Segmentation White Paper* Mike Morgan, PhD First edition: January 19, 2006 Current edition: October 9, 2012 The purpose of this paper is to: Give a brief overview of the history of market segmentation. Provide a statement of the client value proposition and a “Market Segmentation Manifesto.” Illustrate how segmentation addresses the heterogeneity of consumer preferences. Illustrate how segmentation addresses the fact that many dimensions of consumer preference are essentially unobservable and require unique types of modeling to capture them. Describe different kinds of segmentation approaches, along with their limits and advantages. Delineate typical segmentation projects. Describe the characteristics of “good” segmentation solutions, along with benchmark goals for success. Some of the material herein is borrowed or adapted from Wedel and Kamakura’s Market Segmentation, which is, in turn, borrowed from other theorists over the past 40 years in this wonderful, diverse, and still very vibrant field of research and analysis. Thanks very much to Scott Van Manen for his invaluable insights in this endeavor. * 1 Overview of the history of market segmentation In the Beginning was the Market… Marketing can be thought of as a framework of ways in which sellers deliver value to buyers in a market economy. It involves creating wares for sale, communicating to buyers their availability, pricing of the wares, and making them available. The theory and practice of marketing has evolved, sometimes slowly and sometimes by leaps and bounds, ever since the emergence of the village marketplace. The most profound early reflections about what capital-driven markets deliver to consumers were made by John Stuart Mill in the mid-nineteenth century, in his book Utilitarianism (1861). For Mill, consumer value could be measured in “utiles,” or utility, a close relative if not synonym of the word “happiness.” All products and services could be reduced to this single common denominator. All buyers had the same preferences, and all competing goods were commodities. This economic worldview dominated the theory and practice of marketing until the early 20th century. Its last great cultural expression in America was Henry Ford’s mass production and sale of the Model T. What drives advancements in the practice of marketing? Despite advancements in agricultural, manufacturing, and information technologies over the past 150 years or so, the supply side of the market has not necessarily been the primary driver of marketing’s evolutionary process. Granted, it is still largely true that “If you build it, they will come.” But, the supply side alone has not taken us as far as we have so obviously come beyond John Stuart Mill’s rudimentary, commodity market economy, into the age of mass customization. Instead, it is the evolution of consumers – physically (in numbers, shapes, and sizes), psychologically, socially, morally, spiritually, intellectually, ethnically and racially, sexually, and last but not least, financially – that primarily drives advancements in marketing in the modern world. In other words, marketing has evolved mainly by specializing – the wares themselves, the pricing schemes, channels of distribution, and communications – to fit an ever-increasing degree of diversity in the consumer population. Did the remarkable range of consumer diversity we observe today first begin to explode when it became popular to “do your own thing” in the late 1960s? Actually, economists and marketers began to notice key differences among consumer types at the turn of the 20th century. In his book, Theory of the Leisure Class (1899), Thorstein Veblen (who was age 15 when John Stuart Mill died) argued that economic life is not driven by notions of utility, but by differences between the more privileged class – those who control politics, money, and society – and the less privileged who could, at best, emulate the consumption power of the privileged in limited and artificial ways. Without really knowing it, Veblen identified not only the first basis for market segmentation – in socioeconomic terms – but also discovered the earliest strategy for marketers to target their wares more effectively – by enabling ordinary consumers to emulate the wealthy though “conspicuous consumption” of powerfully symbolic products and services. 2 Client Value Proposition – What is the Real Value of Market Segmentation? There is a lot of discussion in this article about a range of theories and practices in marketing segmentation. To help build some perspective, here is a starting point, perhaps a framework of values. Please consider each of these points carefully as well as how they might tie together. Market Segmentation Manifesto o The successful marketer depends on uncovering hidden values in the market – new product opportunities, unmet consumer needs, and sources of demand from the very “edges” of every existing and potential target market. o Innovation, launched from a baseline of deep insight into competitors, products, and consumers, is what ultimately delivers marketing value to consumers and wealth to company stakeholders. o In order to identify market opportunities comprehensively and innovate towards them effectively, the marketer must have a clear view, not only of these hidden opportunity values themselves, but of the sheer range and complexity of such values. o It is towards understanding this range and complexity of hidden values that market segmentation offers its strongest and most unique contribution to the marketer’s competitive arsenal and, ultimately, to consumer happiness. How Market Segmentation addresses the Heterogeneity of Consumer Preferences. Target Markets Don’t Come in Neat Packages… Everyone likes to think in terms of discrete, mutually exclusive categories. He’s an African American, she’s a nuclear scientist, they belong to the Democratic Party, those people are Sinners, we are proud Americans. In real life, we know (and will readily acknowledge if asked) that such mutually exclusive categories are just convenient generalizations. Instead of categories, the universe is made up of continuous dimensions – we slice it up in order to manage it for ourselves, in order to create discrete rules for our beliefs and behaviors. We do this so that we do not become paralyzed by the fact that no single category is ever really distinct once you begin to dig into it. The African American is not purely of African descent (as Malcolm X continually reminded people during his own short life), nuclear science is just a portion of the vast body of physical science and mathematics, Democrats often cross party lines to vote for the other candidate, sometimes sinners are indeed saints, and Americans are not always so proud, at least, not of everything. 3 Nevertheless, we go ahead and slice and dice up our world to make better sense of it. Thorstein’s classification into leisure versus working class, in fact, requires one to draw a line, however arbitrarily, between higher and lower income and wealth ranges. As we know, income and wealth are instead continuous dimensions, upon which we can draw whatever lines we wish to suit our own purposes. If we think of continuous consumer dimensions for a moment, we need a word to describe differences along these dimensions. The word used in this article is “heterogeneity.” Heterogeneity exists along a continuum. It is the opposite of homogeneity, or sameness, among a group of objects, people, or ideas. The rainbow appearing after a thunderstorm, while appearing to consist of distinct colors, is actually a continuous spectrum of light – we just don’t see the spectral shifts from our own distant perspective. Consumer heterogeneity exists along a spectrum – actually, along multiple spectra. It is continuous and usually smoothly distributed. Heterogeneity contains a subtle element of surprise – not what we expect, not a neat package that is easy to think about or easy to act upon. More than anything else, this means that if you don’t like surprises, and if you don’t like heterogeneity and diversity, you should probably stay out of the market segmentation business. Figure A. Homogeneity to Segmentation to Continuous Heterogeneity: Types of Marketing Phenomena Homogeneous Population Commodities Utility Production Market Demand Market clearing prices Product Purchase Volume Consumer Surplus Market requirements Information Mass-Production Isolated individuals Households Market Share Discrete Market Segments Competitive Set Feature Bundles Styles Segment preferences Price tiers Product line(s) Choices Key drivers Awareness (yes/no) Segment offerings Peer Groups Socio-demographics Segment acquisition/retention Continuous Heterogeneity Brand Equity Features/Options Design Individual preferences Price elasticity Product relationships Participation Value-laden dimensions Honesty Customization Word-of-Mouth Family/Life Styles Responsibility/Ownership Figure A illustrates differences between a view of the market as homogeneous, categorically segmented, or continuously heterogeneous. The market segmentation business focuses efforts, traditionally, on the middle paradigm. We try to identify which segments prefer which brands in the competitive set, what particular combinations of 4 features are important, and we attempt to project that segment’s preferences onto a total market opportunity. Homogeneous Population The Homogeneous market Members of the market are pretty much the same. The important thing is to understand the market overall. Discrete Market Segments Members of the market fall into discrete groups, based on similarity of preferences or member characteristics. Continuous Heterogeneity Buyers have different preferences and characteristics along different continua, so even if we have segments, within-group variation is also part of the scheme. 5 An important note is that we cannot yet, with today’s research technologies, always attain the one-to-one marketing paradigm using the “heterogeneous” column in Figure A, however ideal that may seem. While categorical segmentation is, at best, only an approximation, a convenient generalization of “true” individual reality, it does provide marketers with manageable categories. It may be years before we can fully understand the consumer’s true personal, idiosyncratic, emotional, and cognitive relationships with our company’s products. Today, though, we can generalize these relationships to a manageable segment level. Does this mean that we cannot ever capture more heterogeneity in the client’s market than we currently do? Not at all! In fact, this article will help lay out a path to capitalize on new and existing research and statistical technologies, and in doing so to achieve maximum effectiveness to explore, elucidate, and forecast the heterogeneous frontiers of any of our clients’ markets. Here is the primary challenge with respect to consumer heterogeneity. Challenge #1 – To capture the most consumer heterogeneity possible in any market segmentation analysis. o Market segments generally tie together consumers in such a way that their traits, attitudes, and responsiveness are relatively similar within each segment. o However, each segment always retains, to some extent, a range of diverse traits, attitudes, and responsiveness. o Ultimately, market segmentation must account for this diversity – within market segments – in ways that add valuable marketing opportunities. Many dimensions of consumer preference are essentially unobservable and need to be modeled as such. It’s a kind of magic. The purpose of any magic trick is to obscure or hide the true mechanism behind the outcome. In any consumer (or business) market, the magic is already built in – what the buyer really wants is “unobservable” until he or she is presented with the product. Perhaps this is because the buyer is not sure of the desired outcome until the solution is presented (i.e. the case of a new product). Or, perhaps we observe sales volume for an existing product, but the consumer types who buy are hidden beneath this, each buying at a different rate and for different reasons. Only the magician – or the marketing research specialist – can uncover those hidden forces. If the seller doesn’t care about consumer types, the underlying, unobserved dimensions of the market will forever remain unobserved. This is the case with most commodities – who cares if the buyer of pork belly futures is gay, bisexual, or straight, as long as the price is right? 6 On the other hand, suppose a given consumer market is crowded with competitors, each capable of reaching consumers independently with their own magical marketing tools. Each competitor is (or should be) desperate to find those consumers who – with the right messages, product types, distribution channels, and prices – can be uniquely drawn to its own unique offering. If this weren’t the case, marketing would not exist. Fortunately for us marketers, research tools allow any competitor to sample the market and ask consumers what they would like to buy and then develop marketing strategy accordingly. Available data to use in market segmentation. Consumers can be classified into groups based on age, income, race, ethnicity, education, size and type of family, geographic location, and other “observable” factors. In fact, database marketing companies provide lists of (nearly) all consumers in America on all geographic and demographic dimensions. How do they know these things? Because the U.S. Federal Government has itself conducted surveys of the population to obtain the information and made it available to anyone with the tools to aggregate and store the data. Thus, we might say that the most observable information is already obtained, and (as yet) unobserved information is available, at some cost, to any competitor in the market. It’s what we don’t easily see that gives marketing its magical power. Figure B. Easily Observed versus Hard-to-Observe Bases of Segmentation Type of Information Consumer Traits Consumer’s Relationships with Product Easily Observed <-----------> Hard-to-Observe Culture, Geography, Psychographics, Values, Demographics, SocioPersonality, Lifestyle Economics Product Usage, Options, Product Benefits, Styles, and Extras Attribute preferences, Purchased. Behavioral Intentions Market share and volume response to price changes Maximum Willing-to-Pay, price elasticity Consumer’s Relationships with Promotion Market share and volume response to ad campaigns, promotional offerings, etc. Awareness, Consideration, Evaluation based on promotional information Consumer’s Relationships with Distribution Channels Sales volume and cycle per channel. Penetration and repeat purchase rates from channel offerings. Perceptions of Convenience of, Locus of Control within the Channel, Value-Added features of the Channel. Consumer’s Relationships with Price 7 Figure B illustrates differences in marketing-relevant information on the basis of how easily observable such information is. In the academic literature, you might see the term “unobservable heterogeneity” used to describe items in the rightmost column. History behind the idea of modeling unobservable heterogeneity. A handful of econometricians originally built models that “observe” the unobservable data by modeling them with latent variables (more on this below). One such econometrician, James Heckman, won a Nobel Prize for his work in this area in 2003. During the 1970s, it was believed by many social theorists and politicians that the longer a person remained unemployed, the less employable that person would become. If this were true, then social policy called for quick intervention – skills training, job seeking help, etc. – as soon as possible after a person became unemployed. This would be a very expensive social policy, since it would involve selective investment of government funds – and selective about who gets the most help and when. What is the “unobserved” data in this case? It is the actual employability, or skill level, of the worker. Employability was, from the view of social policy, “unobservable.” Heckman found a way to model this dimension, though, using a latent variable approach. He showed that, once you control for employability, the amount of time one is unemployed is unrelated (uncorrelated) with one’s chances of getting a new job. This saved American taxpayers, needless to say, a tremendous amount of money. From the perspective of the marketing services company, we need to keep in mind two important principles: 1. Nearly anything is observable if you throw enough money into the effort! Collecting and modeling the Hard-to-Observe Data, in fact, is one of the major benefits that clients pay for, using survey research. 2. BUT, before you throw all of the client’s money into revealing the unobservable, we should consider modeling alternatives. 3. Modeling – as opposed to collecting – unobserved data is always a second-best solution – IF costs of collecting such data are within the client’s budget. Are there any kinds of unobservable dimensions – or unobserved heterogeneity in the market – that absolutely cannot and will not be observable? There is one – we can call it The Future. We can guess at it, predict it, or even expect it, but no one can actually observe the future until it happens. From the perspective that any market is continually in flux, any market segmentation is obsolete the moment it is completed. However, if the segmentation solution is based on measurements that are relatively stable over time, we can be assured that the solution will remain useful for some time. Estimates of this time vary. Rules of thumb that I subscribe to, based on my own experience over a number of segmentation analyses, are given below. 8 The life of a segmentation solution: o A solution based on demographics – with or without a geographic component, should last three years. o A segmentation scheme based on consumer attitudes or needs should continue to be valid for at least 12 months. o Segmentation based solely on purchase behavior should last at least six months. Do we ALWAYS have to renew a segmentation study periodically? So far, in both theory and practice, we should. We need to uncover new methodologies that take into account dynamic changes and evolution in the consumers’ geo-demographics, attitudes, and behaviors. The first commercial research firm to offer a proven approach to “dynamic segmentation” will indeed make history in the marketing industry. Challenge #2 – Capture more UNOBSERVED consumer heterogeneity than has been previously possible with existing methodologies. o Market segmentation must account for both observed and unobserved (or less easily observed) dimensions of the market. o Tools for revealing these latent aspects of the market, and market segments, need to be applied as vigorously as possible – this is where a company can gain the most competitive advantage in its markets. o Changes in the market – or market segments – over time are intrinsically unobservable, and these are very difficult to measure and take into account. Different kinds of segmentation approaches, limits and advantages, and characteristics of “good” segmentation solutions. Benchmark goals for success. Criteria for a Successful Analysis It is important to have objective standards for assessing the value of any segmentation solution. We would like to set the bar for such products fairly high, since that makes it harder for our competition to match our products at lower prices. Importantly, our criteria for a “good” segmentation need to be quantifiable, so we can both (1) objectively evaluate any given segmentation solution and (2) track our performance over time, by client, by industry, and across our client base. Below is a relatively concise list of criteria for assessing a market segmentation solution: Identifiability – Will we recognize the segments when we see them in the marketplace? Substantiality – Even if we can recognize them in the marketplace, are more favorable target segments large enough to make it worthwhile investing marketing dollars in them? 9 Accessibility – Even if we can recognize them and they are worthwhile to target, can we reach them through our targeting efforts? Stability – Suppose we can find them, target them, and make money in doing so. Will they be around long enough to respond to our client’s offerings? Or, is there a lack of stability in the market that will cause segments to shift attitudinally or behaviorally? Actionability – Assume we have identifiable, targetable, profitable, and stable segments. Will targeting them effectively serve the client’s overall strategic needs? This is not just a question of making a profit, but of how the client positions itself in the market, leverages its core capabilities, and remains competitive – a much harder question for most marketers to answer. Yet, it is actionability that delivers the most vital long-run value to the client’s company. Responsiveness – Now we are at the point at which our chosen segments are indeed strategically the best ones we can target – they have value and can be reached successfully through marketing. But wait! Will they respond to our marketing efforts (price, product design, distribution channels, and marcom)? If this final hurdle is not overcome, all has been for naught. We still have the issue of whether segmentation criteria can be quantified. Some initial, suggested metrics are shown in the table below. These are treated in more depth in a subsequent section of this article. Figure D. Recommended Criteria for Assessing the Value of a Specific Segmentation Solution, with Suggested Performance Metrics Criteria for Segmentation Identifiability Substantiality Accessibility Stability Actionability Responsiveness Some quantifiable measures/metrics Probability that a predicted segment member actually belongs to the segment we predict, from cross-validation tests. Segment size (percentage of the total market belonging to the target segment(s)) and economic value (spending or profitability of the segments). Also, actual ROI on targeting programs. Audience TRPs available, return rate on direct mailings, and other reachability metrics. Cross-tabulation of earlier against later segment classifications. Client satisfaction rating of the solution. Number of sales closed, hands raised, interests piqued, etc. The easiest measurement in the table above, under Substantiality, is segment size – the results of the segmentation analysis include this automatically. Once segment size is determined, and we have data on spending and purchase frequency, we can project the economic value component of Substantiality as well. 10 Next easiest is cross-validation for satisfying Identifiability. This is because standard methods are available using the same software and data as that used to do the segmentation analysis. The rest of the segmentation criteria can be evaluated only after the client has implemented the segmentation and applied it to marketing strategy. In fact, calculating the ROI to assess Substantiality, while not impossible, is more difficult than for any other measure. Quantitative methods leveraging recency, frequency, and monetary value of customers and segments are available but not as yet widely tested. More importantly, can a follow-up research study be conducted to assess the criteria in a proactive way for the client? The use of client satisfaction ratings to assess the actionability of a segmentation solution seems less objective than the other metrics. Regardless, this particular criterion is crucial to the marketing services company, because it constitutes the ultimate value of segmentation. Thus, we all need to maintain constant focus on whether we are meeting the actionability requirement in our segmentation solutions. A subsequent section lays out more comprehensively how to assess the Identifiability and Substantiality criteria for any given segmentation solution. Before that, though, we need to have an overview of segmentation methodologies. Different kinds of segmentation approaches, along with their limits and advantages. The Magic behind the Method – an overview of Segmentation Methodologies Despite the dizzying array of segmentation methodologies, there are really only two main dimensions along which all segmentation methodologies can be positioned. 1. A priori versus Post Hoc. o An a priori segmentation assumes that the market segments are already clearly defined in the data. Examples include segments defined by age categories or by geographic regions. Also included here are segments the client has developed and rely upon for strategic guidance. o A post hoc segmentation assumes that the market segments exist in the population, but are as yet unobserved (see discussion above and Figure B on the issues surrounding observability – unobserved segments are the sources of market heterogeneity). Examples of post hoc segmentation include segments defined in terms of psychographics or product preferences. 2. Predictive versus Descriptive. Every targeted segment has some outcome our client wants to drive through market strategy, whether it is trial of the client’s new product, customer satisfaction, increased spending levels, or some other favorable outcome. 11 o A predictive segmentation solution explicitly incorporates the outcome and drivers in the model. That is, different segments have different drivers of the outcome, and this predicted outcome is sought by addressing the drivers of the target segment(s) via marketing strategy. o In contrast, a descriptive segmentation solution seeks only to identify the most homogeneous subgroups of consumers, with the homogeneity defined by variables of interest to marketers. For example, suppose that consumers are targetable in terms of their values, as reflected in VALS2. A valuable descriptive segmentation may then consist of identifying segments which are homogeneous, within themselves, on the VALS2 variables. It is important to note that the objective of the descriptive segmentation is not necessarily to identify segments from which we might expect a return on marketing investments. Segments providing a favorable ROI, of course, should emerge from the descriptive approach, but that would be more of a serendipitous outcome and perhaps require several different iterations to uncover. Lack of actionability is a frequent criticism aimed at descriptive segmentation solutions. However, it is also true that the descriptive approach is more likely to produce the clearest and most intuitively plausible picture of the market. Why? Because these are customer or prospect types that we can easily see “in the street.” When the emphasis is on prediction, not just description, the segmentation model can group together consumers who tend to respond the same way to the same marketing stimuli. If we classify segmentation methods based on whether they are descriptive or predictive, and a priori versus post hoc, and combine some of the other segmentation dimensions described in this article, we get Figure E. Figure E. Types of Segmentation Methods A Priori Descriptive Banner tabulation, ANOVA, chi-squared tests, t-tests. Predictive Selective cross-tabulation with chi-squared tests, Regression, Logit, Discriminant Analysis. Post Hoc Hierarchical and K-Means cluster analysis, latent class cluster analysis, fuzzy set models, unsupervisedlearning neural networks CHAID/CART, latent class regression, latent class choice models, hierarchical Bayes choice models, supervised-learning neural networks. A priori, descriptive. These methods are classical applications in marketing of what some fondly call “slicing and dicing” of the data. Banner tabs are the staple product of all marketing research studies, and ANOVA, chi-squared, and t-tests enjoy a rich 12 grounding in experience and credibility among end-users. No one will ever argue with you about a t-test (perhaps about the underlying assumptions about the data, but not about the test itself). A priori descriptive analyses have three fairly strict assumptions: 1. Consumer heterogeneity is measurable typically by dividing up the market into discrete, non-overlapping groups. Thus, continuous heterogeneity is usually subsumed into group classification (see Figure A for illustration). 2. Since the rules for grouping, or segmentation, are decided on a priori – in advance of data analysis – those rules, and the data to be analyzed, are easily observable (see Figure B for illustration). 3. Appropriately designed, an a priori segmentation satisfies our segmentation criteria of identifiability, substantiality (with the right cuts), and accessibility, since the segments are easily observable, their sizes can be determined, and their mailing addresses can be identified. It is unclear – some would say unlikely – that a priori, descriptive segments will always be stable, actionable, or responsive. We might have them well-defined, but they could very well be distributed completely randomly on other, more important criteria. This is because a priori, descriptive segments are designed only to describe the market, not predict how well any given segment will respond to marketing stimuli. In order to capture segment value in terms of marketing responsiveness, we need a predictive model. A priori, predictive. These methods also depend on easily observable data and predefined segment classification rules. They are perhaps more valuable to marketing strategy than a priori, descriptive segments. Why? Because they can be measured on Responsiveness. After all, a regression model – or its discrete version, the choice model – is intended to predict some valuable consumer response for the marketer. These tools assume a cause-and-effect relationship between marketing action (product design, pricing, promotions, channel distribution) and consumer outcomes (shopping visits, consideration, brand favorability, purchase). The problem with a priori, predictive segmentation analyses is that they cannot be used to uncover unobserved (or less easily observed) dimensions that, arguably, more strongly drive consumer response than do the observable characteristics. Markets may contain latent segments that cannot be observed in advance of a rigorous research study and data analysis – that’s what marketing research firms make money on, right? Latent or unobservable segments, once identified, always offer marketers the strongest inroads to success (also, see Marketing Manifesto on page 2). It should be noted that, once latent market segments are uncovered, a priori, predictive tools can used to predict segment membership in the population. Our main tools for this purpose are discriminant analysis and logit modeling. 13 Post hoc, descriptive. As noted above, marketing dimensions and segments often need to be teased out of the data – that is, they are not reliably or completely apparent in the observed ratings data, or database variables themselves. Various exploratory techniques, such as factor analysis and K-Means cluster analysis, are often helpful in extracting these latent dimensions and segments. K-Means, in fact, is the method most often used to do post hoc, descriptive segmentation. Respondents are assigned to clusters based on how similar they are on key variables – attitudinal or needs-based, behavioral, and others. The results are (usually) clear-cut, easily described, and intuitively plausible segments. Unfortunately, the post-hoc, descriptive approach suffers from the same problem as the a priori, descriptive approaches – it may describe the segments well, but it also may not predict how well a target segment will respond to the client’s products or marketing strategy. A secondary drawback, unique to post hoc, descriptive approaches, stems from the fact that, prior to investigation, segments are not observable. Unless the post hoc clusters are defined at least in part in terms of observable variables, such as purchase frequency, age, or geographic region, it may be impossible to find them for targeting purposes. This may not matter if the client’s marketing activity is primarily national broadcast advertising, but it makes things quite difficult for direct mail or placing ads on cable TV or the Internet. In addition, unless there is a very clear connection between the needs measured in the study and demand for the client’s products, there is very little guidance from these solutions on what should be offered to the market and what economic value different segments might offer. This means we could fail to meet the crucial segmentation criteria of Substantiality, as well as Identifiability and Responsiveness. These problems place the researcher in a dilemma. On the one hand, she wants to identify segments that differ in terms of enduring and consistent attitudes and needs. This is the best way to make the segments human and thus to think about what the client can say to them. Clustering study respondents on their ratings of benefits accomplishes this goal, and K-Means, when done correctly, accomplishes it superbly. (The analog in database marketing is what we often call “look-alike” models – people with similar needs based on past purchase behavior). However, the more you cluster on attitudes or past purchase behaviors, generally, the more you sacrifice within-segment similarity on observable variables like demographics. Hence, the less reachable your needs-based segments may become. The researcher might say, “The market is ultimately created and defined by needs (or purchase histories). Therefore, my job is to identify these segments as best as possible and worry about targeting them afterwards.” Moreover, in some industries, finding and targeting such segments in the population are not so difficult (automobiles are an example – broadcast advertising potentially reaches everyone). Also, some end-users may desire crisp attitudinal or behavioral segments because they will be used only for ideation and long-term strategic planning, not for targeting the consumer in the street or launching the next brand extension. These uses are exceptions, but they do exist. 14 There are scores of segmentation analyses that, despite their huge samples and high prices, are ultimately never used by the client’s company, because the segments cannot be targeted by concrete marketing actions. Needs or behavioral patterns may be crisp, but, unfortunately, demographics are randomly distributed across segments – in other words, the segments are invisible to targeting efforts. Appropriate messaging to each segment may be as clear as the summer sky, but the messages cannot be matched to media in any cost-effective way. The results are client dissatisfaction with and distrust of large market segmentation studies of any sort. To sum up, identifying latent segments and defining them in terms of attitudes, needs, or behaviors alone, while this may describe the market well, can be deadly for the market segmentation research business. Yet, these still seem to be the methods of choice for most (but not all) commercial research and database marketing firms. Post-hoc, predictive. Some of the problems associated with post-hoc, descriptive segmentation can be addressed by deriving segments based not on who they are, but on how they respond to marketing stimuli. Such segments, though they may still be virtually invisible in the population, will presumably buy the product they are predicted to buy. This helps pinpoint the best new product designs and gauge their market opportunity value. Offering the right product line allows each segment to “self-select” into that offering which best meets its needs. Predicting promotional response works the same way, except we still have the problem of media allocation unless different segments can be identified with different media vehicles. Based on the above discussion of segmentation methods, it seems there are two key areas which are not easily or completely covered by existing frameworks and tools: 1. Predicting segment responsiveness for the purposes of marketing strategy. 2. Identifying and evaluating unobserved (latent) segments in the population. That these two areas are difficult to address is actually very good news for high-quality marketing services vendors – our skills and methods give us a clear advantage in accomplishing them when our competitors cannot! Typical segmentation projects. These include: o Customized research studies based on primary data collection. o Database marketing models. Customized studies. Survey samples produce ratings, rankings, and yes/no responses on a relatively large number of survey items. These items are designed specifically for a single study client, not to be conducted more than once unless they are monthly or quarterly tracking studies. Thus, segments are developed for a specific client in a specific 15 competitive market, to uncover previously unobserved targeting opportunities and to provide the client with competitive advantages and strategic guidance. (See post-hoc descriptive and predictive model definitions above). The segmentation methodologies used here include mainly K-Means clustering (of variables or choice/conjoint model coefficients) and Latent Class Analysis. Database marketing models. When the models are originally developed here for segmentation, we produce groupings of records (customers or prospects) in a database that either (a) look alike, or look like an existing profile, or (b) respond in similar ways to the same marketing strategy. The segmentation methodologies used here include mainly a priori segment definitions (provided by people in Marketing and Insights Integration) and K-Means cluster analysis. However, we have in the past also applied neural network and genetic algorithm models. Although these have not been used specifically for segmentation, they are adaptable for that purpose. Most promising for future applications are neural network segmentation models (self-organizing networks) and Latent Class segmentation models. Sometimes database models are developed to predict membership in pre-existing segments (see a priori, descriptive and predictive model definitions above). Such preexisting segments come from either (a) a client’s customer database or (b) a research study based on primary data. Usually, we assume that such pre-existing segments need to be predicted for some external database, for profiling or targeting purposes, and the quality of our predictions will depend very much on what variables are available in those databases to use in our models. There are no specific methodologies used here, such as Latent Class or K-Means, to derive new segments. Instead, we focus on finding models to predict the segments using discriminant analysis and multinomial logit. Characteristics of “good” segmentation solutions, along with benchmark goals for success. Recap of segmentation issues, bases, and criteria for success. A top-line summary of the material covered so far is: Successful competitive marketing depends on exploiting hidden values in the market, and the practice of market segmentation offers its best contribution when it is leveraged for this purpose. Target markets exist in the population – we don’t make them up. But neither do target markets come in neat packages – it takes expertise to uncover them, which is a key value proposition of the marketing services company. Markets are heterogeneous, and segmentation is but one compromise in how to measure that heterogeneity. Its success will depend on individually customized consumer preferences are. Some dimensions along which market segments can be identified are easily observable in the population of buyers – for example, demographics. Other dimensions are hard to observe and may need to be inferred indirectly from observed patterns of behavior – for example, price sensitivity. 16 Successful segmentation produces valuable target segments for the client. The value of the target segments can be assessed in terms of their: o Identifiability – heterogeneity is accounted for, latent dimensions are revealed. o Substantiality – provides economic value to the client o Accessibility – targetable through existing media vehicles o Stability – will retain predictable profile and behaviors long enough to be targetable o Actionability – clearly suits the client’s unique value propositions in its market. o Responsiveness – will respond favorably to the client’s messages, products, offerings, and pricing. There are two particularly difficult areas in doing an excellent segmentation analysis. These require the marketing services company to accomplish something difficult, but for the successful marketing services vendor, offers unique competitive advantages. 1. Uncovering the latent dimensions and 2. Predicting the marketing responsiveness of the segments. It is important not to consider segmentation performance criteria independently of one another. For example, a segmentation solution may not need to provide much Stability in the target segments if they are clearly Identifiable and are only targeted with a series of short-term promotional offer. All else being equal, however, here are target metrics for each of our segmentation criteria: As mentioned earlier, it is clearly beneficial for to quantify these criteria for evaluating a segmentation solution (see Figure D). This requires looking at, for example, the number of client sales closed from direct mailings, cross-validated segment classification rates, ROI on targeting specific segments, cross-tabulation of earlier against later segment classifications to assess stability, and client ratings of satisfaction with the actionability of the segmentation solution. As we saw above, though, there are only two criteria for which metrics can be assessed before the segmentation solution is implemented in the client’s organization. These are: o Identifiability – Our ability to correctly predict segment membership. Although we cannot yet find out whether we can identify them in the market, we can assess whether the segment assignments in the sample are internally consistent enough to predict segment membership among respondents. o Substantiality – Whether we have derived segments of sufficient size and economic value to make them strategically useful to the client. Quantifying Identifiability. Here we consider correct classification rates when key variables – generally including those used to derive the segments to begin with – are used 17 to predict each respondent’s segment membership. If there is too much random error in these predictions, the segmentation solution at least lacks internal consistency. If the sample is representative of the population, then the segmentation solution also lacks generalizability. Either or both of these can (and should) be the kiss of death for the solution. There are two kinds of correct classification rates to consider: In-Sample and Holdout Sample. 1. In-Sample classification rates are the rates at which the segments can be correctly predicted when data on all respondents are used to predict the segment membership of all respondents. 2. Holdout-Sample classification rates are the rates at which the segments can be correctly predicted when data from one portion of the sample – the Calibration Sample – are used to predict segment membership of the other portion of the sample, the Holdout Sample. Holdout Sample tests are tougher simply because we limit which data can be used to predict a given respondent’s segment. On the other hand, “tough” doesn’t always mean “good.” Any given portion of a single Holdout Sample may or may not be “representative” of the Calibration Sample, even if it was drawn randomly. Statisticians have developed several different approaches to minimize this potential problem, and the most commonly used are those under the heading of “Bootstrap Sampling.” Bootstrap sampling is a way to “sample from the sample” in a way that eventually produces o Holdout Samples that are representative of the entire sample (that is, of the Calibration Sample). There are two easy ways to do bootstrap sampling: o Sampling without replacement – draw several random samples from the original respondent base. o Do not replace these on each draw, but continue to draw until the respondent base is depleted. Then, project segment membership for each of the Holdout Samples separately. o The average classification rates are then used as metrics. o Sampling with replacement – identical to the previous approach, except the Holdout Sample members are put back into the respondent base for each successive draw. o The pluses and minuses of the “with replacement” option are o Technically,it is more valid statistically, since we are drawing from the same universe (respondent base) each time we pull a Holdout Sample. o The minus is that it is possible (even likely) that not all respondents will make it into any of the Holdout Samples with a finite number of draws. The best approach is to draw “without replacement.” This allows us to examine the predictive results for each Holdout Sample individually and to help pinpoint the causes if the correct classification rates for that sample are unusually low compared to others. Here we should note that the standard holdout classification approaches in SAS and SPSS are “one-held-out.” This is a sampling of each respondent in his/her own individual Holdout Sample, without replacement. 18 What size of Holdout Sample is the best? At a minimum, the SAS and SPSS default of one respondent at a time should be (and currently is) used. When the internal validity of the segments is in further question, for any reason, it is wise to conduct a 5% and/or 10%, boot-strapped Holdout Sample test without replacement. The greater the percentage, the tougher the test. 50% should be considered a maximum, though, based on sound statistical theory. 50% should never be considered a standard criterion. Statistical models used for predictive validation. If the segmentation solution was derived using K-Means or other cluster analysis without a latent class component, both In-Sample and Holdout-Sample classification rates are predicted using discriminant analysis or a multinomial logit model. (Note that these are not the same models as were used to identify or derive the segments initially, they are used just to test predictive validity once the segments have been derived.) o If the segmentation was derived using a latent class or other latent variable analysis, the In-Sample classification rates are computed from the posterior probabilities of membership in each of the latent classes. These are provided automatically in Latent Gold®. o Holdout-Sample classification rates are also based on the posterior probabilities of membership, but these need to be computed manually in SPSS or SAS. Note all classification tests, regardless of segment development model that was used, utilize the same Holdout Sample draws. Variables used to predict segment membership in a cross-validation test. So far, no mention has been made of what variables are used to predict segment membership in cross-validation tests. There are two different levels of validation at which specific variables need to be considered: 1. General validation. At a minimum, all the variables originally used to develop the segments are included. Other variables can be included on a judgmental basis. For our purposes below, we call this the “Full Variable Set.” 2. Short form validation. One of the key deliverables in most custom studies is a segment classification model, called a “short form.” This classifier is based on survey items that can be incorporated into future research studies (or even sales calls) to identify the segment membership of new contacts. In this case, only a handful of items can be used, given a limited amount of time to administer them live. Below, we call this limited number of items the “Short Form Variable Set.” Benchmarks for classification rates. Below are some benchmark metrics of predictive performance to support the Identifiability dimension of a segmentation solution. As can be seen, these depend to a large extent on the type of segmentation analysis to be conducted (see previous section for a description of the types seen in the table below). 19 Figure E. Benchmark Correct Classification Rates for Segmentation Models Full Variable Set K-Means/Cluster Analysis In-Sample Holdout-Sample Custom Research Studies 80% 75% Latent Class Analysis In-Sample Holdout-Sample 90% 85% Short Form Variable Set K-Means/Cluster Analysis In-Sample Holdout-Sample 75% 60% Latent Class Analysis In-Sample Holdout-Sample 70% 55% New Segmentation Analysis using Database Variables K-Means/Cluster Analysis In-Sample Holdout-Sample Database Marketing Models 75% 60% Latent Class Analysis In-Sample Holdout-Sample 80% 75% Linkage of Database Variables to Existing Segments (assuming more than 3 or 4 segments) Customer database + Syndicated Syndicated Variables Only Variables In-Sample Holdout-Sample In-Sample Holdout-Sample 75% 60% 60% 45% The classification rates in Figure E are meant to be average, rather than minimum standards. However, average doesn’t necessarily mean “satisfactory,” and we need to work vigorously on improving these standards over time. It can be seen from Figure E that, in general, segments derived using a latent class analysis should have higher correct classification rates in general, although the advantage is higher for custom research studies than for database models. It should also be noted that latent class models are not typically available for database modeling due to the large size of databases and the limitations of processing power in even fast PCs. The one exception to this is the application of latent class analysis to household grocery panel data. However, those models typically do not use more than a handful of variables, compared to hundreds of variables used in the usual database modeling project. Why, then, do we see less capability to predict latent class segments in the short form application? The reason is that the short form, utilizing only a small number of survey items, cannot fully replicate the underlying latent classes that are identified in the original model used to derive the segments. Instead, the short form predictive model assumes that the segments are based just on means of observable (i.e., non-latent) variables. More work needs to be done in this area, since the short form is considered a vital deliverable. 20 One alternative is to replace the typical discriminant model in short forms with something that approximates the latent class structure. Three general conclusions can be drawn from Figure E: o First, latent class segmentation solutions are usually more internally consistent (and hence, more projectable from a representative sample). o Second, although database models are typically less predictive than custom research models, relatively powerful models can be built for new segmentation schemes in the client’s database, as long as client data fields are included in predicting segment membership. o Third, the weakest predictive capability resides in linkage models in which preexisting segments are predicted solely by syndicated data fields (not customer data fields). Unfortunately, most database models tend to fall into this category. Because of the fact that database clients require a lot of linkage modeling, it may be wise to prioritize improving the predictive capability of those models. Based on decades of forecasting model development in academia and commercial research, here is a list of measures we have taken and hope to take in the near future: 1. We have traditionally used logistic regression model scores to identify the most appropriate segment into which to classify a record/respondent. We need not always use linear functions, however, because in SAS both quadratic (non-pooled) and nonparametric classification routines are available. These nonlinear methods tend to provide up to a 10-percentage-point improvement in classification rates. 2. Due to constraints on computer processing speed, we have not explored a number of variable transforms and derived variables which, using solely syndicated data, might provide a better fit to the data and better predictions of segments. These include principal component (factor) scores and polynomial (squared, cubed, Box-Cox transform, or logarithmic) expressions of the existing database variables. This produces our third and final challenge to the practice of market segmentation: Challenge #3: o Nonlinear models, variable transforms, and derived variables would most certainly increase the predictive strength of our linkage models. o How can these be implemented within a reasonable computing and network infrastructure? 21 Summary of Challenges to Market Segmentation, in Theory and Practice. Here are the challenges identified in the above discussion. I’ve abbreviated them to highlight the main points. These are, perhaps, goals to achieve over the next five years. Challenge #1 – How can market segmentation account for diversity within market segments, in ways that add valuable marketing opportunities? Challenge #2 – How can market segmentation better account for both observed and unobserved (or less easily observed) dimensions of the market? This includes especially changes in the market – or market segments – over time. Challenge #3 – how can we utilize nonlinear models, variable transforms, and derived variables to increase the predictive strength of our predictive – especially linkage – models? These are questions even our most advanced theoreticians in marketing science – Michel Wedel, Wagner Kamakura, and Greg Allenby – would have a hard time answering. Nevertheless, we should press forward with potential solutions. Perhaps in partnership with our clients, these questions can and will be aggressively addressed and resolved. 22