To Add Importance Or Not to Add Importance: That Is the Question

Citation:

James R. Bettman (1974) ,"To Add Importance Or Not to Add Importance: That Is the Question", in NA - Advances in Consumer Research Volume 01, eds. Scott Ward and Peter Wright, Ann Abor, MI : Association for Consumer Research, Pages: 291-301.

Advances in Consumer Research Volume 1, 1974    Pages 291-301

TO ADD IMPORTANCE OR NOT TO ADD IMPORTANCE: THAT IS THE QUESTION

James R. Bettman, University of California, Los Angeles

In recent years much work has been done in marketing using multiattribute models, particularly expectancy-value models of the type developed by Rosenberg (1956) and Fishbein (1972). This work is summarized and analyzed in some detail by Wilkie and Pessemier (1972). The general form of these models, as used in marketing, is expressed in equation (1):

(1)    EQUATION

where Aj = attitude toward brand j, ai = the evaluative aspect of attribute i, its goodness or badness, or the importance of attribute i; bij = the strength of the individual's belief that brand j possesses attribute i, or the extent to which brand j possesses attribute i; n = number of relevant attributes for the product class. The use of evaluative aspect and strength of belief has been associated with the Fishbein approach, and importance and extent of possession with Rosenberg (Cohen, Fishbein, and Ahtola, 1972).

In the empirical work to date a rather spirited controversy has arisen over the specific form of the model. In particular, Sheth and Talarzyk (1972) first argued, based on results of Rosenberg (1956), that prediction of attitude is actually better if the evaluative or importance term ai is deleted from the model. That is, they postulated a suppression effect for the ai. They proposed a model involving only the belief or possession scores, as in equation (2).

(2)    EQUATION

Since this initial test there have been studies supporting this notion (e.g., Moinpour and MacLachlan, 1971; Lutz and Howard, 1971), studies supporting the original model (e.g., Hansen, 1969; Lehmann, 1971) and studies finding little difference (e.g., Cohen and Ahtola, 1971) (see Wilkie and Pessemier, 1972, 37-47). Explanations for these differing results have tended to be totally empirical and atheoretical, such as sample heterogeneity and use of nonnormalized ai measures. This is not to deny that measurement problems may be important. However, a theoretical explanation might lead to more insights.

The purpose of this paper is to offer a theory, based on information processing model notions, that specifies when the ai term should improve prediction of attitude, and when it should be deleterious to this prediction. A rough review of the studies investigating the issue to date is then carried out to determine whether the theory is generally upheld. The results show that the theory may have some merit, so finally an empirical study is reported which tests the theory more directly. Important implications of the theory for attitude research are also examined.

INFORMATION PROCESSING THEORY AND EXPECTANCY-VALUE MODELS

The basic assumption to be used in the theoretical argument is to suppose that consumers actually use decision nets in making brand choices. This assumption is borne out by Bettman (1970), Alexis, Haines, and Simon (1968), and Haines (1969). A decision net, such as the hypothetical net shown in Figure 1, can be interpreted as follows: each point in the net represents a test of a particular brand attribute or condition (e.g., is price high?). The arcs out of each node depict the sequence of processing taken, depending upon the outcome of the test. That is, the arcs lead to whichever condition or attribute should be checked next, given the outcome at the present node. Eventually the consumer processes an alternative (brand) down a particular path through the net, given its attribute values, and an accept or reject decision is made. For example, in Figure 1, there ere five attributes or conditions to be checked. Processing would proceed as follows: if attribute 1 is not present, reject, if it is present, check attribute 2, and so on. Once the net is given, the flow of processing is deterministic in that attribute 1 is checked first, etc. Haines (1969) has argued that such decision nets represent the structural aspects of attitude. If a brand passes through the net and is accepted, or purchased, such a view would imply that the attitude for that brand is more favorable than for brands which are rejected. This view is essentially the same as that used in studies such as Bass and Talarzyk (1972), where attitude scores are used to predict preference orders.

FIGURE 1

HYPOTHETICAL DECISION NET WITH FIVE ATTRIBUTES

Thus far it has been assumed that consumers in actuality use decision nets. In addition, an assumption about the structure of the nets used is made. Assume that consumers use simple one branch decision nets such as that shown in Figure 1. That is, there is essentially a single sequence of attributes in the net, rather than a split to multiple branches. Although this assumption seems highly simplifying, many nets reported in the literature have this form or a structure very close to it (Alexis, Haines, and Simon, 1968; Haines, 1969). The crucial property of this kind of decision net for the following discussion is that the order of the nodes becomes unimportant in determining acceptance or rejection of a brand if all attributes are considered, as is assumed for the moment. That is, it is essentially a conjunctive model where all attributes must be present for acceptance. Hence, the order in which they are examined is irrelevant for purposes of prediction. To summarize the arguments to this point, consumers are assumed to use single branch decision nets which examine all relevant attributes of a brand. If a brand is accepted by the decision net it is assumed that the consumer's attitude toward that brand is higher than for rejected brands.

To go through the net, the consumer must decide, for each attribute considered, whether or not the brand being processed possesses that particular attribute. Although the processing order of the attributes is assumed to be fixed and deterministic, this decision about attribute possession is not. The consumer is rarely certain that a brand possesses a particular attribute or quality to the extent desired; rather there is some degree of uncertainty. Let Xij be a binary variable representing attribute i for brand j, where Xij = 1 if brand j possesses attribute i to the desired extent, and Xij = 0 if lt does not. Then the consumer, to decide which arc out of the point representing attribute i should be taken, must subjectively assess Prob (Xij = 1). [This notation means the individual's subjective probability that Xij =1.] If this probability is high enough, then the consumer will follow the branch corresponding to possession of attribute i. Note that this assumes a cut-off point for the subjective probability of possession for each attribute, and hence that Prob (Xij = 1) varies over time and situations for an individual, rather than being fixed. For example, in Figure 1 the consumer must first assess, for a brand j, Prob (Xij = 1). If this is high enough, or above the cut-off point, then the Y branch will be taken out of the first node. Another way of stating this notion is that consumers make attributions of attribute possession to brands, but that the attributions are probabilistic.

Therefore, the higher that Prob (Xij = 1) is, the more likely it is that the consumer will follow the branch in the decision net corresponding to possession of the attribute. For a one branch net like Figure 1, with n attributes, the likelihood that brand j successfully traverses the net and is accepted should then be approximated [It is only approximate for two reasons: 1) it assumes the attribute assessments are independent; 2) since there is a cut-off point for Prob (Xij = 1) above which the consumer will go down the branch for possession, the probability of acceptance is related to the likelihood that Prob (Xij = 1) is above the cut-off point for each attributes rather than directly to Prob (Xij = 1) itself.] by EQUATION Prob (Xij = 1). But the measure bij used in multi-attribute attitude models is closely related to Prob (Xij = 1), and is in fact conceptually almost identical with it in Fishbein's view (Fishbein, 1972, p. 247). Although there is this conceptual linkage, bij in practice is usually not measured on a scale between 0 and 1. However, since the measures used for bij should be highly correlated with the underlying probabilities, the probability of acceptance for brand j should be highly correlated with Ni bij . Then it can be further argued that the probability of acceptance for brand j is at least roughly monotonically related [It is not strictly monotonically related, as it is clear that the sum of the bij can be increased overall by having some bij increase and some decrease. Such a change can lead to either an increase or decrease in the product of the bij. However, if only a single attribute possession score changes, or a B changes are increases or all decreases, then the sum and product are monotonically related, assuming that bij is measured on a non-negative scale. If it is not, then it must be transformed to be non-negative. This argument is supported by Lutz (1973), who found that after a communication designed to influence a single bij was given, the great majority of changes in the other possession score's were in the same direction as the change in that single element.] to, and hence reasonably highly correlated with Ei bij. But this is the attitude model of equation (2). These arguments certainly involve several stages of approximation. Probability of acceptance is approximated by a product of subjective probabilities of possession which is highly correlated with a product of b measures, which itself is correlated with the sum of the same bij measures. This may seem like too great a series of approximations to be acceptable. However, Dawes (1972) points out that empirically, linear models have very high correlations with those nonlinear functions of the same variables with which they are roughly monotonically related, so that at least one stage in the argument is not weak. Since degree of fit is often used to assess attitude models, correlationS arguments are really all that is necessary. What has been accomplished thus far is to argue that is consumers really use decision nets and these nets constitute their attitude structures, but a multi-attribute model is used to describe attitudes instead, then the Ei bij model should fit best. However, two crucial points were used in obtaining this result: 1) for a single branch net, all attributes in the net must be satisfactory, and 2) we assumed all relevant attributes are in the net for each decision.

Suppose the assumption that all relevant attributes are processed each time the consumer uses a decision net for a product class is relaxed. Nakanishi (1972) proposes a different model, called the contiguous retrieval model. In this model, a subset of the relevant attributes are considered. A single branch net is still assumed for each decision, but using only this subset of attributes. Nakanishi further argues that the selection of attributes in the subset is governed by the centrality of the attribute for the particular decision. It can be argued that ai provides a measure related to this concept of centrality. Tigert (1966), for example, found a rank-order correlation of .82 between a measure of attribute importance and the order in which subjects requested brand attribute information in choosing between two alternatives.

For this case where not all attributes are considered, attitude should be related to the sum of the bij for those attributes i which ere in the subset actually used by the consumer. If an attribute were not considered, adding in its bij term would not aid in prediction, but wouLd be adding error to the model. Hence, for the case where only a subset of the attributes are examined, the probability of acceptance and thus attitude may be more closely related to Ei aibij than to Ei bij, since the former expression does include, via the ai, some effects related to selection of the subset of attributes. The bij for those elements not likely to be selected as part of the subset should have small ai values and hence not be counted heavily in the resultant summation. As an extreme example of this argument, suppose attributes 1 and 2 were much more central to the decision for a consumer using the decision net of Figure 1 than the other attributes. If his ai measures on a 0 to 6 scale were 6,6,0,0,0 for the five attributes, then Ei aibij reduces to 6b1j + 6b2j, or 6(b1j + b2j), which is equivalent to the sum of the bij for the attributes in the subset. Although one wouLd certainly not expect She results to be quite this neat in practice, this example illustrates how the ai term can account for attribute selection effects. [For the Fishbein model, bipolar scales are used to assess ai, e.g., -5 (very bad) to +5 (very good). In this case, the arguments above hold with respect to the absolute value of ai rather than ai itself. That is, small absolute values for ai should correspond to those attributes not likely to be used by the consumer.]

Finally, note that all of the above arguments refer to the summative form of the model. However, Sheth (1970) and Cohen and Ahtola (1971) have argued that a disaggregative form should be used. In this case, the arguments above would imply that a disaggregated model using bij should be more closely related to probability of brand acceptance than one using aibij units as variables, regardless of the number of attributes selected, since each bij then receives an individual weight.

The final argument to be advanced concerns the conditions under which one would expect all attributes to be considered. Slovic (1972) argues from a great many research studies that man's limited information processing capacity leads him to apply decision making strategies that reduce cognItive strain. Cognitive strain increases with the number of attributes and as difficulty of attribute evaluation increases. Hence, all attributes would be considered probably only in those situations where small numbers of attributes and relatively simple, straightforward alternatives are considered. For large numbers of attributes or more complex alternatives (i.e., those where there are potentially many attributes and where attributes are difficult to evaLuate), one might argue that information processing demands are higher and attribute selection will occur. Attribute selection has been found in marketing studies, and results are summarized in Wilkie and Pessemier (1972), 18-24. These results indicate that somewhere around 5-6 attributes is a reasonably small number, and that beyond this number attribute selection seems to occur.

In summary, if decision net models were in fact trues then the arguments above lead to the following predictions about the relative fit of the two multi-attribute modeLs:

P1: For < 6 attributes and simple alternatives, Model 2 (EQUATION) should be more appropriate if a summative model is used.

P2: For > 6 attributes or complex alternatives, Model 1 (EQUATION) should be more appropriate if a summative model is used.

P3: If a disaggregative model is used, Model 2 should be superior.

Now these predictions are checked by examining the results of past studies.

RESULTS FROM PREVIOUS STUDIES

By examining the resuLts of past studies on this problem, one can obtain a rough idea of the validity of the predictions. Table 1 outlines several of the studies performed, giving the type of alternatives used and their number (not all brands), the number of attributes measured, whether or not the alternatives are judged complex or simple, whether the models used were summative or disaggregated, which model was better, and whether or not the predictions given above were supported. Only a rough notion can be obtained for severaL reasons: the various studies used vastly different methodologies and measurement techniques; the judgement of attribute complexity is the author's subJective judgement and was not obtained from the subjects themselves; different techniques for analyzing resuLts were used by the various authors; and not all relevant studies are included. [Also, the hypotheses are really based on arguments about processes of individuals, and most of the studies report aggregate models. However, given the nature of most past work in the area, this is unavoidable.] Table 1 shows that on the whole the predictions are upheld. A few specific comments about some of the studies ere in orders however.

It is difficult to interpret the Hansen (1969) resuLts, since t-tests were used to compare chosen and rejected alternatives, rather than correlations. For the 3 attribute hairdryer choice, the Ei aibij model seemed slightly better, but the Ei bij model also performed well. In the Bass, Pessemier, and Lehmann (1972) study, studies were performed using from 5 to 8 attributes. As the number of attributes increased, adding ai terms seemed to improve resuLts. However, since many model variations were performed in their study, the only case where a direct comparison of Models 1 and 2 was possible Prom their stated results was the 5 attribute case. Finally, Churchill's (1972) results showed an interesting pattern in that Models 1 and 2 were more even for 1-5 and 12-19 attributes, whereas for 6-11 attributes Model 2 was much better.

The overall pattern of resuLts is certainLy not uniformly in favor of the predictions. However, there is a relatively large measure of support. To really test the predictions however, a study is needed where the same measurement techniques are used for a range of numbers of attributes and rated complexity of alternatives and the various models are compared, rather than the type of comparison reported here, where almost all the studies use different measures and methodologies. Sheth (1972) areas for the same type of study.

In spite of all the problems with examining these past studies, there is enough support to justify direct empirical investigation. Although a study such as that outlined above is beyond the scope of this paper, a more limited study was carried out to directly examine the arguments-presented.

TABLE 1

RESULTS OF ATTITUDE MODEL STUDIES

AN EMPIRICAL STUDY

Based on the arguments above, if consumers in fact used decision nets for a particular product class rather than a linear compensatory model, then Ni bij should predict individuals' attitudes best. This prediction was examined directly. The indirect evidence of the Ei bij model was used above simply because direct evidence on the product model was not available from previous studies.

The product category used was toothpaste (Bass and Talarzyk, 1972; Cohen and Ahtola, 1971; Sheth and Talarzyk, 1972). The five attributes used for the study were those found relevant in the earlier work: whitening teeth, preventing cavities and tooth decays economical in use, freshening breath, and pleasant tasting. Seven brands were used: Ultra Brite, Pepsodent, Macleans, Crest, Close Up, Colgate, and Gleem. Subjects were 121 graduate and undergraduate students.

For each subject, measures were obtained for attitude toward the act of buying each brand (Aj); the evaluative aspect of attribute i (ai); and beliefs (bij). In response to the criticism of Cohen, Fishbein, and Ahtola (1972) that appropriate measures of model components were not being used, the AB scales developed by Fishbein and Raven (1967) were used to measure Aj, ai, and bij. The AB scales are semantic differential scales, with five adjective pairs utilized in the A scale for measuring ai and Aj (good-bad, clean-dirty, healthy-sick, wise-foolish, and beneficial-harmful) and five in the B scale for measuring bij (possible-impossible, true-false, existent-nonexistent, probable-improbable, and likely-unlikely). The concepts rated ares for examples "Buying Ultra Brite toothpaste for my own use is:" (Aj); "Having white teeth is:" (ai); The statement 'Ultra Brite toothpaste whitens teeth' is:" (bij). Five filler pairs are added for each scale, and the order randomized. Since each pair is scored on a -3 to +3 bipolar scale, the final measures for Aj, ai, and bi;, obtained by summing the scores for the five relevant pairs, range from -15 to +15.

To test the hypothesis that individuals use net models rather than summative models, it was decided to use three models of Aj: Ei aibij (Model 1), Ei bij (Model 2), and Ni (bij + 15) (Model 3). Note that for Model 3, since the b scores can be as low as -15, 15 was added to each bij measure to insure that the product would be non-negative. Such a transformation affects the degree of fit of a multiplicative model. However, it was felt that the transformation used was the simplest one consistent with the objective of making each bij non-negative, which is necessary for the product model to make sense. The analysis of the models for each subject was carried out as follows. Since models of each individual are needed to really test the hypothesis, the normal aggregated approach was not used. Instead, since seven brands were rated by each subject, a regression was run for each subjects for each models using the data for the seven brands as the data points. This is the approach used by Beckwith and Lehmann (1973). Since there were five attributes, all attributes were included in the analysis, in line with the earlier arguments. Data reported by Bass and Wilkie (1973) for toothpaste also support including all five attributes. Since 11 subjects rated all brands the same for all measures, regressions for these subjects could not be run. Hence, the final sample size was 110. For each of these 110 subjects, then, a value of R2 for each of the three models was available.

If individuals use nets, then the arguments in earlier sections would imply that the product model should fit subjects' attitude ratings best. However, the Ei bij model is ambiguous. If it fits best, the arguments above imply that there is some support for decision nets. However, it is still clearly a summative type model. Hence, two sets of analyses were run. One compares Models 1, 2, and 3. The second only includes Models 1 and 3. The model selected as best for each subject is that with the highest R value for that subject.

TABLE 2

INDIVIDUAL SUBJECT ATTITUDE MODEL RESULTS

The results of the first analysis, comparing all three models, are shown in Table 2. Approximately equal numbers of subjects are best fit by Models 1 and 3, with Model 2 best fitting about half as many. Roughly half of the R2 values for the best fitting model are statistically significant. These results show that at least some substantial proportion of the subjects may be using net type models, since Model 3 fits their data best. Since Model 2 is ambiguous to some extent in terms of what type of model it supports, Table 2 also shows the results of comparing only Models 1 and 3. Again, about half of the subjects are best fit by Model 3.

IMPLICATIONS

One must be careful in interpreting any results such as those presented here, since, as Hoffman (1960) pointed out, our regression models are only paramorphic representations of decision rules, and may not accurately represent underlying processes. With this caution in minds we have found that product models, which purportedly support a decision net model of attitude and choice, fit a relatively large proportion of subJects better than the standard multiattribute attitude model. Let us examine how this conclusion was obtained.

In essence, assuming that decision nets are really what consumers use led to some predictions about how various attitude models should perform in certain situations. Of course the decision net models assumed (one branch nets) are very simplistic, and this crucial assumption that consumers use only such nets in many situations has not been directly proven. Certainly such nets are not used in all cases (Bettman, 1970). However, for nets dealing with only one product class, single branch models seem empirically to be appropriate descriptors (Haines, 1969, Alexis, Haines, and Simon, 1968). Since the attitude models are for single product classes, the assumption of single branch nets seems justifiable in this context. In addition to this assumption, the arguments used were rough. For example, it was assumed that only bij affects Prob (Xij = 1). However, it may be that judging that an attribute is satisfactory is different from judging that a brand has some degree of possession of an attribute. The variation in bij across brands or perhaps ai might also be related to the decision a consumer must make as to whether or not a brand is satisfactory for a given attribute.

However, despite all of the problems and cautions, both the analysis of previous studies and the study reported here support the decision net approach as a valid descriptor of consumer choice processes. If this is trues then multi-attribute attitude models may in fact be inappropriate as descriptors of human information processing for some substantial proportion of consumers. They may have predictive values but research aimed at understanding how consumers make choices may be emphasizing the wrong approach in concentrating on these models alone.

REFERENCES

Alexis, Marcus, George Haines, and Leonard Simon. Consumer Information Processing: The Case of Women's Clothing. In Proceedings of The American Marketing Association Fan Conference, 1968, 197-205.

Bass, Frank, Edgar Pessemier, and Donald Lehmann. An Experimental Study of the Relationships between Attitudes, Brand Preference, and Choice. Behavioral Science, 17 (November 1972), 532-41.

Boss, Frank and W. Wayne Talarzyk. An Attitude Model for The Study of Brand Preference. Journal of Marketing Research, IX (February 1972), 93-6.

Bass, Frank, and William Wilkie. A Comparative Analysis of Attitudinal Predictions of Brand Preference. Journal of Marketing Research, X (August 1973), 262-9.

Beckwith, Neil, and Donald Lehmann. The Importance of Differential Weights in Multiple Attribute Models of Consumer Attitude. Journal of Marketing Research, May 1973, 141-5.

Bettman, James. Information Processing Models of Consumer Behavior. Journal of Marketing Research, VII (August 1970), 370-6.

Churchill, Gilbert. Linear Attitude Models: A Study of Predictive Ability. Journal of Marketing Research, IX (November 1972), 423-26.

Cohen, Joel, and Olli Ahtola. An Expectancy x Value Analysis of the Relationship between Consumer Attitudes and Behavior. In D. Gardner, Ed., Proceedings of the Second Annual Conference of The Association for Consumer Research, 1971, 344-64.

Cohen, Joel, Martin Fishbein, and Olli Ahtola. The Nature and Uses of Expectancy-Value Models in Consumer Behavior Research. Journal of Marketing Research, IX (November 1972), 456-60.

Dawes, Robyn. Slitting the Decision Maker's Throat with Occam's Razor: The Superiority of Random Linear Models to Real Judges. Oregon Research Institute Bulletin, Vol. 12, No. 13, 1972.

Fishbein, Martin. The Search for Attitudinal-Behavioral Consistency. In Joel Cohen, Ed., Behavioral Science Foundations of Consumer Behavior New York: The Free Press, 1972, 245-52.

Fishbein, Martin, and Bertram Raven. The AB Scales: An Operational Definition of Belief and Attitude. In Martin Fishbein, Ed., Readings in Attitude Theory and Measurement, New York: John Wiley, 1967, 183-9.

Haines George. Information and Consumer Behavior. Working Paper Series 6901, Graduate School of Management, University of Rochester, 1969.

Hansen, Flemming. Consumer Choice Behavior: An Experimental Approach. Journal of Marketing Research, VI (November 1969), 436-43.

Hansen, Flemming, and Thomas Bolland. Relationship between Cognitive Models of Choice and Non-metric Multi-dimensional Scaling. In D. Gardner, Ed., Proceedings of The Second Annual Conference of The Association for Consumer Research, 19 %, 376-88.

Hoffman, Paul. The Paramorphic Representation of Clinical Judgment. Psychological Bulletin, 57 (March 1960), 116-31.

Lehmann, Donald. Television Show Preference: Application of a Choice Model. Journal of Marketing Research, VIII (February 1971), 47-55.

Lutz, Richard. Cognitive Change and Attitude Change: A Validation Study. Unpublished doctoral dissertation, University of Illinois, 1973.

Lutz, Richard, and John Howard. Toward a Comprehensive View of The Attitude-Behavior Relationship: The Use of MuLtiple-Set Canonical Analysis. Proceedings of The Social Statistics Sections American Statistical Association, 19 X, 215-24.

Moinpour, Reza, and Douglas MacLachlan. The ReLations among Attribute and Importance Components of Rosenberg-Fishbein Type Attitude Model: An Empirical Investigation. In D. Gardner, Ed., Proceedings of The Second Annual Conference of The Association for Consumer Research, 19 n, 365-75.

Nakanishi, Masao. Decision Net Models and Human Information Processing. Paper presented at the ACR/AMA Workshop on Buyer/Consumer Information Processing, Chicago, November 1972.

Rosenberg, Milton. Cognitive Structure and Attitudinal Affect. Journal of AbnormaL and Social Psychology, 53 (November 1956), 367-72.

Sheth, Jagdish. An Investigation of Relationships Among EvaLuative Belief, Affect, BehavioraL Intention and Behavior. Unpublished paper, University of Illinois, 1970.

Sheth, Jagdish. Reply to Comments on The Nature and Uses of Expectancy-Value Models in Consumer Attitude Research. JournaL of Marketing Research, rx (November 1972), 462-6.

Sheth, Jagdish, and W. Wayne Talarzyk. Perceived Instrumentality and Value Importance as Determinants of Attitudes. JournaL of Marketing Research, IX (February 1972), 6-9.

Slovic, Paul. From Shakespeare to Simon: Speculations - And Some Evidence About Man's Ability to Process Information. Oregon Research Institute Monograph, 1972, Vol. 12, No. 2.

Tigert, Douglas. An Investigation of Experimental Techniques to Determine The Relative Importance of Various Product Attributes in The Consumer Decision Making Process. Institute Paper 126, Krannert Graduate School of Industrial Administration, Purdue University, 1966.

Weddle, David and James Bettman. Marketing in The Underground: An Investigation of Fishbein's Behavioral Intention Model. In S. Ward, Ed., Proceedings of the Fourth AnnuaL Conference of the Association for Consumer Research, 1973 (forthcoming).

Wilkie, WilLiam, and Edgar Pessemier. Issues in Marketing's Use of MultiAttribute Attitude ModeLs. Institute Paper 365, Krannert Graduate School of IndustriaL Administration, Purdue University, 1972.

----------------------------------------