Vous êtes sur la page 1sur 13
A New Approach to Market Segmentation Although market segmentation is still considered a major topic of interest by the marketing research community, it might be useful to question some premises of traditional segmentation. PAUL E. GREEN Paul E, Green is S. 8. Kresge Professor of Marketing at The Wharton School, University of Pennsyteansa, Few, if any, topics in marketing have com manded more attention than market segmen- tation, Hardly a journal issue appears that does not include some article on the subject. Almost all marketing texts devote at least @ chapter to the topic, and whole books have been written on segmentation, Techniques for carrying out segmentation have also increased in variety and sophistica- tion. Multiple regression and discriminant analysis were the standard approaches a dec- ade ago. Now they frequently take a back seat to such procedures as Automatic Interaction Detection, factor analysis, cluster analysis, perceptual and preference mapping and conjoint scaling. With the profusion of tech- niques has come a profusion of segmentation bases—perceptual-preference segments, needs- oriented segments, common-problems seg- ments, benefits-secking segments, psycho- graphic segments, and so on. One could easily get the impression that the topic has become too complex for mere mortals to compre- hend. One of the purposes of this article is to argue that such is not the case. I first discuss the role that segmentation can play in the formulation of marketing strategy for either products or services, consumer or industrial. The second part of the article suggests that FEBRUARY 1977 the complexity of segmentation methods is more apparent than real. ‘Two basic ap: proaches to segmentation~a priori and post hoc—are described, and it is shown that all current approaches boil down to one of these types, or possibly a hybrid of the two. A new method is then proposed for examining rela- tionships among alternative bases for segmen- tation that may be under consideration by the researcher. The third part of the article describes a new approach to segmentation. componential segmentation. The characteris tics of this method are described nontechni cally, and the results of an empirical applica tion are reported! SEGMENTATION AND MARKETING Contemporary segmentation studies often col- lect such a welter of different kinds of data that it is difficult to know how the separate facets of the survey are to be integrated. For example, in segmenting the market for auto 1, Componential segmentation has been developed joint ly with J. Douglas Carroll, Bell Laboratorier, and Frank J. Garmone, Drexel University. More details oa the model appear in Paul E. Green, J- Douglas Carroll and Frank J Carmone, “Superordinate Factorial Designs in the Analysis of Consumer Judgments " (Working paper, Univesity of Penn sylvania, September 1975) and Paul E- Green, J. Douglas Caroll ‘and’ Frank J. Carmone, “Design Considerations in Attitude Measurement ” (Paper presented at the Seventh Annual Conference on Attitude Research, February 1976). Copyright © 2001 All Rights Reserved 61 62 insurance policies, data could be collected on the respondent's ‘current insurance—amount, type of coverage, cur rent supplier, previous supplier, recent claims ‘perceptions and evaluations of altemative sup- pliers of insurance Utilities for various combinations of insurance policy characteristics evaluations of new policy concepts insurance needs, Benefits sought, problems en countered with current supplier life style—insurance related and more general aspects ‘communications channel selection ‘demographic and socioeconomic variables. tis banal to say that segmentation studies are but a means to an end. However, one some: times loses sight of the strategic aspects of segmentation research when the incremental costs of tacking on more sections to the questionnaire are so low. At any point in time, the firm markets a particular set of offerings (products or ser- vices) to a particular set of buyers (consumer, commercial or industrial). This all takes place within a set of interconnected environments made up of competitors, distributors and various governmental agencies. Just as obvi- ously, the firm could, in principle, be market- ing other kinds of offerings, to the same or to different kinds of people. Figure 1 is an adaptation of the original Ansoff schema involving four basic competi tive strategies in marketing: market penetra- tion, market development, product develop- ment and diversification? In this adaptation, however, we distinguish between the struc- tural and functional characteristics of offer ings and markets. Structural characteristics of products in- clude physical and chemical attribute levels, packaging, distribution and price. Structural ics of persons refer to their demo- characte graphic and socioeconomic characteristics. While these latter characteristics are expected to change over time, they may be treated as fixed for relatively short time intervals. More- 2, H. Igor Ansoff, Corporate Strategy (New York: McGraw-Hill Book Co 1965). PAUL E. GREEN over, it is reasonable to assume that demo- graphic and socioeconomic variables moderate persons’ choices among offerings but, in turn, are pretty much independent of the firm’s marketing strategies. Functional characteristics of offerings per- tain to the uses or purposes to which the products or services are put. (Included here are the symbolic values that offerings may display ina social-psychological context.) Stretching the analogy a bit, functional char- acteristics of people pertain to the ways in which demographic and socioeconomic char- acteristics shape one’s style of living and one’s perceptions and values in the market place. ‘Two different structural offerings—for ex- ample, a candle and table lighter-may serve the single function of lighting one’s cigarette. Conversely, a single offering—a lemon-lime soft drink—may serve two or more functions, such as a mid-afternoon refreshment or a mixer with one’s favorite bourbon at bedtime. The functions that an offering purports to fill are part of its array of promotional appeals, claims and symbolic characteristics. Develop: ment of its image and market positioning is the job of advertising message design and communications channel selection. By the same token, two different demo- graphic structures could display (within limits) a similar life style, benefit-secking pattern, or brand preference profile. Con- versely, a single type of demographic struc- ture could manifest different goal-seeking and expressive behaviors. Functional-expressive behaviors may, in turn, be arrayed hierarchically in terms of general life style, product-related needs, prob- lem solving and specific brand-related percep- tions and preferences. (In the simplified classi- fication of Figure 1 these are all called functional-expressive.) The original Ansoff schema was repre- sented by a table which did not distinguish between structural and functional characteris: tics of offerings and markets. From the standpoint of strategy formulation, a more BUSINESS HORIZONS eS A New Approach to Market Segmentation FIGURE 1 Alternative Marketing Strategies and Segmentation Research* Present Segments New Segments Structural Functional: Structural | Functional- Character- Expressive Character. Expressive isties Behavior istics Behavior Structural Characteristics Present MARKET MARKET Offerings DEVELOPMENT Functional Symbolic ‘Appeals Structural Characteristics New PRopucT : Offerings DEVELOPMENT DIVERSIFICATION Functional Symbolic Appeals “Adapted from detailed classification might be preferred. For example, in the auto insurance example one might want to distinguish between structural product development in which the actual policy’s characteristics are altered (increased , extended coverage, increased pre- mium), and functional product development in which different appeals for selling the original policy are implemented. In turn, these appeals could be targeted to specific demographic groups, different life style seg- ments, or both. Depending upon the empirical context, the interconnections between structure and function may be loosely or tightly coupled, dene FEBRUARY 197 Igor Ansott, Corporate Strategy (New York: McGraw-Hill Book Cou, 1965). For example, lowering the insurance policy’s premium and offering the policy by mail may markedly change the offering’s functional- symbolic character. Conversely, including a new restriction on claims (as part of the policy's fine print) may have no effect what- soever on the policy’s image. Whatever may be the case, a general problem in strategy formulation is to achieve congruence between structure and functional-symbolic appeals. In particular, one may wish to find a specific structure that allows for a variety of beliew- able functional-symbolic appeals—each one being attractive to a different market seg- ment. Copyright © 2001 All Rights Reserved 63 64 ‘The value of the schema of Figure 1 is the guidance it provides for designing the segmen- tation study in the first place. As a further example, consider the problem of a telephone company wishing to increase revenues for Jong distance usage. Clearly, little can be gained by attempting to increase installed sets; the market is already saturated. Simi larly, little opportunity (except for the ex. ploitation ot picture phones) exists tor drama- tically changing the physical structure of telephones. What can be done involves such strategic aspects as increasing the variety of appeals related to why people call long distance ‘changing the pricing or discount structure for load leveling and increased overall demand appealing to special demographic groups—senior citizens, college students, military service personnel via special rates or gift certificates appealing to special life styles—the highly mobile, gregarious and involved person—via special appeals or channels, such as direct mail ‘The reader can easily think of many examples where the permissible courses of action shape the kind of segmentation study that should be implemented. However, as basic as this point is, many omnibus-type segmentation studies are launched and the data analyzed before the central idea of formulating strategy alterna- tives even comes up. TRADITIONAL APPROACHES In the past, marketing researchers have em- ployed two basic analytical approaches to segmentation: A priori segmentation, in which the re- searcher chooses some cluster-defining de- scriptor in advance, such as respondent's favorite brand. Respondents are then classi- fied into favorite-brand segments and are further examined regarding their differences in other characteristics, such as demographics or product benefits being sought. Post hoc segmentation, in which respon- imilavity dents are clustered according to the PAUL E. GREEN of their multivariate profiles regarding such characteristics as purchasing behavior or atti- tudes. Following this, the segments may then be examined for differences in other charae- teristics, not used in the original profile definition, In post hoc segmentation one does not know the number of clusters or their relative size until the cluster analysis has been completed. ‘As an example of a priori segmentation, one might classify all respondents according to their stated favorite brand of beer. Having done this, some technique such as multiple discriminant analysis might be used to deter- mine if the groups differ in terms of average demographic profiles or life style variables. In post hoc segmentation, we prespecify only the set of variables on which consumers are to be clustered—benefits sought, problems encountered with the product, or whatever. ‘One then takes the consumers’ response pro- files on the whole battery of selected variables and clusters the respondents. Having done this, some technique such as multiple discrim: inant analysis (or simple cross-classification, for that matter) can be employed to see if the various clusters differ with regard to demo- graphics, product usage, and so on. In the preceding beer example, respondents could first be clustered on the basis of the common- ality of their benefit-secking profiles. Having done this, one could then see if the various clusters differed significantly with regard to weekly consumption of beer, brand favorite, respondent age, and so on. In some studies a hybrid of the two approaches is used. For example, respondents could first be grouped according to favorite brand and then a clustering procedure could be employed to see if segments evincing ‘common benefit-seeking profiles appear with- in each of the brand-favorite segments that were found via the a priori approach. Figure 2 shows, in stylized form, hypothe- tical examples of each approach. To illustrate, assume that we have m consumers’ consump: [BUSINESS HORIZONS SESS ERERERERSOSS A New Approach to Market Segmentation FIGURE 2 Traditional Approaches to Segmentation A Priori Segmentation Brand Brond ear eed Segment eee cemeied ifse 8 1f- - x 2]2 4 9 toes s]o 5 12 OS Consumes 412 1 3{ ___ 4 ald 6 alk 2 2 Post Hoe Segmentation Ratings on Product Needs Need Statements Consumers Segment 123 n r234. 123) 4 ar. r 1 ifx -- = 36 lillie 2 2 axe 46) 1 3 Celt entries are a}x - = Consumers 36) 22] ye dissimilarity ets oF ft : Between cae ml 24 3 ™ ml 2 1 x Hybrid Segmentation Needs Brands Statements Brand-Need Segment ABC 1230000, * A B ¢ 1"2]1 2 31 2°3 4 1f4 6 sya 1. vi ee aan 2)2 4 9/2 6! 2 2/2 2/20 5}x ti 3 | 0 15 a2] 3 6! nM sf--|[x 0 “J260207 Consumers 4/12 11 3/4 6! oligo x |S TPIS FEBRUARY 1977 Copyright © 2001 All Rights Reserved 66 tion data (in cases, over some base period) of three brands of beer: A, B and C. Under a priori segmentation we may elect to group people on the basis of the brand that enjoys the highest consumption rate. The second matrix is a simple transformation of the first in which an X appears under the brand for which consumption is highest for each respon- dent. (We note, for example, that the first two respondents are assigned to segment C.) In post hoc segmentation, an extra step computation of a matrix of interperson dis similarity measures—is involved. Then, the actual grouping process is carried out by some type of clustering algorithm. To illustrate, assume that each of the m consumers re- sponds to a set of needs-type attitude seg- ments regarding beer consumption, on a seven-point agree-disagree scale. Under post hoc segmentation, the first matrix shows the original response profiles. This matrix is trans- formed into an mx_m symmetric matrix of dissimilarity measures in which each cell entry measures how disparate each pair of con- sumers is across the whole set of needs-type statements. This second matrix is then sub- mitted to a cluster analysis, yielding, in this illustrative case, four segments. Hybrid segmentation starts out with both data sets. Respondents are first clustered by favorite brand. Following this, separate and independent dissimilarity-measure computa- tions and cluster analyses are carried out within each favorite-brand segment. As shown in Figure 2, different numbers of clusters may emerge and, in general, the needs-type sub- groups found by the hybrid approach differ from those found by the post hoc approach. As can be noted from the preceding example, segmentation is an aggregative pro- cess insofar as matrices of individual data are concerned. Whether a single variable is se- lected for respondent assignment or whether a clustering is performed over a general measure of interperson dissimilarity, the net effect is to reduce the m original respondents to a more manageable number of groups. In the PAUL E, GREEN process of doing this, information about detailed consumption of brands or detailed responses to needs-type statements is (willing- ly) discarded. ‘Two other points are worth making, First, in Figure 2 only persons were grouped into segments. This is generally the way market segmentation proceeds, but it is quite possible to take the obverse point of view and to cluster offerings. For example, in the case of a large number of brands, one could just as easily develop a dissimilarity measure for pairs of brands across persons’ consumption pro- files and cluster those brands that exhibit a relatively high commonality of usage. This may not always be useful, but the point is that a kind of symmetry exists between each facet of the response matrix—one can cluster market offerings just as readily as people. Second, the choice entities need not be products or services in the more narrow sense. They could be political candidates, legislative actions, charitable appeals, home site loca- tions, or whatever. CHOOSING A SEGMENTATION BASE In most segmentation studies, the researcher is faced with a surfeit of variables that could serve as the basis of segmentation, Returning to the insurance policy example, eight differ- ent sets of data were collected, ranging from the characteristics of current policies to the respondent’s demographic-socioeconomic pro- file. Which set is to serve as the “distin- guished” set or segmentation base? Unfortunately, there is no simple answer to this question. Historically, researchers have selected some variable or battery of variables and proceeded from there, For example, if a needs-type attitude segmentation is being done, then responses to the battery of needs- type statements provide the segmentation base. However, suppose one finds that other variables—demographics, product usage, new concept evaluations—-do not show systematic BUSINESS HORIZONS SSS SSS SSEOSSSSOOOO OOOO A New Approach to Market Segmentation profile differences among the needs-type seg- ments? If so, should some other battery of variables be chosen as a base? Relating the choice of segmentation base to strategy considerations, as described car- lier, appears to be the most appropriate approach. Even so, cases can arise in which several candidate bases are tenable choices. What is needed is a method that will: (1) show how closely related various clusterings obtained from alternative bases are; (2) test whether some clusterings are independent of the remaining ones; and (3) find the clustering that exhibits the highest contribution to the ‘al association among the subset of clus- terings that evince significant associations in the first place. Furthermore, even if th researcher has settled in advance on a dis guished set of variables to serve as the segmentation base, he may be interested in finding out how well the clustering obtained from this distinguished set can be predicted by some function of the other clusterings. As it turns out, all of these objectives are interrelated and can be attained through application of a new approach that utilizes multidimensional contingency table analysis. ‘The basic ideas of the method can be briefly described as follows: Each candidate segmentation base (battery of variables) is independently used to cluster the respon- dents into some specified number of groups. ‘The number of clusters so formed may or may not be the same across bases, depending upon the researcher's objectives, Each clustering serves as a categorical variable (where the clusters represent the categories), and a ‘multidimensional contingency table is formed. A variety of tests can then be carried out on the multiway table to see which clusterings are related to which other ones, One clustering can be chosen as the distinguished clustering and a model can be developed for predict jing this clustering from clusterings obtained from other bases. In this way one can ascertain how closely associated some subset of clusterings may be with the distinguished clustering. If desired, clusterings on nnondistinguished sets can be repeated with subsets of the variables that individually evince high association with the distinguished-set clustering. FEBRUARY 197 While practitioners may continue to choose some segmentation base by fiat, the new method can still show them how strongly the clusterings obtained from alternative bases are related to the distinguished base. More- over, if one should select a benefits-seeking base and find that its clustering is indepen- dent of the clusterings of all other candidate bases—while the other bases are mutually associated—then serious questions about the choice could be raised. In brief, the clustering comparison ap- proach can provide new insights into market segment relationships and statistical support for the choice of some specific segmentation base. Moreover, it can do this in terms of the natural output of a segmentation analysis— groupings of consumers—rather than being limited by the linearity assumptions of such techniques as generalized canonical correla- tion? COMPONENTIAL SEGMENTATION Up to this point, our discussion of segmenta- tion has proceeded along more or less tradi- tional lines. It is now appropriate to describe, briefly and nontechnically, a new way to segment markets, called componential seg- mentation, In one sense, componential seg- mentation is disaggregative in nature. Unlike the a priori and post hoc approaches, in componential segmentation persons who evince similar response profiles. are not grouped together. On the contrary, their individual responses are further decomposed into component contributions that are per- son-related (demographics, life styles, and so ‘on) and contributions that are offering-related (such as attribute levels of the product or service). 3. For a more technical Paul E. Green and Frank J Analysis: Alternative, Bases for Market Segmentation,” Journal of Consumer Research (forthcoming, March 1977}, cussion of this approach, see Copyright © 2001 All Rights Reserved 67 68 ‘An Example ‘A firm engaged in the marketing of credit cards was interested in developing a new card that would maintain current patron privileges but would also exhibit attractive benefits to potential retailing establishments that might wish to honor the card. Interest centered on the establishments’ utilities for various cost- benefit profiles. Five credit card attributes were selected for study: discount rate—24% versus 6% speed of payment—one day after receipt of week's vouchers versus ten days after receipt of week's vouchers credit authorization—computer terminal versus tolhfree telephone number marketing support to establishment—0.1% of an- ‘ual billings versus 0,75% of annual billings rebate on annual growth in billings—none versus @ 15% rebate on billings exceeding quota. Insofar as the preceding factors are con- cerned, the discount factor refers to the percentage of the establishment’s billings on the credit card that is deducted by the credit card company for its service. Speed of pay- ment refers to how quickly the company pays (less its service discount) the establishment after receiving the establishment's vouchers. The credit authorization procedure consists of two alternatives: a company-supplied com- puter terminal versus a manual system involv- ing credit-check calls on a toll-free telephone number. Marketing support refers to the company’s payment of local advertising on behalf of the establishment. The rebate proce- dure pertains to a plan in which the company rebates, to the establishment, 15% of the company’s charges on all billings exceeding the establishment’s quota, usually set at 1.25 of the previous year’s billings level. As noted, each of the above factors appears at two levels, so that a full factorial design would entail 25 = 32 combinations. Ordinarily, one would make up thirty-two credit card combinations and have retailers evaluate them on some type of scale, such as a one to eleven rating scale, ranging from highly PAUL E, GREEN undesirable to highly desirable. Conjoint scal- ing could then be used to estimate each respondent’s utility for each level of each of the five factors.* However, by the application of certain experimental designs—called ortho- gonal arrays—one can reduce the labor of respondent evaluation from thirty-two card descriptions to only sixteen, while still retain- ing the ability to estimate all utilities.’ Orthogonal arrays represent a type of frac- tional factorial design that is a generalization of the more familiar latin square design, (“Orthogonal” means that all effects are mutually uncorrelated.) ‘The Componential Design In the usual approach to conjoint scaling, a sample of retail establishments would be chosen and respondents asked, among other things, to evaluate the hypothetical credit cards. Segments may then be formed on the basis of commonality of utility functions. In componential segmentation, however, cone applies the same design principles to the selection of respondents—namely, fractional factorial designs, such as orthogonal arrays~as are applied to the construction of the stimu- lus cards. In the present study, four establish- ment-type factors were believed to be impor- tant background variables for segmentation: type of establishment—lodging: retailing; restau iM Ge of establishment within type—small (below median in sales volume) versus large establishment's current favorite card—card type Ayeard type Bs card type C ‘number of credit cards honored by establish- ‘ment™three o fewer versus four or more cards. In this case we would have 3 x 2x 3.x 2= 36 combinations if a full factorial were used. Johnson, “Trade-O1! Analysis of Consumer Values ‘of Marketing Research (May 1974), pp. 121-127. 5. See references in footnote I for technical details regarding orthogonal arrays. [BUSINESS HORIZONS errr A New Approach to Market Segmentation However, by employing orthogonal array principles, a design involving only binations can be prepared. One of the basic features of componential segmentation is that respondents exhibiting specific profiles are sought from some much larger frame of potential respondents in order to find the particular set of background profiles dictated by the design. If full factorials were employed in design: ing both stimuli and respondents, the re: sponse matrix would consist of thirty-six respondent types. times thirty-two stimulus types, a total of 1,152 data points. By the use of orthogonal array principles, this size is reduced to only 9x 16 = 144 data points. In practice, of course, one would obtain several replicates of each candidate respondent pro- file so as to examine within-profile homo- geneity and provide higher stability in the parameter estimates of the model. Applying the Model Componential segmentation is a model in which a series of decomposition techniques are applied to the basic 9x16 matrix of averaged profile evaluations of the sixteen stimulus cards. The basic objective is to decompose the response matrix into separate parameter values for each of the levels of the five attributes comprising the stimulus cards and each of the levels of the four profile characteristics describing the respondents. Componential segmentation (COSEG) «1 ploys more or less standard statistical tech- niques, such as analysis of variance, to carry out the first of the sequence of decomposi- tions. In the process of doing this, a useful measure is computed that shows how “seg- mentable” the market is, with respect to the stimuli and background variables used in the study. The segmentability index is simply a measure of how large the interaction sum of, squares (persons x stimuli) is, relative to the total variation in the data, In the present FEBRUARY 1977 problem, the effect of the interaction term was quite small (about 8% of the total), suggesting that this market was not highly segmentable. The COSEG model develops parameter values for both stimulus components (the credit cards) and respondent profile charac- teristics, In the former case the parameter values represent utilities for the levels of the five components making up the stimulus cards. In the latter case the parameters are saliences that indicate how much each profile characteristic contributes to variation in the evaluative responses. No attempt is made here to describe the model in detail. Rather, a brief description of the results of the COSEG analysis is offered. Figure 3 shows the output for the stimulus part of the response matrix. In the case of credit card utilities, the solid lines show that discount rate is, by far, the major evaluative component. (It is not surprising that respon- dents prefer smaller discount payments to larger ones.) Speed of payment and credit authorization procedure are also important components, while marketing support and rebate show virtually no variation at all. Figure 4 presents respondent results, a unique output of COSEG. The solid lines indicate the contribution that each back- ground variable makes to differences in over- all stimulus evaluations. As noted, type of favorite card displays the highest salience or contribution to the overall evaluation. However, COSEG also develops, sary, a set of interaction parameters, denoted by the selective dotted lines in Figures 3 and 4. In this case, only discount rate in Figure 3 and type of establishment and type of favor- ite card in Figure 4 displayed interac effects. A specific interpretation of the int action effect is that restaurants whose fav card is type A, when evaluating credit cards with the small discount rate, evince a higher utility than would be implied by the sum of the main effects (solid lines) for this combina- tion. Copyright © 2001 All Rights Reserved 69 70 FIGURE 3 Stimulus Utilities Obtained from the COSEG Model PAUL E. GREEN Usiities 10 10 10 0 0 —=_ 0 1.0 rol. . a0l. 1 2a o% 1Day 10Days Computer Telephone Discount Rate Speed of Payment Credit Authorization Lor Lor o = Oe 1.0L 1 10 1 0.1% 0.78% None 1% Marketing Support Main effect parameters Interaction parameters (Stimulus mode) Use of COSEG The main value of COSEG lies in its ability to make predictions about the relative ratings of any of the thirty-two stimulus cards by any of the thirty-six possible respondent types, as based on parameters developed for the small subsets of each. For example, suppose a researcher wished to learn how a large restau- rant, honoring three or fewer cards, whose Rebate favorite card was type A, would evaluate six different credit cards, each of which is com- posable from the five basic stimulus factors, shown earlier. COSEG provides that prediction, using parameters estimated from the calibration sample. In practice, errors are associated with the estimation of COSEG’s parameters. A measure of this overall error can be obtained (by standard statistical procedures) for the 9x16 matrix that was used for parameter estimation in the first place. Then, having estimated the standard deviation of prediction errors, one could estimate various segments? BUSINESS HORIZONS EE A New Approach to Market Segmentation FIGURE 4 Respondent Saliences Obtained from the COSEG Model Saliences Lor 107 o 10LI L 1 ol L Lodging Retailing __ Restaurants ‘Small Large Type of Establishment Size of Establishment 1.0- Lo “10 L L 1.0L 1 a B c Type of Favorite Card Main effect parameters ——— Interaction parameters (Respondent mode) evaluations of alternative combinations of credit card service levels by Monte Carlo simulation. Moreover, it is a relatively simple matter to weight the choices of specifie segments in accordance with their relative importance in the population of interest. The typical kind of output from these Monte Carlo runs entails a cross-classification of each segment’s fre- FEBRUARY 197 smber of Different Cards quency of choices for each stimulus alterna- tive under evaluation. Hence, for a specific respondent segment one could find the fr quency of first choices received by each credit card composite. Similarly, for each credit card cone could find the frequency across segments in which it receives first choice. In short, once the parameters and overall prediction error of COSEG are estimated, it is relatively straight forward to run various sets of Monte C: trials involving alternative strategies of in- terest to the marketing manager. lo Copyright © 2001 All Rights Reserved 72 Contribution of COSEG The value of COSEG would appear to lie as much in the point of view that underlies its construction as in the specific model itself. Componential segmentation focuses on the building blocks of segments rather than on specific groups that may emerge from a particular type of clustering procedure (or other means for developing segments from a specific data base). If the interaction effects are small, segmentation may not be warranted with respect to the stimulus domain under study. If the interaction effects are large, segmentation could be of major interest. In the illustrative example we considered a case in which both respondents and stimuli were capable of being decomposed into sub- ordinate factor levels. Two other cases are of interest. where: (1) respondents, but not stimuli, are decomposable; and (2) stimuli, but not respondents, are decomposable. A number of segmentation studies— benefits secking, problem-oriented segmenta- tion, life styles, brand preferences—are carried out in which respondents can be selected according to specific background profiles but the stimuli are ordinarily not easily decom posable into subordinate components. For example, in needs segmentation studies con- sumers are asked to indicate how strongly they agree or disagree with each of a set of items that characterize various needs that a product might satisfy. Ordinarily, these items cannot be described as factorial combinations of more basic attributes. COSEG can still be applied to this type of problem, although the leverage for using it decreases. Illustrations of the obverse case frequently appear in the context of concept testing or general studies involving conjoint scaling where respondents (not chosen according to a factorial design) evaluate designed stimuli that are constructed factorially. Again, COSEG could be applied to the response data and post hoc examination made of the relation. ship between concept ratings and various PAUL E. GREEN background variables. Moreover, Monte Carlo simulations could be set up in either case. However, COSEG seems most useful cases where both respondents and stimuli can be further decomposed into subordinate com- ponents and where the advantages of ortho- gonal arrays (and similar kinds of fractional factorials) can be fully realized in parameter estimation. I suspect that a number of mar- keting problems could be profitably formu- lated in this way. The high efficiency of COSEG is based on the use of fractional factorial designs that, in tum, permit uncorrelated estimation of re- spondent and stimulus effects, As such, COSEG is particularly useful in the design of stimuli and the selection of respondents. However, cases will often arise where ortho- gonal estimation of respondent background characteristics is not feasible. If so, the procedures described here can still be used to make nonorthogonal estimates of the para- meter values of interest, albeit with lower statistical precision. Thus, even if one cannot select respondents according to orthogonal- array desiderata, the principles of COSEG can still be applied to obtain nonorthogonal esti- mates of respondent-type subordinate factors. ‘Some Caveats ‘As the reader has probably surmised by now, componential segmentation is not a variant of market segmentation in its usual sense. Some major differences should be noted COSEG considers the question of whether or not a market is segmentable to be an empirical matter, capable of being answered statistically for the background variables and stimuli under study. No groups or clusterings in the usual sense are formed. Rather, COSEG develops para- meters for predicting how any segment com- posed from the basic background levels evaluate similarly composed stimuli. BUSINESS HORIZONS A Naw Approach to Market Segmentation The size of the segment must be estimated from the larger data bank from which the specific respondent profiles (needed for the orthogonal arrays) are drawn. A number of caveats should be listed in considering the COSEG approach: To be most effective, COSEG requires the ability to screen a large number of possible profiles for the particular subset that meets the design conditions. Ongoing panels where respondent background variables exist in com- puterized form are most useful for this approach, COSEG does not employ a random sam ple. Profiles selected to meet the design characteristics are not necessarily “typical” of the population as a whole. Moreover, it is one thing to design stimuli and quite another to select respondents from a larger universe. In the former case one can design out (or randomize over) unwanted variables. In the latter case one gets all of the respondent, not just the profile of interest When the desideratum of profile ortho- gonality must be relaxed, the high statistical efficiency of COSEG diminishes. Moreover, interpretation of correlated parameter values becomes more difficult substantively The preceding caveats clearly call for the collection of additional data for internal cross validation of the COSEG model’s predictions. This should be @ sine qua non for using the model. SSF] The main objective of this article BM 5 oon abonte nome rather_basie beliefs: * Segmentation works best when the study design is tied in with market strategy FEBRUARY 1977 * The traditional Ansoff table can be expanded to produce a table that distin- guishes between strategies at the structural level versus those at the functional level. This expanded table can serve as a kind of check list for strategy development within various segments. ‘* While a myriad of segmentation bases exist, only two fundamental approaches—a priori and post hoc (and a hybrid of the two)—have been traditionally applied. # It is not always clear what subset of segmentation variables should be chosen as the distinguished set. However, procedures have recently been developed to examine the degree of association among alternative se mentation bases. Moreover, if some base is chosen by fiat, the same approach can be used to find a functional relationship between its clusters and those produced by employing other bases. * Componential segmentation represents a new approach to segmenting markets, one that emphasizes the prediction of new respon- dents’ evaluations of new stimuli (for exam- ple, new product offerings) from 4 relatively small sample of factorially selected respon- dents and factorially designed stimu * In the COSEG approach, the segmen- tability of a market can be measured for the specific background variables and stimuli un- der study. * Given the fact that COSEG estimates a relatively large number of parameters on the basis of a relatively small number of calibra: tion responses, it is essential that all COSEG studies include an internal validation step where a control sample of responses are held out of the analytical stage and later predicted by the model. Copyright © 2001 All Rights Reserved 3

Vous aimerez peut-être aussi