The Moderating Effect of Product Knowledge on Multi-Attribute Attitude Model Predictions

Richard L. Oliver, University of Kentucky
[ to cite ]:
Richard L. Oliver (1975) ,"The Moderating Effect of Product Knowledge on Multi-Attribute Attitude Model Predictions", in NA - Advances in Consumer Research Volume 02, eds. Mary Jane Schlinger, Ann Abor, MI : Association for Consumer Research, Pages: 295-306.

Advances in Consumer Research Volume 2, 1975      Pages 295-306

THE MODERATING EFFECT OF PRODUCT KNOWLEDGE ON MULTI-ATTRIBUTE ATTITUDE MODEL PREDICTIONS

Richard L. Oliver, University of Kentucky

[Richard L. Oliver is Assistant Professor in the Department of Business Administration, University of Kentucky. He thanks Professor Jagdish N. Sheth of the University of Illinois for his helpful comments and the University of Kentucky Research Foundation for providing partial support for this research.]

The effect of product familiarity on instrumentality motel predictions was assessed by adding "don't know" (DK) alternatives to the attribute instrumentality set of a new product attitude study. Subjects were classified into subsamples based on the number of DK responses and predictor-criterion correlations were then calculated for each subsample separately. Results obtained with two instrumentality models showed that the correlations were positively related to product knowledge although a reversal of this trend was observed at the high DK extreme. Further analysis indicated that number of DKs could be predicted from product familiarity variables and product specific confidence. The results imply that nonrational explanations for attitude formation may be required when the assumption of product familiarity is not tenable

The predictive ability of various multiple attribute attitude motels including those of Rosenberg (1956), Fishbein (1967), and others reported in Cohen, Fishbein, ant Ahtola (1972) is fairly well established (Bass & Wilkie, 1973; Wilkie & Pessemier, 1973). In fact, the current status of these models is such that poor results can usually be traced to methodological problems (Fishbein, 1971). This has served to shift the academic interest in multi-attribute motels from simple attempts to predict overall affect to discovering conceptual improvements that will increase the accuracy of the motel's predictions (e.g., Bass & Wilkie, 1973; Craig & Ginter, 1974; Ahtola, 1973). One implication of this shift in emphasis is that poor results are interpreted as meaning that the researcher has misconceptualized his model or has committed a methodological error. The purpose of this paper is to demonstrate that low correlations between affect and summated instrumentality times value terms may result from factors other than concept or method.

The various multi-attribute attitude formulations used in consumer behavior are essentially rational models of consumer preference. Given a brand j and r. salient attributes, a consumer is presumed to base his overall affect for that brand on the summated products of the degree to which brand j possesses attribute i (instrumentality) and his evaluation or rated importance of i. Thus the formulation is compensating in that a high or positive instrumentality times importance product for one attribute can "balance" a low or negative product for another attribute. An underlying assumption of rational choice is implicit in this formulation. A consumer who considers a brand to possess important attributes should not dislike that brand nor should he prefer another brand that lacked or blocked these same attributes.

It appears critical, then, that the consumer be able to make instrumentality assessments. This presumes that he is aware or believes he is aware of the degree to which brands possess salient attributes.. It follows that if he cannot respond to instrumentality questions, his brand preference will be based on factors other than the terms in the instrumentality times value formulation. Sheth (1973) has explicitly accounted for this possibility in a recent paper. Drawing on the work of Hull (1952) and Howard (1965), Sheth posited a habit or conditioning construct that would explain preference when instrumentality beliefs were not operative. Specifically, a consumer may be conditioned to respond to an object in terms of affect without making a rational assessment or mentally multiplying anything.

A second explanation for judgments of affect in the absence of underlying supporting beliefs may be found in the literature on personality response sets (Jackson & Messick, 1958; McGee, 1962). Jackson and Messick (1958) have distinguished between the content and personal style of an individual's response and argue that style will be pervasive when the subject cannot respond to the content of the question. Although a number of response styles have been identified including social desirability, acquiescence, and deviation, attempts to control for these influences will depend to some extent on knowing the subject's personality which, of course, is no mean task.

The potential impact of response style on attitude model predictions is not easily predicted and may have an effect opposite to that argued here if subjects respond to both affect and instrumentality on style. This will serve to artificially inflate the predictor-criterion correlations when no attitude structure may exist at all. However, when affect is based on style due to an inability to assess content and the subject is given an opportunity to indicate that he does not know the answers to instrumentality questions, low or zero correlations between affect and instrumentality predictions would be expected.

Unfortunately, information relating to the issues raised here cannot be obtained from previous studies of multi-attribute attitude predictions because researchers have generally neglected the possibility that subjects may not have been able to respond to instrumentality items. [Mazis and Klippel (1973) provide a notable exception. They restricted their study to subjects who had used all brands in a product category. It will be suggested later in this paper, however, that usage alone provides only one dimension of product familiarity.] This is not to say that researchers have avoided the issue; rather they have allowed the respondent to ambiguously interpret scale midpoints. To clarify this point, the following examples are presented to illustrate the more common operational measures of instrumentality.

The semantic differential. A number of studies (Sampson & Harris, 1970; Scott & Bennett,1971; Churchill, 1972) have used the traditional 7-point bipolar adjective scale as shown below.

Bad taste :_:_:_:_:_:_:_: Good taste

Because no scale values are typically shown, the subject is permitted to interpret the mid-range category on an individual basis.

The integer scale. Other studies (Lehmann, 1971; Hughes & Guerrero, 1971; Sheth & Talarzyk, 1972) have used the following type of scale:

Low  1  2  3  4  5  6  7 High

Again, each respondent interprets the midpoint, category "4," on an individual basis. come researchers (Sheth & Talarzyk, 1972) have used only six integers thereby forcing the subject to respond with a positive or negative evaluation.

In both types of scales above, no explicit provision is made for a "don't know" response. While some subjects may have interpreted the midpoint as such, it is this author's opinion that the "4" response would be used as an "average" or "neither" category. As failure to include a "don't know" alternative may result in an answer when none exists (Boyd & Westfall, 1972), it appears critical that the basis for these answers be known.

Failure to explicitly include a "don't know" alternative in the instrumentality set of multi-attribute attitude studies has a number of implications other than those previously discussed. First, subjects who do not have the information needed to respond to a particular instrumentality item may answer in a fashion consistent with their overall attitude or on the basis of the response pattern of earlier belief items. This phenomenon would be predicted by a number of cognitive consistency theories (McQuire, 1972) and will serve to artificially inflate the predictor-criterion relationship.

A second implication concerns the unknown consequences of comparing results across subjects who are basing their preferences on different numbers of different attributes. This issue reduces to one of combining subjects who determine their overall affect for an object on the basis of a large set of attributes with those who use one or two as though both groups of individuals were equivalent and homogeneous. With few exceptions (Churchill, 1972; Craig & Ginter, 1974), authors have assumed that all subjects use the same number of "salient" attributes.

The problem above is further compounded if instrumentality items are scored 1 to 7 rather than -3 to +3. If a respondent selects the scale midpoint to indicate "don't know," the latter scoring procedure will correctly preclude that attribute from contributing to the subject's aggregate instrumentality times value score. The former scoring procedure, however, implicitly includes that item as though it were scored "neither" or "average." There is more to this issue than is reported here; interested readers are referred to Schmidt (1973).

In an attempt to provide answers to some of the questions raised in the introduction, this study was conducted with two purposes in mind. First the effect of adding "don't know" alternatives to instrumentality items on predictions made by a multi-attribute attitude model was investigated. Second, an effort was made to show that the number of "don't know" responses could be predicted by two constructs, product familiarity and product specific confidence. It was hoped that this information would suggest a perspective for the role of product familiarity in models of consumer preference.

METHOD

Sample

Three groups of undergraduate students were asked to participate in. study of attitudes toward Ford's new Mustang II shortly after its introduction. The total sample was comprised of a "captive" classroom convenience sample (n=164), a "man-on-the-street" convenience sample (n=105), and a quota sample controlled for sex (n=142). Due to the subsample sizes needed in the analysis and because the distinctions between the three samples would not be expected to affect the results, the three groups were combined into a total sample of 405 usable replies. Sixty-three percent of the sample was male.

Measures

In an attempt to reduce the effect of cognitive consistency, the variables were measured on a three-page questionnaire in the following manner. An index of overall affect, a self-rating of product specific confidence, and product familiarity questions were included on page 1. These measures were followed by attribute importance scales on page 2 and instrumentality items on page 3. It was hoped that positioning the importance scales between the affect and instrumentality measures would help to interfere with the subjects' recall processes and thus discourage them from responding to the belief measures on the basis of overall affect.

To obtain an evaluation of affect, subjects were asked to check how appealing the new Mustang II was to them on a seven point scale ranging from "Very unappealing" through "Neither" to "Very appealing." In a similar manner, product specific confidence was measured on a seven point scale ranging from "Very unconfident" to "Very confident." Finally, product familiarity was obtained from the following question:

Are you familiar with Ford's new Mustang II? Yes [ ] No [ ]

If yes, how do you know about it? (Check all that apply.)

[ ] Have seen or heard ads for it.

[ ] Have talked to someone that owns one.

[ ] Have seen it in a dealership or on the street.

[ ] Have ridden in or driven one.

Fifteen product attributes [The attributes selected for the study were roominess, quietness, price, acceleration, safety, popularity, manufacturer's reputation, luxury, fuel economy, handling, comfort, styling, warranty, construction, and luggage space. Factors for which no information existed at the time of the study (e.g., maintenance cost, resale value) were purposely excluded.] were selected as important to students in car buying decisions from research summaries provided by Ford Motor Company and from an earlier study by Richmond, Krafe, and Hubbard (1973) on students' reactions to the Ford Pinto. Because all attributes were somewhat important (i.e., they would generally be rated as "good" on an evaluative dimension), attribute importance was measured on a relative basis (Scott & Bennett, 1971; Mazis & Klippel, 1973). That is, the subjects were asked to scale the importance of each attribute when compared to all others in an effort to discourage them from rating every attribute as important. Scale values ranged from "Not important "(O) through "Less important" (1) to "More important" (5).

Instrumentality items were constructed using a seven point bipolar adjective scale with the midpoint set apart from the two adjacent categories and explicitly labeled "Don't Know." Moreover, the scales were anchored at the poles with good (+) and bad (-) adjectives so that no neutral term was used as a scale endpoint. Two typical instrumentality items are shown below:

             (-)     Don't Know    (+)

Cramped :_:_:_:_:_:_:_:_:_:_:_: Roomy

     Noisy :_:_:_:_:_:_:_:_:_:_:_: Quiet

Note that no "neither" category was included. If a student did not check "Don't Know," it was assumed that he could evaluate the Mustang II on that attribute as at least either slightly good or slightly bad. [To the extent that this was not true, the results may have been affected. Future researchers may wish to include both "don't know" and "neither" categories.] The scale values were scored from -3 to +3. A "Don't Know" response was scored zero.

Analysis

Subjects were placed in subsamples on the basis of the number of instrumentality "don't know" (DK) responses checked. Simple correlations were then calculated between the appeal scores and (a) summated instrumentality times importance terms and (b) summated instrumentality terms alone for each subsample of DK categories. These correlations were then treated as dependent variables (Jones, 1968) to determine if the subsamples could be drawn from a population with a common correlation.

To investigate the determinants of a "don't know" response set, number of DKs was used as a dependent variable and regressed on familiarity, confidence, and sex of the respondent. Because familiarity was measured dichotomously and by exposure category, two regressions were run. In the first, number of DKs was regressed on confidence, overall familiarity (yes = 1, no = 0) and sex (male = 1, female = 0). The second regression was similar to the first except that the single familiarity variable was replaced with the four exposure categories coded in dummy variable format.

RESULTS

The mean number of instrumentality DK responses over 15 attributes was 7.57. Thus, the average respondent was only knowledgeable in half of the attribute areas. This should immediately call to question the implicit assumption of product knowledge presumed in many consumer behavior studies. Of the 405 subjects, 51 (12.6%) checked "Don't Know" for every attribute. The remaining 354 subjects were distributed over the other DK categories with subsample ns ranging from 6 to 38.

To overcome the effect of sampling variations in the smaller DK categories, they were aggregated into groups of two. Respondents in the subsample who responded with a positive or negative evaluation for every instrumentality item (zero DKs) were considered unique and analyzed separately. The other seven subsamples included those with 1 and 2, 3 and 4, ..., and 13 and 14 DK responses. The group responding with 15 DKs would yield predictor-criterion correlations of zero and was not analyzed in this section of the study. The subsample us are shown in Table 1 along with correlations between appeal and (a) E (instrumentality x importance) and (b) z instrumentality scores.

TABLE 1

CORRELATIONS BETWEEN APPEAL AND TWO INSTRUMENTALITY MODELS AS A FUNCTION OF THE NUMBER OF DON'T KNOW RESPONSES

It is apparent from-the results that product knowledge clearly had a moderating effect on the results obtained with the two instrumentality models. When product knowledge was complete or almost complete (i.e., 0-4 DKs), the correlations obtained were on the order of .6 or higher. However as the number of DK responses increased beyond that point, a rather consistent decline in the magnitude of the correlations occurred for both models up to the last DK category where a reversal of this trend occurred. The high correlations obtained for the group of subjects who were knowledgeable in only one or two attribute areas may have been an artifact of the small subsample n or may indicate that, when only one or two pieces of information are known about an object, one's attitude is formed rather consistently with this limited information.

To determine if the correlations obtained could be interpreted as being from different populations (e.g., knowledgeable and unknowledgeable subjects), a x2 test of independence was calculated on the z transforms for both models (Jones, 1968; Snedecor & Cochran, 1967). With seven degrees of freedom, a x2 of 14.07 is needed for significance at the .05 level. Correlations for the instrumentality times importance and instrumentality models yielded x2 values of 13.50 and 13.21 respectively. While these values are significant at the .10 level, the reader may wish to draw his own conclusion as to the acceptance or rejection of a common population correlation.

In order to determine if the number of DK responses could be predicted, DK was used as a dependent variable and regressed on product specific confidence, product familiarity, and sex. The correlation matrix for this set of variables is shown in Table 2 where it can be seen that all independent measures were negatively correlated with the criterion ranging in magnitude from -.20 to -.35. Thus, the number of DK responses was greater for females, subjects who rated themselves as having little confidence in their ability to judge automobiles, and subjects who had had little or no exposure to the Mustang II.

TABLE 2

INTERCORRELATIONS BETWEEN NUMBER OF DON'T KNOW RESPONSES AND THE INDEPENDENT VARIABLES (N=405)

Because the intercorrelations between independent variables were generally moderate, an attempt was made to improve the degree to which the number of DK responses could be predicted through multiple regression. To achieve this, the dependent variable was stepwise regressed on two models: (a) confidence, overall familiarity, and sex, and (b) confidence, the four familiarity categories, and sex. The results reported in Table 3 show the "best" regression equation using the "Maximum R2 Improvement" technique (Barr & Goodnight, 1972).

TABLE 3

RESULTS OBTAINED WHEN NUMBER OF DON'T KNOW RESPONSES WAS REGRESSED ON THE INDEPENDENT VARIABLES: TWO MODELS

The results show that the two models explained 13% and 19% of the criterion variance respectively. Product familiarity was clearly the best overalL predictor although confidence played an independent contributing role in both models. Sex did not contribute any variance beyond that already explained probably due to the fact that it was correlated with confidence. The standardized 3-weights show that overall familiarity played a greater role in determining DK responses than did confidence although when familiarity was viewed in terms of specific types of exposure, the 6-weights for all significant predictors were similar.

Surprisingly three of the four familiarity variables made independent contributions to the second regression model. Word-of-mouth, product usage, and media exposure were negatively related to the number of DK responses although word-of-mouth appeared to have a slightly greater effect on the criterion. The remaining familiarity measure, product observation, was not included in the regression equation due to its high correlation with product usage. The fact that three of the four measures entered the model is significant in that it suggests that the influence of product information sources may be additive. It also implies that exposure to different kinds and types of information is necessary for full brand comprehension, a concept implicit in Howard and Sheth's (1969) taxonomy of significative, symbolic, and social inputs to the buyer decision process.

DISCUSSION

The results of this study demonstrate that low predictor-criterion correlations obtained using multi-attribute instrumentality models may not result from poor conceptual work or methodological flaws. An alternative explanation, that of product or brand unfamiliarity, may be equally tenable. This effect has been concealed in previous studies because researchers have generally failed to include a "don't know" alternative in the instrumentality set.

Thus, indirect but preliminary support has been shown for an extended model of affect which incorporates response style tendencies or conditioning terms (Sheth, 1973). In the absence of complete or nearly complete product knowledge, consumers may respond to affect on the basis of an unspecified but habitually determined response set. While Sheth did not investigate the effect of conditioning in his study, future researchers are advised to more fully specify their models by including an operational measure of habit, particularly when the assumption of product familiarity across subjects is suspect. Howard (1965), for example, has drawn on the work of Hull (1952) and suggested that habit may be measured by the absolute number of reinforced purchases. Further elaboration on this point will ultimately lead to a discussion of stochastic learning models; interested readers are referred to Massy, Montgomery, and Morrisson (1970).

The attempt to predict product knowledge was somewhat encouraging. While the variance explained in the two regression models was not exceptionally high, the results clearly show that the number of DK responses to instrumentality items can be predicted from product familiarity and product specific confidence. Had better measures of the predictors been obtained, the results may have been more convincing. As the reader may recall, all familiarity measures were coded in dummy variable fashion. More precise measures of media recall (Lucas & Britt, 1963) or of the nature and content of word-of-mouth (Arndt, 1967), for example, may have improved upon the relationships obtained here. Similarly, a more construct-oriented approach to the measurement of product specific confidence such as that used in perceived risk investigations (Cox, 1967) may also have increased the significance of the findings.

Hopefully this preliminary study on the influence of product knowledge in attitude model predictions will encourage researchers to extend their research beyond the current vogue of rational cognitive multiplication models to include other factors explaining nonrational or habitual preference.

REFERENCES

Ahtola, O. T. The vector model of preferences: An alternative to the Fishbein model. Faculty working paper, University of Florida, 1973.

Arndt, J. Role of product-related conversations in the diffusion of a new product. Journal of Marketing Research, 1967, 4, 291-295.

Barr, A. J. & Goodnight, J. H. Statistical analysis system. Raleigh, N.C.: North Carolina State University. 1972.

Bass, F. M. & Wilkie, W. L. A comparative analysis of attitudinal predictions of brand preference. Journal of Marketing Research, 1973, 10, 962-269.

Boyd, H. W., Jr. & Westfall, R. Marketing research. (3rd ed.) Homewood, Ill.: Richard D. Irwin 1972.

Churchill, G. A. Linear attitude models: A study of predictive ability. Journal of Marketing Research, 1972, 9, 423-426.

Cohen, J. B., Fishbein, +4. & Ahtola, O. T. The nature and uses of expectancy-value models in consumer attitude research. Journal of Marketing Research, 1972, 9, 456-460.

Cox, D. F. Risk taking and information handling in consumer behavior. Cambridge, Mass.: Harvard University, 1967.

Craig, C. S. & Ginter, J. L. An investigation of the effect of the dependent measure and attributes selected on the predictive power of the multiattribute attitude model. Proceedings, Marketing Educators' Conference, American Marketing Association, 1974.

Fishbein, .M. A behavior theory approach to the relations between beliefs about an object and the attitude toward the object. In  Fishbein (Ed.), Readings in attitude theory and measurement. New York: Wiley, 1967. PD. 389-399.

Fishbein, M,. Some comments on the use of "models" in advertising research. In Seminar on "Translating advanced advertising theories into research reality." Amsterdam: ESOMAR, 1971. Pp. 297-318.

Howard, J. A. Learning and consumer behavior. In. J. A. Howard (Ed.), Marketing theory. Boston, Mass.: Allyn and Bacon, 1965. Pp. 100-112.

Howard, J. A. & Sheth, J. N. The theory of buyer behavior. Nev York: Wiley, 1969.

Hughes, G. D. & Guerrero, J. L. Testing cognitive models through computer-controlled experiments. Journal of Marketing Research, 1971, 8, 291-297.

Hull, C. L. A behavior system. New Haven, Conn.: Yale University Press, 1952.

Jackson, D. N. & Messick, S. Content and style in personality assessment. Psychological Bulletin, 1958, 55, 243-252.

Jones, M. B. Correlation as a dependent variable. Psychological Bulletin, 1970, 70, 69-72.

Lehmann, D. R. Television show preference: Application of a choice model. Journal of Marketing Research, 1971, 8, 47-55.

Lucas, D. B. & Britt, S. H. Measuring advertising effectiveness. New York: McGraw-Hill. 1963.

Massy, W. F., Montgomery, D. B., & Morrisson, D. G. Stochastic models of buying behavior. Cambridge, Mass.: M.I.T. Press, 1970.

Mazis, M. B. & Klippel, R. E. Instrumentality theories and consumer attitudes: Comparing alternative models. Proceedings. Fourth Annual Conference, Association for Consumer Research, 1973.

McGee, R. K. Response style as a personality variable: By what criterion? Psychological Bulletin, 1962, 59, 284-295.

McQuire, W. J. The current status of cognitive consistency theories. In J. B. Cohen (Ed.), Behavioral science foundations of consumer behavior. New York: Free Press, 1972. Pp. 253-274.

Richmond, D., Kraft, A., & Hubbard, C. Some attitudinal shifts of the automobile purchaser. Akron Business and Economic Review, 1973, 4 (2), 54-58.

Rosenberg. M. J. Cognitive structure and attitudinal affect Journal of Abnormal and Social Psychology, 1956, 53, 367-372.

Sampson, P. & Harris, P. A user's guide to "Fishbein." Journal of the Market Research Society, 1970, 12, 145-166.

Schmidt, F. L. Implications of a measurement problem for expectancy theory research. Organizational Behavior and Human Performance, 1973, 10, 243-251.

Scott, J. E. & Bennett, P. D. Cognitive models of attitude structure: Value importance is important. In F. C. Allvine (Ed.), Relevance in marketing: Problems, research, action. Chicago: American Marketing Association, 1972. Pp. 346-350.

Sheth, J. N. A field study of attitude structure and attitude-behavior relationship. Faculty working paper #116, University of Illinois, 1973.

Sheth, J. N. & Talarzyk, W. W. Perceived instrumentality and value importance as determinants of attitudes. Journal of Marketing Research, 1972, 9, 6-9.

Snedecor, G. W. & Cochran, W. G. Statistical methods (6th ed.). Ames, Iowa: Iowa State University Press, 1967.

Wilkie, W. L. & Pessemier, E. A. Issues in marketing's use of multiattribute attitude models. Journal of Marketing Research, 1973, 10, 428-441.

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