Advances in Consumer Research Volume 2, 1975 Pages 375-388
MULTI-ATTRIBUTE ATTITUDE MODELS: A COMPARATIVE ANALYSIS
Morris B. Holbrook, Columbia University
James M. Hulbert, Columbia University
[The authors wish to express their gratitude to Georgi Zahariev and Bernardo Cohen for their help in the early stages of the project and to Harry Steinberg, for his invaluable assistance in later stages. They also wish to thank the Faculty Research Fund of Columbia University Graduate School of Business for its support of the project.]
[Morris B. Holbrook is a doctoral candidate at the Columbia University Graduate School of Business. James M. Hulbert is Associate Professor at the Columbia University Graduate School of Business.]
Existing research on attitude structure is summarized and a series of hypotheses derived. These hypotheses are tested on data gathered in the context of the 1972 Presidential election. Results suggest that overall attitude may be predicted quite well using beliefs and evaluations, with a limited role for salience in choosing appropriate attributes.
A consistent view of attitude has long eluded behavioral scientists. Indeed, recent research has tended to complicate, rather than simplify, the underlying issues. In this paper, we attempt to move closer to a resolution of these issues. First, we briefly review relevant literature summarizing different schools of thought in a series of models. From this analysis we derive a set of hypotheses which are then tested on a common, purposefully collected data base.
RESEARCH ON THE UNDERLYING STRUCTURE OF ATTITUDE
Several psychologists (Rosenberg, 1956, 1960, Fishbein, 1963, 1965, 1967B, 1967C) have proposed a model of attitude toward an object (Ao) as a function of two components (C1 and C2) where C1 is a set of cognitions (beliefs or perceptions) concerning the object and C2 the degree of affect (the evaluation or amount of satisfaction) associated with each cognition. In Fishbein's words, these models reflect:
The hypothesis that an individual's attitude toward any object is a function of his beliefs about the object and the evaluative aspects of those beliefs. (Fishbein, 1967B, p. 395)
Some attempt has often been male (Rosenberg, 1956; Fishbein, 1963, 1965) to select the n most salient cognitions, where 'salient' cognitions are assumed to be those which determine Ao (Fishbein, 1967B, p. 395, Cohen, Fishbein, & Ahtola, 1972, p. 457). Ao has then been postulated to be a function of the sum of the crossproducts of the n C1 and C2 components. Or, algebraically: EQUATION
Sheth (May, 1972) has pointed out at least three crucial assumptions of such models: (1) two factors (C1 and C2) are needed to predict Ao; (2) one C should be used to 'weight' the other in a multiplicative relationship, (3) the n weighted products should be summed without further weighting to form an overall prediction of Ao (for other statements of these assumptions, cf. Cohen, Fishbein & Ahtola, 1972, p. 117, and Talarzyk, 1972, p. 466).
Researchers working in the Rosenberg-Fishbein tradition have tested such models and found them to be satisfactory predictors of the criterion (Rosenberg, 1956; Bither & Miller, 1969; Hansen, 1969; Fishbein, 1963, 1965; Triandis & Fishbein, 1963; Fishbein & Hunter, 1964; Anderson & Fishbein, 1965; Sampson & Harris, 1970). Recently, however, several marketing applications of two-component models have concluded that an unweighted model of the form EC1 works about as well as (Sheth, May, 1972; Farley, et al., undated; Scott & Bennett, 1971; Cohen & Ahtola, 1971; Lehmann, 1971; Beckwith & Lehmann, 1973) or even better than (Sheth & Talarzyk, 1972; Moinpour & MacLachlan, 1971) the two-component EC1.C2 version.
A troubling question thus arises: a well-established psychological finding has been repeatedly contradicted by studies in the marketing literature. Have psychologists erred in their formulation of the two-component model, or have marketing researchers failed to test the EC1.C2 model in a theoretically appropriate manner?
This question was addressed by several articles in the Journal of Marketing Research (Cohen, Fishbein & Ahtola, 1972; Bass, 1972; Sheth;-November, 1972; Talarzyk, 1972). In our view, however, this debate failed to deal with the problem in its fullest scope. Only a few of the relevant studies were discussed. In addition, the ways in which C1 and C2 have been measured in the studies cited above lead to the conclusion that there are three, not two, determinants of attitude being measured and that all three of these components are implicit in the Rosenberg-Fishbein formulations. These three components may be defined as follows:
(1) BELIEF (Bi): the perceived extent to which some concept or object (o) is related to some other object i, some attribute i, some value i, or some goal i, (Fishbein, 1967A, p. 259; 1967B, p. 389)--as measured, for example, by a subjective probability or perceived instrumentality scale,
(2) EVALUATION (Ei): the degree of favorability of affect-i.e., the attitude toward that other object i, attribute i, value i, or goal i (Fishbein, 1967A, p. 260)--as measured, for example, by a series of bipolar adjectival scales heavily loaded on the evaluative dimension of the Semantic Differential (Fishbein & Raven. 19671;
(3) SALIENCE (Si): the degree of importance of Bi and its associated Ei in determining the overall attitude (Ao) toward the concept or object (o) (Fishbein, 1967B, p. 396)--as measured, for example, by a bipolar important-not important scale or by a ranking of the n object i's. attribute i's value i's, or goal i's.
Obviously, when the three basic components B, E, and S, are reduced to a two-component EC1.C2 model, different researchers may define those components differently. Indeed, examination of the studies cited above reveals that they have included at least four different ways of operationalizing C1 and C2. It follows that these researchers were virtually certain to have reached different conclusions precisely because they were testing different models. We now examine these models individually.
The Rosenberg Model: EQUATION
Rosenberg (1956) proposed a two-component model in which C2 was called 'value importance' and was viewed as the importance of a value i as a source of satisfaction. Cohen, Fishbein, and Ahtola (1972), however, argued that "despite Rosenberg's use of the term 'value importance', Vi is not a measure of importance...but a measure of satisfaction or evaluation" (p. 456). In contrast, Sheth (November, 1972) quoted Rosenberg (p. 463) indicating that the latter was interested in the importance of values as defined above. Sheth concluded that "more or less importance of a valued state does not seem to mean the same thing as the-evaluation of that valued state" (p. 463) and charged that "In their enthusiasm to relate other theories to Fishbein's, the authors have misunderstood and misinterpreted Rosenberg" (p. 462). There is some truth on both sides of this exchange, though neither seems to have recognized that Rosenberg's formulation of his second component relates to both the evaluation and the salience dimensions. Thus he confounds the components we labelled Ei and Si within the same measure, which we designate F(Ei, Si). The resulting model of the form EB.F(E,S) has been supported in a number of studies, some of which suggested that EB.F(E,S) was a better predictor than perceived instrumentality alone (Rosenberg, 1956, 1960; Bither and Miller, 1969; Hansen, 1969).
The Fishbein Model: EQUATION
Fishbein has tested a two-component model in which the C1 and C2 components correspond quite closely to what we have called Bi and Ei. A number of studies have reported good prediction with the EBE model (Fishbein, 1965; Sampson and Harris, 1970), while comparative research has supported the EBE model over competitors (Fishbein, 1963; Triandis and Fishbein, 1963; Fishbein and Hunter, 1964).
The Columbia Model: EQUATION
A number of marketing studies have used two-component models of the form EB.S. to predict Ao. We call this model the 'Columbia Model T since its first application was apparently in a study of instant breakfast carried out at Columbia in the mid-1960's. It omits the Ei component, assuming implicitly that attributes are evaluated positively (with similar degree of affect).
Such applications have generally found that weighting Bi by the Si component contributes little to prediction of Ao (Scott and Bennett, 1971; Hansen and Bolland, 1971; Sheth, May 1972). When B and BS scores have been used in regressions to predict the criterion, the St component has also been found to contribute little or nothing (Cohen & Ahtola, 1971; Sheth, May 1972; Farley, Howard and Weinstein, undated).
The Purdue Model: EQUATION
Several researchers--most of them associated in one way or another with Purdue--have employed measures of the first component which pertain to how satisfactory (Bass & Talaryk, 1972; Moinpour & MacLachlan, 1971; Kraft, et al., undated) or how ideal (Lehmann, 1971;
Bass, et al., 1972; Beckwith & Lehmann, 1973) a brand is with respect to a given attribute (i). Cohen, Fishbein, and Ahtola (1972) pointed out that such measures of C1 actually combine aspects of Bi and Ei into one component [which we designate F(B,E)], concluding that the Purdue researchers have "proposed a new model which might be termed an 'adequacy-importance T model" (p. 456).
A number of studies have reported that EF(B,E) models predict about as well as EF(B,E).S (Bass and Talarzyk, 1972; Sheth and Talarzyk, 1972, Beckwith and Lehmann, 1973), thus reinforcing the Columbia findings that S-weights do not offer significant improvement in the prediction of overall affect.
Some Recent Tests of Three-Component Models
The usefulness of S-weights is also challenged by recent psychological studies using all three components (B,E, and S) to predict Ao. Hackman and Anderson (1969) defined 'relevance' of a belief as "the degree to which a particular belief is important to a subject in evaluating an attitude object..." (p. 56). In predicting students' attitude toward a teacher who was rated on various attributes (i) (in turn rated on six-point scales for Ei and Si) no improvement of the EB.E.S over the EB.E model was found. A similar conclusion was reached in a study by Anderson (1970). He concluded that "the Fishbein formula, which considers the 'strength of belief T times the 'affect associated with that belief' does not benefit from the inclusion of another score, relevance of belief." (p. 46)
In summarizing studies using the EB.E. model which have, in one way or another, taken S-weights into account, Fishbein and Ajzen (19721 state the following rather free interpretation:
It has sometimes been argued that each piece of information should also be given a weight for its importance, salience, or relevance.... Despite the intuitive plausibility of this position, recent studies that have obtained measures of importance or relevance in addition to belief strength have consistently found that adding these weights to an expectancy-value model attenuates the prediction of attitude (p. 509, see also Cohen, Fishbein, and Ahtola, 1972, pp. 458-549).
Conclusions on Predictive Power of 1-, 2-, and 3-Component Models
Not surprisingly we may conclude that the performance of various models in predicting Ao depends upon how their components are defined in the first place. In brief, when models of the form EB are compared with models of the form EB.E (Fishbein) or EB.F(E,S) (Rosenberg), the second component almost invariably contributes to the prediction of Ao (especially where the attributes are not all equally favorable). When, on the other hand, the comparison is between models of forms EB vs. EB.S (Columbia), EF(B,E) vs. EF(B,E).S (Purdue), or EB.E vs. EB.E.S (Three-Component Model), the addition of S-weights seems to contribute little predictive power.
This failure of S-weights in the Columbia, Purdue, and Three-Component Models may be explained in part by several important considerations that have sometimes been overlooked in practice. First, we could question the Columbia Model's implicit assumption that Ei is more or less equally positive for all attributes used (cf. Bither & Shuart, p. 7). Second, most studies using the Columbia and Purdue models have consciously attempted to discover and use only the most salient attributes. If such attempts were successful, the inclusion of more or less equally important S-weights would clearly do little to improve prediction of Ao. Another problem in using S-weights arises if first-component scores are closely inter-related. If such multicollinearity exists separate S-weights can contribute little to prediction. Beckwith and Lehmann (1973) found such intercorrelations in their data, as did Farley, et al. (undated). An even more serious challenge to S-weights arises if the first and second components are somehow closely related. This type of redundancy would seem especially likely for the Purdue Model EF(B,E).S since there is strong a priori reason for thinking that Ei (the favorability of an attribute) might be strongly related to Si (its importance). Finally, there has been a variety of objections to the methods used to measure the S-component (Sampson and Harris, 1970; Scott and Bennett, 1971; Alpert, 1971; Schendel, 1971). Perhaps most seriously, Sheth and Talarzyk (1972) argue that the whole procedure of rating attribute salience for a product class may be inappropriate if salience weights differ between brands, as indeed they might if the brands are not perceived as substitutes or if each brand is evaluated in terms of its 'strong points.'
HYPOTHESES ABOUT THE PERFORMANCE OF 2- AND 3-COMPONENT MODELS
We may summarize the considerations raised above by setting forth the following hypotheses as a way of reconciling the findings reported in the psychological and marketing literature.
First, we suggest that a model preferable conceptually to any of the two-component EC1.C2 versions discussed above is the three-component model of the form EB.E.S which explicitly treats the B,E, and S components separately (cf. Hackmann & Anderson, 1970). Furthermore in testing such a model, attributes (i = 1,2...n) should be consciously chosen to display variation between attributes and across individuals in the Bi, Ei, and Si ratings. In other words, attributes should be selected not all of which are certainly possessed by the attitude object, not all of which are unequivocally favorable, and not all of which are of paramount salience for everyone in determining Ao.
Given the existence of these kinds of data, we might then cheek our interpretation of the empirical findings discussed above against the following predictions: (1) In accord with empirical findings supporting the Rosenberg Model, EB.(E.S) should perform better than EB in predicting Ao. (2) In accord with results supporting the Fishbein Model, EB.E. should perform better than EB. (3) In accord with repeated tests of the Columbia Model, the performance of the EB.S model should be quite low and should not be damaged by omitting the S-weights. (4) In accord with tests of the Purdue and Three-Component Models, when B or E and S are highly correlated and/or Var(B) or Var(E) and S are highly correlated, LB.E should perform about as well as E(B.E).S in predicting Ao (5) With any given model, predictive power can always be improved by using multiple regression rather than straight summation (neglecting the effects of a loss in degrees of freedom), but the above conclusions concerning the relative performance of B vs. BES, B vs. BE, B vs. BS, and BE vs. BES scores should still apply.
We note that, in reviewing the debate between Cohen, Fishbein and Ahtola; Bass; Sheth; and himself, Talarzyk (1972) suggested "the need for research to verify the similarities and differences between applications of models across areas of study" (p. 467). Although the study to be described below was designed and data collected before the appearance of Talarzyk's suggestion, we feel that it answers, in part, his call for research bearing on the relationship between the Rosenberg, Fishbein, Columbia, and Purdue Models.
AN EMPIRICAL TEST
The above hypotheses--originally developed by Holbrook (1972)-were tested in the context of the 1972 presidential election. Just prior to the voting, 99 eligible student voters were asked to rate twenty-two attributes on (a) Ei ("how desirable you personally feel each attribute is in a presidential candidate" on a ten-point scale from "very undesirable" (scored -4.5) to "very desirable" (scored +4.5)) and (b) Si ("how important you personally feel each attribute is in making your overall evaluation of a candidate" on a ten-point scale from "very unimportant" (1) to "very important" (10)). Immediately afterwards, the candidates--George McGovern and Richard Nixon-were rated on Bi for each attribute (in randomized order) according to "how likely you personally feel it is that possesses each attribute" (using a ten-point scale from "very unlikely" (scored -4.5) to "very likely" (scored +4.5)). Then Ao and voting intention (Io) scores were obtained for each candidate on ten-point scales running from "dislike very much" (1) to "like very much" (10) and from "definitely will not" (1) to "definitely will" (10).
Informal inspection of the Bi, Ei, and Si scores suggested that we had succeeded in our attempt to include attributes representing a meaningful range of variation along these dimensions. For McGovern, mean B scores ranged from -3.17 (condones wiretapping) to 3.85 (favors a reduction in the national defense budget); their standard deviation varied from 1.23 (supports civil rights) to 3.00 (has a competent running mate).- Similarly, for Nixon, B ranged from -3.89 (wants amnesty...) to 3.40 (believes in law and order), with SD of B varying from 1.31 (wants amnesty...) to 3.28 ( . . . competent running mate). Mean E's ranged from -2.54 (condones wiretapping) to 4.20 (has good decision-making ability) with SD's of 0.96 (is honest) to 3.15 (wants amnesty...). Similarly mean S's ranged from 1.64 (attends church regularly) to 8.37 (...decision-making ability) with SD's of 1.33 (...decision-making ability) to 3.47 (condones wiretapping). (See Table 1.)
Correlation and regression analyses were used to test the hypotheses. Results are presented in Table 2.
Our first hypothesis is supported by the significant improvement of the EB.(E.S) over the EB model in predicting Am (.735 vs. 490, Z = 2.88, p = .002), Im (.691 vs. .428, Z = 2.70, p = .003), An (.785 vs. .612, Z = 2.51, p = .006), and In (.777 vs. 613, Z = 2.33, p = .01). Similarly, in support of our second hypothesis, the differences are of virtually the same magnitude and significance for comparisons between EB.E and EB models in predicting Am (.742 vs. .490, Z = 2.88, p = 002), Im (.700 vs. .428, Z = 2.82, p = .002), An (.796 vs. .612, Z = 2.71, p = .003), and In (.784 vs. .613, Z = 2.34, p = .01). Clearly, then, both the Rosenberg-like and the Fishbein two-component versions are--as usual--better predictors than a naive one-component form.
BELIEFS, SALIENCIES AND EVALUATIONS
CORRELATIONS OF VARIOUS MODELS WITH ATTITUDE (Ao) AND INTENTION (Io)
Turning to those models which use S-weights in connection with a first component represented by B or (B.E), we find that--contrary to our third hypothesis--the EB.S model does offer some partially significant tendency toward improvement over EB in predicting Am (.648 vs. .490, Z = 1.62, p = .053), Im (.578 vs. .428, Z = 1.40, p = .081), An (.699 vs. .612, Z = 1.10, p = .136), and In (.693 vs. .613, Z = .97, p = .166). This mixed disinformation of our hypothesis may be explained by the high correlations between E and S (a) for individuals across attributes (rES = .618, p < .001), (b) for attributes across individuals (rES = .519, p < .001), and (e) for Ei and Si across attributes (rES = .83, p < .001). These strong relations between E and S may have meant that S was in effect acting as a proxy for E in the EB.S model thereby imparting to it some of the superiority of the EB.E over the EB model. At any rate, this high degree of redundancy between E and S establishes the conditions under which we raised the fourth hypothesis. In accordance with this hypothesis, we find that EB.E performs about as well as S(B.E).S in predicting Am (.742 vs. .735, not significant), Im (.700 vs. .691, n.s.), An (.796 vs. .785, n.s.), and In (.784 vs. .777, n.s.). These results support the contention that, when E and S are highly correlated, the EB.E.S does not improve upon the EB.E model.
Our final hypothesis was supported in part by the improvement over the EB model of the five-variable stepwise regression using B scores to predict Am (.81 vs. .49, Z = 4.04, p < .001), Im (.80 vs. .428, z = 4.36, p < .001), An (.78 vs. .612, z = 2.30, p = .011) and In (.78 vs. .613, Z = 2.30, p = .011). Similarly, stepwise regression with the BS scores tended to improve over the EB.S model in predicting Am (.82 vs. .648, Z = 2.61, p = .005), Im (.80 vs. .578, Z = 2.99, p = .001), An (.78 vs. .699, Z = 1.22, p = .111), and In (.79 vs. .693, Z = 1.53, p = .063). (In these comparisons Z-values are only approximate though some attempt has been made to allow for the loss of degrees of freedom in the stepwise regression.)
In the ease of the BE and BES scores, the summative models performed well enough so that the improvements gained from multiple regression were marginal at best (and in no ease significant, even at p = .10). Nor was there any support for our anticipation of a difference in favor of regression models using BE or BES scores. Regressions with B scores performed fully as well as any competing models.
In summary, we might conclude that, for our data, the EB model performs less well than the EB.E or EB.E.S models, with the EB.S model falling somewhere in between, perhaps because S serves as a proxy for E. In addition, maximal predictions can be obtained in stepwise regression using as few as five B scores.
IS THERE A ROLE FOR SALIENCE?
What then, if anything is left for the role of salience (S)? In our data, the sizeable redundancy between E and S obviated any major contribution of the latter variable. It should be possible, however, to remove the redundancy between E and S by conjuring up attributes which are highly important, yet highly unfavorable (dishonesty, inability to make good decisions, opposition to civil rights). Presumably such salient attributes--though now expressed negatively-would still be important as determinants of Ao.
Indeed, it appears that the most useful role of salience might lie in selection of a limited set of attributes to use in predicting Ao. (The findings of Fishbein and his colleagues discussed earlier are not inconsistent with this conclusion. Some of their best predictions were obtained with attributes selected for their general or idiosyncratic salience). We noticed, for example, that the first one or two attributes to enter the stepwise regressions (decision-making and civil rights for McGovern, candidness and civil rights for Nixon) were also among the highest in saliency. In fact, when the individual correlations between B scores and Ao for each attribute were correlated with the attribute's S-weights, r's were found of .58 (p = .002) for McGovern and .44 (p = .022) for Nixon. This relation between the salience of a belief and its ability to predict Ao, however, might (in light of the close relation between E and S) be interpreted as nothing more than a reflection of the almost tautological relation between a belief scale's E-value and its relation to Ao (i.e., rrB.AE = .81 for McGovern and .78 for Nixon) due to the tendency of these correlations to shift from negative to positive as Ei varies between negative and positive valences. In an attempt to clear up some of this ambiguity introduced by the directionality of Ei, we therefore looked at (B.E)i scores for which higher values should unambiguously predict Ao. The correlation between BE and Ao increased dramatically with an attribute's salience (S)(rrBE,AS = .74 for McGovern and .56 for Nixon--both significant at p = = .003). By contrast, the relation of these correlations to E itself fell to r = .47 for McGovern and .56 for Nixon. In addition, when rBEA was regressed on S and E together, only S attained significant regression coefficients (t = 4.27, p < .001 for McGovern, t = 2.53, p < .02 for Nixon). In our view, this latter finding warrants the tentative conclusion that, when the direction of E is taken into account (eg., by weighting Bi by Ei or simply using only attributes with positively evaluated E's), the relation between B and Ao increases in strength with attribute salience (S). More specifically, it appears that good predictions of An using B scores could be obtained using only those few attributes for each individual which are (a) positively evaluated and (b) highly salient.
This conclusion could not be tested directly since many subjects assigned a great number of attributes the maximum S and E ratings. As a partial test, however, we selected the five attributes (numbers 3, 11, 15, 17 and 18) which were on the average judged highest in both S and E and were among the seven lowest attributes in Var (E). Using only these five attributes in a summative model (EB), predictions of Ao were obtained (r = .70 for McGovern and .71 for Nixon--with 94 usable questionnaires) which rivaled those of the full EB.E.S model with all 22 attributes (r = .735, Z = .08, n.s., and .785, Z = 1.28, n.s., respectively). Moreover, when only the n attributes of these five which were (a) above the average in E and (b) at the highest level of S for any given subject were used in a summative model (EB/n) to predict Ao, slightly better predictions were obtained for both McGovern (r = .730) and Nixon (r = .769). The average number of attributes (n) used in these predictions was only 2.7 (median = 3). These results suggest, then, that--using as few as the three most salient (positively evaluated) attributes for each individual in a EB model--predictions of Ao may be achieved that are as good as those obtained from the full array of twenty-two attributes in the EB.E.S version, which in turn performs about as well as a model using B, BE, or BES scores in stepwise regression. (The differences between the EB/n and five B score regression models are not significant at p = .108. We may conclude, then, that in the data described above individual salience weights are useful in that they may be used to specify for each respondent a very limited number of, say, three positively evaluated attributes whose B scores may be used in a simple summative model which predicts Ao virtually as well as models using far more information yet attaining only marginal improvements in performance.
These findings corroborate the results of Wilkie and Weinreich (1973). They compared the performance of a variety of summative models and found that when the order of inclusion of weights was the same for all subjects, maximum correlation was attained with five attributes. When the order of inclusion was determined idiosyncratically, best fits occurred with only three attributes.
Attribute models of attitude structure have received a great deal of marketing research effort in recent years. By now, a number of useful generalizations appear finally to have emerged. In this connection, the results of this study are strikingly similar to those reported by Wilkie and Weinreich (1973), moreover they relieve some of the concerns expressed by those authors.
First, since the EB predictions based on an aggregate selection procedure (high S and E, low Var (E)) were not significantly lower than the EB/n predictions based on idiosyncratic attribute selection, it appears that, for applied marketing research, prior selection of a relatively small set of attributes based on highest Y's (or some similar criterion) can be relied upon to result in good prediction.
Second, where the analyst is interested in the consumer's belief structure (eg. as a variable intervening between message exposure and affect), the results suggest that selection of the three idiosyncratically highest-S attributes provides a satisfactory measure.
Third, the findings of this study and those of Wilkie and Weinreich suggest that the Purdue and Columbia practices of using a limited number of salient attributes is justified using the pragmatic criterion of prediction of Ao (though in the Columbia case this conclusion only applies if attributes with positive affect are used).
Finally, the present study used a large set of twenty two attributes, varying widely in B, E and S. The similarity of our results to theirs should alleviate Wilkie and Weinreich's expressed concern over their use of only seven attributes.
Alpert, M.I. Identification of determinant attributes: A comparison of methods. Journal of Marketing Research, 1971, 8, 184-191.
Anderson, L. R. Prediction of negative attitude from congruity, summation and logarithmic formulae for the evaluation of complex stimuli. Journal of Social Psychology, 1970, 81, 37-48.
Anderson, L. R. & M. Fishbein. Prediction of attitude from the number, strength and evaluative aspect of beliefs about the attitude object: A comparison of summation and congruity theories. Journal of Personality and Social Psychology, 1965, 3, 437-433.
Bass, F. M. Fishbein and brand preference: A reply. Journal of Marketing Research, 1972, 9, 461.
Bass, F. M., E. A. Pessemier & D. R. Lehmann. An experimental study of relationships between attitudes, brand preference and choice. Behavioral Science, 1972, 17, 532-541.
Bass, F. M. & W. W. Talarzyk. An attitude model for the study of brand preference. Journal of Marketing Research, 1972, 9, 42-96.
Beckwith, N. E. & D. R. Lehmann. The importance of differential weights in multiple attribute models of consumer attitude. Journal of Marketing Research, 1973, 10, 141-145.
Bither, S. W. & S. J. Miller. A cognitive theory view of brand preference. Proceedings, American Marketing Association, Fall, 1969, 280-286.
Bither, S. W. & A. H. Shuart. Uses and misuses of expectancy attitude models in marketing. Pennsylvania State University, unpublished paper, undated.
Cohen, J. B. & O. T. Ahtola. An expectancy X value analysis of the relationship between consumer attitude and behavior. Proceedings, Second Annual Conference Association for Consumer Research, 1971, 344-364.
Farley, J. U., J. A. Howard & D. Weinstein. The relationship of liking and choice of an alternative to its attributes and their saliency. Columbia University, unpublished paper, undated.
Fishbein, M. An investigation of the relationships between beliefs about an object and the attitude toward that object. Human Relations, 1963, 16, 233-249.
Fishbein, M. Prediction of interpersonal preferences and group member satisfaction from estimated attitudes. Journal of Personality and Social Psychology, 1965, 1, 663-667.
Fishbein, M. A consideration of beliefs and their role in attitude measurement In M. Fishbein (Ed.), Readings in attitude theory and measurement. New York: John Wiley and Sons, 1967a.
Fishbein, M. A behavior 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: John Wiley and Sons, 1967b.
Fishbein, M. Attitude and the Prediction of behavior. In M. Fishbein (Ed.), Readings in attitude theory and measurement. New York: John Wiley and Sons, 1967c.
Fishbein, M. & I. Ajzen. Attitudes and opinions. In P. H. Mussen and M. R. Rosenzweig, (Eds.), Annual review of psychology. Palo Alto: Annual Review, 1972. Pp. 487-544.
Fishbein, M. & R. Hunter. Summation versus balance in attitude organization and change. Journal of Abnormal and Social Psychology, 1-554, 69, 505-510.
Fishbein, M. & B. H. Raven. The AB scales: An operational definition of belief and attitude. In M. Fishbein (Ed.), Readings in attitude theory and measurement. New York: John Wiley and Sons, 1967.
Hackman, J. R. & L. R. Anderson. The strength, relevance and source of beliefs about an object in Fishbein's attitude theory. Journal of Social Psychology, 1968, 76, 55-67.
Hansen, F. Consumer choice behavior: An experimental approach. Journal of Marketing Research, 1969, 6, 436-443.
Hansen, F. & T. Bolland. The relationship between cognitive models of choice and non-metric multidimensional sealing. Proceedings, Second Annual Conference, Association for Consumer Research, 1971, 376-388.
Holbrook, M. B. The definition, measurement and prediction of attitude. Columbia University, unpublished paper, 1972.
Kraft, F. B., D. H. Granbois & John O. Summer. Brand evaluation and brand choice: A longitudinal study. Indiana University, unpublished paper, undated.
Lehmann, D. R. Television show preference: Application of a choice model. Journal of Marketing Research, 1971, 8, 47-53.
Moinpour, R. & D. L. MacLachlan. The relationship among attribute and importance components of Rosenberg-Fishbein type attitude models: an empirical investigation. Proceedings, Second Annual Conference, Association for Consumer Research, 1971, 365-375.
Rosenberg, M. J. Cognitive structure and attitudinal affect. Journal of Abnormal and Social Psychology, 1956, 53, 367-372.
Rosenberg, M. J. An analysis of affective-cognitive consistency. In C. I. Hovland and M. J. Rosenberg (Eds.), Attitude organization and change. New Haven: Yale university Press, 1960.
Sampson, P. & P. Harris. A user's guide to Fishbein. Journal of the Market Research Society, 1970, 12, 145-166.
Schendel, D. E., W. L. Wilkie, and J. M. McCann. An experimental investigation of attitude importance. Proceedings, Second Annual Conference, Association for Consumer Research, 1971, 404-416.
Scott, J. E. & P. D. Bennett. Cognitive models of attitude structure: Value importance is important. Proceedings, American Marketing Association, 1971, 346-350.
Sheth, J. N. Measurement of attitude scores from beliefs and importances. College of Commerce and Business Administration. University of Illinois. Faculty working paper no. 50, 1972.
Sheth, J. N. Reply to comments on the nature and uses of expectancy-value models in consumer attitude research. Journal of Marketing Research, 1972, 9, 465-467.
Sheth, J. N. & W. W. Talarzyk. Perceived instrumentality and value importance as determinants of attitudes. Journal of Marketing Research, 1972, 9, 6-9.
Talarzyk, W. W. A reply to the response to Bass, Talarzyk and Sheth. Journal of Marketing Research, 1972, 9, 465-467.
Triandis, H. C. & M. Fishbein. Cognitive interaction in person perception. Journal of Abnormal and Social Psychology, 1963, 67, 446-453.
Wilkie, W. L. & R. P. Weinreich. Effects of the number and type of attitudes included in an attitude model: More is not better. Krannert Graduate School of Industrial Administration, Purdue University. Working paper no. 385, 1973.