Extending the Extended Model: Some Comments

Turning to the determinants of attitude, two very different types of questions have been raised. Ryan and Etzel were concerned with the question of how one identified salient beliefs, and Lutz raised some specific questions about the measurement of beliefs, as well as the combination rule underlying the expectancy-value model. Further, as I'm sure you all know, in an earlier paper (Ahtola, 1975) Olli also raised questions about belief measurement; in particular, the problem of distinguishing between belief strength and belief content.

There is unfortunately, relatively little I can say about the paper by Ryan and Etzel. Although I agree that the problem of identifying salient beliefs is a crucial one, I don't really see that their paper contributes very much to our current body of knowledge. What these investigators have shown is that when beliefs are elicited about different brands in the same product class, different beliefs may be salient for the different brands, and further, that the order (or frequency) of elicitation is not related to some independent estimate of the importance of the elicited attribute (or attribute dimension). Neither of these findings are new or exciting. Anthony Cowling (1973), as well as other investigators in Great Britain, has previously shown that different beliefs, or if you like, different attributes or attribute dimensions, are associated with different brands in the same product field. Similarly, Kalman Kaplan and I (Kaplan & Fishbein, 1969) provided evidence that judgments of attribute importance were not related to order of elicitation or frequency of elicitation of an attribute. In our recent book, however, Icek Ajzen and I (Fishbein & Ajzen, 1975) suggested that an attribute's importance might be reflected in the strength of a person's belief that the product has the attribute, and/or in his evaluation of that attribute. That is, it seems reasonable to assume that in many cases the more important the attribute, the more information a person may have about it, and thus the more certain he should be that the product has (or does not have) the attribute in question. Similarly, in many cases, the more important the attribute, the more likely the evaluation of the attribute will be extreme (or polarized). However, people can hold strong beliefs about attributes they view as unimportant, and they ma}, evaluate unimportant attributes very favorably or unfavorably. Thus, we would not expect a strong relation between either belief strength and importance or evaluation and importance. However, the importance of a given attribute may be reflected in the absolute value of the b_ie_i score associated with the attribute in question. Since either b_i or e_i (or both) may be at an extreme when the attribute is important, the absolute value of the b_ie_i score should be higher when the attribute is important than when the attribute is unimportant. Please note that if this hypothesis is correct, it would provide strong evidence that "importance" is taken into account by the expectancy-value formulation, and we could do away with importance once and for all. On the other hand, if the hypothesis is not correct, it does not mean importance should be incorporated in the expectancy-value formulation. We already know that this will not improve prediction (see e.g., Fishbein & Ajzen, 1975, Ch. 6). Further, underlying my use of the expectancy-value formulation is the assumption that a person's salient beliefs are the "important" beliefs since they determine the attitude. It's for this reason that it is necessary to first identify salient beliefs if one wants to understand the determinants of a given attitude. I would like to see Ryan and Etzel reanalyze their data in the light of these considerations.

Although the Ryan and Etzel paper has not contributed anything new to our body of knowledge, I don't want to simply dismiss it offhand. I do think it's very important to recognize that the beliefs that determine a person's attitude toward one brand in a product class may be very different from the beliefs that determine his attitude toward another brand in the same product class. Although I have not thought this out in any great detail, findings like those of Ryan and Etzel may have important implications for models and approaches that assume that choice behavior is determined by a comparison of two or more brands along a common set of "relevant" or "important'' dimensions. I might buy A rather than B because I think buying A is better than buying B but if the reasons underlying these two attitudes are very different, how meaningful is it to ask respondents to compare the two brands (or the two behaviors) on a set of common scales? I don't know the answer to this question but I think it is one possible direction that Ryan and Etzel could take.

I think the real question is not one of whether the belief measure should be an associative or probability measure, but rather a question of why both Rosenberg and I have used a bipolar belief measure if we're in agreement that belief strength or probability is what we're after. The answer to this question is relatively straightforward, and it's largely pragmatic.

I think it would be fair to say that both Rosenberg and I assumed that if a person believed that a given object did not have a negative attribute or if a given policy did not lead to (or blocked) a negative consequence this would imply a favorable attitude toward the object or policy. If beliefs were scored on a straight probability basis, (i.e., from 0 to 1), or on any unipolar scale, they could not reflect this double-negative. We both chose to measure our beliefs on a bipolar scale so that we could capture this double-negative property.

In retrospect, I have come to realize that the double-negative assumption is not necessary. More specifically, viewing beliefs as only associative rather than as associative or disassociative, forces one to reconsider the definition of the outcome or attribute associated with the object of belief. For example, consider the belief "X is not sweet". I can either view this as a belief that disassociates X and sweet or as a belief that associates X and "not sweet". Similarly, the belief that "X blocks Y" can either be viewed as a disassociation between X and "Y" or an association between X and "blocks Y". Viewing all beliefs as associative eliminates the need for a double-negative assumption, since the contribution of the belief to attitude will always be reflected in the evaluation of the associated attribute. Thus, as I indicated above, if the content of a person's belief is known, there is no question about how that belief should be measured -- I would want to measure the strength of that belief on a unipolar probability scale and the evaluative aspect of that belief on a bipolar evaluative scale. These measurements will tell me whether the belief is contributing positively or negatively to the overall attitude. For example, if someone believes that "Performing X blocks Y" and he thinks that "blocking Y" is "bad", I know this belief contributes negatively to this attitude. Similarly, if I know a person believes X leads to Y and that he positively evaluates Y, I know that this belief is contributing positively to his overall attitude. The problem comes when I don't know the content of the person's belief, i.e., when I don't know whether he believes that "performing X leads to or blocks Y". Consider the person described above who believes "X blocks Y" -- if I simply asked him to indicate how probable it was that "X leads to Y" he would indicate a low probability. Further, if I asked him to evaluate Y, he would evaluate it positively. Thus, if I scored the probability judgment from zero to one, I would infer that this belief was contributing positively to his attitude (e.g., .20 x +2 = .40). However, we have already seen that his belief is really that X blocks Y and he thinks that blocking Y is bad. Thus, this belief should contribute negatively to his attitude. By treating the belief dimension as a bipolar dimension (e.g., by subtracting .5 from the obtained probability) the negativity is captured (i.e., .2 - .5 = -.3 x +2 = -.6). What this procedure does not do is take into account the fact that the person's evaluation of (blocking Y) may not be the opposite of his evaluation of (Y). In our recent book, Icek Ajzen and I (Fishbein & Ajzen, 1975) discussed these problems in detail, and we suggested that when one could not assume symmetry in evaluation, the appropriate procedure would be to measure both of these beliefs. Olli has carried this even further by suggesting that we ask the respondent to provide us with all of his beliefs along the attribute (or outcome) dimension. That is, we should ask people how probable it is that "Behavior X leads to Y", that "Behavior X is unrelated to Y", that "Behavior X blocks Y", etc. He wants to identify all possible positions on a given dimension, measure belief strength with respect to each position, and then measure the evaluation of each of these possible outcomes. That is, we should also ask them to evaluate "leads to Y", "does not lead to Y", 'blocks Y", etc.

To return to the main point however, the problem is not in knowing how to define belief, but in arriving at a measurement procedure that best represents the theory. Both Rosenberg and I relied on bipolar measures of belief and evaluation. As Rich's work shows, this does correspond quite well to the cognitive algebra used by subjects. To summarize briefly then, if you don't know the content of a person's belief and if you can only measure one belief per attribute dimension (or if you only want to measure one belief) the most appropriate thing to do is to treat the belief dimension as bipolar. However, if you know the content of a person's belief, then you should use a unipolar measure. Whether the benefits obtained from asking more than one question per attribute dimension (i.e., whether following a procedure such as the one Olli suggested) will outweigh the costs (in subject time and cooperation as well as in instrumentation) is an empirical question that can only be answered by additional research.

Rich raised one other question about the expectancy-value model -- he reopened the old problem of adding versus averaging. While time does not permit a complete discussion of this problem, let me just say that although the Bettman, Capon and Lutz (1975) studies are an improvement upon earlier studies of adding vs. averaging, they still fail to provide an adequate test for resolving the adding-averaging controversy. At this point in time I'm still convinced that the relevant data supports an additive model, and while I am not willing to completely rule out the possibility that there may be averaging on some occasions or in some situations, I will stick to adding until valid data suggests otherwise. For a more complete discussion of this problem, see Fishbein and Ajzen (1975, Ch. 6).

More important, I think Mike is using a viable approach to attack some of the real problems with the normative component and its determinants. In most of my early work with the model, motivation to comply contributed little, if anything, to the understanding of intentions. At first, it seemed to me that this might have to do with the level at which motivation to comply was measured. Mike's studies measuring motivation to comply at three different levels of specificity is an important first step in untangling this concept. Unfortunately, his data suggest that the level at which one measures motivation to comply really doesn't seem to matter. While I still think this problem has to be resolved, I now think the major problem resides in the instrument we have used to measure motivation to comply. Theoretically, I have assumed that if a person believed that a given referent didn't want him to engage in some behavior and if the person were motivated not to comply with that referent, this would increase his intention to perform the behavior. Thus, as in the case of belief and evaluation we have measured normative beliefs and motivation to comply on bipolar scales (e.g., from +3 to -3) in order to capture this double negativity. As Mike pointed out, it's only recently that we have come to realize that this bipolar scoring is probably inappropriate for a scale that ranges from "I want to" to "I want not to do what referent X thinks I should do. This wording is clearly unipolar, and consistent with this, when we went hack and rescored the motivation to comply measure from 1 to 7 or 0 to 6, we found that motivation did contribute something to prediction, although its contribution was still relatively small. Rich's data also demonstrates that, at least in its present form, the motivation to comply measure should have been scored in a unipolar fashion since people treat it unipolarly in their cognitive algebra. Unfortunately, this does not capture the underlying theoretical assumption, and I personally feel that as a next step we should measure motivation to comply on a true bipolar scale. A scale like "I want to do - I want to do the opposite of what referent X thinks I should do" would be more in keeping with our underlying assumptions and may give motivation to comply an opportunity to contribute more to prediction. If nothing else, this type of scale will at least lead to a multiplicative cognitive algebra. I will return to this problem below.

First however, let me turn to the more specific suggestions that were raised. Although Olli suggested doing away with a general normative concept, he did want to retain the sum of normative beliefs weighted by motivation to comply. However, he again tries to make a distinction between belief content and belief strength, and he proposes the use of a vector model for this component as well as for the attitudinal component. Although the vector model may have some use on the attitudinal side, I do not think that this will be the case on the normative side. The reason for this is that despite Olli's elaborate argument, I do not think there is a problem of belief content when it comes to normative beliefs. The distinction between belief strength and belief content is important on the attitudinal side, for even if I know a person's belief that "X leads to Y", I cannot predict his belief that "X blocks Y". I can, however, predict his belief that "X does not lead to Y." That is, P(X) = 1 - P(X). On the normative side, all we are interested in is whether the person believes that a given referent thinks he should [P(x)] or should not [P(X)] engage in a particular behavior. Further qualifications such as "Referent A thinks I always should, or Referent A thinks I sometimes should engage in the behavior" will be reflected in the probability measure P(X). What I'm trying to point out is that I will obtain essentially the same information whether I use a "should-should not" scale or a "probable-improbable" scale, as long as I treat the probability scale as bipolar. Olli's argument that my two different measures imply two different concepts leads him to make the same mistake Rich made earlier -- he's reifying operations into concepts without paying attention to theory.

Olli was correct however, when he pointed out that, in at least one study, I did use an operation which did, in fact, define a different concept. Asking someone to indicate what another person expects him to do, is not the same as asking him what that other person thinks he should do. Mistakes like this are serious ones, and here differences in measurement procedures do change the meaning of the concept. But notice I used the word mistakes, and I would like to take this opportunity to make a general point. We all make mistakes, and in trying to develop the model I did, once, make the mistake of measuring normative beliefs in terms of expectancies. But I realized this was a mistake, I realized that expectations are not the same as shoulds. It has been over seven years since that study was conducted, and in every paper since then, I have defined normative beliefs in terms of should . Similarly, as I pointed out earlier, many people have tried to make an issue out of the fact that Rosenberg measured instrumentality on a "block-attain" scale while I measure belief on a probability scale. But Rosenberg used that scale in 1956; in his later work, he, too, referred to beliefs in terms of probability. I think that if someone is going to criticize another person's work, or compare the work of two people, the least the critic can do is consider the person's current work. One purpose of research is to enable us to learn and modify our thinking. I don't feel we should be accused of inconsistency because of early mistakes that have been corrected in more recent work.

To return to the problem at hand, however, Olli had one other suggestion or comment -- namely he argued that the motivation to comply measure was not tapping all of the bases of social influence and he suggested some alternative measures. Olli may be right but at this stage he's merely speculating, and I think it's a serious mistake for anyone to suggest changes in someone else's measurement procedures without first providing empirical evidence that such changes are warranted. As I said earlier, I too think that motivation to comply has been badly measured, but I think that one is much more likely to resolve problems by doing the kind of work that Mike Glassman is doing than by introducing unsupported speculations into the literature. Unfortunately, I think this same criticism applies to many of Rich's suggestions. Since Rich has made several different suggestions, let me try to discuss them one by one.

Further, it's interesting to note that, at this stage, Rich's own data invalidate his objection. That is, as I pointed out above, the objection is only valid if the measure of motivation to comply can take on negative values. Rich's data, as well as Mike Glassman's, point out that in all our research to date, motivation to comply has been measured as a unipolar construct. Rich, however, obviously does believe that motivation to comply can take on negative values. Indeed, he suggested that the subjective norm would be better defined in terms of a global motivation to comply. That is, by assuming that b_j (normative beliefs) was analogous to b_i (instrumental beliefs) and m_j was analogous to e_i, Rich arrived at a purely motivational description of the normative component and he suggested that SN could be measured by a scale such as: "With respect to performing behavior X, I want to -- I want not to do what other people think I should do." Unfortunately, as Rich himself points out, although he views this concept as bipolar, it is not directional with respect to the behavior. While a scale such as I want to -- I want to do the opposite" will better capture the bipolarity of the motivational concept, it will still not be directional with respect to the behavior, and despite his claim to the contrary, I really don't see how this nondirectional concept could contribute to prediction of a directional intention.

Rich's alternative suggestion for improving the measure of the subjective norm was that it might be better defined as a "social attitude". This suggestion is just plain wrong. Dividing beliefs about performing a given behavior into two arbitrary categories will not solve any problems. In fact, what a procedure of this type will do is provide two different estimates of exactly the same attitude. Thus, rather than reducing the multicollinearity between the attitudinal and normative components, this procedure will actually increase it. More important, such a procedure would eliminate any consideration of normative pressure. Since Rich would merely have two different estimates of the person's attitude toward performing the behavior he would have no normative component at all. On the basis of a great deal of previous research in many different areas, I can assure you that, despite the fact that different types of beliefs may be used to estimate the attitude, the two estimates will be very highly correlated.

The Relative Weights of the Two Components. The final set of problems relates to the meaning of the weights attached to the attitudinal and normative components as factors influencing intentions. Olli argues that since these weights are assumed to vary across individuals and behaviors, they are not parameters but variables, and thus multiple regression procedures for estimating the weights are inappropriate. Rich also questions the use of a regression approach as a means of estimating weights, hut on somewhat different grounds. He points out that since SN and A_B are usually highly correlated, their multicollinearity raises serious questions about the reliability and validity of the obtained regression weights.

While there is some truth in what they say, I'm not quite sure what to do about it. One thing I do know, however, is that I do not want to ask people to tell me how much importance they place on each component in determining their intentions. There is already more than enough data coming from studies of cue utilization to demonstrate that people cannot provide accurate estimates of the relative weights they place on different cues in arriving at judgments.

Further, I must admit I am not terribly concerned about the "multicollinearity problem". That is, I think in most cases in the "real world" there are "true" correlations between subjective norms and attitudes toward behaviors, and I think that any attempt to eliminate this correlation would be inappropriate. My current position is that I view the weights as theoretical parameters (parameters can vary), and correctly or incorrectly, I see multiple regression as one way of estimating these parameters. Despite the many problems associated with the multiple regression procedure, we have been able to show that weights obtained through multiple regression analyses do covary in accordance with our theoretical predictions. For example, consistent with expectations, SN was found to be more important (received a higher regression weight) than A_B in cooperative situations, while it was less important than A_B in competitive situations. Similarly as predicted, authoritarians did tend to place more weight on the normative component than did egalitarians. Given findings like this, I feel that despite some obvious inadequacies, the use of multiple-regression procedures to estimate the relative weights of the two components has yielded useful and meaningful results and in the absence of a viable alternative I will continue to use this procedure. Let me make it clear however, that I would really like to find some independent method for estimating or measuring the weights for each individual with respect to any given behavior. Unfortunately, I don't know how this can be done at the present time, but I do think this is a very important area for research.

Before concluding, there is one additional point I would like to make. In the course of the session, both Mike and Olli referred to my concern with specificity. Olli said that if I was concerned with specificity I should measure attitudes toward "my performance" Of a behavior rather than "performing" the behavior per se, and Mike argued that to be consistent with my concern with specificity he feels that motivation to comply should be measured at the most specific level. Since I agree with Olli and disagree with Mike, it should be clear that my concern is not with specificity. However, many other people have also criticized the model for being too specific, and they have questioned the value of a "specific" concept such as an attitude toward an act. I would like to take this opportunity to point out that I am not now, nor have I ever been, concerned with specificity. What I am concerned with is correspondence. The model is a general model, and it is designed to account for people's intentions to perform any act. From a theoretical point of view, it makes no difference whether we are trying to predict a very specific intention (e.g., the intention to buy Crest toothpaste the next time I go shopping at my local A & P) or a very general intention (e.g., the intention to buy toothpaste). Once an intention has been identified, however, the model does state that all other measurements must correspond to that intention if one wants to obtain satisfactory prediction.

If I'm interested in a consumer's intentions to "buy toothpaste", then, according to the model, I must measure his or her attitude toward "buying toothpaste" and his or her subjective norm with respect to "buying toothpaste". Similarly, if I want to know the beliefs underlying this attitude or the subjective norm, I should be assessing his or her salient beliefs about "buying toothpaste" or his or her normative beliefs about "buying toothpaste". On the other hand, if I'm interested in a consumer's intention to "buy Crest toothpaste the next time I go shopping at my local A & P", the model says I must measure attitudes toward "buying Crest toothpaste the next time I go shopping at my local A & P" the subjective norm concerning "buying Crest toothpaste the next time I go shopping at my local A & P", etc. To put this somewhat differently, behavioral criteria and intentions can be described in terms of four elements: the action, the target the action is directed at, the situation or context in which the action occurs, and the time at which the action occurs. All four elements can vary along dimensions of specificity. For example, the target may be a specific brand (e.g., Wheaties) a relatively specific product class (e.g., dry cereals), a more general product class (breakfast cereals), a fairly global product class (food), or it can be left unspecified. The model doesn't care what level of specificity is adopted. However, once a behavioral criterion has been identified (i.e., defined in terms of the four elements), the intention must correspond directly to that criterion. Similarly, measures of beliefs, attitudes, and subjective norms must also correspond to the intention and the behavior. It is only under conditions of correspondence that maximal prediction can be expected. The actual level of specificity of the elements one deals with should be determined by the problem or behavior you are investigating.

REFERENCES

O. T. Ahtola, "The Vector Model of Preferences: An Alternative to the Fishbein Model," Journal of Marketing Research, 12, 1975, 52-59.

I. Ajzen and M. Fishbein, "Attitude-behavior Relations: A Theoretical Analysis and Review of Empirical Research", unpublished manuscript, 1975.

J. R. Bettman, N. Capon, and R. J. Lutz, "Information Processing in Attitude Formation and Change", Communication Research, 2, 1975, 267-278.

A. Cowling, A. "Use of Elicitation Technique for Producing Dimensions of Brand Choice", Sixteenth annual conference papers, Market Research Society of Great Britain, 1973.

M. Fishbein and I. Ajzen, Belief, Attitude, Intention and Behavior: An Introduction to Theory and Research (Boston: Addison-Wesley, 1975)

C. A. Insko, R. R. Blake, R. B. Cialdini, and S. A. Mulaik, "Attitude Toward Birth Control and Cognitive Consistency: Theoretical and Practical Implications of Survey Data", Journal of Personality and Social Psychology, 16, 1970, 228-237.

K. J. Kaplan and M. Fishbein, "The Source of Beliefs, Their Saliency and Prediction of Attitude", Journal of Social Psychology, 78, 1969, 63-74.

V. Kothandapani, "Validation of Feeling, Belief, and Intention to Act as Three Components of Attitude and Their Contribution to Prediction of Contraceptive Behavior," Journal of Personality and Social Psychology, 19, 1971, 321-333.

M. J. Rosenberg, "Cognitive Structure and Attitudinal Affect," Journal of Abnormal and Social Psychology, 53, 1956, 367-372.

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