Return on Marketing

8/2/2019 Return on Marketing

1/20

Return on Marketing / 109Journal of Marketing

Vol. 68 (January 2004), 109127

Roland T. Rust, Katherine N. Lemon, & Valarie A. Zeithaml

Return on Marketing:Using Customer Equity to Focus

Marketing StrategyThe authors present a unified strategic framework that enables competing marketing strategy options to be tradedoff on the basis of projected financial return, which is operationalized as the change in a firms customer equity rel-ative to the incremental expenditure necessary to produce the change. The change in the firms customer equity isthe change in its current and future customers lifetime values, summed across all customers in the industry. Eachcustomers lifetime value results from the frequency of category purchases, average quantity of purchase, andbrand-switching patterns combined with the firms contribution margin. The brand-switching matrix can be esti-mated from either longitudinal panel data or cross-sectional survey data, using a logit choice model. Firms can ana-lyze drivers that have the greatest impact, compare the drivers performance with that of competitors drivers, andproject return on investment from improvements in the drivers. To demonstrate how the approach can be imple-mented in a specific corporate setting and to show the methods used to test and validate the model, the authorsillustrate a detailed application of the approach by using data from the airline industry. Their framework enableswhat-if evaluation of marketing return on investment, which can include such criteria as return on quality, return on

advertising, return on loyalty programs, and even return on corporate citizenship, given a particular shift in cus-tomer perceptions. This enables the firm to focus marketing efforts on strategic initiatives that generate the great-est return.

Roland T. Rust is David Bruce Smith Chair in Marketing, Director of theCenter for e-Service, and Chair of the Department of Marketing, Robert H.Smith School of Business, University of Maryland ([email protected]). Katherine N. Lemon is Associate Professor, Wallace E. CarrollSchool of Management, Boston College (e-mail: [email protected]).Valarie A. Zeithaml is Roy and Alice H. Richards Bicentennial Professorand Senior Associate Dean, Kenan-Flagler School of Business, Universityof North Carolina, Chapel Hill (e-mail: [email protected]). Thisresearch was supported by the Marketing Science Institute, University ofMarylands Center for e-Service, and the Center for Service Marketing atVanderbilt University. The authors thank Northscott Grounsell, RicardoErasso, and Harini Gokul for their help with data analysis, and they thankNevena Koukova, Samir Pathak, and Srikrishnan Venkatachari for theirhelp with background research.The authors are grateful for comments andsuggestions provided by executives from IBM, Sears, DuPont, GeneralMotors, Unilever, Siemens, Eli Lilly, R-Cubed, and Copernicus. They alsothank Kevin Clancy, Don Lehmann, Sajeev Varki, Jonathan Lee, Dennis

Gensch, Wagner Kamakura, Eric Paquette, Annie Takeuchi, and seminarparticipants at Harvard Business School, INSEAD, London BusinessSchool, University of Maryland, Cornell University, Tulane University, Uni-versity of Pittsburgh, Emory University, University of Stockholm, Norwe-gian School of Management, University of California at Davis, and Mon-terrey Tech; and they thank participants in the following: AmericanMarketing Association (AMA) Frontiers in Services Conference, MSI Cus-tomer Relationship Management Workshop, MSI Marketing Metrics Work-shop, INFORMS Marketing Science Conference, AMA A/R/T Forum, AMAAdvanced School of Marketing Research, AMA Customer RelationshipManagement Leadership Program, CATSCE, and QUIS 7.

The Marketing Strategy Problem

Top managers are constantly faced with the problem ofhow to trade off competing strategic marketing initia-tives. For example, should the firm increase advertis-

ing, invest in a loyalty program, improve service quality, or

none of the above? Such high-level decisions are typicallyleft to the judgment of the chief marketing or chief executiveofficers, but these executives frequently have little to basetheir decisions on other than their own experience and intu-ition. A unified, data-driven basis for making broad, strate-gic marketing trade-offs has not been available. In this arti-cle, we propose that trade-offs be made on the basis ofprojected financial impact, and we provide a framework thattop managers can use to do this.

Financial Accountability

Although techniques exist for evaluating the financial returnfrom particular marketing expenditures (e.g., advertising,direct mailings, sales promotion) given a longitudinal his-tory of expenditures (for a review, see Berger et al. 2002),the approaches have not produced a practical, high-levelmodel that can be used to trade off marketing strategies ingeneral. Furthermore, the requirement of a lengthy historyof longitudinal data has made the application of return oninvestment (ROI) models fairly rare in marketing. As aresult, top management has too often viewed marketingexpenditures as short-term costs rather than long-terminvestments and as financially unaccountable (Schultz andGronstedt 1997). Leading marketing companies considerthis problem so important that the Marketing Science Insti-tute has established its highest priority for 20022004 asAssessing Marketing Productivity (Return on Marketing)and Marketing Metrics. We propose that firms achieve thisfinancial accountability by considering the effect of strategicmarketing expenditures on their customer equity and byrelating the improvement in customer equity to the expendi-ture required to achieve it.


2/20

110 / Journal of Marketing, January 2004

1For expositional simplicity, we assume throughout much of thearticle that the firm has one brand and one market, and therefore weuse the terms firm and brand interchangeably. In many firms,the firms customer equity may result from sales of several brandsand/or several distinct goods or services.

Customer Equity

Although the marketing concept has reflected a customer-centered viewpoint since the 1960s (e.g., Kotler 1967), mar-keting theory and practice have become increasinglycustomer-centered during the past 40 years (Vavra 1997, pp.68). For example, marketing has decreased its emphasis onshort-term transactions and has increased its focus on long-term customer relationships (e.g., Hkansson 1982; Stor-backa 1994). The customer-centered viewpoint is reflected

in the concepts and metrics that drive marketing manage-ment, including such metrics as customer satisfaction(Oliver 1980), market orientation (Narver and Slater 1990),and customer value (Bolton and Drew 1991). In recentyears, customer lifetime value (CLV) and its implicationshave received increasing attention (Berger and Nasr 1998;Mulhern 1999; Reinartz and Kumar 2000). For example,brand equity, a fundamentally product-centered concept, hasbeen challenged by the customer-centered concept of cus-tomer equity (Blattberg and Deighton 1996; Blattberg, Getzand Thomas 2001; Rust, Zeithaml, and Lemon 2000). Forthe purposes of this article, and largely consistent with Blat-tberg and Deighton (1996) but also given the possibility ofnew customers (Hogan, Lemon, and Libai 2002), we definecustomer equity as the total of the discounted lifetime valuessummed over all of the firms current and potentialcustomers.1

Our definition suggests that customers and customerequity are more central to many firms than brands and brandequity are, though current management practices and met-rics do not yet fully reflect this shift. The shift from product-centered thinking to customer-centered thinking implies theneed for an accompanying shift from product-based strategyto customer-based strategy (Gale 1994; Kordupleski, Rust,and Zahorik 1993). In other words, a firms strategic oppor-tunities might be best viewed in terms of the firms opportu-nity to improve the drivers of its customer equity.

Contribution of the Article

Because our article incorporates elements from several liter-ature streams within the marketing literature, it is useful topoint out the relative contribution of the article. Table 1shows the contribution of this article with respect to severalstreams of literature that influenced the return on marketingconceptual framework. Table 1 shows related influential lit-erature streams and exemplars of the stream, and it high-lights key features that differentiate the current effort fromprevious work. For example, strategic portfolio models, asLarrech and Srinivasan (1982) exemplify, consider strate-gic trade-offs of any potential marketing expenditures. How-

ever, the models do not project ROI from specific expendi-tures, do not model competition, and do not model thebehavior of individual customers, their customer-level brandswitching, or their lifetime value. Our model adds to the

strategic portfolio literature by incorporating thoseelements.

Three related streams of literature involve CLV models(Berger and Nasr 1998), direct marketingmotivated modelsof customer equity (e.g., Blattberg and Deighton 1996; Blat-tberg, Getz, and Thomas 2001), and longitudinal databasemarketing models (e.g., Bolton, Lemon, and Verhoef 2004;Reinartz and Kumar 2000). Our CLV model builds on theseapproaches. However, the preceding models are restricted tocompanies in which a longitudinal customer database exists

that contains marketing efforts that target each customer andthe associated customer responses. Unless the longitudinaldatabase involves panel data across several competitors, nocompetitive effects can be modeled. Our model is more gen-eral in that it does not require the existence of a longitudinaldatabase, and it can consider any marketing expenditure, notonly expenditures that are targeted one-to-one. We alsomodel competition and incorporate purchases from com-petitors (or brand switching), in contrast to most existingmodels from the direct marketing tradition.

The financial-impact element of our model is foreshad-owed by two related literature streams. The service profitchain (e.g., Heskett et al. 1994; Kamakura et al. 2002) and

return on quality (Rust, Zahorik, and Keiningham 1994,1995) models both involve impact chains that relate servicequality to customer retention and profitability. The return onquality models go a step farther and explicitly project finan-cial return from prospective service improvements. Follow-ing both literature streams, we also incorporate a chain ofeffects that leads to financial impact. As does the return onquality model, our model projects ROI. Unlike other mod-els, our model facilitates strategic trade-offs of any prospec-tive marketing expenditures (not only service improve-ments). We explicitly model the effect of competitionanelement that does not appear in the service profit chain orreturn on quality models. Also different from prior research,

our approach models customer utility, brand switching, andlifetime value.Finally, we compare the current article with a recent

book on customer equity (Rust, Zeithaml, and Lemon 2000)that focuses on broad managerial issues related to customerequity, such as building a managerial framework related tovalue equity, brand equity, and relationship equity. The bookincludes only one equation (which is inconsistent with themodels in this article). Our article is a necessary comple-ment to the book, providing the statistical and implementa-tion details necessary to implement the books customerequity framework in practice. The current work extends thebooks CLV conceptualization in two important ways: Itallows for heterogeneous interpurchase times, and it incor-porates customer-specific brand-switching matrices. In sum-mary, the current article has incorporated many influences,but it makes a unique contribution to the literature.

Overview of the Article

In the next section, on the basis of a new model of CLV, wedescribe how marketing actions link to customer equity andfinancial return. The following section describes issues inthe implementation of our framework, including dataoptions, model input, and model estimation. We then present


3/20

Return on Marketing / 111

TABLE1

ComparingtheReturnonMarketingModelwithExistingMarketingModels

Strategic

Brand

Trade-

ROI

CanBe

NetPresent

Switching

offsofAny

Modeled

Explicitly

Appliedto

Valueof

Modeledat

Typeof

Marketing

and

Models

Calculation

Most

Revenuesand

Customer

Statistical

Model

Exempla

rs

Expenditures

Calculated?

Competition?

ofCLV?

Industries?

Costs?

Level?

Details?

Strategic

Larrechand

Yes

No

No

No

Yes

Yes

No

Yes

portfolio

Srinivasan

(1982)

CLV

Bergera

nd

No

No

No

Yes

No

Yes

No

Yes

Nasr(1998)

Direct

Blattbergand

No

Yes

No

Yes

Yes

Yes

No

Yes

marketing:

Deighto

n

customer

(1996);

equity

Blattberg,

Getz,an

d

Thomas(2

001)

Longitudinal

Bolton,

Lemo

n,and

Yes

Yes

No,unless

Yes

No

Yes

No,unless

Yes

database

Verhoef(2004);

paneldata

paneldata

marketing

Reinartzand

Kumar(20

00)

Serviceprofit

Heskettetal.

No

No

No

No

No

No

No

Yes

chain

(1994);

Kamakuraet

al.(2002)

Returnon

Rust,Zahorik,

No

Yes

No

No

Yes

Yes

No

Yes

quality

and

Keiningham

(1994,1

995)

Customer

Rust,Zeith

aml,

Yes

Yes

Yes

Yes

Yes

Yes

No

No

equitybook

andLem

on

(2000)

Returnon

Currentpa

per

Yes

Yes

Yes

Yes

Yes

Yes

Yes

Yes

marketing


4/20


an example application to the airline industry, showing someof the details that arise in application, in testing and validat-ing our choice model, and in providing some substantiveobservations. We end with discussion and conclusions.

Linking Marketing Actions toFinancial Return

Conceptual Model

Figure 1 shows a broad overview of the conceptual modelthat we used to evaluate return on marketing. Marketing isviewed as an investment (Srivastava, Shervani, and Fahey1998) that produces an improvement in a driver of customerequity (for simplicity of exposition, we refer to an improve-ment in only one driver, but our model also accommodatessimultaneous improvement in multiple drivers). This leadsto improved customer perceptions (Simester et al. 2000),which result in increased customer attraction and retention(Danaher and Rust 1996). Better attraction and retentionlead to increased CLV (Berger and Nasr 1998) and customerequity (Blattberg and Deighton 1996). The increase in cus-tomer equity, when considered in relation to the cost of mar-

keting investment, results in a return on marketing invest-ment. Central to our model is a new CLV model that incor-porates brand switching.

Brand Switching and CLV

It has long been known that the consideration of competingbrands is a central element of brand choice (Guadagni andLittle 1983). Therefore, we begin with the assumption thatcompetition has an impact on each customers purchase

decisions, and we explicitly consider the relationshipbetween the focal brand and competitors brands. In con-trast, most, if not all, CLV models address the effects ofmarketing actions without considering competing brands.This is because data that are typically available to directmarketers rarely include information about the sales or pref-erence for competing brands. Our approach incorporatesinformation about not only the focal brand but competingbrands as well, which enables us to create a model that con-tains both customer attraction and retention in the context ofbrand switching. The approach considers customer flowsfrom one competitor to another, which is analogous tobrand-switching models in consumer packaged goods (e.g.,Massy, Montgomery, and Morrison 1970) and migrationmodels (Dwyer 1997). The advantage of the approach is thatcompetitive effects can be modeled, thereby yielding a fullerand truer accounting of CLV and customer equity.

When are customers gone? Customer retention histori-cally has been treated according to two assumptions (Jack-son 1985). First, the lost for good assumption uses thecustomers retention probability (often the retention rate inthe customers segment) as the probability that a firms cus-tomer in one period is still the firms customer in the fol-lowing period. Because the retention probability is typicallyless than one, the probability that the customer is retaineddeclines over time. The implicit assumption is that cus-

tomers are alive until they die, after which they are lostfor good. Models for estimating the number of active cus-tomers have been proposed for relationship marketing(Schmittlein, Morrison, and Columbo 1987), customerretention (Bolton 1998), and CLV (Reinartz 1999).

The second assumption is the always a share assump-tion, in which customers may not give any firm all of theirbusiness. Attempts have been made to model this by amigration model (Berger and Nasr 1998; Dwyer 1997).The migration model assigns a retention probability as pre-viously, but if the customer has missed a period, a lowerprobability is assigned to indicate the possibility that thecustomer may return. Likewise, if the customer has been

gone for two periods, an even lower probability is assigned.This is an incomplete model of switching because itincludes purchases from only one firm.

In one scenario (consistent with the lost-for-goodassumption) when the customer is gone, he or she is gone.This approach systematically understates CLV to the extentthat it is possible for customers to return. In another scenario(consistent with the migration model), the customer mayleave and return. In this scenario, customers may be eitherserially monogamous or polygamous (Dowling and Uncles

FIGURE 1Return on Marketing

Marketing investment

Driver improvement(s)

Improvedcustomer perceptions

Increasedcustomerattraction

Increasedcustomerretention

Increased CLV

Increasedcustomer equity

Cost ofmarketinginvestment

Return on marketing investment


5/20


2It is also possible to model the share-of-wallet scenario that iscommon to business-to-business applications by using the conceptof fuzzy logic (e.g., Varki, Cooil, and Rust 2000; Wedel andSteenkamp 1989, 1991).

3The Pfeifer and Carraway (2000) Markov model considers onlyone brand and does not capture brand switching. Its states pertainto recency rather than brand.

1997), and their degrees of loyalty may vary or even change.We can model the second (more realistic) scenario using aMarkov switching-matrix approach.2

Acquisition and retention. Note that the brand-switchingmatrix models both the acquisition and the retention of cus-tomers. Acquisition is modeled by the flows from otherfirms to the focal firm, and retention is modeled by the diag-onal element associated with the focal firm. The retentionprobability for a particular customer is the focal firms diag-

onal element, as a proportion of the sum of the probabilitiesin the focal firms row of the switching matrix. Note that thisimplies a different retention rate for each customer firmcombination (we show the details of this in a subsequentsection). This describes the acquisition of customers whoare already in the market. In growing markets, it is alsoimportant to model the acquisition of customers who arenew to the market.

The switching matrix and lifetime value. We propose ageneral approach that uses a Markov switching matrix tomodel customer retention, defection, and possible return.Markov matrices have been widely used for many years tomodel brand-switching behavior (e.g., Kalwani and Morri-

son 1977) and have recently been proposed for modelingcustomer relationships (Pfeifer and Carraway 2000; Rust,Zeithaml, and Lemon 2000). In such a model, the customerhas a probability of being retained by the brand in the sub-sequent period or purchase occasion. This probability is theretention probability, as is already widely used in CLV mod-els. The Markov matrix includes retention probabilities forall brands and models the customers probability of switch-ing from any brand to any other brand.3 This is the featurethat permits customers to leave and then return, perhapsrepeatedly. In general, this returning is confused with ini-tial acquisition in other customer equity and CLVapproaches. The Markov matrix is a generalization of the

migration model and is expanded to include the perspectiveof multiple brands.To understand how the switching matrix relates to CLV,

consider a simplified example. Suppose that a particularcustomer (whom we call George) buys once per month,on average, and purchases an average of $20 per purchase inthe product category (with a contribution of $10). Supposethat George most recently bought from Brand A. Supposethat Georges switching matrix is such that 70% of the timehe will rebuy Brand A, given that he bought Brand A lasttime, and 30% of the time he will buy Brand B. Suppose thatwhenever George last bought Brand B he has a 50% chanceof buying Brand A the next time and a 50% chance of buy-

4Actually, Georges CLV also depends on word-of-mouth effects(Anderson 1998; Hogan, Lemon, and Libai 2000), because Georgemay make recommendations to others that increase Georges valueto the firm. To the extent that positive word of mouth occurs, ourCLV estimates will be too low. Similarly, negative word of mouthwill make our estimates too high. Although these two effects, beingof the opposite sign, tend to cancel out to some extent, there willbe some unknown degree of bias due to word of mouth. However,word-of-mouth effects are notoriously difficult to measure on apractical basis.

ing Brand B. This is enough information for us to calculateGeorges lifetime value to both Brand A and Brand B.4

Consider Georges next purchase. We know that he mostrecently bought Brand A; thus, the probability of him pur-chasing Brand A in the next purchase is .7 and the probabil-ity of him purchasing Brand B is .3. To obtain the probabil-ities for Georges next purchase, we simply multiply thevector that comprises the probabilities by the switchingmatrix. The probability of purchasing Brand A becomes(.7 .7) + (.3 .5) = .64, and the probability of purchasingBrand B becomes (.7 .3) + (.3 .5) = .36. We can calcu-late the probabilities of purchase for Brand A and Brand Bas many purchases out as we choose by successive multipli-cation by the switching matrix. Multiplying this by the con-tribution per purchase yields Georges expected contributionto each brand for each future purchase. Because future pur-chases are worth less than current ones, we apply a discountfactor to the expected contributions. The summation of theseacross all purchase occasions (to infinity or, more likely, toa finite time horizon) yields Georges CLV for each firm.Note that if there are regular relationship maintenanceexpenditures, they need to be discounted separately and sub-tracted from the CLV.

The bridge of actionability. We assume that the firm canidentify expenditure categories, or drivers (e.g., advertisingawareness, service quality, price, loyalty program) thatinfluence consumer decision making and that compete formarketing resources in the firm. We also assume that man-agement wants to trade off the drivers to make decisionsabout which strategic investments yield the greatest return(Johnson and Gustafsson 2000). The drivers that are projected to yield the highest return receive higher levels ofinvestment. Connecting the drivers to customer perceptionsis essential to quantify the effects of marketing actions at theindividual customer level. Therefore, it is necessary to havecustomer ratings (analogous to customer satisfaction rat-

ings) on the brands perceived performance on each driver.For example, Likert-scale items can be used to measure eachcompeting brands perceived performance on each driver;perceptions may vary across customers.

The firm may also want to assemble its drivers intobroader expenditure categories that reflect higher-levelresource allocation. We refer to these as strategic invest-ment categories. For example, a firm may combine all itsbrand-equity expenditures into a brand-equity strategicinvestment category, with the idea that the brand manager isresponsible for drivers such as brand image and brandawareness.


6/20


5To the extent that heterogeneity in the regression coefficientsexists, the state dependence effect will likely be overestimated(Degeratu 1999; Frank 1962). This would result in underestimationof the effects of the customer equity drivers, which means that theeffect of violation of this assumption would be to make the projec-tions of the model more conservative. However, it has been shownthat our approach to estimating the inertia effect performs betterthan other methods that have been proposed (Degeratu 2001).

Modeling the Switching Matrix

Thus, the modeling of CLV requires modeling of the switch-ing matrix for each individual customer. Using individual-level data from a cross-sectional sample of customers, com-bined with purchase (or purchase intention) data, we modeleach customers switching matrix and estimate model para-meters that enable the modeling of CLV at the individualcustomer level.

The utility model. In addition to the individual-specific

customer-equity driver ratings, we also include the effect ofbrand inertia, which has been shown to be a useful predic-tive factor in multinomial logit choice models (Guadagniand Little 1983). The utility formulation can be conceptual-ized as

(1) Utility = inertia + impact of drivers.

To make this more explicit, Uijk is the utility of brand kto individual i, who most recently purchased brand j. Thedummy variable LASTijk is equal to one if j = k and is equalto zero otherwise; X is a row vector of drivers. We thenmodel

(2) Uijk = 0kLASTijk + Xik1k + ijk,

where 0k is a logit regression coefficient corresponding toinertia, 1k is a column vector of logit regression coeffi-cients corresponding to the drivers, and ijk is a random errorterm that is assumed to have an extreme value (double expo-nential) distribution, as is standard in logit models. The coefficients can be modeled as either homogeneous or het-erogeneous.5 For the current exposition, we present thehomogeneous coefficient version of the model. In a subse-quent section, we build and test alternative versions of themodel that allow for heterogeneous coefficients. The modelcan also be estimated separately for different marketsegments.

The individual-level utilities result in individual-levelswitching matrices. Essentially, each row of the switchingmatrix makes a different assumption about the most recentbrand purchased, which results in different utilities for eachrow. That is, the first row assumes that the first brand wasbought most recently, the second row assumes that the sec-ond brand was bought most recently, and so on. The utilitiesin the different rows are different because the effect of iner-tia (and the effect of any variable that only manifests withrepeat purchase) is present only in repeat purchases.

Consistent with the multinomial logit model, the proba-bility of choice for individual i is modeled as

6To simplify the mathematics, we adopt the assumption that acustomers volume per purchase is exogenous. We leave the mod-eling of volume per purchase as a function of marketing effort as atopic for further research.

(3) Pijk* = Pr[individual i chooses brand k*, given that brand j

was most recently chosen] = exp(Uijk*)/

Thus, the individual-level utilities result in individual-levelswitching matrices, which result in an individual-level CLV.

Brand switching and customer equity. To make the CLVcalculation more specific, each customer i has an associatedJ J switching matrix, where J is the number of brands, withswitching probabilities p

ijk, indicating the probability that

customer i will choose brand k in the next purchase, condi-tional on having purchased brand j in the most recent pur-chase. The Markov switching matrix is denoted as Mi, andthe 1 J row vector Ai has as its elements the probabilitiesof purchase for customer is current transaction. (If longitu-dinal data are used, the Ai vector will include a one for thebrand next purchased and a zero for the other brands.)

For brand j, dj represents firm js discount rate, fi is cus-tomer is average purchase rate per unit time (e.g., three pur-chases per year), vijt is customer is expected purchase vol-ume in a purchase of brand j in purchase t,6 ijt is theexpected contribution margin per unit of firm j from cus-tomer i in purchase t, and Bit is a 1 J row vector with ele-ments Bijt as the probability that customer i buys brand j inpurchase t. The probability that customer i buys brand j inpurchase t is calculated by multiplying by the Markovmatrix t times:

(4) Bit = AiMit.

The lifetime value, CLVij, of customer i to brand j is

(5) CLVij = t/fivijtijtBijt,

where Tij is the number of purchases customer i is expectedto make before firm js time horizon, Hj (e.g., a typical timehorizon ranges from three to five years), and B

ijtis a firm-

specific element ofBit. Therefore, Ti = int[Hjfi], where int[.]refers to the integer part, and firm js customer equity, CEj,can be estimated as

(6) CEj = meani(CLVij) POP,

where meani(CLVij) is the average lifetime value for firm jscustomers i across the sample, and POP is the total numberof customers in the market across all brands. Note that theCLV of each individual customer in the sample is calculatedseparately, before the average is taken.

It is worth pointing out the subtle difference betweenEquation 5 and most lifetime value expressions, as in directmarketing. Previous lifetime value equations have summed

over time period, and the exponent on the discounting factorbecomes t. However, in our case, we are dealing with dis-tinct individuals with distinct interpurchase times (or equiv-alently, purchase frequencies fi). For this reason, we sum

( )10

+= djt

Tij

exp( ).U ijkk


7/20


7If standard marketing costs (e.g., retention promotional costs)are spent on a time basis (e.g., every three months), they may eitherbe discounted separately and subtracted from the net present valueor be assigned to particular purchases (e.g., if interpurchase time isthree months, and a standard mailing goes out every six months,the mailing cost could be subtracted on every other purchase).

8We should also note that the expression implies that the firstpurchase occurs immediately. Other assumptions are also possible.

over purchase instead of time period.7 The exponent t/fireflects that more discounting is appropriate for purchase t ifpurchasing is infrequent, because purchase t will occur fur-ther into the future. If fi = 1 (one purchase per period), it isclear that Equation 5 is equivalent to the standard CLVexpression. If fi > 1, the discounting per purchase becomesless than the discounting per period, to an extent that exactlyequals the correct discounting per period. For example, forfi = 2, the square root of the periods discounting occurs ateach purchase.8

We can also use the customer equity framework toderive an overall measure of the companys competitivestanding. Market share, historically used as a measure of acompanys overall competitive standing, can be misleadingbecause it considers only current sales. A company that hasbuilt the foundation for strong future profits is in better com-petitive position than a company that is sacrificing futureprofits for current sales, even if the two companies currentmarket shares are identical. With this in mind, we definecustomer equity share (CES, in Equation 7) as an alternativeto market share that takes CLV into account. We calculatecustomer equity share for each brand j as

ROI

Effect of changes. Ultimately, a firm wants to know thefinancial impact that will result from various marketingactions. This knowledge is essential if competing marketinginitiatives are to be evaluated on an even footing. A firm mayattempt to improve its customer equity by making improve-ments in the drivers, or it may drill down further to improvesubdrivers that influence the drivers (e.g., improving dimen-sions of ad awareness). This requires the measurement of

customer perceptions of the subdrivers about which the firmwanted to know more.A shift in a driver (e.g., increased ad awareness) pro-

duces an estimated shift in utility, which in turn produces anestimated shift in the conditional probabilities of choice(conditional on last brand purchased) and results in a revisedMarkov switching matrix. In turn, this results in animproved CLV (Equations 4 and 5). Summed across all cus-tomers, this results in improved customer equity (Equation6). We assume an equal shift (e.g., .1 rating points) for allcustomers, but this assumption can be relaxed if appropriate,because our underlying modeling framework does notrequire a constant shift across customers.

( ) / .7 CES CE CEj j kk=

Projecting financial impact. It is often possible to devotea strategic expenditure to improve a driver, but is that invest-ment likely to be profitable? Modern thinking in financesuggests that improved expenditures should be treated ascapital investments and viewed as profitable only if the ROIexceeds the cost of capital. Financial approaches based onthis idea are known by such names as economic value-added (Ehrbar 1998) or value-based management (Cope-land, Koller, and Murrin 1996). The increased interest ineconomic value-added approaches has attracted more atten-

tion to ROI approaches in marketing (Fellman 1999).The discounted expenditure stream is denoted as E, dis-

counted by the cost of capital, and CE is the improvementin customer equity that the expenditures produce. Then, ROIis calculated as

(8) ROI = (CE E)/E.

Operationally, the calculation can be accomplished by usinga spreadsheet program or a dedicated software package.Note, though, that even ifCE is negative, the ROI expres-sion still holds.

Implementation IssuesCross-Sectional Versus Longitudinal DataOur approach requires the collection of cross-sectional sur-vey data; the approach is similar in style and length to thatof a customer satisfaction survey. The survey collects cus-tomer ratings of each competing brand on each driver. Othernecessary customer information can be obtained either fromthe same survey or from longitudinal panel data, if it isavailable. The additional information collected about eachcustomer includes the brand purchased most recently, aver-age purchase frequency, and average volume per purchase.The logit model can be calibrated in two ways: (1) byobserving the next purchase (from either the panel data or a

follow-up survey) or (2) by using purchase intent as a proxyfor the probability of each brand being chosen in the nextpurchase.

Obtaining the Model Input

The implementation of our approach begins with managerinterviews and exploratory research to obtain informationabout the market in which the firm competes and informa-tion about the corporate environment in which strategicdecisions are made. From interviews with managers, weidentified competing firms and customer segments; chosedrivers that correspond to current or potential managementinitiatives; and obtained the size of the market (total number

of customers across all brands) and internal financial infor-mation, such as the discount rate and relevant time horizon.In addition, we estimated contribution margins for all com-petitors. If there was a predictable trend in gross margins forany firm in the industry, we also elicited that trend. Fromexploratory research, using both secondary sources andfocus group interviews, we identified additional drivers,which we reviewed with management to ensure that theywere managerially actionable items. On the basis of thecombined judgment of management and the researchers, we


8/20


reduced the set of drivers to a number that allows for a sur-vey of reasonable length. The drivers employed typicallyvary by industry.

Estimating Shifts in Customer Ratings

The calculation of ROI requires an estimate of the ratingshift that will be produced by a particular marketing expen-diture. For example, a firm may estimate that an advertisingcampaign will increase the ad awareness rating by .3 on afive-point scale. These estimates can be obtained in several

ways. If historical experience with similar expenditures isavailable, that experience can be used to approximate theratings shift. For example, many marketing consulting firmshave developed a knowledge base of the effects of market-ing programs on measurable indexes. Another way, analo-gous to the decision calculus approach (Blattberg andDeighton 1996; Little 1970), is to have the manager supplya judgment-based estimate. The manager may reflect uncer-tainty by supplying an optimistic and a pessimistic estimate.If the outcome was favorable for the optimistic estimate, butunfavorable for the pessimistic estimate, the outcome wouldbe considered sensitive to the rating shift estimate, indicat-ing the need for more information gathering. Another lim-

ited cost approach is to use simulated test markets (Clancy,Shulman, and Wolf 1994; Urban et al. 1997) to obtain a pre-liminary idea of market response. Finally, the marketingexpenditure can be implemented on a limited basis, usingactual test markets, and the observed rating shift can bemonitored (e.g., Rust et al. 1999; Simester et al. 2000).

Calibrating the Data

It is typical in many sampling plans to have respondentswith different sampling weights, wi, correcting for varia-tions in the probability of selection. We can use the samplingweights directly, in the usual way, to generate a sample-based estimate of market share, which we denote as

MSsample. If the sample is truly representative, MSsampleshould be equal to the actual market share, MStrue. To makethe sample more representative of actual purchase patterns,we can assign a new weight, wi,new = (MStrue/MSsample) wito each respondent, with market shares corresponding to thatrespondents most recently chosen brand, which will correctfor any sampling bias with respect to any brand. The impliedmarket share from the sample will then equal the actual mar-ket share.

If purchase intent rather than actual purchase data isused, the application must be done with some care. Previousresearchers have long noted that purchase-intention subjec-tive probabilities occasionally may be systematically biased(Lee, Hu, and Toh 2000; Pessemier et al. 1971; Silk andUrban 1978). We assume that the elicited purchase inten-tions, pij, of respondent i purchasing brand j in the next pur-chase need to be calibrated. In general, we assume that thereis a calibrated probability, p*ij, that captures the true proba-bility of the next purchase. These probabilities can be cali-brated in two possible ways. First, if it is possible to followup with each respondent to check on the next purchase, wecan find a multiplier Kj for each brand that best predictschoice. (We set Kj for the first brand arbitrarily to one, with-out loss of generality, to allow for uniqueness.) The Ks can

be quickly found using a numerical search. If p*ij is thestated probability of respondent i choosing brand j in thenext purchase, the calibrated choice probability is p*ij =Kjpij/kKkpik. Second, if checking the next purchase is notpossible, it is still possible to calibrate the purchase inten-tions by making an approximating assumption. Assumingthat the market share (as the percentage of customers whoprefer a brand) for each brand in the near future (includingeach respondents next purchase) is roughly constant, weemploy a numerical search to find the Kjs (again setting

K1 = 1) for which MStrue = meani(p*ij).

Model Estimation

Principal components regression. In this application, asin customer satisfaction measurement, multicollinearity isan issue that needs to be addressed (Peterson and Wilson1992). For this reason, we adopt an estimation approach thataddresses the multicollinearity issue. Principal componentsregression (Massy 1965) is an approach that combats multi-collinearity reasonably well (Frank and Friedman 1993), yetit can be implemented with standard statistical software.Principal components regression is a two-stage procedurethat is widely known and applied in statistics, econometrics,

and marketing (e.g., Freund and Wilson 1998; Hocking1996; Naik, Hagerty, and Tsai 2000; Press 1982). Principalcomponents multinomial logit regression has been used suc-cessfully in the marketing literature, leading to greateranalysis interpretability and coefficient stability (e.g., Gess-ner et al. 1988).

The idea is to reduce the dimensionality of the indepen-dent variables by extracting fewer principal components thatexplain a large percentage of the variation in the predictors.The principal components are then used as independent vari-ables in the regression analysis. Because the principal com-ponents are orthogonal, there is no multicollinearity issuewith respect to their effects. In addition, eliminating the

smallest principal components, which may be essentiallyrandom, may reduce the noise in the estimation. Because theprincipal components can be expressed as a linear combina-tion of the independent variables (and vice versa), the coef-ficients of the independent variables can be estimated as afunction of the coefficients of the principal components, andthe coefficients (after the least important principal compo-nents are discarded) may result in better estimates of the dri-vers effects. Estimation details are provided in Appendix A.

Importance of customer equity drivers. The results fromthe model estimation in Equation 2 provide insight intowhich customer equity drivers are most critical in the indus-try in which the firm competes. When examining a specificindustry, it is useful to know what the key success factors arein that industry. Ordinarily, this might be explored by calcu-lating market share elasticities for each driver. However, thatapproach is not correct here, because the drivers are inter-vally scaled rather than ratio scaled. This means that it isincorrect to calculate percentages of the drivers, as is neces-sary in the calculation of elasticities. Moreover, our focus iscustomer equity rather than market share. To arrive at theimpact of a driver on customer equity, we need to determinethe partial derivative of choice probability, with respect tothe driver, for each customer in the sample. That is, if a dri-


9/20


ver were improved by a particular amount, what would bethe impact on customer choice and, ultimately, on CLV andcustomer equity? Appendix A provides details of these com-putations and significance tests for the drivers.

An Example ApplicationData and Sampling

Survey items. We illustrate our approach with data col-lected from customers of five industries. We assume threestrategic investment categories: (1) perceived value (Para-suraman 1997; Zeithaml 1988), (2) brand equity (Aaker andKeller 1990), and (3) relationship management (Andersonand Narus 1990; Gummeson 1999). The three categoriesspan all major marketing expenditures (Rust, Zeithaml, andLemon 2000). We drew heavily on the relevant academicand managerial literature in these areas to build our list ofdrivers and ensured that the drivers could be translated intoactionable expenditures. The resulting survey containedquestions pertaining to shopping behavior and customer rat-ings of each driver for the four or five leading brands in themarkets we studied. In addition, several demographics ques-tions were asked at the end of the survey. We selected indus-

tries (airlines, electronics stores, facial tissues, grocery, andrental cars) that represented a broad set of consumer goodsand services. To save space, we present the details for theairline industry analysis only; however, our approach wassimilar across the other four industries. The complete list ofthe survey items used in our analysis of the airline industryappears in Appendix B.

Population. We obtained illustrative data from two com-munities in the northeastern United States: an affluent smalltown/suburb and a medium-sized city that adjoins a largercity. Respondents were real consumers who had purchasedthe product or service in the industry in question during theprevious year. Demographic statistics suggest that the sam-ple is representative of similar standard metropolitan statis-tical areas in the United States, with the exception of gener-ally high levels of education and income. For example, inthe small town (with a population of approximately 20,000),the average age of the respondent was 47, the averagehousehold had two adults and one child, the average house-hold income was $91,000, and the average years of educa-tion was 17. In the larger city, the average age was somewhatlower (39 years), the average household had two adults andone child, the average household income was $70,000, andthe average years of education was similar to that of thesmall town.

Sampling. We obtained respondents from three randomsamples. The first sample, drawn from the city population,answered questions about electronics stores and rental carcompanies. The second sample, also drawn from the citypopulation, addressed groceries and facial tissues. The thirdsample, drawn from the small town, focused on airlines.Potential respondents were contacted at random byrecruiters from a professional market research organization(by telephone solicitation or building intercept). The screen-ing process consisted of two criteria: (1) the respondent hadpurchased from the industry in the past 12 months, and (2)

9Mean substitution can result in biased estimates, but in ourjudgment, the additional effort of employing a more sophisticatedmissing values procedure (e.g., data imputation) was not justifiedin this case, given the relatively low percentage of missing values.

10This decision was based solely on the researchers best judg-ment. Our model does not require this.

the respondent had a household income of at least $20,000per year. Respondents agreed to participate and received $20compensation for completing the questionnaire. In the elec-tronics stores and rental cars survey, 246 consumers wereapproached: 153 were eligible, 144 cooperated, and 7 weredisqualified, resulting in a total of 137 total surveys com-pleted. In the groceries and facial tissues survey, 177 con-sumers were approached: 124 were eligible, 122 cooperated,and 4 were disqualified, resulting in a total of 118 surveyscompleted. In the airline survey, 229 consumers were

approached: 119 were eligible, 105 cooperated, and 5 weredisqualified, resulting in a total of 100 surveys completed.

Data collection and preliminary analysis. Data werecollected in December 1998 and January 1999 at the firmsoffices in each location. The respondents came to the facil-ity to complete the pencil-and-paper questionnaire, whichtook about 30 minutes. They were then thanked for their par-ticipation and compensated. In addition, we obtained aggre-gate statistics on the small town and city (e.g., percentage ofpopulation that uses rental cars, average spent at grocerystore) from secondary sources and used them in subsequentanalysis. For purposes of financial analysis, we used localpopulation and aggregate usage statistics for predominantly

local industries (electronics stores and groceries) andnational statistics for predominantly national industries (air-lines, facial tissues, and rental cars). Although our randomsamples may not be fully representative of U.S. users, weextrapolated to the national market for national industries toshow the type of dollar magnitudes that can arise given alarge population. Because our examples are illustrative,truly precise dollar estimates are unnecessary.

Data were cleaned to eliminate obvious bad cases andextreme outliers. Because listwise deletion of cases wouldhave resulted in too many cases being removed (even thoughonly a relatively small percentage of responses were missingfor particular items), we employed mean substitution as our

missing data option for all subsequent analyses.9 Becausewe suspected that the relationship drivers would affect pri-marily repeat purchasers, we collected relationship itemsonly for the brand most recently purchased.10 We mean-centered the relationship-related drivers for the cases inwhich the brand considered was the previously purchasedbrand, and we set them equal to zero for the cases in whichthe brand considered was different from the previously pur-chased brand. This enabled the pure inertia effect to beseparated from the relationship effect of the drivers.

Choice Model Results

Principal components analysis results. We reduced the

dimensionality of the predictor variables in each industry byconducting a principal components analysis. We used aneigenvalue cutoff of .5, which we judged to provide the best


10/20


11The 1.0 eigenvalue cutoff (Kaiser 1960) is typically employedin marketing, but it is just one of many possible cutoff criteria (fortwo alternatives, see Cattell 1966; Jolliffe 1972). As Kaiser (1960,p. 143), who proposed the 1.0 cutoff, points out, by far [the] mostimportant viewpoint for choosing the number of factors [is] psy-chological meaningfulness. In other words, the cutoff should bechosen such that the results are substantively meaningful, which isour justification for using the particular cutoff level that we chose.

trade-off between parsimony and managerial usefulness.11The airline analysis began with 17 independent variables,and we retained 11 orthogonal factors. Table 2 shows theloadings on the rotated factors. The resulting factor structureis rich. All the factors are easily interpretable. The few neg-ative loadings are small and insignificant; they are zero forall practical purposes. All drivers load on only one factor,and many (e.g., inertia, quality, price, convenience, trust,corporate citizenship) load on their own unique factor.

There is some degree of discrimination among the value,

brand, and relationship strategic investment categories inthat drivers in the three strategic action categories of differ-ent drivers do not correlate highly on the same factors. Aswe expected, the strategic investment categories, value,brand, and relationship are not unidimensional. The driversthat constitute the categories can be grouped for managerialpurposes as managers consider them, but drivers in a partic-ular strategic investment category may be quite distinct inthe customers mind.

Logit regression results. Using the resulting factors asindependent variables, we conducted multinomial logitanalyses, using the analysis we described previously. Table3 shows the coefficients that arise from the multinomial

logit regression analysis, highlighting the significant factors.Using Equations A1A9, we converted the factor-levelresults to the individual drivers. The resulting coefficients,standard errors, and test statistics are shown in Table 4. All

12This nested chi-square was also insignificant in the other fourindustries that we studied (electronics = 6.79, facial tissues = 1.00,grocery = 5.82, and rental cars = .28).

the drivers are significant and have the correct sign, butsome drivers have a larger effect than others. The mostimportant drivers span all three strategic investment cate-gories. In addition to the drivers, inertia has a large, signifi-cant impact (.849,p < .01). Among the value-related drivers,convenience has the largest coefficient (.609), followed byquality (.441); for brand-related drivers, direct mail infor-mation has the largest impact (.638), followed by ad aware-ness (.421) and ethical standards (.421). The loyalty pro-gram (.295) and preferential treatment (.280) are the most

important relationship-related drivers.

Model Testing and Validation

We tested and validated the core choice model in several dif-ferent ways. We tested for brand-specific effects, hetero-geneity of response, a more general covariance matrix, andthe reliability of the coefficient estimates.

Brand-specific effects. The model in Equation 2 assumesthat there are no brand-specific effects. We tested the valid-ity of this assumption by including brand-specific constantsin the model of Equation 2. Testing the significance of themore complicated model can be accomplished through theuse of a nested-likelihood-ratio chi-square test (in the airlineapplication, this involves three degrees of freedom, reflect-ing a number of brand-specific constants that is equal to thenumber of brands minus one). The resulting nested modelcomparison was not significant (23 = .977), from which weconclude that brand-specific constants are not required.12

Heterogeneity of response. It is reasonable to suspectthat there may be unobserved heterogeneity of response

TABLE 2Factor Loadings: Airline Industry

Driver F1 F2 F3 F4 F5 F6 F7 F8 F9 F10 F11

Inertia .013 .004 .033 .038 .015 .116 .024 .984 .029 .043 .002

Quality .097 .058 .174 .076 .147 .212 .014 .049 .068 .904 .083Price .044 .007 .128 .054 .078 .023 .039 .030 .975 .059 .034Convenience .078 .068 .219 .161 .043 .830 .066 .163 .018 .260 .015

Ad awareness .031 .130 .038 .938 .010 .022 .048 .011 .074 .101 .058Information .216 .077 .248 .656 .322 .299 .207 .125 .038 .058 .016Corporate citizenship .011 .122 .150 .093 .880 .001 .256 .021 .077 .137 .006Community events .021 .100 .188 .042 .226 .051 .921 .026 .041 .011 .029Ethical standards .016 .044 .605 .105 .458 .266 .028 .034 .104 .109 .218

Image fits my personality .098 .112 .878 .107 .069 .092 .203 .058 .110 .142 .081

Investment in loyalty program .921 .044 .090 .032 .007 .060 .018 .014 .003 .137 .103Preferential treatment .898 .087 .082 .002 .022 .071 .029 .032 .007 .077 .104Know airlines procedures .708 .232 .022 .116 .029 .166 .058 .033 .010 .069 .240Airline knows me .681 .309 .073 .059 .001 .356 .012 .061 .136 .075 .219Recognizes me as special .214 .851 .069 .077 .036 .042 .138 .004 .044 .118 .092Community .175 .876 .065 .031 .166 .035 .015 .001 .036 .042 .129Trust .246 .227 .179 .069 .041 .003 .031 .006 .038 .091 .889

Notes: Loadings greater than .5 are shown in bold.


11/20


TABLE 3

Logit Regression Results: Airline Industry

Independent Variable Coefficient Standard Error b/s.e. p

F1 .325** .122 2.65 .008F2 .031 .118 .26 .795F3 .421* .210 2.00 .045F4 .459 .285 1.61 .107F5 .212 .228 .93 .352F6 .331* .159 2.08 .037

F7 .081 .279 .29 .771F8 .633** .100 6.36 .000F9 .034 .164 .21 .835F10 .184 .147 1.26 .209F11 .033 .115 .29 .772

Log-likelihood = 98.46Chi-square (11 degrees of freedom) = 69.246**

*p< .05.**p< .01.

TABLE 4

Driver Coefficients: Airline Industry

Driver Coefficient Standard Error b/s.e.

Inertia .849 .075 11.341*

Quality .441 .041 10.871*Price .199 .020 9.858*Convenience .609 .093 6.553*

Ad awareness .421 .099 4.242*Information .638 .082 7.819*Corporate citizenship .340 .045 7.617*Community events .170 .024 6.974*Ethical standards .421 .053 7.901*

Image fits my personality .390 .050 7.874*

Investment in loyalty program .295 .027 10.956*Preferential treatment .280 .026 10.857*Know airlines procedures .238 .027 8.779*Airline knows me .249 .041 6.108*Recognizes me as special .167 .017 9.771*Community .151 .016 9.412*Trust .203 .023 8.725*

*p< .01.

across the respondents. That is, expressed in terms of Equa-

tion 2, the s may be different across respondents. To testthis, we employed a random coefficients logit model (Chin-tagunta, Jain, and Vilcassim 1991) in which we permittedthe driver coefficients to be distributed as an independentmultivariate normal distribution. The log-likelihoodimproved from 97.58 to 93.24, which is an insignificantimprovement (211 = .8.68). Therefore, we conclude that therandom coefficients logit formulation does not produce abetter model and that it is not worthwhile in this case tomodel unobserved heterogeneity in the parameters.

Correlated errors. Another way the independence from

irrelevant alternatives property can be violated is if the errorterms in Equation 2 are correlated. For example, it is possi-ble that people who prefer American Airlines more than themodel predicted will systematically dislike Southwest Air-lines more than the model predicted. To address this issue,we turned to a multinomial probit model (Chintagunta1992). In this model, the error terms in Equation 2 areassumed to be normally distributed rather than extremevalue, and they are permitted to have a general covariancematrix.


12/20


FIGURE 2

Distribution of CLV: American Airlines

0%

10%

20%

30%

40%

50%

60%

70%

Relativ

eFrequency

0 9 9 1 00 1 99 2 00 2 99 3 00 3 99 4 0 0 49 9 5 00 +

CLV ($)

Our original logit model is no longer a constrained ver-sion of the more complicated model, so we cannot do thenested likelihood-ratio chi-square test. However, by com-paring the general multinomial probit model with a multin-omial probit model in which the error terms are assumed tobe independent, we can address the issue of whether model-ing the more general covariance matrix is useful. Thisresults in a nested test. We found that the uncorrelated errorsversion of the model resulted in a log-likelihood of 98.46(slightly worse than the multinomial logit log-likelihood)

and that the more general model produced a log-likelihoodof 97.57. The improvement is insignificant (23 = .82). Wealso can compare the general multinomial probit model withthe original multinomial logit model by using the Akaikeinformation criterion. The improvement from 97.58 to97.57 does not compensate for the additional three esti-mated parameters (we would need a log-likelihoodimprovement of at least 3.0), suggesting that the generalmultinomial probit model is not better than our multinomiallogit model. Thus, we conclude that the more generalcovariance matrix is not warranted.

Coefficient reliability. Given our relatively small samplesize (96 usable data points after the data are cleaned), wewere unable to pursue split-half tests or complete holdoutsamples. However, to further understand the reliability ofour model estimates, we randomly split our sample intothree parts (A, B, and C) and estimated our model on AB,AC, and BC. The mean range (and median range) of thecoefficient estimates across the 11 factors was .14; that is, onaverage, the swing between the largest and smallest coeffi-cient estimate across the three samples was .14, whichcomes out to about .6 standard errors, on average. Thus, themodel appears to produce reasonably stable coefficientestimates.

CLV

Using Equation 5, we calculated CLV for American Airlinesfor each respondent in our airline sample. To operationalizethe equation, we assumed a time horizon of three years, adiscount rate of 10%, and a contribution margin of 15%. The15% figure was approximately equal to the average operat-ing margin for the industry for the five years preceding thesurvey, according to annual reports of the four firms westudied (since our study, airline industry operating marginshave declined). We also based our contribution margin fig-ures in the other four industries on financial data fromannual reports. To extend the CLV figures to the firms U.S.customer equity, we used U.S. Census data to determine thenumber of adults in the United States (187,747,000), and we

then combined this with the percentage of U.S. adults whowere active users of airline travel (23.3%), yielding a totalnumber of U.S. adult airline customers of 43,745,051. Toapproximate the total customer equity, we multiplied thisnumber by the average CLV across our respondents. Notethat though we used average CLV to project customerequity, we calculated CLV at the individual customer levelfor each customer in the sample.

Customer loyalty and CLV. Some insights can beobtained from examining American Airlines CLV distribu-

tion. For example, Figure 2 shows the distribution of CLVacross American Airlines customers. The $0$99 categoryincludes more than 60% of Americans customers, and the$500-plus category includes only 11.6% of customers, indi-

cating that the bulk of Americans customers have low CLV.Figure 3 also indicates that Americans customers are fickle.Almost half of Americans customers have a 20% or lessshare-of-wallet (by CLV) allocated to American. Only10.5% give more than 80% of their CLV to American. Thispercentage shows dramatically that the vast majority ofAmericans customers cannot be considered monogamouslyloyal. Figure 4 shows a startling picture of the percentage ofAmericans customer equity that is contributed by each CLVcategory. The $0$99 category, though by far the largest(more than 60% of Americans customers), produces lessthan 10% of Americans customer equity. In contrast, the$500-plus CLV category, though only 11.6% of Americans

customers, produces approximately 50% of Americans cus-tomer equity.

Comparison with the lost-for-good CLV model. Previ-ously, we proposed that some models of CLV that do notaccount for customers returning systematically underesti-

FIGURE 3

Distribution of CLV Share (Share of Wallet):American Airlines

0%

10%

20%

30%

40%

50%

RelativeFrequency

0%20% 21%40% 41%60% 61%80% 81%100%

CLV


13/20


0

10

20

30

40

50

Percentage(%)of

CustomerEquity

0 9 9 1 00 1 99 2 00 2 99 3 00 3 99 4 00 4 99 5 00 +

CLV ($)

FIGURE 4

Percentage Customer Equity by CLV Category:American Airlines

mate CLV and customer equity (for an exception, see Dwyer1997). Using the airline sample, we explored the degree to

which this was true. The lost-for-good model is simply aconstrained version of our switching model, such that allprobabilities of switching from another brand to the focalbrand are zero. In other words, to calculate the results, weconsidered only the customers who were retained from thefirst purchase. When the customer chose another brand, wegave a probability of zero to any further purchase from thefocal brand. For American Airlines, our brand-switchingmodel gives a customer equity of $7.303 billion. Withoutaccounting for switching back, the estimated customerequity declines to $3.849 billion. Thus, the lost-for-goodmodel provides a systematic underestimation of customerequity that, in this case, is an underestimation of 47.3%.

Customer equity and the value of the firm. It has beensuggested (Gupta, Lehmann, and Stuart 2001) that customerequity is a reasonable proxy for the value of the firm. Ouranalysis of American Airlines provides some support for this

idea. Multiplication of Americans average CLV ($166.94)by the number of U.S. airline passengers (43,745,051)yields a total customer equity for American of $7.3 billion.Given Americans opening share price for 1999 ($60) and itsnumber of shares outstanding at that time (161,300,000)(AMR Corporation 1999), we calculate a market capitaliza-tion of $9.7 billion. Because our projection ignores profitsfrom international customers and nonflight sources ofincome, our customer equity calculation is largely compati-ble with Americans market capitalization at the time of the

survey.

Projected Financial Return

Our framework enables the financial impact of improvementefforts to be analyzed for any of the usual marketing expen-ditures. For example, American Airlines recently spent areported $70 million to upgrade the quality of its passengercompartments in coach class by adding more leg room. Issuch an investment justified? To perform an analysis such asthis, we estimated the amount of ratings shift and the costsincurred in effecting the ratings shift. We then used the rat-ings shift to alter (for each respondent) the focal brandsutility, switching matrix, and CLV (see Equations 25),

which, when averaged across respondents and projected tothe size of the population (see Equation 6), resulted in arevised estimate for the firms customer equity. In this way,and using the discount rate and contribution margin we dis-cussed previously, we analyzed the recent American Airlinesseating improvement. We used the $70 million cost figurereported by the company.

If we assume that the average for the item that measuresquality of the passenger compartment (a subdriver of qual-ity) increases by .2 rating points on the five-point scale, ouranalysis (see Table 5) indicates that customer equity willimprove by 1.39%, resulting in an improvement in customerequity of $101.3 million nationally, or an ROI of 44.7%,

which indicates that the program has the potential to be alarge success. Table 5 shows the results of similar analysesfrom the other four industries. For example, a $45 millionexpenditure by Puffs facial tissues to improve ad awareness

TABLE 5

Projected ROI from Marketing Expenditures

Percentage DollarImprovement Improvement

Company Area of Geographic Amount in Customer in Customer Projected(Industry) Expenditure Region Investment Improved Equity Equity ROI

American Passenger United States $70 million .2 rating point 1.39% $101.3 million 44.7%(airlines) compartment

Puffs Advertising United States $45 million .3 rating point 7.04% $58.1 million 29.1%(facial tissues)

Delta (airlines) Corporate United States $50 million .1 rating point 1.68% $85.5 million 71.0%ethics

Bread & Circus Loyalty Local market $100,000 .5 rating point 7.04% $87,540 12.5%(groceries) programs in two measures


14/20


13To conserve space, we show the details only for the airlineexample, but the details of the other industry examples are similar.

by .3 rating points would result in a $58.1 million improve-ment in customer equity and an ROI of 29.1%.13

It is even possible to measure the financial impact of cor-porate ethical standards or corporate citizenship. For exam-ple, if Delta spent $50 million to improve customers per-ceptions of Deltas ethical standards by .1 rating points, thiswould project to a customer equity improvement of $85.5million (a 1.68% increase). Such findings may cause someairlines to reconsider practices such as canceling flights thatare not full in order to be profitable.

Not all investments will project to be profitable. Forexample, suppose that the grocery store Bread & Circusdecides to spend $100,000 in the local retail area to improveits loyalty program ratings across two measures by .5 points.The projected benefit is not enough to justify the expendi-ture, and the ROI is 12.5%.

The preceding examples illustrate only some of the mar-keting expenditures that can be evaluated by means of thecustomer equity framework. Any marketing expenditure canbe related to the drivers of customer equity, measured, andevaluated financially. This capability enables a firm toscreen improvement ideas either before application or aftera test market has nailed down the expected degree of

improvement.Model Sensitivity

The preceding analyses are based on point estimates, buthow sensitive is the ROI model to errors of estimation ormeasurement? Sensitivity to errors of estimation can be ana-lyzed by considering the sampling distribution of x.Appendix A shows how to construct confidence intervals forx. Then, by applying the end points of the confidence inter-val to the ROI model, it is possible to analyze the sensitivityof ROI to estimation error. In general, this error is of moreconcern on the low side, because overoptimism may resultin inappropriate expenditures. With this in mind, we suggest

calculation of a coefficient, x, which will be greater thanthe true value only 5% of the time. Assuming that there is alarge n, this is calculated as

(9) x =x 1.645(standard error ofx).

Then, x can be used to produce conservative projectionsof the customer equity change and the ROI. This can bedone by inserting x directly into the customer equity calcu-lations. For example, if we calculate a conservative estimateof customer equity impact for the American Airlines exam-ple in Table 5, we obtain a $93.9 million increase in cus-tomer equity, or a 1.29% increase. This would result in a34.2% ROI, indicating that even a conservative estimateshows a quite favorable return.

Sensitivity to errors of measurement can be addressed byconsidering the sampling distribution of the sample mean. InEquation A5 in Appendix A, unlike the case in regressionanalysis, the level of a variable affects the extent to which achange in the variable affects choice and thus utility, CLV,and customer equity. By evaluating the end points of the

confidence interval for the sample mean of a variable to beimproved, we can thus obtain a confidence interval for theROI that will result from a shift in that variable. We per-formed this analysis for the Delta Airlines corporate ethicsexample in Table 5. A 95% confidence interval for the meanon the corporate ethics variable was 3.346 .188, whichresults in a 95% confidence interval for corporate ethicsimprovement of $83.1 million/$87.8 million and a 95% con-fidence interval for ROI of 66%/76%.

If the projected rating shift results from a test market, the

sampling distribution of the rating shift can also beemployed to generate a confidence interval. Under theassumption that the sources of error are independent, whichis not unreasonable, it would then be straightforward to sim-ulate an all-inclusive confidence interval for ROI, incorpo-rating errors in the model coefficient estimate, estimatedsample mean, and estimated shift that are based on anassumption of a multivariate normal distribution withorthogonal components.

Discussion and Conclusions

Contributions to Theory and Practice

We make several contributions to marketing theory andpractice. First, we identify the important problem of makingall of marketing financially accountable, and we build thefirst broad framework that attempts to address the problem.We provide a unified framework for analyzing the impact ofcompeting marketing expenditures and for projecting theROI that will result from the expenditures. This big-picturecontribution extends the scope of ROI models in marketing,which to date have focused on the financial impact of par-ticular classes of expenditure and have not addressed thegeneral problem of comparing the impact of any set of com-peting marketing expenditures. Our work is the first serious

attempt to address this issue in its broadest form: the tradingoff of any strategic marketing alternatives on the basis ofcustomer equity. Marketing Science Institute member com-panies have identified this research area as the most impor-tant problem they face today.

Second, we provide a new model of CLV, incorporatingthe impact of competitors offerings and brand switching;previous CLV models have ignored competition. We alsodiscount according to purchase rather than time period. Pre-vious CLV models have been limited to the consideration ofpurchases made in prespecified time units, which is realisticfor some businesses (e.g., subscription services, sports sea-son tickets) but not for others (e.g., consumer packaged

goods). By discounting according to purchase, at the indi-vidual level, our model is more widely applicable. Theapproach set out previously considers customer equity forthe entire relevant competitive set, which has two advan-tages over existing approaches. First, this approach consid-ers the expected lifetime value of both existing customersand prospective customers, thereby incorporating acquisi-tion and retention (for the focal firm and competitors) in thesame model. Second, by explicitly considering competitiveeffects in the choice decision, it is possible to use the model


15/20


to consider the impact of competitive responses on thefirms customer equity.

Third, we provide a method for estimating the effects ofindividual customer equity drivers, testing their statisticalsignificance, and projecting the ROI that will occur fromexpenditures on those drivers. We present a principal com-ponents multinomial logit regression model for estimatingthe Markov brand-switching matrix, and we separate the dri-ver effects from the inertia effect. The identification andmeasurement of key drivers has been a process widely and

successfully employed in the fields of customer satisfactionmeasurement and customer value management (e.g., Gale1994; Kordupleski, Rust, and Zahorik 1993). We extend thisidea to customer equity. By doing so, companies can answerquestions such as, Should we spend more on advertising, orshould we improve service quality? and Which will havea bigger effect?

Fourth, customer equity provides a theoretical frame-work for making the firm truly customer centered, and it isapplicable to a wide variety of market contexts and indus-tries. Basing strategic investment on the drivers of customerequity is an outside-in approach that directly operationalizesthese fundamental marketing concepts. In other words, the

customer equity approach provides a means of makingstrategic marketing decisions inherently information driven,which is consistent with the long-term trends of decreasingcosts for information gathering and information processing.

Fifth, application of the customer equity framework isconsistent with practical management needs. The resultsprovide insight into competitive strengths and weaknessesand an understanding of what is important to the customer.By contrasting the firms customer equity, customer equityshare, and driver performance with those of its competitors,the firm can quickly determine where it is gaining or losingcompetitive ground with respect to the value of its customerbase. In addition, the model results include the distribution

of CLV across the firms customers, the distribution of CLVshare (discounted share-of-wallet) across the firms cus-tomers, and the percentage of the firms customer equityprovided by the firms top X% of customers. Collectively,the information gives useful information about how to seg-ment the firms customers on the basis of importance.Finally, the mathematical infrastructure of our frameworkcan be implemented by means of widely available statisticalpackages and spreadsheet programs, and we have conductedall the analyses by using only standard, commercially avail-able software packages.

Limitations and Directions for Further Research

In this article, we have developed and illustrated a practicalframework for basing marketing strategy on CLV and cus-tomer equity. As with any new endeavor, there is much workyet to be done. Specifically, we have determined seven keyareas for further research. First, the effects of marketdynamics on customer equity should be examined. Forexample, if the market is rapidly expanding or rapidlyshrinking, an assumption of stable market size is inappro-priate. In such markets, it would be necessary to model thechanging size of the market and relate that to customer

equity. This also implies the explicit modeling of a birth anddeath process for customers in the market. New-to-the-world products and services and markets in which firms areexpanding globally are examples of contexts in which webelieve this will be particularly important.

Second, our model assumes that there is one brand orproduct in the firm and does not consider cross-sellingbetween a firms brands or products. We believe that themodel we have described provides a solid foundation forfirms to understand what drives customer equity in a given

brand or product category. However, because many firmshave multiple offerings and hope to encourage customercross-buying of these products, it will be important to under-stand the influence of the drivers of customer equity on cus-tomer cross-buying behavior and to incorporate the impactof cross-selling on customer equity. This is particularlyimportant for firms that rely on customer cross-buyingbehavior for long-term customer profitability (e.g., financialservice firms).

Third, we adopt the assumption that a customers vol-ume per purchase is exogenous. An extension of thisresearch would permit volume per purchase to vary as afunction of marketing effort. For example, it will be impor-

tant to understand whether marketing efforts that may resultin forward buying (e.g., short-term price discounts) have along-term effect on customer equity.

Fourth, there is a need to develop dynamic models ofCLV and customer equity. Traditional models of CLV havebeen adopted from the net-present-value approach in thefinance literature. Understanding how the value of the firmscustomers (and overall customer equity) changes over timewill enable managers to make even better marketing invest-ments. There is also an opportunity to develop richer mod-els of CLV that incorporate a deeper understanding of con-sumer behavior.

Fifth, there is an opportunity to relate customer equity to

corporate valuation (Gupta, Lehmann, and Stuart 2001).This should involve the evaluation of corporate assets, lia-bilities, and risk, as well as the estimated customer equity.Sixth, applications of this framework and further empiricalvalidation of its elements would be useful, especially acrossdifferent cultures. For example, in what kinds of cultures arevarious drivers more important or less important, and why?Seventh, although our model incorporates competition, itmakes no provision for competitive reactions. An extensionof this work might involve a game theoretic competitivestructure in order to understand the effects of potential com-petitive reactions to the firms intended improvements in keydrivers of customer equity.

Summary

We have provided the first broad framework for evaluatingreturn on marketing. This enables us to make marketingfinancially accountable and to trade off competing strategicmarketing investments on the basis of financial return. Webuild our customer equity projections from a new model ofCLV, one that permits the modeling of competitive effectsand brand-switching patterns. Customer equity provides aninformation-based, customer-driven, competitor-cognizant,


16/20


and financially accountable strategic approach to maximiz-ing the firms long-term profitability.

Appendix AEstimation Details

Principal Components Regression

The independent variables for the principal componentsanalysis are all the drivers and the LAST variable. The vec-

tor Xijk denotes the original independent variables for eachcustomer i by previously purchased firm j by next-purchasefirm k combination. Treating the customer by firm combi-nations as replications, we extract the largest principal com-ponents ofXijk and rotate them using varimax rotation tomaximize the extent to which the factors load uniquely onthe original independent variables, thereby aiding manager-ial interpretability. The vector Fijk denotes the rotated factor.These form the independent variables for our logit regres-sion, which we describe subsequently.

Expressing Equation 2 in terms of the underlying factorsleads to the following:

(A1) Uijk

= Fijk

+ i,

where is a vector of coefficients.From factor analysis theory, it is known that the factors

are linear combinations of the underlying variables Xijk. Inother words, there exists a matrix A for which Fijk = XijkA.However, the idea of the principal components analysis wasto discard the potentially muddling effects of the leastimportant components. Denoting A* as the subvector ofAthat corresponds to the reduced factor space (discarding theprincipal components that do not meet the eigenvalue cutoff)and * as the estimated that corresponds to the reducedspace, Equation A1 can be expressed as

(A2) ijk = (XijkA*)* = Xijk(A**),

where ijk is the estimated utility, which means that * =A** can be the estimated coefficient vector. In otherwords, the coefficients ofXijk are obtained by multiplyingthe regression coefficients obtained from the logit regressionon the factors by the factor coefficients that relate the driversto the factors.

Logit Estimation

Usually in multinomial logit regression, the observed depen-dent variable values are ones and zeroes, corresponding tothe purchased brand (1 = brand was purchased, 0 = brandwas not purchased). This will be the case if the next pur-chase is observed from a longitudinal panel or follow-upsurvey. However, if purchase intent is used as a proxy fornext purchase, the dependent variable values will be propor-tions that correspond to the stated (or calibrated) purchaseintention probabilities. This does not create any difficulties.From Equation 9, we have Uijk = Fijk* + i, after discard-ing the principal components that did not meet the cutoff.Using the laws of conditional probability, we can expressthe likelihood of a particular parameter vector * givenrespondent is observed next purchase (or purchase inten-tion) vector p*i as

14Actually, convergence is faster, because the probabilities ofnext purchase are given directly, so the law of large numbers doesnot need to be evoked with respect to the dependent variable.

15For convenience, we suppress the j subscript for D, P, and U,because for any customer i in the sample, j is fixed.

where Yij equals one if customer i chooses brand j andequals zero otherwise, the likelihoods on the right side arethe usual 01 logit likelihood expressions obtained as in

Equation 3, and p*ij is the element ofp*i that corresponds tofirm j. The resulting likelihood for the sample is then theproduct of the individual likelihoods across the respondents.It is easily shown that with this adjustment in the likelihood,the standard logit regression maximum likelihood algo-rithms can be employed (Greene 1997, p. 916, 1998, pp.520, 524). The same adjustment of the likelihood does notaffect the derivation of the asymptotic distribution of theregression coefficients14 (as is evident in McFaddens [1974,pp. 13538] work), which means that the usual chi-squarestatistics, as given in standard logit software such asLIMDEP, can still be employed, even if the p*ij vector is notall zeroes and ones.

From Equation 3, it is easily shown that the partial deriv-ative of probability of choice with respect to utility, forrespondent i and firm k, is15

Then, from Equation A2 we have

(A5) Pik/Xijk = Pik/Uik Uik/Xijk = (A**)Dik

= (A**) p*ik(1 p*ik).

This equation shows how each customers brand-switching matrix will change given a change in any driver(or changes in more than one driver). This result is nonlin-

(A4) D = P / U

= exp(U ) exp(U )

exp(U ) exp(U )

exp(U ) exp(U )

exp(U ) exp(U )

= p* exp(U ) exp(U )

ik ik ik

ik*

k*ik

ik ik*

k*

ik ik*

k*

ik ik*

k*j k

ikij ik

j

[ ]

[ ] }

= [ ]

22

exp(U )

= p* (1 p* ).

ik*

k*

ik ik

(A3) L( | ) = L( |Y = 1)P(Y = 1)

= L( |Y = 1)p*

ij ij

ij ij

* *

*

p*ij

J

j

J

=

=

1

1

,


17/20


ear and implies diminishing returns for any driver improve-ment. By applying the altered switching matrix in Equation4, reestimating CLV by using Equation 5, and aggregatingacross customers by using Equation 6, we find the impact oncustomer equity.

The relative importance of each driver, measured as theimpact of a marginal improvement in the driver on utility,can also be addressed as a proportion of the total marginalimpact summed across all drivers. In other words, theimportance of driver x is the per-unit amount that it con-

tributes to utility, and the relative importance is that amountexpressed as a percentage.

where C is the set of retained principal components, Acx isthe factor coefficient relating driver x to factor c, and c isthe logit coefficient corresponding to factor c.

In addition, we address the statistical significance of thedrivers. The coefficients, c, as estimated by the logit model,are distributed asymptotically normally, and mean and vari-ance are estimated and reported by standard logit regressionsoftware. If the estimated logit coefficient and variance ofthe estimate for factor c are gc and 2c, respectively, andx = cAcxc is the coefficient estimator for driver x, then, ifwe assume that the gcs are distributed independently, x isa linear combination of independent normal distributionsand thus is also normally distributed. Specifically:

which results asymptotically in the following z-test for x:

which is easily calculated from the results of the principalcomponents analysis (for A2cx) and logit analysis (for x and2c).

Computational Issues in Estimating CLV

If the time horizon is long or the customers frequency ofpurchase is high, there may be many purchases expectedbefore the time horizon, which increases computation con-siderably. Therefore, it is useful to make a simplifyingapproximation that can speed up the computation. In prac-tice, the expected purchase probabilities, Bijt, approachequilibrium and change little after about 15 purchases. Thisenables us to employ the approximation that the purchaseprobabilities do not change after 15 purchases. If Ti 15, wecan estimate CLVij as in Equation 5. However, if Ti > 15, we

(A9) z = ,x

1

2 Acx

c

c2 2

(A8) Standard error of =x Acxc

c2 2

1

2 ,

( ) ),

( )

) ( ) ,**

A

A

A

c

c

C

c

c

C

cx

c

C

c

x Sd

6

7

100

1

1 1

Importance = (A and

Relative importance

(A

cx

cx

=

= =

=

16We could also use the equilibrium probabilities. In practice,there is little difference between the two.

can simplify the calculations. Let CLVij(15) denote the life-time value of customer i to firm j in 15 purchases, as calcu-lated by Equation 5, and let CLVij*(Ti) and CLVij*(15)denote the lifetime values that would occur (through Ti pur-chases and 15 purchases, respectively) if the purchase prob-abilities were constant and equal to Bij,15.16 The expectedlifetime value of the purchases beyond purchase 15 can beapproximated as CLVij*(Ti) CLVij*(15). This is helpfulbecause CLV* can be viewed as a net present value of anannuity, a

Return on Marketing

Documents

Transcript of Return on Marketing