach," Journal of Marketing Research, 6 (November 1969), 436-43.

52 Hays, William. Statistics. New York: Holt, Rinehart and Winston, 1963. Jones, Marshall B. "Correlation as a Dependent Variable," Psychological Bulletin, 70 (January 1970), 69-72. Katz, Daniel. "The Function Approach to the Study of Attitudes," Public Opinion Quarterly, 24 (1960), 163204. Kirk, Roger E. Experimental Design: Procedures for the Behavior Sciences. Belmont, California: Brooks! Cole, 1969. Klippel, R. Eugene and Stewart W. Bither. "Attitude Data in Allocation Models," Journal of Advertising Research, 12 (April 1972), 20-24. Lawler, E. E. "A Correlation-Causal Analysis of the Relationship Between Expectancy Attitudes and Job Performance," Journal of Applied Psychology, 32 (1968), 462-68. Lawler, E. E. and R. W. Porter. "Antecedent Attitudes of Effective Managerial Performance," Organizational Behavior and Human Performance, 2 (1967), 122-42. Levy, Sidney J. "Symbols for Sale," Harvard Business Review, 37 (July-August 1959), 117-24. Lewin, Kurt. A Dynamic Theory of Personality. New York: McGraw-Hill, 1935. Lewin, Kurt, Tamara Demob, L. Festinger, and Pauline S. Sears. "Level of Aspiration," in J. McV. Hunt, ed., Personality and the Behavioral Disorders, New York: Ronald, 1944. Marineau, Pierre D. Motivation in Advertising. New York: McGraw-Hill, 1957. Mazis, Michael B. "Decision-Making Role and Information Processing," Journal of Marketing Research, 9 (November 1972), 447-50. "Cognitive Tuning and Receptivity to Novel

THE JOURNAL OF CONSUMER RESEARCH Information," Journal of Experimental Social Psychology, 9 (July 1973), 307-19. Nakanishi, Masao and James R. Bettman. "Attitude Models Revisited: An Individual Level Analysis," Journal of Consumer Research, 1 (June 1974). Peak, Helen. "Attitude and Motivation," in Marshall R. Jones, ed., Nebraska Symposium on Motivation. Lincoln: University of Nebraska Press, 1955, 149-89. Rosenberg, Milton J. "Cognitive Structure and Attitudinal Affect," Journal of Abnormal and Social Psychology, 53 (November 1956), 376-82. Rotter, Julian B. Social Learning and Clinical Psychology. New York: Prentice-Hall, 1954. Smith, M. Brewster. "Personal Values as Determinants of a Political Attitude," Journal of Psychology, 28 (1949), 477-86. Snedecor, George W. Statistical Methods. Ames, Iowa: Iowa State University Press, 1962. Talarzyk, W. Wayne. "An Empirical Study of an Attitude Model for the Prediction of Individual Brand Preference for Consumer Products," unpublished doctoral dissertation, Purdue University, 1969. Tolman, E. C. Purposive Behavior in Animals and Men. New York: Century, 1932. Wilkie, William, John M. McCann, and David L. Reibensten. "Halo Effects in Brand Belief Measurement: Implications for Attitude Model Development," S. Ward and P. Wright, eds., Advances in Consumer Research: Volume 1, Urbana, Illinois: Association for Consumer Research, 1974, 280-90. Wilkie, William and Edgar A. Pessemier. "Issues in Marketing's Use of Multi-Attribute Models," Journal of Marketing Research, 10 (November 1973 ), 428-41.