Dimensional Validity, Consistency of Preference and Product Familiarity: an Exploratory Investigation of Wine Tasting

J. Wesley Hutchinson, University of Florida
Brady Farrand, IBM
ABSTRACT - This exploratory study tests the psychological validity of the dimensions uncovered by INDSCAL analysis of judgments about the similarity of the tastes and smells of ten white wines. Products, like wines, that can be blended into mixtures afford the opportunity for an "intrinsic" test of dimensional validity. Specifically, the coordinates of a binary mixture should, for each valid dimension, fall between the coordinates of its components. The results of this study demonstrate the logic and the utility of this approach. In addition, the results suggest that, even for experienced wine tasters, consistency of preference is attained at the expense of richness of psychological structure.
[ to cite ]:
J. Wesley Hutchinson and Brady Farrand (1982) ,"Dimensional Validity, Consistency of Preference and Product Familiarity: an Exploratory Investigation of Wine Tasting", in NA - Advances in Consumer Research Volume 09, eds. Andrew Mitchell, Ann Abor, MI : Association for Consumer Research, Pages: 398-401.

Advances in Consumer Research Volume 9, 1982      Pages 398-401

DIMENSIONAL VALIDITY, CONSISTENCY OF PREFERENCE AND PRODUCT FAMILIARITY: AN EXPLORATORY INVESTIGATION OF WINE TASTING

J. Wesley Hutchinson, University of Florida

Brady Farrand, IBM

ABSTRACT -

This exploratory study tests the psychological validity of the dimensions uncovered by INDSCAL analysis of judgments about the similarity of the tastes and smells of ten white wines. Products, like wines, that can be blended into mixtures afford the opportunity for an "intrinsic" test of dimensional validity. Specifically, the coordinates of a binary mixture should, for each valid dimension, fall between the coordinates of its components. The results of this study demonstrate the logic and the utility of this approach. In addition, the results suggest that, even for experienced wine tasters, consistency of preference is attained at the expense of richness of psychological structure.

Which aspects of wines are salient to people? How do these aspects combine to form an overall impression of a wine relative to other wines? These are important questions in the sensory evaluation of wines, in marketing research, and in the psychology and neurophysiology of taste and smell in general. Most approaches to these problems assume the validity of certain traits and dimensions (e.g., sweetness, acidity, etc.) and investigate human sensitivities to them or human abilities to rate them accurately. The techniques of similarity scaling avoid such assumptions and are one way to "uncover" traits and dimensions in a relatively unbiased manner. This study is designed to determine the feasibility of such techniques in the domain of wine perception. These initial efforts have been quite encouraging.

The main purpose of similarity scaling is to "fit" well-specified mathematical models to psychological data. Typically, measures of pair-wise similarity (e.g., confusions or direct ratings of similarity) are modeled as distances in a multidimensional space, or, alternatively, in a hierarchical tree. The dimensions of the spaces and the structures of the trees are determine, directly by the data. Neither the method of data collection, nor the methods of analysis, impose any prior constraints on the resulting structures (other than that they be spaces or trees). That is, these methods "uncover" the traits and dimensions which best characterize the data (in a statistical sense). The trouble with these methods is that they will fit everything in the data, including random error. However, wines permit the construction of stimuli which have built-in checks for "real" vs. "random" structure under certain circumstances.

It is fairly reasonable to assume that wines may be characterized by some set of physical variables, even though most of them are still unknown (Lichine, 1981). If it is assumed thai psychological dimensions are simple monotonic functions [In particular, the simple monotonic function should result in an n-component conjoint additive structure in which each component corresponds to a physical variable (see Krantz, Luce, Suppes and Tversky, 1971).] of these physical variables, then a mixture of two wines should have values on each dimension that are between the values of its constituent wines. Note that this betweeness condition applies intra-dimensionally and does not imply that a mixture will have coordinates that fall on the line segments connecting the constituents in the multidimensional psychological space. INDSCAL analysis of proximity data, in principle, yields rotationally unique dimensions (see below). Thus, the betweeness condition for mixtures provides a qualitative test of the validity of those dimensions given the assumptions.

In this study, the application of similarity scaling techniques to wine perception, in addition to its obvious significance for marketing (cf. Green and Carmone, 1972), has also shown great promise for addressing important general issues in the theory of psychological similarity. The simple devise of using wines and mixtures of wines as stimuli in the same experiment proved to be a powerful tool for testing the psychological validity of dimensions uncovered by the INDSCAL procedure. Unexpectedly, the results also suggested some interesting interactions between product familiarity and consistency of preference.

METHOD

Stimuli

Five white wines and five binary (50% 50%) mixtures of these original wines were used as stimuli. They were served from liter carafes which were randomly assigned numbers for the first evening and letters for the second evening. The wines were served in glassware which was similarly numbered on the first evening and lettered on the second evening. Each subject had 10 glasses containing approximately one ounce each of the 10 wines. Glasses were refilled as necessary, although this was infrequent. Paul Masson Vineyards donated all wines. They were: Chablis, Chardonnay, Chenin Blanc, Johannisberg Riesling and Rhine Castle. The mixtures were: Chablis/Chardonnay, Chardonnay/Johannisberg Riesling, Johannisberg Riesling/ Chenin Blanc, Chenin Blanc/Rhine Castle, and Rhine Castle/ Chablis.

Subjects

Subjects were 12 unpaid volunteers from the Stanford community, varying in wine tasting experience from complete novices to serious hobbyists. The small sample size was a result of the exploratory nature of the study and the time consuming nature of the task.

Procedure

At the beginning of the first evening, all subjects were given a brief questionnaire which was designed to determine approximately their level of experience with wine evaluation (i.e., product familiarity). When all questionnaires had been completed, score sheets and rating scale instructions were handed out and explained. The score sheets listed all 45 unordered pairs of the numbers (letters on the second evening) which corresponded to all unordered pairs of the ten wines. Next to each pair were three columns. The first was for rating the similarity of smells the second was for rating the similarity of taste, and the third was for rating the degree to which they preferred one of the two wines (which they were to circle) over the other. The order of presentation of wine pairs was randomized individually for each subject.

Subjects were instructed to first smell each of the ten wines, and then to taste each of the ten wines. Expectoration after each sip was encouraged, but not required, and an empty cup was provided for that purpose. Water and bread were also provided for cleansing the palate. After this initial acquaintance with the wines, subjects were instructed to begin the pair-wise similarity ratings according to the ordering on their score sheets. They were to complete only the first 23 pairs. When the pair-wise comparisons were completed, subjects were instructed to list the numbers of the wines in the order of their overall preference.

At the beginning of the second evening the instructions were repeated in a brief version. Subjects smelled and tasted each wine and then began the last 22 pair-wise comparisons on their score sheets. These sheets contained pairs of letters, rather than numbers, and the correspondence to the previous evening's wines was unknown to the subjects. Upon completion of the similarity and preference judgments, subjects listed the letters of the wines according to their overall preferences.

RESULTS

Subjects were ranked according to the level of experience evident in their questionnaires and according to the level of consistency evidenced by the correlation of their final preference order on the first evening with that on the second evening (i.e., test-retest reliability). It turned out that median splits of these two rankings of subjects allowed for division of the subjects into four groups of three:

1) HH -- High consistency, high experience

2) HL -- High consistency, low experience

3) LH -- Low consistency, high experience

4) LL -- Low consistency, low experience

Similarity of taste data

ADDTREE analYsis. ADDTREE (Sattath and Tversky, 1977) is a computer program which fits an additive (hierarchical) tree to proximity data. Small distances between items in the tree correspond to high degrees of similarity between the items. Conversely, large distances correspond to low similarity (or high dissimilarity). In the figures presented here, distances are represented in the horizontal branches only. The vertical branches reflect the structure of the tree- which items are groups with each other. As the tree proceeds from left to right, increasingly smaller subgroups are defined until the terminal nodes, representing single items, are reached. The ADDTREE analysis of the mean data from all 12 subjects is presented in Figure l.

FIGURE 1

ADDTREE SOLUTION FOR SIMILARITY OF TASTE DATA (R=.954)

The correlation between the similarity and distances in the tree is .954.

The tree divides the wines into three basic groups: (1)the sweet wines, (2) the dry, or perhaps "sharp", wines, and (3) the Rieslings. The sweet group forms a "comb" which is indicative of a tree solution to data which has a strong unidimensional component--in this case, sweetness. Proceeding down the "comb" are Rhine, Rhine/Chenin, Rhine/ Chablis, and Chenin Blanc. This is exactly the order of percent residual sugar in the wines.

It is beyond the scope of this report to extensively discuss the ADDTREE solution for the four subgroups; however, it should be pointed out that the mixtures are generally grouped with at least one of their components for all except the LL group.

INDSCAL analysis. INDSCAL (Carroll and Wish, 1974) is a computer program which fits similarity data to distances in a multidimensional Euclidean space. Items are assumed to be points in a space. Similarity data from several groups are fit to the same coordinates; however, each group is assumed to weight the dimensions differently. That is, if one dimension is particularly important to a given group then the data should require that a large weight be given to that dimension for that group. One can think of this as a common group stimulus space whose dimensions are expanded and contracted in order to best describe each individual group. This model allows for a rotationally unique determination of dimensions. (Generally, Euclidean distances are invariant under rotation of the axes.)

Since the ADDTREE solution for the LL group revealed a lack of structure, only the HH, HL, and LH groups were submitted for INDSCAL analysis. A 6-dimensional solution was obtained; however, only two dimensions satisfied the betweeness conditions described above, so a new, 3-dimensional, solution was obtained. (The purpose of the third dimension was to "absorb" random error.)

TABLE 1

INDSCAL DIMENSIONAL WEIGHTS FOR TASTE DATA

FIGURE 2

PLOT OF DIMENSION 1 VS. % SUGAR

The first dimension was obviously sweetness. The plot of these coordinates vs. percent sugar is given in Figure 2 (r = -.964). This dimension accounted for most of the variance and was given the highest weight for all subgroups. However, the high consistency groups received higher weights for this dimension than the low consistency group. This suggests that consistency was achieved by attending to the "easy" dimension.

The second dimension, although satisfying the betweeness condition quite well, did not correspond to any readily available physical parameters (i.e., either acidity or alcohol content). Since the betweeness condition is a rather strong test of dimensionality, some confidence in the psychological validity of the second dimension is in order. This dimension contrasts the Chablis and Chardonnay on the one hand with the Chenin Blanc on the other. Perhaps the term "sharpness" captures this distinction. The INDSCAL solution had correlations of .92, .90, and .91 for the HH, LH, and BL data, respectively. Figure 3 depicts the INDSCAL solution for HH weights.

FIGURE 3

INDSCAL SOLUTION FOR TASTE DATA (HH)

FIGURE 4

INDSCAL SOLUTION FOR TASTE DATA (LS)

Notice how the "sweetness" dimension is contracted and the "sharpness" dimension is expanded relative to the EH solution. HL weights were similar to HH weights, differing slightly in the direction of the LH. One plausible interpretation of these results is that members of the LH group were experienced enough to differentiate a second dimension; however 9 attending to two dimensions made the preference judgments more variable from night to night. Evidently, both high consistency groups were attending mainly to the "sweetness" dimension. The conditions under which there will be such a trade-off between the richness of perceptual representation and the consistency of preferences, and the psychological processes that mediate the trade-off are interesting problems for future research. Also, it remains to develop a unitary theory of product familiarity that accounts for other reported effects (e.g., Johnson and Russo, 1981) as well as those reported here

Similarity of smell data

The results of the smell data were very similar to those for the taste data in terms of the conclusions that they suggested. The ADDTREE solution is shown in Figure 5. Note the structure is the same as for the taste data, except that Chenin Blanc is now grouped with the dry wines.

FIGURE 5

ADDTREE SOLUTION FOR SIMILARITY OF SMELL DATA (R= .860)

Again, group by group ADDTREE analysis revealed a lack of consistent structure for the LL group, so data from that group was omitted from INDSCAL analysis. Table 2 lists dimensional weights for the HE, HL and LH groups.

TABLE 2

INDSCAL DIMENSIONAL WEIGHTS FOR SMELL DATA

The unweighted group stimulus space is presented in Figures 6 and 7.

FIGURE 6

INDSCAL SOLUTION FOR SMELL DATA (1 VS. 2)

Again, Dimension 1 satisfies the betweeness condition, is essentially "sweetness", and was most heavily weighted by the HH and HL groups. Dimension 2 was most heavily weighted by the LH group and is similar to the "sharpness" taste dimension; however, Chenin Blanc has switched poles and there is one (rather large) violation of the betweeness condition.

FIGURE 7

INDSCAL SOLUTION FOR SMELL DATA (1 VS. 3)

Dimension 3 was most heavily weighted by the EL group, ant, although it accounts for over 12% of the total variance, it fails the betweeness condition. Dimension 3 might best be considered a pseudo-dimension. The smell data, therefore, generally support the conclusions drawn from the taste data.

SUMMARY AND CONCLUSIONS

Overall, these results provide good evidence that the betweeness condition afforded by mixable products is a useful test of the psychological validity of dimensional representations. For these particular wines, sweetness was the dominant attribute, although experienced subjects were clearly sensitive to a few others. This could be partially due to the fact that these wines did not differ greatly on physical dimensions other than sweetness. INDSCAL analysis revealed an interesting interaction between product familiarity and consistency of preference. Specifically, it seems that sophisticated wine tasters become inconsistent in their preferences when, through applying their knowledge, they base their preferences on a two-dimensional rather than a uni-dimensional attribute structure.

Mainly, and in keeping with the exploratory nature of the experiment, these results suggest that further research in these areas will be rewarding.

REFERENCES

Carroll, J. D., ant Wish, M. (1974), "Models ant Me for Three-way Multidimensional Scaling. In D. H. Krantz, R. C. Atkinson, R. D. Suce ant P. Suppes, (eds.) Contemporary Developments in Mathematical Psychology, Vol. II (San Francisco: Freeman).

Green, P. E., ant Carmone, F. J. (1972), Marketing Research Applications of Nonmetric Scaling Methods. In A. R. Romney, R. N. Shepard, ant S. B. Nerlove (eds.), Multidimensional Scaling: Theory and Applications in the Behavioral Sciences. Vol. II (New York: Seminar Press).

Johnson, E. J., and Russo, J. E. (1981), "Product Familiarity ant Learning New Information" in R. Monroe (ed.) Advances in Consumer Research. Vol.VII, Association for Consumer Research.

Krantz. D. H., Luce. R. D., Suppes. P., and Tversky, A. (1971), Foundations of Measurement. Vol. I (New York: Academic Press).

Lichine, A. (1981), New Encyclopedia of Wines and Spirits (New York: Alfred A. Knopf, pp. 38-51).

Sattath, S., and Tversky, A. (1977), "Additive Similarity Trees Psychometrika, 42, 319-345.

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