The Family Resemblance Approach to Understanding Categorization of Products: Measurement Problems, Alternative Solutions, and Their Assessment

ABSTRACT - How consumers categorize products, and how to measure the extent to which they perceive a particular product to be a member of a category is an issue of interest to both academic and applied researchers. The study examines the family resemblance approach to measuring category membership. Although family resemblance is perhaps the most widely used and cited measure of how attribute sharing relates to typicality, scrutiny of the procedures for its computation recommended by Rosch and Mervis (1975) suggests several ways the measure might be improved or modified. Some of these alternative methods have been used in the literature, but their relative performance has not been assessed. The study compares five alternative methods of computing family resemblance, and finds similarity in results for some, but not other, measures. The results have implications for both academic and applied students of consumer behavior.

Citation:

Don Saunders, Steve Tax, and James Ward (1991) ,"The Family Resemblance Approach to Understanding Categorization of Products: Measurement Problems, Alternative Solutions, and Their Assessment", in NA - Advances in Consumer Research Volume 18, eds. Rebecca H. Holman and Michael R. Solomon, Provo, UT : Association for Consumer Research, Pages: 84-89.

Advances in Consumer Research Volume 18, 1991      Pages 84-89

THE FAMILY RESEMBLANCE APPROACH TO UNDERSTANDING CATEGORIZATION OF PRODUCTS: MEASUREMENT PROBLEMS, ALTERNATIVE SOLUTIONS, AND THEIR ASSESSMENT

Don Saunders, Arizona State University

Steve Tax, Arizona State University

James Ward, Arizona State University

Kym Court, Arizona State University

Barbara Loken, University of Minnesota

ABSTRACT -

How consumers categorize products, and how to measure the extent to which they perceive a particular product to be a member of a category is an issue of interest to both academic and applied researchers. The study examines the family resemblance approach to measuring category membership. Although family resemblance is perhaps the most widely used and cited measure of how attribute sharing relates to typicality, scrutiny of the procedures for its computation recommended by Rosch and Mervis (1975) suggests several ways the measure might be improved or modified. Some of these alternative methods have been used in the literature, but their relative performance has not been assessed. The study compares five alternative methods of computing family resemblance, and finds similarity in results for some, but not other, measures. The results have implications for both academic and applied students of consumer behavior.

Understanding why a product will be perceived as a member of a particular category is an important issue for consumer researchers and practitioners alike. To the researcher, the question raises basic issues of how consumers perceive categories and judge whether and to what extent an item is like other category members. To the practitioner, such issues are also relevant in a variety of ways. For example, the manufacturer of a sporty looking compact car may wonder whether consumers will tend to compare the car to higher priced vehicles positioned as true sports cars or to lower priced vehicles positioned primarily as compact cars. The answer to this question could influence many aspects of marketing strategy such as market segmentation, advertising (e.g., what attributes to push, what competitors to compare to), and pricing.

One approach to understanding the determinants of product categorization that has been applied by a number of researchers (Nedungadi and Hutchinson 1985, Sujan 1985, Ward and Loken 1986, Solomon 1988) is the family resemblance approach initially developed in psychology by Rosch and colleagues (Rosch and Mervis 1975, Mervis and Rosch 1981). Theoretically, this approach is based upon the idea that category membership is a matter of degree. In this view, most categories include more and less prototypical members. The most prototypical members are those that people tend to think of as the best, truest examples of the category. For example, people might perceive such vehicles as a Ferrari, Jaguar, or Corvette to be highly prototypical sports cars. People may also tend to call a number of other cars "sports cars," but they may tend to regard these as less good, true members of the category.

In a number of studies, Rosch and colleagues have shown that more prototypical members of a category tend to share more attributes with other members of the category than less prototypical members (Mervis and Rosch 1981). Thus, in this approach, degree of attribute sharing, or "family resemblance," determines prototypicality. Rosch and Mervis (1975) developed a procedure for measuring family resemblance that has been widely cited and used in the psychology literature.

According to Rosch and Mervis (1975), family resemblance should be measured as follows. First, the researcher develops a list of category members, perhaps by eliciting their names from subjects in a pretest. Then subjects are asked to list the attributes they believe each category member possesses. Usually, subjects are given a minute or two to list attributes for each category member. Next, the researcher develops a category member by attribute matrix. All attributes mentioned by one or more subjects are listed on the right side of the matrix, and category members are listed at the top. The researcher then goes through the matrix, and checks, for each category member, those cells that correspond to an attribute that at least one subject has noted that a category member possesses. After this task is completed, Rosch and Mervis suggest that the researcher should review the matrix, and credit any member that clearly and obviously possesses an attribute with the attribute although the attribute may not have been mentioned for the product by any subject. The researcher should also take away the credit of an attribute to any member that clearly and obviously does not possess the attribute. Upon completion of these judgements, family resemblance scores can be computed for the category members. The scores are computed by first counting the number of products that share a particular attribute in the final matrix. Each product sharing the attribute is then credited with a score equal to the number of products possessing the attribute. The larger the number sharing the attribute, the larger the score. If the attribute is unique to only one product, the product is given a score of 1. This scoring is done for every attribute listed in the matrix. Finally, the scores for each product are computed by adding the scores that were assigned for each attribute. The larger the number of attributes a member shares with other category members, and the more widely these attributes are shared with other category members, the higher the family resemblance score.

Studies that have applied the family resemblance approach to better understand the determinants of typicality in product categories suggest that the method usually produces scores that correlate highly with alternative measures of category membership, such as typicality ratings, and also yields managerially useful data on what attributes contribute more or less to an item's perception as a member of a category. For example, a study of the prototypicality of snack foods by Ward and Loken (1986) found that the consumers studied rated apples as rather prototypical snack foods. The family resemblance approach revealed that apples shared many attributes with other snack foods such as potato chips and peanuts. These included being round, crunchy, crisp, divisible into pieces, easily eaten "finger food", appropriate for many occasions, readily transportable, and liked by many people. While such attributes are not necessarily determinant in the sense of causing people to view apples as snack foods, they could prove very useful to apple marketers designing an ad to promote the use of apples as snack foods.

Despite the utility of the method, close scrutiny suggests some problems and some alternative methods of computing the measure (Loken and Ward 1987). One problem is that unique attributes increase the family resemblance score by one. This seems counter-intuitive, in that a measure of attribute sharing should not increase to the extent that a member has unique attributes, not shared by any other category member. Another seeming problem in the method is that an attribute need be mentioned by only one subject to enter the matrix. Once there, judges may perceive the attribute to be shared by many other members, and credit them with the attribute. This could be a problem because the resulting scores may reflect less the attributes salient to subjects than the logic of the judges. Yet another problem is the subjectivity involved in judging whether members do or do not have attributes. For example, a subject might list the attributes "sweet," "salty," and "coated" for M&M peanut candies but not apples. Judging whether apples should be credited with these attributes seems necessarily subjective and difficult.

Perhaps as a response to these potential problems, researchers have over time tried a number of modifications to the family resemblance procedure. In the consumer behavior literature, Sujan (1985) used only attributes mentioned by two or more subjects to compute family resemblance. Loken and Ward (1987) criticized the family resemblance approach, and proposed an alternative "attribute structure" measure. In the psychology literature, Barsalou (1985) introduced a measure of family resemblance based upon the average rated similarity of category members to one another and not attribute lists. Malt and Smith (1984) examined the issue of whether attributes contribute independently to perceived typicality (as assumed by the family resemblance procedure) or whether the correlation among attributes in a category also influences judged typicality. Tversky (1977) has proposed alternative ways of computing family resemblance including giving positive weight to common features and negative weight to distinctive features, and accounting for frequency of mention. Although his work did not focus on a comparison of his methods with the Rosch et. al. approach, his data suggest that accounting for frequency of mention may improve the correlation of family resemblance with typicality.

Although a number of alternative methods of computing family resemblance have been tried in the literature, no study that we are aware of has attempted to systematically vary methods of computing family resemblance on the same data set to see if 1) the scoring of unique attributes, 2) experimenter judgment, and 3) accounting for frequency of mention produce family resemblance scores that differ in their correlations with typicality, the traditional Rosch method, and one another. The purpose of this study is to attempt to address these questions. These issues are important for two reasons. First, the family resemblance procedure is a widely used method for studying category structure in both the psychology and marketing literatures. Future researchers should be interested in whether alternative procedures yield the same or different results. Second, the results of the study will help researchers assess the comparability of past family resemblance data.

METHODOLOGY

Stimuli Development

To develop stimuli for the study, a sample of 25 undergraduate students were asked to list members of the category "types of food that people eat at their evening meal." The subjects were reminded to list types of food, not brands, and were asked to list the types in the order they were thought of. Members of other categories were also elicited as part of the pretest.

Once the data were collected, "production ranks" were computed for types of foods using the 20 types that were most frequently mentioned by subjects. A score of 20 was assigned to the first food mentioned, 19 to the second, and so on. The foods with the 20 highest production ranks were then chosen for the stimuli (shown in Table 1) for the next phase of the study.

Attribute Lists

In the next part of the study, subjects' perceptions of the attributes possessed by category members were collected. Subjects were 15 undergraduate students who volunteered to participate in the study during scheduled class periods. None of these students had participated in the first part of the study.

The procedure followed methods recommended by Rosch and Mervis (1975). Each of the 20 category members was printed at the top of a page. Subjects each received a packet of the 20 randomly ordered members.

TABLE 1

TYPES OF FOODS EATEN AT AN EVENING MEAL

Subjects were asked to list the attributes possessed by each item for a minute and a quarter. The instructions, adapted from Rosch and Mervis (1975) read, 'This is a simple experiment to find out the characteristics and attributes that people feel are common to and characteristic of different kinds of ordinary, everyday objects .... At the top of each page is listed the name of one object. For each page, you'll have a minute and a quarter to write down all of the attributes of that object you can think of. But try not to just free associate. For example, if "bicycles" just happens to remind you of your father, don't write down "father." Each subject listed attributes for all 20 category members.

Prototypicality Ratings

The subjects were also asked to rate the prototypicality of the category members on three 010 point scales with endpoints very typical--very atypical, very good example--very poor example, and very representative--very unrepresentative. The instructions for the scales, once again consistent with Rosch and Mervis (1975), asked subjects to rate how good an example each category member was of the category. Subjects were cautioned not to confuse typicality with frequency of encounter or liking, using virtually the same wording as Rosch. The complete prototypicality rating instructions are shown in the Appendix. To develop an overall typicality score, the scores for each of the three scale measures (typicality, representativeness, and goodness-of-example) were summed across all subjects.

Alternative Measures of Family Resemblance

The first step in computing the alternative measures of family resemblance was to create a brand by attribute matrix showing all the attributes listed by one or more subjects for each product. Attributes were written along the right side of the matrix and products along the top. Some degree of judgment enters the creation of the matrix, because subjects often use different words that appear to mean the same attribute. Thus, the attributes that subjects list are subject to a content analysis prior to being included in the matrix. However, this analysis attempted to stay very close to what subjects said, and attempted to minimize the aggregation of disparate comments into single categories. Each alternative measure used the data in the resulting matrix as a starting point.

The first measure, FR1, computed family resemblance for the members of the category in the way recommended and used by Rosch and Mervis (1975) and most often adopted in other studies. Two of the researchers, acting as judges, first examined the applicability of each attribute to each product, as previously explained. In each case, the researchers, relying upon their own knowledge of the stimuli, decided whether the attribute might be possessed by the product or not. If the product, in their judgement, clearly possessed the attribute, the product was credited with the attribute even though no subject had actually listed the attribute for the product. If the product clearly and obviously did not possess the attribute, the attribute was deleted for the product. Although the researchers attempted to be as objective as possible, these judgements were often rather subjective. If the judges disagreed about whether a product had an attribute, a third researcher resolved the dispute.

Once the entire matrix was reviewed, family resemblance scores were computed for the products. First, the number of products possessing a particular attribute was counted, and this count was assigned as a score (weight) to each of the products. Thus, if eleven products were credited with a particular attribute, each received a score of 11. Next, the scores were summed across attributes for each product.

The second measure of family resemblance, FR2, was computed using the matrix resulting from the judges' review and the same procedure as FR1. However, this measure deleted any attributes credited to only one or two products from the computation of the family resemblance scores. As noted earlier, increasing a measure of attribute sharing for attributes unique to one or two members of a category seems to be counter to the logic of a measure that should increase to the extent that a member shares attributes with other members.

The third measure of family resemblance, FR3, was computed by relying only upon the attributes listed by the subjects. Thus the original product by attribute matrix, unchanged by experimenter judgment, was used as input data for the computation of these scores. FR3 was intended to provide insight into whether the judgments made earlier improve correlations with typicality. If they do, one might conclude that either the judges bias the measures in the expected direction, or improve the scores by in effect "reminding" subjects to accurately describe the stimuli. If judgment does not improve the correlations over the raw data, one might suggest that this time-consuming and perhaps problematic aspect of the procedure be dropped. FR4 was computed like FR3, only attributes unique to one or two products were not scored, following the procedure used for FR2.

FR5 was computed to introduce a new factor into the measure, the number of subjects who mentioned an attribute. In this procedure, the number of products that shared a particular attribute was first computed, and this score was assigned as a weight to each product having that particular attribute. The original matrix, and not the matrix modified by experimenter judgement, was used as input. For example, as explained earlier, if eleven products shared an attribute, each product possessing the attribute received a score of eleven. Next, a further weight was applied to the data. The number of subjects who mentioned the attribute was counted (e.g., nine), and then the first weight was multiplied by the number of subjects (e.g., 11 X 9 = 99). The logic for this procedure was that if more subjects mentioned an attribute, and more products shared the attribute, then it should contribute more to the perceived typicality of the category member.

RESULTS

Prior to analysis of the data, the mean typicality score for each type of food was computed across subjects (after reverse scoring so that correlations with other measures would be positive), and the median production rank was also computed across subjects. Family resemblance scores were computed for each type of food as a member of its category in the five ways described above. The n for all correlations is 20, the number of products in the category.

The resulting correlations are shown in Table 2. Supporting the validity of the typicality rating procedure, the correlation between typicality and production rank, found by past studies to be highly positive (e.g., Ward and Loken 1986, Mervis and Rosch 1981), was .63, p < .05.

The principle question addressed by the study is whether the five methods of computing family resemblance result in the same or different results. FR1 versus FR2, and FR3 versus FR4 compare the effect of not increasing the family resemblance score for products with attributes shared by none or only one other product. The correlation of FR1 with FR2 was .99, and the correlation of FR3 with FR4 was .98. These results strongly suggest that, at least in the present data set, increasing the family resemblance score by "1" for unique attributes is not a significant problem. Furthermore, each member of the two pairs of measures correlates about .40 with typicality. These correlations of the four resemblance measures with typicality are all in the expected positive direction and are all significant at the .10, but not the .05 level.

The correlations of FR1 versus FR3, and FR2 versus FR4 address whether experimenter judgement in adding or deleting attributes significantly influences the results. FR1 and FR3 correlated .89, and FR2 and FR4 correlated .91 with one another. Once again, using judged data versus raw data resulted in scores that correlate highly, and have comparable correlations to typicality (about .40, as noted earlier).

The family resemblance measure weighted by frequency of mention across subjects (FR5) was correlated slightly but not significantly more with typicality than the other measures (.47, significant at p < .05). However, this measure was poorly correlated with the other family resemblance measures (r = .33 to .19). This last result seemed puzzling, but may be because the measure introduces another factor into the family resemblance score, akin to familiarity or frequency of instantiation. These factors have been shown to have an influence on typicality that is independent of attribute-based measures of category structure such as family resemblance (Barsalou 1985). In other word, weighting by frequency of mention may have actually reduced the measure's relationship to family resemblance (as suggested by the low correlations with other measures) but may have added an additional factor that compensatorily raised FR5's correlation with typicality.

DISCUSSION

The family resemblance measure is a widely used method of studying categorization in both the psychology and consumer research fields. However, the approach recommended by Rosch and colleagues, although widely adopted, raises questions about whether alternative computational procedures might yield better, or at least different, results.

TABLE 2

CORRELATION MATRIX

The results of the present study indicate that the family resemblance measure seems to be relatively robust to scoring unique attributes or not, and using experimenter judgement or not. These findings suggest that users of the family resemblance procedure may be more confident that using the procedures recommended by Rosch does not contribute significant bias to their results. From another perspective, the results suggest that the use of judgment to decide whether products have attributes may not be an essential part of the procedure. Since this phase of the measure requires two judges to independently review the matrices and perhaps a third to resolve disagreements, its elimination might save significant amounts of time and labor, as well as alleviate whatever anxiety researchers have about introducing their own judgement into the data. This simplification might be of particular interest to practitioners who wish to use the family resemblance approach, since time may often be a more critical factor to them in choosing a method than to academic researchers.

Finally, FR5, the method that weighted for frequency of attribute mention across subjects, exhibited an interesting pattern of correlating slightly more highly with typicality than the other resemblance measures, but correlating poorly with these measures themselves. Tversky (1977) has recommended weighting by frequency of mention as a means of improving the correlation of resemblance measures with typicality. The present results do not contradict his recommendation, but it was argued that weighting by frequency of mention may create a "hybrid" measure incorporating attribute-based aspects of category structure and another factor, perhaps similar to familiarity. These suggestions are speculative, and seem worth pursuing with a larger and more varied set of measures including familiarity and other measures of category structure. Our findings are further qualified by the use of only one category for the comparison of the five measures. The category used, "types of food that people eat at their evening meal", is perhaps more ad hoc and diverse than the types of categories frequently studied by consumer researchers. Therefore, a demonstration of these same findings for other sorts of categories would be worthwhile to pursue in future research. Further confirmation of our results across a larger variety of categories would increase confidence in their generality.

APPENDIX

INSTRUCTIONS FOR RATING PROTOTYPICALITY

REFERENCES

Barasalou, Lawrence (1985), "Ideals, Central Tendency, and Frequency of Instantiation as Determinants of Graded Structure in Categories," Journal of Experimental Psychology: Learning, Memory, and Cognition, 11, 629-654.

Loken, Barbara and James Ward (1987), "Measures of Attribute Structure Underlying Product Typicality," in Advances in Consumer Research, Vol. 14, eds. Paul Anderson and Melanie Wallendorf, Provo, UT: Association for Consumer Research, 22-26.

Malt, Barbara and Edward Smith (1984), "Correlated Properties in Natural Categories," Journal of Verbal Learning and Verbal Behavior, 23, 250269.

Mervis, Carolyn and Eleanor Rosch (1981), "Categorization of Natural Objects," Annual Review of Psychology, 32, 89-115.

Nedungadi, Prakash and J. Wesley Hutchinson (1985), "The Prototypicality of Brands: Relationships with Brand Awareness, Preference, and Usage," in Advances in Consumer Research, Vol. 12, eds., Elizabeth Hirschman and Morris Holbrook, Provo, UT: Association for Consumer Research, 498-503.

Rosch, Eleanor and Carolyn Mervis (1975), "Family Resemblance: Studies in the Internal Structure of Categories," Cognitive Psychology, 7, 573-605.

Solomon, Michael (1988), "Mapping Product Constellations: A Social Categorization Approach to Consumption Symbolism," Psychology and Marketing, 5 (3), 233-258.

Sujan, Mita (1985), "Consumer Knowledge: Effects on Evaluation Strategies Mediating Consumer Judgement," Journal of Consumer Research, 12 (June), 31-46.

Tversky, Amos (1977), "Features of Similarity," Psychological Review, 84 (July), 327-352.

Ward, James and Barbara Loken (1986), "The Quintessential Snack Food: Measurement of Product Prototypes," in Advances in Consumer Research, Vol. 13, ed. Richard Lutz, Provo, UT: Association for Consumer Research.

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