A Discussion of &Quot;Methodological Developments&Quot;

James L. Ginter, Ohio State University
[ to cite ]:
James L. Ginter (1979) ,"A Discussion of &Quot;Methodological Developments&Quot;", in NA - Advances in Consumer Research Volume 06, eds. William L. Wilkie, Ann Abor, MI : Association for Consumer Research, Pages: 589-591.

Advances in Consumer Research Volume 6, 1979      Pages 589-591

A DISCUSSION OF "METHODOLOGICAL DEVELOPMENTS"

James L. Ginter, Ohio State University

INTRODUCTION

The common trait of the three papers in this session on "Methodological Developments" is that they all explore relevant questions in the use of scaling methods in marketing research. The first two papers [2,1] investigate potential problems with common applications of scaling methods, while the third paper [3] proposes a model for estimation of market share from scaling results. Each of the papers is based on a different method of measuring attitudinal constructs, with multidimensional scaling, Thurstone Case V, and conjoint analysis being used. The contribution of each of these efforts is explored through the following questions and comments.

"CHANGES IN CONSUMER PERCEPTIONS: THE IMPACT OF TESTING CONDITION ON PERCEPTIONS OF BRANDED PRODUCTS"

The primary conclusion drawn from this research [2] is that measured perceptions of products, based on multidimensional scaling of perceived differences, tend to be less sensitive to changes in the stimulus set than had been expected. The expectation of change in perceptual structures was based on: 1) the prior research finding that the underlying structure of the perceptual space for a set of products can be related to the characteristics of the stimuli, and 2) the implication that perceptual structure should be altered if the set of attributes represented by the stimuli changes.

The experimental treatments consisted of three different sets of brands: 1) a "standard" set of seven non-diet brands and one diet brand, 2) the previous Set of stimuli with one non-diet brand replaced with another diet brand, and 3) a set consisting of four of the "standard" set non-diet brands, the "standard" set diet brand, and diet analogs of three of the non-diet brands. Similarity-dissimilarity ratings for all pairs of brands were based on sensory evaluations and on use of brand names.

The following questions and comments on this research appear to be relevant.

1) The justification presented for the measure of change in perceptual structure used in this study is that if the perceptual structure is stable, spatial relationships between stimuli should remain unchanged. Operationally, however, the authors rely on the reflection of this logic, i.e., if spatial relationships are unchanged, then perceptual structure is stable. The strength of this latter statement is questionable. If, for example, a new, independent and important, dimension were added to the perceptual space through inclusion of a new stimulus, the former brands may be perceived uniformly on this new dimension and their spatial relationship with each other remain unchanged. The authors' investigation of changes in dimensional salience may address this issue, but the consideration of two-dimensional solutions in each case could tend to suppress the observed effects of additional attributes being considered.

2) It was expected that the substitution of three diet brands for three non-diet brands would induce a change in the space to include a diet/non-diet dimension. It should be noted, however, that the diet/non-diet dimension was already represented in the "standard" set of stimuli. Therefore, the experimental treatments may not have added a new dimension but, instead, altered the distribution of stimuli on an existing dimension. In light of the stated reasons for expecting change, i.e., alteration of the set of attributes represented, one might expect no change in perceptual structure.

3) Changes in the stimulus set did not result in differences in spatial relationships or dimensional saliences of the configurations based on brand evaluations. The corresponding configurations based on sensory evaluations did exhibit difference, however (r = .41). The stability of configurations based on brand evaluations could be explained by the comments under (2), above. That is, the inclusion of Tab (a recognized diet brand) in the standard set had already introduced the diet/non-diet dimension, and substitution of more diet products did not affect the dimensionality. For the results with sensory evaluations, one could conjecture that the diet dimension was not detected in the "standard" set because only one of the eight brands was a diet drink, but increasing the number of diet drinks to four enabled the respondents to identify this dimension. If this were the case, the multiple substitution did add a previously undetected perceptual dimension, and the perceptual configuration changed significantly, as expected.

4) The reported results do not appear to support the general conclusion that perceptual structure is unchanged by alterations in the stimulus set. While configurations based on brand evaluations remained stable, those based on sensory evaluations did not (Table 2).

The authors have identified a question which should be of great interest to researchers scaling perception on the basis of paired comparisons and is worthy of further attention. The concern expressed in this review about the strength of the experimental treatment could be addressed through an additional analysis with Tab deleted from the "standard" stimulus set. The divergence in results based on brand and sensory evaluations also indicates a need for greater understanding of the effects of data collection method under various conditions. The authors' inclusion of both forms of presentation in this research is praiseworthy.

"COVARIANCE BIAS OF THURSTONE CASE V SCALING AS APPLIED TO CONSUMER PREFERENCES AND PURCHASE INTENTIONS"

This research [1] documents the existence of a possible problem in the use of the Thurstone Case V method of generating an unidimensional preference scale for a group of respondents. The covariance addressed in this research is the relationship between preferences expressed for two objects. The necessary assumption that these covariances are equal for all pairs of objects may be violated when data from a group of respondents are scaled. The primary reason for the violation is the possible homogeneity of perception combined with heterogeneity of preference. If objects A and B are seen as very similar by all respondents, there is likely to be high covariance in their preference ratings since a person's preference rating for one of them is likely to be similar to the rating for the other. Objects which are seen as dissimilar may have negatively correlated or uncorrelated preference ratings. Therefore, one could not expect the covariance for all pairs of objects to be equal.

In the "Conceptual Analysis" section, the authors appear to argue that covariance is related to the proportion of times one object is chosen over another through an example in which one object would dominate a similar one because their momentary values would tend to move together. While this example may apply to scaling of data collected on several occasions from one respondent, it is not clear that it applies to analysis of data from several respondents on one occasion each. Two similar objects may have high covariance across a set of respondents, but it is not clear why this similarity and covariance should lead to the choice probability for this pair being any different than that for two dissimilar objects. In other words, if the preference levels for two similar objects are highly correlated across respondents, why should this lead one to expect that either would be preferred over the other with probability greater than .5? This reasoning is an important part of the paper, since it led to the expected nature of the effects of similarity on scale results.

The relationship between difference in transformed observed and predicted proportions and a measure of similarity was investigated. The following comments are related to this set of analyses.

1) The conceptualization of the problem was based on heterogeneity of preferences. It is a strength of this work that the expected effects were investigated on two data sets with different patterns of preference heterogeneity. While heterogeneity (as measured by average subject correlation) was not linearly related to the correlation of bias (zjk - ^zjk) and similarity, the direction of results supports the authors' contentions.

2) For each data set, subjects were divided into preference segments on the basis of "rank order correlations between individual subjects and a group scale." Additional information about this procedure would have been useful. It is not clear how the group scales can be developed without first identifying the individuals in the groups.

3) The measure of dissimilarity was based upon the physical characteristics of the objects (examples given as weave, pattern, lining, and variety of color). Each characteristic was coded as a dummy variable indicating its presence or lack thereof in each object. This set of dummy variable values appears to be a rather weak representation of the characteristics of an object, however. There may have been considerable variation in the colors and number of colors offered, for example. The objects by attributes matrix of dummy variable values was then submitted to principal components analysis. For the drapery data, the four attributes were reduced to two components explaining 79.4 percent of the variance. The concept of variance may take on a different meaning with 0-1 data, and the proper interpretation of the percent of variance explained by the components retained is not obvious. The Euclidian distance between designs on the components was used as the measure of similarity. For the above reasons, the measure appears to be "noisy." It is difficult to argue that the similarity biased results in a positive direction, and to the extent that the measure was noisy, the results were artificially weakened. It is possible, therefore, that a more precise similarity measure would have led to even stronger results.

4) The authors' point that subjective dissimilarity data would be preferable is a good one. The results may have been even stronger with such a measure.

5) The discussion of implications for relative versus "absolute" affective values is somewhat confusing. The authors seem to be saying that covariance bias will result in inaccuracy in the precise location of points on the scale but that rank orders are not likely to be affected.

The importance of this paper is related to the extent to which Thurstone Case V scaling is actually used. The authors have documented the existence of a problem and suggested means of reducing its impact. It would have been useful to have the size of the effect of covariance bias shown in a table comparing scale values for all subjects and for the segments. Such a presentation would have demonstrated the importance of the problem, as well as its existence.

The authors argue that covariance bias applies also to models using the logistic function or the arcsin transform distribution. Many recent research efforts have made use of these functions (the logistic is frequently used on conjoint analysis, for example), and future efforts to determine the effects of covariance bias in these models should be of substantial interest.

"MARKET SHARE ESTIMATES BASED ON CONJOINT ANALYSIS OF CONCEPTS"

This paper [3] proposes a logistic model for estimating market shares for alternative concepts whose overall utilities have been specified through conjoint analysis. The model is deterministic at the individual level, in that it is assumed that each person will choose the concept with highest overall utility to that person. The problem is viewed as stochastic at the aggregate level because of the heterogeneity of utilities across individuals. The model computes the probability that each concept is preferred to all others through use of the logistic distribution and translates these probabilities into market share estimates.

The following comments are related to the model and discussion presented in the paper.

1) In their introductory section, the authors discuss two types of factors that influence consumer preferences. The first consists of attributes and attribute level utilities as modeled in conjoint analysis. The second type of factor is external to the product itself and consists of use context, needs satisfied, institutional constraints, etc. It is commendable that the authors have considered factors in the choice process which are external to the constructs normally considered in models of preference. The perspective offered, however, seems to consider the product-related and "environmental" factors as affecting preference simultaneously. It is possible to view the determination of preference as a two-level process in which preference is determined by attribute utilities which, in turn, depend upon the environmental factors. The two views differ in that in one case the environmental factors are considered as additional explanatory variables for preference, and in the other they are considered as explanatory variables for variation in attribute utilities. The implication of the latter view for measurement in marketing research is that it suggests taking attitudinal measures within environmental scenarios. The alternative is to take global attitudinal measures (i.e., across environmental settings) and add environmental factors to these in order to explain preference.

2) The authors point out that one of the assumptions necessitated by use of the logistic model is that the correlation between difference scores (i.e., (Ui - Uj) and (Ui-Uk), across individuals) is equal to one-half for all combinations of pairs. It is then shown that this assumption holds when the Thurstone Case V assumptions hold (si2 = sj2 and rij = rik for all i, j, and k). The previous paper addressed these Thurstone Case V assumptions and concluded that they may not hold for preference data from a group of individuals. The research reported by Huber and Sewall [1] would therefore open to question the use of the logistic model suggested in this piece.

3) The model is described as providing a practical vehicle for estimating market shares that would be captured. It is based on an individuals by concepts matrix of overall utilities, and the fundamental assumption is that each individual will choose that alternative for which his utility is highest. It appears, however, that the use of an aggregate stochastic model is not necessary, and to the extent that the model does not fit the data, predictive accuracy is lost. A simpler approach would be to simply conduct a "simulation" in which choice is predicted individually for each person on the basis of his utilities. This procedure requires no more data since it is just a comparison of the entries in each row of the utility matrix. Market share can then be predicted as the proportion of individuals for whom each concept has the highest utility. One of the advantages of the simulation approach is that market segmentation implications can be explored directly through consideration of other descriptors of the individuals. This can be done either by dividing respondents into groups on an a priori basis and estimating market shares for the individual groups or by comparing the respondents who prefer the different concepts on other explanatory variables. The use of an individual simulation also allows the researcher to weight respondents by usage level.

One implication derived from the model is that a concept with a lower mean utility and larger variance may capture a larger market share than a concept with higher mean utility and smaller variance. This finding precludes use of aggregate mean utilities of concepts to predict market share, since correlations in utilities are not considered. It does not, however, appear to preclude the individual by individual type of prediction suggested above.

One of the strengths of this paper is the comparison of the logistic model with other models. This type of discussion is very useful to a field in which many research efforts appear on the surface to be unrelated.

CONCLUSIONS

These three papers reflect positively on this conference in that they investigate issues of pragmatic importance in marketing research. Although they may not be elaborate or comprehensive enough in their present states to be published in major Journals, they do represent a contribution by helping to focus interest on specific problems of which current users of these methods may not be aware. They demonstrate one of the most important benefits of this conference in that developing streams of research and focused investigations of specific issues may be exposed and discussed among individuals with common interests. This process may be much more beneficial to the development of research methods in consumer behavior than the alternatives of little exposure or publication in specialized journals read by only a fraction of those potentially interested.

The problem area of developing and refining methods of measuring and modeling attitudinal constructs and using the results to forecast market behavior is filled with specific issues. Therefore, it is difficult to suggest a general direction for future research. If one takes a bit broader perspective, however, a rather fundamental question seems appropriate. Most perception and preference modeling efforts are directed at what could be termed "global" constructs, e.g., preference for brand X. Are researchers missing a possible opportunity for greater understanding and predictive ability by paying insufficient attention to the purchase/usage context? One could hypothesize that the usage context will play an important role in affecting the determinants of behavior for many products, including those studied in these papers. A person's preference for sheets, for example, may be affected by whether they are to be used in the master bedroom, children's room, or a guest bedroom. At present, however, it appears that little is known about methods of identifying purchase/ usage contexts of general importance and incorporating these contexts into measurement methods. A better understanding of the importance of context and means to operationalize it could suggest modification in method, lead to increased explanatory and predictive power (and, therefore, understanding), and hold implications for creative brand management strategy. Although these comments are not intended to deprecate the types of research efforts reported in this session, they do represent a call for a broader perspective and increased attention to a set of potentially important explanatory factors.

REFERENCES

Joel Huber and Murphy A. Sewall, "Covariance Bias of Thurstone Case V Scaling as Applied to Consumer References and Purchase Intentions," Advances in Consumer Research, Vol. VI.

James McCullough, Charlene S. Martinsen, and Linda Sceurman, "Changes in Consumer Perceptions: The Impact of Testing Conditions on Perceptions of Branded Products," Advances in Consumer Research, Vol. VI.

James B. Wiley and Robert Bushnell, "Market Shares Estimates Based on Conjoint Analysis of Concepts," Advances in Consumer Research, Vol. VI.

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