Commercial Use of UPC Scanner Data: Industry and Academic

0732-2399/99/1803/0247/$05.001526-548X electronic ISSN

Marketing Science � 1999 INFORMSVol. 18, No. 3, 1999, pp. 247–273

Commercial Use of UPC Scanner Data:Industry and Academic Perspectives

Randolph E. Bucklin • Sunil GuptaAnderson School, University of California at Los Angeles, 110 Westwood Plaza, Los Angeles, California 90095,

[email protected] School of Business, Columbia University, New York, New York 10027, [email protected]

AbstractThe authors report the findings from an exploratory inves-tigation of the use of UPC scanner data in the consumer pack-aged goods industry in the U.S. The study examines the prac-titioner community’s view of the use of scanner data andcompares these views with academic research. Forty-one ex-ecutives from ten data suppliers, packaged goods manufac-turers, and consulting firms participated in wide-ranging, in-person, interviews conducted by the authors. The interviewssought to uncover key questions practitioners would like toanswer with scanner data, how scanner data is applied tothese questions, and the industry’s perspective regarding thesuccess that the use of scanner data has had in each area.

The authors then compare and contrast practitioners’views regarding the resolution of each issue with academicresearch. This produces a 2 � 2 classification of each ques-tion as “resolved” or “unresolved” from the perspectives ofindustry and academia. Along the diagonal of the 2 � 2,issues viewed as unresolved by both groups are importanttopics for future research. Issues deemed resolved by bothgroups are, correspondingly, of lower priority. In the off-diagonal cells, industry and academics disagree. These topicsshould be given priority for discussion, information ex-change, and possible further research.

Practitioners reported that scanner data analysis has hadthe most success and been most widely adopted for decisionmaking in consumer promotions (i.e., coupons), trade pro-motions, and pricing. For example, logit and regression mod-els applied to scanner data have revealed very low averageconsumer response to coupons which has directly led to re-duced couponing activity. Managers also reported high lev-els of comfort with and impact from analyses of trade pro-motions and price elasticities. While industry views most ofthe issues in these areas to be resolved, academic research

raises concerns about a number of practices in common com-mercial use. These include price threshold analysis and tradepromotion evaluation using baseline and incremental sales.

In product strategy, advertising, and distribution manage-ment, practitioners reported that the use of scanner data hashad more limited development, success, and impact. In thecase of new product decisions, scanner data use has beenslow to develop due to the inherent limitations of historicaldata for these decisions and a heavy reliance on traditionalprimary research methods. In advertising, scanner data iswidely analyzed with models, but confusion among practi-tioners is very high due to controversies about methods (e.g.,what level of data aggregation is best) and conflicting results.In distribution and retail management, scanner data use hastremendous potential but a mixed track record to date. Thus,practitioners view the use of scanner data as unresolved formost issues in product strategy, advertising, and distribu-tion. This view is largely, though not entirely, consistent withacademic research, which has only begun to address manyof the key questions raised by practitioners.

In light of the large number of unresolved issues andmixed record of scanner data use to date, the authors offer aseries of specific recommendations for immediate and long-term research priorities that are likely to have the greatestimpact on commercial utilization of UPC scanner data. Top-ics of immediate priority include price thresholds and gaps,baseline and incremental sales, base price elasticity, compet-itive reactions, measurement of advertising effects, manage-ment of brand equity, rationalization of product assortments,and category management. Long-term priorities include agreater emphasis on profitability versus sales or marketshare, developing prescriptive models versus descriptivemodels, and the need for industry standards.(Scanner Data; Marketing Research; Marketing Models; ResearchPriorities)

COMMERCIAL USE OF UPC SCANNER DATA:INDUSTRY AND ACADEMIC PERSPECTIVES

248 Marketing Science/Vol. 18, No. 3, 1999

Table 1 Participating Companies

Company LocationNumber ofParticipants

Information Resources, Inc. Chicago, IL 7Nielsen North America Schaumburg, IL 5Kraft General Foods, Inc. Glenview, IL 6Procter & Gamble, Inc. Cincinnati, OH 6The Quaker Oats Company Chicago, IL 3Nestle U.S.A. Glendale, CA 3PepsiCo Purchase, NY 2McKinsey and Company New York, NY 2Hudson River Group Valhalla, NY 3Media Marketing Assessment (MMA), Inc. Westport, CT 4

1. IntroductionA major theme of academic research in marketing hasbeen the development of substantive understandingand methodological tools that will help managersmake better decisions. Assessments of the actual im-pact of new approaches on the practice of marketinghave, however, been conducted only on an occasionalbasis (see, e.g., the studies on conjoint analysis byCattin and Wittink 1982, Wittink and Cattin 1989, andWittink et al. 1994). While published articles frequentlyoffer managerial implications and often suggest waysin which new models and data analyses may be put tocommercial use, it is unfortunate that the literaturelacks extensive reporting on the perspectives of prac-titioners—even in areas in which there is significantongoing academic activity.

Research on UPC scanner data has been activelypursued by marketing academics since the mid-1980sand many dozens of papers have been published inthe major marketing journals since then. Clearly, thebody of inquiry collectively known as scanner researchhas established itself as a major subfield within themarketing science discipline. Much of the intendedcontribution of the publications in this area has beento give practitioners better tools for understandingtheir markets (e.g., approaches to analyze consumerresponse and market segmentation) and for makingmarketing mix decisions (e.g., setting prices and de-termining promotion spending). Despite the earnestdesire of many academics to contribute to the practicaluse of scanner data in marketing (see, e.g., Neslin etal. 1994), our understanding of the commercial use ofscanner data—and the views practitioners hold re-garding its usefulness for decision making—is limited.In this paper, we take a step towards closing this gap.

Our primary objective is to report on the commercialuse of UPC scanner data and to provide a comparisonof industry practice with academic research. In so do-ing, we seek to determine where practitioners and ac-ademics agree, where they disagree, and why that isso. We will focus our report on a series of key issuesthat practitioners identified as especially important tothem with respect to the use of scanner data. For eachof these issues (e.g., trade promotion evaluation), wereport on industry practice then compare these prac-tices with those recommended by academic research-ers. The agreement (or disagreement) between practice

and academia leads naturally to a 2 � 2 classificationof each issue: “resolved” from the perspective of bothindustry and academia, “unresolved” in the eyes ofboth, or a disagreement in either direction.

Issues that both groups view as unresolved are im-portant topics for future research because of the poten-tial for significant impact among both academics andpractitioners. Issues deemed resolved by both are, cor-respondingly, of lower priority. In the off-diagonalcells, industry and academics disagree. Issues seen asresolved by one side but unresolved by another maybe good candidates for mutual discussion and infor-mation exchange (leading to possible further research).For example, practitioners may be unaware of prob-lems identified by academic researchers or, if aware,have good reasons for either viewing the issue as re-solved or perceiving the proposed methods as toocumbersome or impractical. Similarly, academics mayhave elevated certain specific problems to importancelevels well beyond what is merited by their practicalconsequences.

Study ApproachOur approach is exploratory. The views of practition-ers have been gathered based on in-depth interviewsconducted, in person, with managers from a sample of10 firms actively engaged in the use and/or analysisof UPC scanner data. Table 1 presents a listing of par-ticipating firms and the number of individuals withineach firm that were interviewed.1 We sought to cover

1We wish to thank the Marketing Science Institute for providing as-

BUCKLIN AND GUPTACommercial Use of UPC Scanner Data

Marketing Science/Vol. 18, No. 3, 1999 249

the perspective of packaged goods manufacturers,2 thetwo major syndicated data suppliers (Information Re-sources, Inc. (IRI) and ACNielsen), and third-partyconsultants. Consulting firms are important in theanalysis of scanner data due, at least in part, to therecent downsizing of in-house marketing research ca-pabilities at many consumer products firms. Both IRIand Nielsen also offer extensive consulting andcustom-study capabilities and therefore should be con-sidered as consultants as well as data suppliers. Welimited the scope of our interviews to the perspectivesof packaged goods manufacturers and the consultingfirms and data suppliers that serve them directly asclients. Retailers were not included among the inter-view participants. Thus, the focus of our paper is onthe commercial practices with respect to scanner datause for and by manufacturers.

We used in-depth personal interviews to generate aninteractive discussion with each participant. Unlike thestudy of the commercial use of a single method (e.g.,conjoint analysis), our study investigated the generalpattern of use of an extended family of models, tools,and methods that have been applied to secondary datawith common characteristics. Consequently, we ex-pected that each interview would take a potentiallydifferent path, with participants providing substantialinformation regarding specific scanner applicationswith which they were familiar and little, if any, infor-mation about others. Since one of our objectives wasto investigate areas of disagreement, we needed an ap-proach that permitted detailed follow-up questionsand probing. The interviews were conducted from fall1995 to spring 1996. All interviews were conductedwith both authors present (the single exception beingthe interview with Nestle USA). In the summer of 1997,all participating firms were invited to comment on apreliminary draft of the study report. The input fromtheir responses was then incorporated. We comple-ment the information from our interviews with results

sistance in securing appropriate contacts at many of the participatingorganizations. Of the firms we contacted, only one declined to par-ticipate in the study.2At each packaged goods company participating, we requested par-ticipation from individuals representing brand management, in-house market research and analysis, and field sales.

from a recent survey-based study by Davidson andStacey (1997). This telephone survey, conducted in1996, queried managers from 56 packaged goods firmsabout the relative importance that their firms attachedto a series of 15 marketing issues and the extent towhich scanner data was used to address those issues.

We organize our discussion following six areas ofthe marketing mix: pricing, trade promotion, coupons,advertising, product strategy, and distribution/retailmanagement. For each mix element, we discuss the keyissues raised by practitioners, how they use scannerdata to address these issues, and how these approachescompare with academic research. This leads to the clas-sification of each issue into one of the four cells in our2 � 2 matrix.3 In our conclusion, we summarize ourfindings, discuss the factors we believe underlie dis-agreement between academics and practitioners, andoffer suggestions for both immediate and long-termresearch projects.

2. PricingPractitioners considered pricing decisions especiallyimportant and scanner data are very widely used inthis domain of the marketing mix. The importance ofpricing to managers is due to the significant and usu-ally immediate impact of price changes on sales andprofits, the ability to implement price changes quickly,and the potential for strong reactions from consumers,retailers, and competitors. Davidson and Stacey (1997)also found that practitioners put pricing at or near thetop of the list of business issues. Pricing ranked thirdin overall importance among the 15 marketing issuesin their survey and was cited as “extremely important”by 78% of respondents. The same respondents alsorated pricing as the marketing issue for which scannerdata analysis and modeling was used more than anyother single topic. Sixty-two percent said scanner datamodeling and analysis were used “heavily” to addresspricing issues.

3The degree to which a given issue is seen as resolved by eitheracademics or practitioners lies on a continuum. We adopt the 2 � 2classification for expositional ease and it is not meant to imply thateach issue appearing in a given cell is precisely the same in terms ofits resolution.



Key Questions and IssuesIn our interviews, practitioners identified two key pric-ing issues that they use scanner data to address (orexpect to be able to use scanner data to address). Theseare (1) determining price elasticities, and (2) analyzingprice thresholds and determining optimal price gapswith competing brands.

Price Elasticity. Managers are interested in know-ing whether a given brand is relatively sensitive (elas-tic) or insensitive (inelastic) to changes in price. In ad-dition to national-level elasticity figures, managersoften expressed a desire to know if price sensitivityvaries across regional markets or retail accounts and,if so, how. Practitioners also told us they believed thatbase or regular-price elasticities differ from elasticitiesassociated with temporary price reductions (TPRs).They want to know both elasticities in order to makedecisions about list prices versus price discounts.

Price Thresholds and Price Gaps. If a brand has aprice elasticity of �2, a 10% change in its price shouldchange its sales by 20%. But many managers believethat such effects are unlikely to occur unless the pricechange crosses a threshold. In other words, responseto price changes can be “sticky” over certain ranges.Managers want to know where the threshold pricepoints lie so they can fine-tune price setting (e.g., setthe maximum possible price without triggering a no-ticeable loss in share or sales). Managers are also in-terested in knowing how sales or share will change asthe “gap” (i.e., absolute difference in price) betweentheir brand and competing brands widens or narrows.The purpose is to determine how much of a price pre-mium, for example, Marlboro can charge over a pri-vate label before the price difference begins to cause aserious erosion in Marlboro’s market share.

Price Elasticity

Industry Practices. A typical approach to addressprice elasticity questions is to run a time-series cross-section regression on store, account, or market-levelscanner data. (We will cover the issue of aggregationbelow.) At IRI and Nielsen, for example, the data setused for the analysis typically includes approximately100,000 observations covering about 2,000 stores over

a one year time period. A typical regression takes thefollowing form (e.g., Wittink et al. 1988):

ln(UNITSALES) � a � b ln(PRICE) � b FEAT0 1 2

� b DISP � b FEAT • DISP3 4

� b TPR � b SPECPACK5 6

� b STORE � b WEEK � ERROR. (1)7 8

In this equation, the log of the unit sales for a brand orstock-keeping unit (SKU) is a function of the log ofprice and a series of other independent variables. Here,PRICE refers to the actual shelf price of the brand orSKU. (Note that the price coefficient can be directlyinterpreted as an elasticity because of the log-log for-mulation.) Promotion activity is captured by dummyvariables (or indices) for feature (FEAT) and display(DISP) activity, as well as whether or not a temporaryprice reduction (TPR) is being offered.4 Additionalvariables include an indicator for special packages(SPECPACK), store dummy variables to control fordifferences in sales volumes across stores (STORE),and weekly indicator variables (WEEK), to control forseasonality or other special events (e.g., holidays suchas the Fourth of July or Thanksgiving).

Competitive effects (not shown in Equation (1)) canalso enter the model in the form of price or promo-tional activity for a judgmentally selected set of brandsor SKUs. When the regression analysis is done at theSKU level, which is often the case, the complexity ofthe modeling problem grows substantially. For exam-ple, including the price and promotion activity for allcompeting SKUs as independent variables would rap-idly become unwieldy. A common compromise is toinclude major competitive brands (aggregated acrosstheir SKUs) and own-company SKUs “above” and “be-low” (e.g., one size up and one size down) the targetSKU to control for potential cannibalization effects.Competitive promotions are sometimes also rolled into

4In this analysis, the price elasticity refers to the shelf price, whoseshort-run variation is primarily driven by temporary price reduc-tions. In Equation (1), TPR is an indicator variable that controls forthe effect of announcing a temporary price reduction via a “shelf-talker” next to the displayed item. We later take up the specific ques-tion of whether or not to separate the shelf-price variable into regularprice and temporary price discounts.



more aggregate measures (e.g., FEAT, DISP, and TPRmay be combined into one variable called promotion).The analysis usually assumes that price and promotioncoefficients are the same across all stores, although theequation can be modified to account for known store-specific effects by including additional terms. Due tothe robust properties of multiple linear regression andthe very large number of observations used, the anal-ysis is almost always conducted using ordinary leastsquares (OLS). IRI and Nielsen also said that they hadvirtually no problems with incorrectly signed esti-mates of own and cross-price elasticities.

In spite of reportedly high degrees of usage andcomfort with price elasticity analysis, managers raiseda number of issues about the practices in common use.These included: (1) the need for—and difficultieswith—analysis and models to take into account the id-iosyncrasies of demand and competition in a givencategory, (2) the appropriate level of analysis (i.e.,store, account, market, national) for the elasticity esti-mation, and (3) problems in estimating base-priceelasticity.

A widely held belief among practitioners is thatmodels applied to scanner data must be specificallytailored for the particular industry and product cate-gory involved. Kraft executives told two cautionarytales to illustrate their concerns with the use of stan-dardized regression models for pricing decisions:

Idiosyncratic Factors. A few years ago, Kraft ex-perienced declining sales in salad dressings. A regres-sion on historical data suggested that reducing pricewould significantly boost sales. Looking at the cate-gory dynamics in more depth, it turned out that theentire salad dressing category had declined over theprevious year because flooding in California had dra-matically reduced the supply of lettuce and raised itsprice. This cut lettuce consumption and led to a dropin demand for salad dressing. Further analysis re-vealed that there was also a significant time lag in thisentire dynamic (i.e., between the flood, the lettuce sup-ply and prices, and the consumption of salad dressing),making it easy to misinterpret regression results. Thestory highlights the need to include potential idiosyn-cratic factors in regression models, in this case a strongcomplementarity effect.

Competitive Reaction and Time Horizon. A cate-gory manager was told by market research that theprice elasticity for one of his products was �2. Theproduct, however, still had strong sales in spite of sev-eral recent price increases. On further investigation, itturned out that competition had also been raisingprices, following a short lag. Kraft executives used thestory to highlight the question of how to interpret thepredictive power of price elasticity. For example, overwhat time frame should an estimate be consideredvalid and, moreover, is it useless once competitionreacts?

In addition to the complexity of tailoring pricingmodels to incorporate idiosyncratic factors and com-petitive reaction, another concern raised by some prac-titioners (especially consultants) is that the typicalanalysis can become disconnected with the actual pric-ing process. Specifically, a regression that providesprice elasticity estimates using data across all stores isan analysis that fits best with a brand manager makinga pricing decision in a centralized fashion (e.g., a na-tionwide price discount of 10%). Except in occasionalinstances (e.g., an across-the-board price cut of 20% byPost cereal), pricing decisions also require attentionand management at the account level. This, in turn,leads to a demand for price sensitivity estimates foreach account rather than a single estimate that is thesame for all. If models are estimated at the accountlevel, however, brand price effects may be overstatedif shifts in store patronage are not held constant. Aprice reduction implemented at one chain but not atothers in the same market may attract sales for thatbrand from competing stores. (In the data, the ex-pected change in brand volume would be augmentedby borrowed sales from other stores.) This can lead tolarger estimates of price elasticity at the account-levelthan would be the case had all accounts implementedthe price reduction at the same time. As some man-agers explained, they are therefore unsure whether theresulting elasticities pertain more to the store or to thebrand.

When estimating elasticity, many practitioners esti-mate a single term (using shelf price as in Equation(1)), but others believe that base and TPR price elastic-ities should be estimated separately because (a) they



Figure 1 Industry and Academic Perspectives on Commercial Use ofScanner Data

are different, and (b) they represent two separate pric-ing decisions. While estimating elasticities for shelfprices or TPRs did not provoke much debate, estima-tion of base-price elasticity was controversial. Somecompanies (e.g., IRI) feel comfortable using cross-sectional data (under a no-promotion condition) fromseveral different markets to estimate base-price elastic-ity. Others question this approach because the modelassumes the same base-price elasticity to hold for allmarkets when, in fact, the market-level base-price elas-ticities may be very different. Still others (e.g., consul-tants at McKinsey) believe that there may be insuffi-cient natural variation in base prices to accuratelygauge its elasticity with scanner data. Instead of scan-ner data, they advocate using surveys or choice exper-iments to determine base-price elasticities.

Academic Perspectives. With respect to the basicdetermination of a brand’s own shelf-price elasticity atthe national level, regression analysis on large samplesof store-level data is likely to be sufficiently robust sothat estimates of these elasticities obtained in this man-ner are likely to possess good properties. A key caveat,however, is the problem of aggregation. As Christenet al. (1997) have shown, aggregating over store-leveldata to the market level can lead to biases in parameterestimates when the model is nonlinear, as in the log-log form of Equation (1). The same problem can ariseif data are aggregated over SKUs (i.e., to form com-posites or brand-level items) when the movement ofprice and other marketing variables differs acrossthem. Given this caveat, we consider that the estima-tion of shelf-price elasticities is largely resolved fromthe perspective of both academics and practitionersand, accordingly, list it in the upper-left quadrant ofFigure 1.

On the other hand, the determination of elasticitiesat the account-level and the estimation of cross-priceelasticities are not likely to be handled well by currentcommercial practice. Recent academic research onprice elasticity estimation has emphasized two majorthemes: (1) there are strong benefits to using prior in-formation (e.g., Bayesian methods) to improve elastic-ity estimates, especially in cases where sample infor-mation is limited (e.g., at the individual account orstore level), and (2) price elasticities, particularly inpackaged goods, have many empirical generalizations.

Published academic studies on store-level scannerdata that have used classical regression approacheshave had a mixed record of producing elasticity esti-mates with correct signs and stable properties, espe-cially for cross-effects. While Foekens et al. (1994) re-ported no incorrect signs for either own- or cross-priceelasticities, problems have been reported in a numberof other cases (e.g., Cooper 1988, Carpenter et al. 1988,Blattberg and Wisniewski 1989). Recognizing theeconometric difficulties (e.g., multicollinearity, limitednumber of observations) often associated with obtain-ing good estimates for own- and cross-price elasticitiesfor a single account or regional market, researchershave developed some promising new approachesbased on Bayesian methods. Working in this type ofdata setting, researchers have refined methods to im-pose prior structure or “shrinkage” on the estimates inorder to obtain own- and cross-price elasticities withgood properties (e.g., Allenby 1990, Blattberg andGeorge 1991, Montgomery 1997). Academic research-ers have emphasized the study of regional or account-level elasticity estimation (versus practitioner empha-sis on national cross-sections) in part because the datasets available for academic research typically coveredonly one, or perhaps two, regional markets.

Academic studies also have revealed many regular-ities in brand price elasticities for packaged goods.



Such “empirical generalizations” can be used as a sup-plement to regression-based estimates or to providepriors for shrinkage methods. For example, a meta-analysis by Tellis (1988) found price elasticities to av-erage �2.5. Ehrenberg and England (1990) reportedthat price elasticities for each of four brands across fivecategories were equal to �2.6 and that this effect sizeheld regardless of whether price changes were large orsmall. Other researchers have found that price elastic-ities are related to brand characteristics (e.g., Bolton1989, Narasimhan et al. 1996, Bell et al. 1999). For ex-ample, Bell et al. (1999) reported that 81% of the vari-ance in brand price elasticities can be explained byreadily identifiable category, brand, and consumer fac-tors, with category factors (e.g., necessity item, stora-bility) accounting for most of the differences in priceelasticities.

With respect to cross-price elasticity estimation, wenote that (1) the regressions used in practice do notinclude estimates for all cross effects, and (2) cross-effect terms are often created by aggregating over bothpromotional variables as well as SKUs. Moreover, wenote that the potential problems are likely to worsen ifconventional regression is used to estimate cross-effects at the account-level. With the availability of newestimation approaches, many of these difficulties canbe avoided. Latent class analysis (e.g., Kamakura andRussell 1989) can be used to estimate elasticities forclusters of stores or accounts that share common levelsof market response parameters. Empirical Bayes meth-ods (e.g., Montgomery 1997) use sample-wide infor-mation to “shrink” estimates of account-specific elas-ticities and cross-elasticities, greatly reducing theoccurrence of incorrectly signed elasticies or nonsen-sical estimates. Bucklin et al. (1998) have shown howcross-price elasticities correspond, via a scaling factor,with brand-switching probabilities. This might be usedto take advantage of managers historical comfort withbrand-switching data to produce better estimates forown- and cross-price elasticities. Thus, in the estima-tion of account-specific elasticities and cross-price elas-ticities, academic researchers can offer a number ofnew methodologies and useful findings that should behelpful in industry practice. We therefore classify theseissues in the lower left quadrant of Figure 1.

Academic research has only just begun to address

the specific concerns practitioners raised about com-petition, confounding store and brand elasticities, andbase-price elasticity. Empirical reaction functions (e.g.,Bresnahan 1997, Leeflang and Wittink 1996, Jedidi etal. 1999) can be used to assess whether or not predic-tions of price response will be seriously invalidated bysubsequent competitive effects. The empirical reactionfunctions can also be used to assess the nature andextent of competitive reaction in the marketplace. Forexample, Leeflang and Wittink (1996) report that com-petitors generally overreact to price changes. Game-theoretic work also has shed substantial light on theforms of competitive pricing behavior that are com-monly found in consumer packaged goods categoriesand their implications (e.g., Rao et al. 1995, Lal andPadmanabhan 1995). Recently, researchers have usedthe industrial organization paradigm to estimate de-mand and reaction functions simultaneously (Kadiyaliet al. 1999).

Concerns that brand-price elasticities at the account-level may be inflated due to store switching effectsmight be addressed with models estimated on paneldata. This is because panel data contains informationon the specific store visits of each panelist. Thus, if sodesired, it is possible to hold the effect of store switch-ing constant (see, e.g., Bucklin and Lattin 1992).

The academic literature reports conflicting resultsregarding whether or not base-price elasticities differfrom separately estimated promotional-price elastici-ties. For example, Guadagni and Little (1983) reportthem to be the same while Blattberg and Neslin (1990)show large differences. Part of the confusion stemsfrom the difficulty in completely separating theattention-getting aspect of a temporary price reduction(e.g., the effect of the shelf talker) from the price changeitself (e.g., Inman et al. 1990). Intuitively, for price re-ductions where shoppers are cued that the change istemporary, elasticities should exceed base-price elas-ticities when the discounts induce consumers to stock-pile and switch brands. We note that base-price elas-ticity estimates from store-level data also might beimproved by the use of Bayesian methods, but re-searchers have not yet explicitly addressed this prob-lem. Due to the nascent academic attention to theseissues or to conflicts in the literature, we classify themas unresolved from both perspectives (see Figure 1).



Price Thresholds and GapsTo investigate price thresholds and price gaps withcompeting brands, industry practice is to use a simpleand intuitive procedure known as sales velocity analysis.This procedure essentially creates a cross-tabulation ofsales versus price points across stores. It consists of thefollowing steps:

1. For all stores of a common format (e.g., drugstores), collect information on total store sales, sales ofthe target brand, and its nonpromoted price. (The anal-ysis can be repeated for promoted prices and for otherstore formats.)

2. Aggregate store and brand sales data across allstores that charge the same nonpromoted price for thetarget brand.

3. Create a sales rate or sales velocity measure asunit sales of the brand per million dollar store sales.This accounts for differences in brand sales due to dif-ferences in overall store sales.

4. Plot the cross-sectional data on sales rate and non-promoted prices. Visually inspect the plot and makean inference about price thresholds.

A similar analysis also can be conducted for pricegaps. In that case, sales rates are plotted against thevarious price gaps between the target brand and agiven competing brand found in the data. The objec-tive in both cases is to use scanner data as a “naturalexperiment” to help find the best specific price pointsfor a brand.

At PepsiCo, executives recounted the story of howthis analysis helped them. A number of years ago, bothCoca-Cola and Pepsi were losing market share to pri-vate label brands. Management tried different tacticsto halt the share loss, but none succeeded. Finally, con-sultants suggested a price-gap analysis on scannerdata. The study revealed price points that were effec-tive at holding share against private labels. These wereimplemented nationwide and the share erosionstopped.

Practitioners told us that this type of analysis forstudying thresholds and gaps is widely used becauseof its simplicity, ease of use and understanding, andthe ability to conduct many such studies in a short pe-riod of time. Though sales velocity analysis is appeal-ing to many, executives at Nestle were more cautious.

They noted that they found the method useful for di-agnostic purposes but believed that, ultimately, priceanalysis needs to be based on within-store changes thattake place over time.

Academic Perspectives. While most practitionersexpressed high levels of comfort with the use of salesvelocity analysis to determine thresholds and gaps, de-termining price thresholds (and gaps) is actually a dif-ficult conceptual and statistical problem. We thereforebelieve that current industry practice has some majorlimitations. These limitations include the use of cross-sectional data for causal inference (e.g., is it reasonableto combine East Coast data with West Coast data?) andthe lack of statistical analysis (e.g., is $2.29 really athreshold?).

Academic researchers have long recognized thatprice response functions need not be smooth or evenmonotonic (e.g., Gabor and Granger 1964, Monroe1990). Thus, there is agreement with practitioners onthe existence of price thresholds as an empirical phe-nomenon. Academics, however, have taken differentapproaches to the problem of estimating price thresh-olds. One stream of research approaches the problemat the individual level, working from the notion thatindividuals have reference prices against which theyevaluate current prices when making a decision (e.g.,Thaler 1985). Price threshold effects occur because alatitude of price acceptance (or region of price insen-sitivity) develops around the reference point. Empiri-cally, researchers have found latitudes of price accep-tance in both scanner panel data (e.g., Kalyanaram andLittle 1994) and in controlled experiments (e.g.,Kalwani and Yim 1992). In the study by Kalyanaramand Little (1994), for example, the latitude of price ac-ceptance was found to be symmetric about the refer-ence point and to average approximately 1.5 times thestandard deviation in the brand shelf price. Anotherstream of work attempts to determine price thresholdsfrom store-level data (or aggregations across stores). Apromising approach was recently proposed byKalyanam and Shively (1998), who use Bayesian meth-ods (specifically Gibbs sampling) in combination withspline regression to estimate irregular price responsefunctions on aggregate-level data. Nevertheless, the es-timation problem is challenging in part because each



consumer (and therefore each store) may have differ-ent zones of price insensitivity.

With respect to price gaps, academic studies haveextensively investigated the asymmetric nature ofprice competition (as captured by cross-price elastici-ties) between brands in different “price tiers” (e.g., apremium national brand versus private label or privatelabel versus generic). Blattberg and Wisniewski (1989),in a study using store-level scanner data, showed thatprice cuts on higher price, higher quality, brands hadgreater effects on the sales of lower price, lower qualitybrands than vice versa. These effects have also beendocumented in academic studies based on panel data(e.g., Kamakura and Russell 1989) and modeled in de-tail with store-level data (Sethuraman 1996). Academ-ics have also investigated some of the implications ofprice gaps using game-theoretic models. For example,Raju et al. (1990) related the magnitude of price gapsbetween brands to the equilibrium level of price pro-motion activity within a product class. In sum, the ac-ademic study of price gaps has largely emphasized thecompetitive implications that may be associated withprice gaps, as opposed to the determination of the op-timal price gap.

While research has validated the existence of pricethresholds (i.e., zones of price insensitivity) and thenature of asymmetries that may exist across price tiers,it has not yet begun to produce methods that can di-rectly provide managers with the answers they expectto be able to obtain from scanner data (e.g., what is theoptimal price gap between my brand and a particularrival brand). Given the widespread industry use andcomfort with sales velocity analysis, we therefore clas-sify this issue as one that practitioners view as resolvedbut academics deem unresolved and place it in theupper-right quadrant of Figure 1.5

3. Trade PromotionTrade promotions typically account for the largest por-tion of the marketing budget at most consumer prod-uct companies (approximately 40–50%). Practitioners

5This is a particularly important research issue because the existenceof price thresholds implies that price elasticities are not constant.Thus, inferences about price elasticities made from regression anal-yses (as in Equation (1)) are subject to this limitation.

therefore had a keen interest in understanding theoverall effectiveness of trade spending as well as therelative effectiveness of the various components (i.e.,TPR, feature, display, etc.). The importance of tradepromotion evaluation is underscored by the resultsfrom the Davidson and Stacey (1997) survey. They re-port that managers ranked trade funds managementas the most important marketing issue they confront(first among 15), and that they used scanner data anal-ysis extensively for this issue. Specifically, trade fundsmanagement ranked third (among 15) in practitioners’reported use of quantitative analysis and modeling,with 54% of respondents indicating heavy usage.

Key QuestionsFor the managers we interviewed, the critical issue intrade promotions is gauging the effectiveness of atrade deal. Two questions that practitioners seek to an-swer with scanner data are:

(1) What is the baseline sales volume (i.e., whatwould sales be if no promotions take place)?

(2) What is the incremental volume6 (i.e., “lift”) dueto a trade promotion?The effectiveness of a trade deal is then assessed by themagnitude of the lift generated by the promotion rela-tive to the baseline sales level. While this type of anal-ysis focuses on sales response, some managers also areinterested in extending the assessment of trade pro-motion effectiveness to profitability. Baseline and in-cremental volumes are, of course, essential inputs tosuch a profitability analysis.7

Industry PracticesBoth ACNielsen and IRI generate estimates of baselinesales using volumes in nonpromoted weeks and an ex-ponential smoothing algorithm (i.e., more recentweeks are given greater weight). In some cases, ad-justments are made to the baseline estimate usingcross-sectional data (i.e., actual sales in similar storesin the same time period). Although there are some

6Incremental volume, or lift, is defined as the difference betweenactual volume and estimated baseline volume.7Strictly speaking, TPR, display and feature are actions taken by re-tailers which often result from trade promotions. Therefore, manu-facturers should assess the effectiveness of trade promotions by con-sidering incremental sales and retailer pass-through. In this study wefocus only on the issue of incremental sales.



technical differences in the two approaches used byACNielsen and IRI (see Abraham and Lodish 1993 fora description of IRI’s approach), they are conceptuallyvery similar.8

Trade promotions also can be analyzed with regres-sions on store-level data. This approach essentially fol-lows the regression template of Equation (1). In tradepromotion analysis, however, the simple exponentialsmoothing procedure is preferred to regression be-cause baseline estimates are readily obtained everyweek for each market, as well as for every SKU ofevery brand. Running regressions at this level is notonly cumbersome, but also tends to generate unstableparameter estimates. Thus, practitioners have a strongpreference for the simplicity and robustness of thesmoothing algorithm for obtaining estimates of base-line sales.

Managers and consultants we interviewed had ahigh degree of comfort with the analysis of trade pro-motions. Baseline estimates from the smoothing algo-rithm that are embedded in the on-line data from IRIand ACNielsen are so well accepted that managers of-ten view them as “data” rather than estimates. Tradepromotion effectiveness is then analyzed by examingthe lift associated with the specific characteristics of thetrade deal (e.g., extent of feature, display, and TPR ac-tivity obtained at the retail level). Managers then studywhat types or combinations of deals appear to be mosteffective, ranking them according to the relative lift(i.e., incremental volume divided by baseline volume)that they generate. Trade dealing can then be adjustedto emphasize the types of deals that generated stronglift over the deals that did not.

Many practitioners consider trade promotion anal-ysis with scanner data to be an industry-wide successstory. Consultants at McKinsey said that trade pro-motion analysis has added “hundreds of millions ofdollars to the bottom lines of U.S. packaged goodscompanies.” Nestle executives echoed this view, re-porting widespread use and impact. Interestingly, theypointed out that the impact has come not so much from

8The technicalities have, upon occasion, led to concern at some clientcompanies about discrepancies in baseline and incremental volumeestimates from the two data suppliers.

changes in trade promotion spending, but from “re-directed” spending, with increases in some areas bal-anced by decreases in others. Although trade promo-tion analysis is very useful for marketing managers inbudget allocation decisions, it is potentially even morepowerful in the hands of the sales force, where it canbe used in negotiations with retail accounts. This hasled to increased interest in obtaining incremental vol-ume information at the account (or even store level).

Some companies and consulting firms told us thatthey are extending trade promotion analysis beyondincremental sales to incremental profit (e.g., Kraft andMcKinsey specifically indicated that this was an areaof emphasis for them). Nevertheless, most companieswere focused on sales. The most common explanationfor the emphasis on sales versus profitability was thedifficulties in assessing the true cost of trade promo-tions. The difficulties arise because promotions vary onmany dimensions (case allowances, off-invoice allow-ances, scanback pricing, etc.) and because of time lagsbetween invoice and payment. Unless the companyhas designed a careful accounting system (such as ac-tivity based costing), most find it difficult to apportionrelevant costs to promotions. This cost allocation prob-lem is further exacerbated when a product is divertedin the channel from one retail account to another.These difficulties in assessing cost may explain whytrade promotions continue to thrive even though somestudies (e.g., Abraham and Lodish 1990) suggest thatalmost 84% of trade promotions are not profitable.

While most of the emphasis on trade promotion ef-fectiveness involves the short-run effect of trade deals,an issue that also arose is whether or not trade pro-motions have long-term effects. On this question, mostpractitioners responded that they do not find post-promotion dips in their data. This leads many to con-clude that promotions do not have long-term negativeeffects.

Academic PerspectivesAcademic researchers have had a significant role in thedevelopment of industry procedures for evaluating theeffects of trade promotions. The work of Abraham andLodish (1993) is cited as especially influential (e.g., spe-cifically by executives at IRI). Thus, much of the day-to-day use of baseline and incremental sales analysis



has been developed with the participation of academicresearchers. At the same time, however, this researchalso has recognized many of the limitations inherentin the conventional approach (see, e.g., the discussionof limitations in Abraham and Lodish 1993, pp. 250and 268). These limitations include the inability of thetraditional baseline/incremental method to take intoaccount forward-buying by loyal consumers (i.e., pur-chase acceleration or stockpiling), store switching byconsumers, the cannibalizing effects of sister brand orsize promotions, and promotions by competitors.

When we asked managers directly about the limi-tations of the baseline/incremental approach, mostagreed that baseline estimates can have serious prob-lems and that they are probably biased downwards.Downward bias in baseline sales implies that the liftattributed to trade promotions is overstated. This hasimportant implications for assessing the profitability oftrade promotion because the reclassification of salesvolume from incremental to baseline can have a largenegative effect on the economics of a trade deal. De-spite these issues, managers told us that they rely onthe baseline/incremental approach because (1) it is thebest readily available method, and (2) they believe thatthe relative effectiveness of different promotions andpromotion vehicles (e.g., TPR, feature and display) isunlikely to change significantly even if the missing fac-tors were to be accounted for. This means that tradedeals are effectively sorted by relative performance ver-sus evaluated on their absolute performance. Relativeperformance information makes it possible to reallo-cate trade funds among deal types but does not pro-vide the necessary information for determining the op-timal total amount of trade promotion spending. Thus,while there has been enormous positive impact fromtrade promotion analysis (e.g., the “redirecting” ofspending mentioned by Nestle executives), we believethat additional gains in trade promotion productivitycan come from using approaches that better assessprofitability.

Much of the academic research on trade promotionhas studied market response to promotion variables atthe individual-level using scanner panel data. This re-search has shown that promotions induce consumersto accelerate their purchases and stockpile quantity,

not just switch brands (e.g., Neslin et al. 1985). Re-searchers have also decomposed brand sales elasticityinto components due to switching, acceleration, andquantity (e.g., Gupta 1988), finding that brand switch-ing effects are the largest source of sales response, butthat significant sales response is often due to purchaseacceleration and/or stockpiling. Bell et al. (1999) de-composed elasticities for 173 brands in 13 product cate-gories. On average, brand switching accounted for 75%of total sales elasticity, with a low of 49% and a highof 94% among the categories in their sample.9 Thesefindings establish that a substantial portion of whatappears to be “incremental” sales may be “borrowed”from the future. Unfortunately, the widely used base-line/incremental methods do not make thisdistinction.

While some practitioners recognized the limitationsin the baseline/incremental method, they were alsoquick to point out severe limitations to panel datastudies. Sampling problems are one reason for the re-luctance to rely on panel data analysis. For example,many practitioners told us that market shares com-puted from panel data sometimes move in oppositedirections from the shares computed from store-leveldata. The sampling issue involves both bias (e.g., arescanner panelists representative of the population?) aswell as size (e.g., volume and share estimates for smallbrands or SKUs can be quite shaky). In an importantstudy of the representativeness of panel data, Gupta etal. (1996) reported that the demographic characteristicsof a sample of panelists in two data sets did signifi-cantly differ from the population and that the brandmarket shares in the panel data differed from the storedata. Nevertheless, the price elasticities estimated fromthe panel and the store-level data differed by only 5–7%.10 This suggests that panel data may be able to pro-vide key insights into the effectiveness of marketingactivity despite the sampling issues. More recently,Silva-Risso et al. (1999–this issue) showed that tradepromotion productivity could be improved in a field

9A recent study by Van Heerde et al. (1999) uses store-level data anddistributed lag models to infer that 4–24% of current sales effect isborrowed.10As the authors of the study are careful to point out, this finding isbased on the assumption that purchase observations are drawn fromthe scanner panel data following what they call a purchase selectionapproach and not a household selection approach.



application using results from a decision support sys-tem based solely on panel data. Clearly, further re-search is needed to take advantage of the ability ofpanel data to distinguish incremental from borrowedsales while overcoming sampling problems.

Many industry executives believe that promotionscan increase consumption substantially, especially forimpulse items such as snack food. Pepsi executivestold us that 50% of their incremental volume may comefrom increased consumption. Some recent academicwork has shown this phenomenon to exist in an ex-perimental setting (e.g., Wansink and Deshpande1994) and it was incorporated into models estimatedon panel data (Ailawadi and Neslin 1998). A weaknessof scanner data is that it tracks only purchases andprovides no specific information on consumption or at-home pantry holdings. This is an important researchtopic since it affects how much of the incremental sales“bump” represents net new sales of the brand (due toconsumption increases) versus sales borrowed fromthe future (due to pantry loading). Most recently, Bellet al. (1999) reported evidence consistent with con-sumption increases for 4 out of 13 grocery categoriesthey studied. Specifically, promotions were associatedwith probable consumption increases in bacon, potatochips, soft drinks, and yogurt.

As noted above, the absence of a dip in sales follow-ing a promotion has led many managers to believe thattrade promotions have no (negative) long-term effects.The academic perspective is, however, that the lack ofa post-promotion dip per se need not suggest anythingabout the long-term effect of a promotion. Based onsimulations, Neslin and Stone (1996) showed that post-promotion dips in sales data will not occur unless con-sumers are very sensitive to at-home inventory lev-els—i.e., it is normal not to observe such dips. Over thelonger term, if consumers come to expect promotions,overall baselines may become depressed. Again, thereneed not be any post-promotion dips in the data. Al-though academic work on long-term effects is limited,initial studies have found that promotions do havenegative long-run consequences. For example, Mela etal. (1997) studied data for one product categorystretching over an eight-year period. They found thatconsumers became more sensitive to promotions as

promotional activity increased. Using a varying pa-rameter model, Foekens et al. (1999) show that the netincremental sales from a dynamic model can be sig-nificantly smaller than the incremental sales based ona static model.

Both practitioners and academics agree that scannerdata and trade promotion have been an “industry-wide success story.” Perhaps because the positive im-pact of scanner data has been so strong in trade pro-motions, practitioners expressed some surprise whenwe questioned the conventional approaches in wide-spread use. Nevertheless, we believe that the use ofscanner data in trade promotions is not resolved. Thisis primarily because current methods are likely tooverstate both the real sales response to promotion aswell as the profitability of those sales. Accordingly wehave placed this topic in the upper-right quadrant ofFigure 1. In closing, we note that while trade promo-tion spending has been reallocated among deal typesand by retail account, the overall spending on tradepromotion has remained at high levels. This may re-flect the difficulties involved in assessing the true prof-itability of these marketing activities.

4. Consumer PromotionManagerial use of scanner data in consumer promo-tions is focused primarily on understanding the effectsof coupons. Among the managers we interviewed,coupon analysis ranked below pricing and trade pro-motion in importance, but not necessarily in the ex-pectations for how scanner data could aid decisionmaking. Davidson and Stacey (1997) also report thesame pattern. In their study, consumer promotionevaluation ranked seventh out of 15 issues in terms ofimportance, but fourth in the use of scanner data anal-ysis and modeling. The major issues in consumer pro-motions are similar to those in trade promotions: un-derstand incremental sales and profits due to coupons.At the broadest level this will help managers decidewhether to increase or decrease spending on coupons.Some detailed decision issues involve the face valueand frequency of coupon drops, expiration dates, andthe potential advertising role of coupons (e.g., how im-portant is coupon design and copy).



Industry PracticesIn our interviews, we found that practitioners em-ployed three approaches to the analysis of coupons: (1)judgment, (2) regression, and (3) logit models. Withrespect to the first, managerial judgment, several man-agers expressed high levels of comfort in relying onjudgment as the primary means to decide on couponrelated issues. They also noted that they want to beable to include such judgments along with other ana-lytical techniques. For example, the average purchasecycle of the product (which can be measured withscanner panels or by survey) is considered a drivingfactor in deciding coupon frequency. As one managerexplained, it does not make sense to him to drop cou-pons every 30 days if the average purchase cycle ofyour product is two months. Competitive intensity ofcoupons is also an important factor in deciding couponface value and frequency. Similarly, expiration datesof coupons are a function of purchase cycle, the ur-gency that managers wish to create, and logistics fac-tors for coupon clearance and payment to relevantparties.

A simple analytical tool used by some practitionersis to treat coupons as another promotion variable inthe store-level regression model and assess its effec-tiveness in driving sales (see Equation (1)). A limitationof this approach is that it is unable to distinguish whencoupons are used primarily by loyal consumers or byswitchers (panel data is required for that). Thus, whilethe method is useful for obtaining general results oncoupon response, it has limitations for assessing bor-rowed versus incremental sales.

A more sophisticated approach used by some com-panies (via custom consulting studies conducted by IRIand others) is the household-level logit model to assessincremental sales due to coupons (Guadagni and Little1983, Little 1994). This approach, known at IRI asCouponScan, fits a logit (or nested logit) model topanel data and then simulates brand sales to determinewhat the sales would have been if coupons had notbeen dropped.11 Incremental sales are then computed

11Despite the claims of some academic researchers that probit is asuperior choice model to logit, probit models remain unappealingto practitioners because of the complexities associated with their es-timation. Said one executive at IRI, “Probit versus logit was never adebate here; probit would be too costly.”

as the difference between actual and simulated sales.Thus, the principle is analogous to that of the baselinesales approach used in trade promotion. The major dif-ference is that baseline sales estimates are producedfrom the logit model based on panel data, not an ex-ponential smoothing algorithm based on store-leveldata.

By setting up the analysis in the baseline/incremen-tal fashion, the CouponScan approach gets around theproblem of the lack of information on coupon avail-ability. While data are usually collected on coupon re-demption, i.e., we know which panelists redeemed acoupon, there is usually no information on when apanelist had a coupon available but decided not to useit. Executives at Procter and Gamble expressed a veryhigh level of comfort with the logit modeling approachused in CouponScan. Said one, “If we get a result outof the logit model that sounds weird, we will challengethe data first.”

We found a strong and consistent response from allparties about coupons: they are generally regarded asrelatively ineffective. Managers at Quaker reportedthat the studies they commissioned on marketing mixeffectiveness (e.g., by ACNielsen) for their products re-vealed that coupons had the lowest impact of any mar-keting mix element. They also noted that their brands’heaviest users were also the heaviest redeemers of cou-pons. Consultants at the Hudson River Group reportedthat they found in their work an average elasticity forcoupon spending of 0.07 (based on regression analysisof aggregate-level data). They noted that this placescoupon response below that of media advertising (see,e.g., Assmus et al. 1984, Lodish et al. 1995). As one ofthe consultants put it, “We love to shoot coupons.” AtMcKinsey, consultants called coupons, “a lousyinvestment.”

Several factors were cited as contributing to thestrong conclusions about coupons. These were low re-demption rates, the large proportion of coupon re-demptions by loyal consumers as revealed by paneldata analysis, low levels of lift as revealed by regres-sion and logit modeling (especially when comparedwith trade promotion), lack of enthusiasm from retail-ers in handling coupons, and the cost of clearing andpayments. The consensus regarding the relative inef-fectiveness of coupons has led many packaged goods



manufacturers to reduce their spending on coupons.Executives at both Quaker and Procter and Gamble,for example, said that scanner-based findings led man-agement to cut coupons. Indeed, P&G recently con-ducted tests in upstate New York where they elimi-nated the use of coupons (e.g., Narisetti 1996). AtNestle, coupons are now used primarily in associationwith new product launches. Still others have moved tomore targeted use of coupons (e.g., the check-out cou-poning system from Catalina Marketing, Inc.). In sum,the use of scanner data has had a major impact on howmanagers view coupon promotions in the packagedgoods industry. This is because the data have clearlyshown that coupons tend to have relatively small ef-fects on sales and because coupons are redeemedheavily by loyal users. Our discussions surfaced littlein the way of unresolved issues from practitioners.

Academic PerspectivesMany of the issues in evaluating the response to andprofitability of coupon promotions parallel those intrade promotion. Both require that sales be decom-posed into those incremental to the promotion versusthose that would have been obtained without the pro-motion (i.e., baseline). The CouponScan system, for ex-ample, is engineered to do exactly this. Researchershave also stressed the need to determine the propor-tion of coupon redemptions that deliver incrementalsales to the brand versus providing a discount to buy-ers who would have purchased the brand without thecoupon (e.g., Blattberg and Neslin 1990). Analogous totrade promotion, there are major implications for theprofitability of couponing. For example, Neslin (1990)reports that for the category he studied only about 40%of coupon redemptions can be classified as incrementalsales for the brand. Based on typical coupon cost esti-mates, he argues that a much higher proportion isprobably needed in order for couponing activity to beprofitable. This reinforces the need to examine couponpromotions with panel data (as practitioners are in-deed doing) to assess the extent of redemptions madeby loyal buyers. We note that most of these studieshave evaluated coupon profitability assuming that reg-ular shelf prices remain constant. If coupons are usedas a price discrimination device (Narasimhan 1984), itmay be desirable to increase regular price while in-creasing the frequency and/or depth of coupons.

Another issue that has begun to be examined by ac-ademic researchers involves the potential for couponsto have an advertising effect on brand sales. Srinivasanet al. (1995), for example, found that coupons in free-standing inserts can have advertisement value to someconsumers. The potential for this effect to improve theeconomics of a coupon promotion has been incorpo-rated into a model of coupon profitability by Leoneand Srinivasan (1996).

While there are clearly opportunities for additionalresearch in coupons (e.g., the advertising effect of cou-pons, refining market response and profitability mod-els), the perspectives of practitioners and academicsare better aligned with respect to coupons than mostother areas of the marketing mix. In particular, it iscritical that practitioners have recognized that a sig-nificant portion of the temporary sales increase that acoupon may generate is likely to represent borrowedsales. Accordingly, we classify the evaluation of cou-pons as largely resolved from both the academic andpractitioner perspectives.

5. AdvertisingAdvertising is extremely important to practitioners inpackaged goods, but the expectations for how scannerdata can help improve its effectiveness are more mod-est. Much of this comes from the role that creative fac-tors play in advertising strategy and the widespreaduse of survey-based testing and tracking measures.The Davidson and Stacey (1997) study reports that ad-vertising effectiveness is ranked fifth in importance butninth in the use of scanner data analysis and modelingamong the 15 issues they investigated. Thus, the rela-tive importance of advertising as a business issue isgreater than the current use of scanner data to helpaddress it.

The practitioners we interviewed highlighted threekey questions that they would like to be able to addresswith scanner data: (1) the short-term effect of advertis-ing on brand sales, (2) the long-term effect of advertis-ing, and (3) the indirect effects of advertising (e.g., theeffects on the health or equity of a brand). Answers tothese questions have implications for advertisingspending, allocation of resources between promotionand advertising, and evaluation of advertising content.



Short-Term Effect of AdvertisingTwo broad approaches are used by practitioners to as-sess advertising effectiveness: an experimental ap-proach and a regression-based approach. An exampleof the experimental approach is the split-cable method,where the media weight, creative, and scheduling allcan be manipulated across different households in acity and their response measured through sales datacaptured by scanners. A detailed discussion is pro-vided in Lodish et al. (1995), who describe a meta-analysis of 389 split-cable experiments conducted byIRI. The second method for analyzing the effectivenessof advertising is to include advertising spending mea-sures or Gross Rating Points (GRPs) as independentvariables in the regression model as described in Equa-tion (1).

Considerable disagreement exists among practition-ers regarding short-run advertising effect sizes. Someexecutives we interviewed stated that their analysesshowed no short-term effect of advertising on sales(e.g., PepsiCo). To these managers, such results wereneither surprising nor problematic since they believedthe purpose of advertising is to positively impact thebrand in the long run. Another view holds that adver-tising is a “sacred cow,” not to be sacrificed at any cost,and that any advertising is an absolute good (e.g.,P&G). A middle-ground view taken by many othersholds that advertising has a small, but significantshort-term impact on sales. Several practitioners toldus that they have found elasticities to average about0.10 to 0.12 for mature products and 0.20 to 0.40 fornew products. A study by Frito-Lay found the adver-tising impact to be significantly greater for smallbrands, and for brands that have some “news” to offer(Riskey 1997).

Academic studies are generally consistent with themiddle-ground—small but significant short-run elas-ticities. For example, based on a meta-analysis, Assmuset al. (1984) found average short-term advertising elas-ticities for established products to be about 0.15. Basedon 389 real-world split-cable TV ad experiments,Lodish et al. (1995) concluded that the average adver-tising elasticity is 0.05 for established products and0.26 for new products. We emphasize that these num-bers are averages. For example, although the averageelasticity for established products in the Lodish et al.

study was 0.05, they also found that these effects werenot statistically significant for 67% of the products. Inother words, while academic studies tend to agree onthe average short-term effect sizes of advertising, thereis still considerable uncertainty regarding specific prod-ucts or campaigns.

This conflict is represented by several studies.Lodish et al. (1995) are convinced that increasing ad-vertising budgets in relation to competitors does notincrease sales in general. In contrast, some academicstudies (e.g., Pedrick and Zufryden 1991) as well asindustry studies (e.g., research by ASI Market Researchand Media Marketing Assessment Inc. (MMA)) showstrong evidence of advertising weight or GRP on sales.It is difficult to say whether these conflicting findingsarise due to (a) differences in methodology (split-cablefor Lodish et al., logit model for Pedrick and Zufryden,and regression-based models for ASI and MMA), (b)differences in product categories studied, or (c) differ-ences in the level of aggregation—both time (e.g., Lod-ish et al. use annual data, while Pedrick and Zufrydenuse weekly data) and geographic aggregation (e.g.,Pedrick and Zufryden use panel data, while MMA typ-ically uses market-level data).

A major source of controversy is the appropriatelevel of aggregation at which to study advertising ef-fects with regression analysis. Many companies, suchas Nielsen and IRI, argue in favor of store- or account-level analysis. They claim that this is consistent withthe analysis conducted to assess the effectiveness ofprice and promotions (which vary by stores/accounts).Consequently, it is also appropriate to assess advertis-ing effects after controlling for sales effects due to priceand promotions. Others (MMA in particular) arguestrongly in favor of market-level analysis. They sug-gest that since advertising decisions are made at themarket level—not at the store or account level—differ-ences in advertising weight or copy can only be ob-served at the market level.

While Nielsen and IRI have store- and account-leveldata, third-party consultants such as MMA usuallyconduct their analysis on the market-level data towhich they are given access. Consequently, it is diffi-cult to disentangle the self-interest of the respectiveparties from the technical merits of the approaches.Though their paper is not about advertising per se,



Christen et al. (1997) have shown that the aggregationof store-level data to the market level can create biasesin estimated effect sizes of marketing variables (e.g.,feature and display). A careful comparative study ofthe estimation of advertising effect sizes with store-level versus market-level data would be of substantialinterest to the wide variety of practitioners who comeinto contact with this issue.

Another controversy involves experimental ap-proaches to assessing advertising effect sizes. Recently,John Philip Jones (1995) proposed an approach to mea-sure the effect of advertising using a method calledSTAS (Short Term Advertising Strength). Using AC-Nielsen single-source panel data, the STAS procedurefirst estimates baseline and stimulated STAS. The base-line STAS is a brand’s share of purchase occasionsamong households who have not been exposed to tele-vision advertising for it during the seven days beforepurchase. The stimulated STAS is the brand’s shareamong households who have had at least one televi-sion ad exposure for the brand in the seven-day period.The difference between the baseline and stimulatedSTAS is a measure of the short-term impact of adver-tising. The STAS procedure has been criticized byLodish (1997) for lack of random assignment of house-holds to advertising conditions. He advocates the useof a split-cable approach, such as IRI’s BehaviorScan.This continuing controversy (see, e.g., Jones 1998 andLodish 1998) provides another example of the impor-tance of resolving the measurement issues pertainingto the assessment of advertising effects.

Long-Term Effects of AdvertisingThree issues arise in the study of long-term effects ofadvertising: (a) how large is the long-term effect, i.e.the effect size (if the long-term effect exists), (b) howlong does the advertising effect last, i.e. the duration,and (c) what are the sources of long-term effect. Al-though most practitioners agree that advertisingshould have a long-term effect on sales, few know howto measure it. Managers told us that they consider thelong-run effects of advertising so difficult to quantifythat they make little attempt to measure them. Someresearch companies, such as Millward Brown Inter-national, are an exception. Based on its client work andproprietary model, Millward Brown suggests that the

long-term effect of advertising is about two times theshort-term effect, and in some cases these effects couldbe as large as eight times the short-term effect (Hollis1997, Scott and Ward 1997).

Academic studies have typically used a distributed-lag (e.g. Koyck) model to assess the short- and long-term impact of advertising on sales. Assuming a geo-metric decay of advertising effect, the Koyck modelcan be written as:

S � � � bA � kS � � (2)t t t�1 t

where St is the sales at time t, At is advertising at timet, b is the short-term effect of advertising, and k is thelag coefficient or carry-over effect of advertising. It isthen easy to show that in this model, the long-termeffect of advertising is b/(1 � k).

Assmus et al. (1984) conducted a meta-analysis on128 models reported in 22 studies and reported an av-erage k of 0.468, which suggests the long-term effect tobe 1/(1 � 0.468) � 1.88 times the short-term effect (b).Using four product categories, Givon and Horsky(1990) estimated k to be between 0.25 to 0.82 with anaverage of 0.59. Based on three categories, DeKluyverand Brodie (1987) estimated k to lie between 0.50 and0.80 with a mean of 0.70. Using an eight-year stretchof panel data, Mela et al. (1997) estimated the long-term effect to be between 1.5 to 2 times the short-termeffect. Collectively, all these academic studies suggestthe long-term effect to be between 1.3 to 5.6 times theshort-term effect with the average close to 2.0.

Lodish et al. (1995) used an experimental approachto assess the long-term effect of advertising. They con-ducted 42 in-market split-cable experiments on estab-lished brands where the advertising weight waschanged for one year, and its sales effect (compared toa control cell) was monitored over the next two years.They found that, on average, the long-term effect (i.e.,two years following the test year) of advertising onsales was about two times its short-term (or test year)effect. In sum, all the academic studies corroborate theMillward Brown work, showing that, on average, thelong-term effects of advertising are about twice theshort-term effects. These effects, however, could be sig-nificantly larger for some brands and products.12

12Millward Brown finds that, in general, large brands have smallshort-term but large long-term effects.



The second issue related to the long-term effect ofadvertising is the duration or length of the effect. Thereare conflicting results in this area. In its split-cable ex-periments, Frito-Lay found that most of the effect ofadvertising occurs within six months (Riskey 1997). Incontrast, Millward Brown finds the long-term effectsof advertising have a half-life of well over six months(Hollis 1997). Within academia there is also little con-sistency on this issue. The classic study by Clarke(1976) found the 90% duration interval of advertising(i.e., the length of time it takes to realize 90% of theeffect of the advertising) to last 6–9 months. But morerecent studies by Lodish et al. (1995) and DeKimpe andHanssens (1995) both indicate that advertising effectsmay last over several years.

Lastly, there are few studies devoted to understand-ing and quantifying the sources of long-term effects ofadvertising. Nevertheless, there is some consistency inthe results among the work that has addressed thisissue. For example, both academics (Lodish et al. 1995)and practitioners (von Gonten and Donius 1997) agreethat the long-term effects of advertising are more at-tributable to increases in consumer buying rates (i.e.,more sales to existing customers) than to higher pen-etrations of the brand into nonusers.

Indirect Effects of AdvertisingMany managers we interviewed said that they, andespecially their advertising agencies, focus on a num-ber of dependent variables other than sales when as-sessing advertising effectiveness. For example, manyexecutives feel that the role of advertising is not to in-fluence sales directly but to create awareness and actas a reminder at the time of purchase. The actual salesare then a function of price, promotion, product qual-ity, and distribution. Since our focus is scanner data,which does not include scores on awareness, recall,and persuasion, we will not discuss most of these in-direct effects.

One indirect effect that has been examined withscanner data is the effect of advertising on brand eq-uity. While the long-term effect of advertising maycapture some aspects of this effect on brand equity, itis far from a complete representation. There is an on-going debate within both industry and academia abouthow to measure brand equity and advertising’s effect

on it. We found many executives concerned aboutbrand equity issues but few had any sound method-ology to address them.13 Some executives suggestedusing the ratio of own-price to cross-price elasticity asa measure of brand equity (in this approach one caneither make brand equity relative to each competingbrand, or use cross-price elasticity of a generic brandas a benchmark). A few firms offer models of brandequity (e.g., BrandBuilder by The NPD Group) but theacceptance of these models and approaches is lim-ited.14 Academic research on this topic is also far fromconclusive. Scanner-based research has considered the“intercept” in the logit model as a possible surrogatefor brand equity (Kamakura and Russell 1993). Usingthis surrogate, Jedidi et al. (1999) found that advertis-ing has a significant positive impact on brand equitywhile long-term promotion has a negative impact onbrand equity. While these preliminary results are in-tuitively appealing, more research is clearly needed.

Another indirect effect of advertising is on consum-ers’ current and future purchase behavior. For exam-ple, many academic studies have shown that nonpriceadvertising reduces consumers’ price sensitivity (Kauland Wittink 1995). Practitioners with whom we dis-cussed this issue also believe this to be true and severalcompanies rely on analysis by Nielsen and IRI to mea-sure this effect. A remaining issue is how this indirecteffect should be captured when attempting to assessthe overall (i.e., both direct and indirect) effectivenessof advertising.

Summary for AdvertisingThere is remarkable consistency between the academicand “middle-ground” industry views regarding theaverage size of short-run advertising elasticities (0.05–0.15 for mature products and 0.20–0.40 for new prod-ucts). Moreover, most of the academic as well as theindustry studies on the subject agree that, on average,the long-run effect of advertising is approximately two

13Corroborating this, Davidson and Stacey (1997) report that mea-surement and management of brand equity ranked ninth out of 15marketing issues in importance, but fourteenth out of 15 in use ofscanner data analysis and modeling.14As is often the case with proprietary industry models, it is impos-sible for outsiders to evaluate them.



times the short-run effect. Both practitioners and aca-demics also agree that ad copy makes a significant dif-ference in the impact of advertising on sales, the sourceof long-run effects is due more to increasing buyingand not increased penetration, and that advertising de-creases price sensitivity. Nevertheless, practitionersview the assessment of all advertising effects withscanner data to be highly unresolved, perhaps due tothe difficulties involved in measuring them in theirown specific businesses. This is highlighted by the con-troversies surrounding both the regression-based ap-proaches (i.e., the appropriate level of data aggrega-tion) and other approaches (i.e., experimental as insplit-cable versus nonexperimental as in STAS). Thereis also some debate about the correlation between re-call and persuasion scores and brand sales, the dura-tion of advertising effects, and how to capture the im-pact of advertising on brand equity. Accordingly, weplace the average size of advertising effects in thelower-left quadrant of Figure 1 (resolved for academ-ics, unresolved for practitioners), while the specificmeasurement issues appear in the lower-right quad-rant (unresolved for both groups).

6. ProductFor the managers we interviewed, the most importantissue in product policy and strategy is the rationali-zation of product lines and limiting product prolifer-ation. As companies added new SKUs to their lineswithout deleting older items, significant product pro-liferation occurred to the point where many categories,such as toothpaste, now have as many as 1,200 sepa-rate UPC items. Several studies have shown (e.g.,Cummings et al. 1990) that assuming the 80/20 rule(i.e., 80% of the profit is accounted for by 20% of theitems), such product proliferation not only adds sig-nificant cost to the system but it also creates confusionamong consumers. In addition, retailers are demand-ing that manufacturers rationalize product lines in or-der to obtain and hold supermarket shelf space. Recentindustry trends, such as the movement to EfficientConsumer Response (ECR), also highlighted this issuefor most of the practitioners we interviewed.

Corroborating the views of the practitioners we in-terviewed, the Davidson and Stacey (1997) study

ranked product assortment/SKU optimization sixth inimportance and fifth in the use of scanner data analysisand modeling. Their study also included new productintroductions as an issue, and it ranked very close tothe assortment/SKU optimization issue: fourth in im-portance and sixth in the use of scanner analysis andmodeling. Many of the issues involved in new productintroduction (especially for line and brand extensions)overlap extensively with product assortment and SKUoptimization. With respect to launching new products,many practitioners said that they rely on traditionalsurvey-based methods for concept testing.

Industry and Academic PerspectivesA few industry studies have been conducted recentlyto examine the impact of SKU reduction on categorysales. The Food Marketing Institute reported that re-ducing the number of SKUs in six test categories inthree retail chains resulted in no significant loss in cate-gory sales (Kurt Salmon Associates 1993). A similarresult was reported by Progressive Grocer (Krum 1994).The executives we interviewed shared the belief thatmany categories could be trimmed of SKUs with littleeffect on sales. Nevertheless, there remains a reluc-tance to cut SKUs in categories that are highly com-petitive or where supermarket shelf space is scarce.

Academic studies that have addressed this issuehave reached roughly the same conclusion. Dreze et al.(1994) customized shelf facings of SKUs while deletingapproximately 10% of the less popular items. Theyfound that this change resulted in a 4% increase in sales.Based on lab experiments, field tests, and surveys,Broniarczyk et al. (1998) concluded that eliminating asmany as 50% of the items had no significant impact onconsumers perceptions or purchase behavior as longas category shelf space was held constant and mostconsumers could find their favorite items. On the otherhand, a number of studies have argued that consumersdemand a wider assortment of varieties when they buylarge quantities of a product (Simonson 1990,Simonson and Winer 1992, Walsh 1995). Behaviorally,this phenomenon results from consumers’ uncertaintyabout future consumption preferences. If this were alarge effect in the marketplace, it would provide acounterargument for retaining large product assort-ments. Examining the phenomenon in the yogurt cate-gory, Bucklin et al. (1998) show that the size of assort-ment effect is modest—even with a 50% increase in



total quantity purchased. Taken together, these studiesare consistent with the industry perspective that a re-duction in the number of SKUs is unlikely to affectcategory sales.

Given the agreement on the need to rationalize prod-uct assortment and reduce the number of SKUs, thenext question is how to decide which items to elimi-nate. Our discussion with the practitioners suggestedthat they follow a simple, and somewhat naive, pro-cedure of deleting, say, the bottom third (in terms ofsales or profits) of the items in a category. Some re-tailers have adopted this approach unless they obtainspecific guidance from manufacturers. Since this ap-proach ignores market structure issues (i.e., where dosales go when an item is deleted), it does not neces-sarily provide the retailer or manufacturer with thebest reduced assortment of items.

One promising approach to this problem is due toFader and Hardie (1996). Their approach combines thefeatures of the logit model with certain aspects of con-joint analysis and is currently in widespread use at IRI.Traditional logit models (e.g., Guadagni and Little1983) estimate a constant for each item, brand-size, orSKU. This model has severe limitations when the num-ber of SKUs is very large, a situation which is nowtypical in most packaged goods categories. Most aca-demic studies have dealt with this problem by eithereliminating low share items or by aggregating overthem in some fashion. But this approach is often in-compatible with the needs of practitioners becausetheir focus is on understanding and enhancing the po-sition of specific SKUs. Fader and Hardie (1996) sug-gest a solution by characterizing an SKU as a combi-nation of its product attributes (e.g., brand name,package size, flavor, variety, type, etc.). This respeci-fication significantly reduces the complexity of thelogit model for the product category. For example, ifthere are eight characteristics, each with two levels(e.g., package size: large and small), then the tradi-tional logit model will estimate 28 � 1 � 255 itemconstants. (Indeed, many of these may not be estimableif certain combinations do not exist in the marketplace.) The Fader and Hardie approach will need toestimate only eight parameters and, as in conjoint anal-ysis, also can forecast the share for a new combination

of product characteristics. Although this approach as-sumes that the product characteristics only have maineffects, and researchers may want to test this assump-tion, the approach is noteworthy because it is one ofthe first to be able to use scanner data to provide directguidance on product strategy.

Traditional approaches to new product develop-ment and testing (e.g., concept testing) still dominatethe decision-making process in the new products do-main. The Fader and Hardie (1996) model representsan important step in bringing the use of scanner dataand models into product policy and strategy. Never-theless, because scanners provide behavior and not in-tentions data, it may not be possible to forecast thepotential share of a truly new concept with scannerdata alone. In sum, we found little disagreement be-tween the academic and practitioner perspectives onproduct-related issues. Both agree that reducing thenumber of SKUs in a category may not damage sales,and that more study is needed on the best way to goabout optimizing a particular assortment.

7. Distribution and RetailManagement15

Executives in our survey raised three key issues re-garding distribution and retail management. Theseare, in order of importance, (a) category management,(b) account-specific insights, and (c) managing shelfspace. The most significant of these issues for manu-facturers and retailers is category management. Thisconcept has gained substantial favor since the KurtSalmon Associates (1993) study on efficient consumerresponse. (The study recommended four major areasfor improvement, one of which focused on categorymanagement.) The primary implication of categorymanagement is a shift in retailers’ attitudes from abuyer orientation (where money is made based on howthe product is bought), to a buyer and merchandiserorientation (where the focus is on category profitabil-ity). Davidson and Stacey (1997) report that categorymanagement was the second most important market-ing issue for the practitioners they surveyed, with 77

15Note that we did not interview retailers. This discussion focuseson manufacturers’ perspectives on how to create “win-win” strate-gies with retailers.



percent indicating that it was extremely important.Scanner data use, however, is only moderate with theissue ranking seventh out of 15 on this dimension.

A second key issue is the desire of the sales force foraccount-specific insights. Simply focusing on nationalor even market-level price elasticities is consideredmeaningless by most sales people. Skepticism aboutmarket-level results comes from the strong belief that“my market or my store is different and unique.”Davidson and Stacey (1997) reported that micromar-keting/store cluster analysis ranked twelfth out of 15issues in importance, but was eighth in terms of theuse of scanner data analysis and modeling.

With respect to the issue of shelf space, manufactur-ers and consultants both noted that retailers are grow-ing more interested in the competitive market struc-ture of products (e.g., tree-structured representationsof consumer switching behavior) which can help thembetter organize their displays. Indeed, manufacturerssay that retailers often expect them to provide this in-formation and related analyses. For example, retailerswant to know whether it is better to group all CocaCola and all PepsiCo soft drink SKUs together, versusgrouping the diet and nondiet SKUs together. Theanalyses needed to answer these questions must oftenbe conducted at the disaggregate level (either withscanner panel data or by survey). In the Davidson andStacey (1997) study, shelf-space allocation was rankedeleventh in overall importance and in the use of scan-ner analysis and modeling.

Category ManagementMost manufacturers perceive retailers’ sophisticationin category management to vary widely and that someretailers still focus on cost and gross margins. Never-theless, our interviewees reported that retailers arelooking closely at the role of a category in the successof a store, its traffic, customer penetration, and growth.This leads to questions about the role that a particularmanufacturer’s brand has in enhancing the overallprofitability of the category. At Procter and Gamble,executives said that they welcomed the change becauseit gives sales people and retailers an objective way todevelop win-win strategies. The fact that most P&Gbrands are market leaders—or near-leaders—in theircategories helps to reinforce this view.

Even though executives consider category manage-ment to be an important issue, few told us that theyrely on quantitative models to address this topic. Cur-rent practice is characterized by the simple approachof comparing winners and losers. In this approach, amanufacturer attempts to convince a retailer to makeits brand the “category captain” by contrasting thecategory sales of a store where its brand is the “cate-gory captain,” with another store where its brand isnot the category captain.

Several academic studies have explicitly addressedthe category management issue. Studies that decom-pose brand sales into category volume and brandshare/choice are also useful in addressing categorymanagement issues (e.g., Bell et al. 1999, Dillon andGupta 1996). Hoch et al. (1994) conducted field exper-iments in 26 product categories in a grocery chain with86 stores and found that while an every day low price(EDLP) strategy provides a small win for manufactur-ers (average 3% increase in unit sales), it reduces re-tailer profits by 18%. As retailers continue to move to-ward category management, these strategic issues arelikely to grow in importance for manufacturers as theyattempt to design “win-win” strategies. The decisionsupport system developed by Silva-Risso et al. (1999–this issue), for example, takes one step in this directionby optimizing the manufacturer’s promotion planssubject to the fulfillment of “win-win” constraints im-posed by the retailer.

Micromarketing and Account-Specific InsightsManufacturers report that their sales people want de-centralized decision making because they are closer tothe market and know the idiosyncrasies of each ac-count. Procter and Gamble is using an approach whereheadquarters decides on national and internationalstrategies (e.g., the budget allocation across advertis-ing versus promotion), but leaves micro decisions toindividual sales teams. Managers said this could bethought of as follows: the responsibility of the mar-keting department is to manage baseline sales (e.g.,brand equity), while the responsibility of the salesforce is to manage incremental sales. Implementingthis approach required equipping the sales force with,and training it in the use of, more advanced analyticaltools.

Similar to category management, current practice in



account-specific management typically involves thecomparison of winners and losers to search for causalfactors. For example, a company may match consumerand store demographics to see what types of consum-ers shop where, which stores grew, which stores didnot grow, and what they did differently in terms ofmerchandising, store layout, etc. Another example ofthis simple analysis includes an SKU productivitycomparison (across SKUs and across stores), whereshelf audit data are used to assess sales per square foot,share of shelf for own brand, “fair share” of shelf, andover/underdeveloped SKUs. These approaches pro-vide account-specific recommendations (e.g., accountA should promote more of item X) and also aid in thedesign of win-win strategies. Some manufacturers(e.g., P&G) are providing their retail accounts withanalyses of how the promotion of their brands can addto the store’s overall sales based on the type of buyerthat the brand attracts to the store. For the sales force,the appeal of these approaches is in their simplicity,ease of understanding and communicability to retail-ers, and their ability to provide major insights for largenumber of SKUs and stores in a short period of time.

A second approach is an innovative procedure pio-neered by practitioners. To develop better promotionplans for each retail account, one manufacturer, withthe help of a data supplier and an outside consultingfirm, designed a decision support system based on acreative combination of panel and store-level data.This approach consisted of five steps:

1. Use panel data to create preference segments ofconsumers. For example, 30% of consumers may beloyal to brand A, 20% may be very price sensitive andfrequently switch among brands, 10% may be loyal tobrand B, and so on.

2. Use panel data to find where consumers shop.3. Assess the percent of each type of consumer (e.g.,

loyal to A) that makes up a store profile. For example,store 1 may have a consumer profile of 10% loyal to A,5% loyal to B, and 85% switchers, whereas store 2 mayget mostly brand loyal consumers.

4. Cluster stores based on the similarity of their con-sumer preference profile. For example, one cluster ofstores may consist mainly of loyal consumers, anothercluster of stores may consist mainly of switchers, andso on.

5. Conduct analysis for each cluster of stores sepa-rately. For example, a pricing analysis would be doneseparately for cluster one, cluster two, etc. The implicitassumption is that stores in the same cluster will havea similar response function since the brand preferenceprofile of consumers is similar within each cluster.

Note that a store-by-store analysis may be infeasibledue to the limited number of observations for each in-dividual store, while an aggregate analysis may maskimportant differences across stores. Thus, the aboveprocedure provides one way of overcoming theseproblems. An alternative approach, advanced by ASIMarket Research, uses “modified pooled” modelswhere the data are pooled across accounts or marketsunless the data suggests otherwise (Richardson andDratfield 1997). Markets that deviate from the pooledresults are reported separately. This approach also at-tempts to strike a balance between one model for allstores/markets versus one model for each store.

Significant progress has been made in the academicliterature on this issue. For example, using weeklyscanner data for 18 categories for a chain of 83 super-markets, Hoch et al. (1995) correlate store-specific priceelasticities to consumer demographics and competitivevariables. They find that consumer demographics ex-plain a large proportion of the variance in price elas-ticities across stores. These results can be used to de-sign more effective everyday and promotional pricingstrategies that exploit store-level differences in pricesensitivity. Although academic studies have not spe-cifically looked at the issue of store segmentation, ap-proaches that employ latent class analysis for con-sumer segmentation could be used to segment storesbased on store-specific price and promotion respon-siveness (e.g., Wedel and Kamakura 1998). Similarly,Bayesian models could also be used to address theproblem of determining elasticities for a specific store.

Managing Shelf SpaceBecause of requests from retailers for informationabout “how consumers shop the category,” the class oftechniques known in the academic literature as marketstructure analysis is attracting renewed attention frompractitioners. The analysis seeks to reveal the under-lying substructure of a product category, usually interms of the degree of brand and/or item substitut-ability. Most approaches involve panel-data based



models such as Hendry (e.g., Kalwani and Morrison1977), tree models (e.g., Rao and Sabavala 1981), oroverlapping clusters obtained via latent class analysis(e.g., Grover and Srinivasan 1987, Kamakura andRussell 1989). Both IRI and Nielsen are well acquaintedwith the academic literature and methods in this areaand regularly provide custom studies for their clientson this topic.

Interestingly, these developments coincide with theintroduction of frequent shopper programs by majorsupermarket chains in the U.S. Indeed, some practi-tioners foresee that a key benefit to retailers will be thepanel-level information on transactions generated bythese programs and, hence, the ability to better under-stand the shopping behavior of consumers in theirstores. In the meantime, manufacturers reported thatretailers are increasingly using such measures as DPP(direct product profitability) and sales/profit persquare foot to allocate shelf space in their stores. As incategory management, comparing winners and losersis the most widely used approach.

Summary for Distribution and RetailWhile academics have recently begun to develop mod-els to address category management (e.g., Silva-Rissoet al. 1999—this issue), these are in the early stages andhave not yet found extensive use in practice. We clas-sify category management as unresolved from bothperspectives. Academics have made good progress inthe area of account-specific management or micromar-keting but, apart from a few innovative approaches,practitioners are generally using a simple comparisonof winners and losers. Consequently, we classify mi-cromarketing as resolved from the academic perspec-tive but unresolved from the practitioner perspective.In the area of shelf-space management, both industryand academia are in agreement on the benefits of usingmarket structure analysis, and we therefore classify itsuse as resolved from both perspectives.

8. ConclusionOur goals have been to investigate, distill, and presentthe practitioner community’s perspective on the use ofUPC scanner data and to contrast this with the per-spectives represented in the academic marketing lit-erature. For each of the issues raised by managers, we

reviewed the academic literature to assess whether ornot academics have developed models and/or analyt-ical procedures that have or readily could settle theissue. This leads to a 2 � 2 classification of the variousissues as resolved or unresolved from the perspectiveof both academics and practitioners (see Figure 1).Given the desire that many marketing academics havefor their research to acquire a broad impact, we hopethat seeking out and reporting back “the voice of thepractitioner” and comparing it to the academic mar-keting literature will help in guiding future research,both in the short and long run. We now turn to a sum-mary of what we believe to be the most important im-mediate and long-term research needs in this area.

Immediate Research NeedsWith respect to immediate needs for more academicresearch, we follow the unresolved items on the right-hand side of Figure 1 and propose the following:

1. Price Thresholds and Gaps. Academic re-searchers could extend existing models of price thresh-olds that are based on store-level and panel-level datato address practitioners’ questions regarding optimalprice points and price gaps versus competing brands.For example, this could be done by incorporating re-sponse models into decision support systems.

2. Baseline and Incremental Sales. Research isneeded to develop simple, robust models that will pro-duce better estimates of promotional sales that aretruly incremental for the manufacturer, not borrowedfrom the future, from another store, or from a sisterbrand. These models will also need to take into accountthe effect that promotions may have on consumptionin some product categories (e.g., stocking up on sodameans drinking more soda).

3. Base Price Elasticity. Studies are needed to de-termine whether or not base-price elasticities can bereliably estimated from scanner data and, if so, whatis the extent of natural price variation or number ofobservations needed. These estimates could also becompared with and tested against those obtained fromsurvey-based methods such as discrete choice analysis.

4. Competitive Reactions. Omission of competi-tive reactions may lead to biased estimates of market



response to a change in price (or other marketing vari-ables). Two potential approaches to this problem areto build game-theoretic models or to incorporate em-pirical reaction functions into models of marketresponse.

5. Advertising Measurement. Research isneeded to resolve aggregation issues in assessing ad-vertising effects. Determining the best approach to as-sess advertising effects with scanner data could resolvesome of the methods and data aggregation debate inthis area. The vast majority of practitioners we inter-viewed expressed high levels of frustration with theinability to get clean, consistent answers about adver-tising effectiveness from scanner data. As one managerat Nestle put it, “Somebody needs to solve this.”

6. Brand Equity. More research is needed intowhat can be learned about brand equity from the anal-ysis of scanner data. Choice models, for example, couldbe extended to capture consumer valuation of productfeatures as opposed to brand names, following the ap-proach developed by Fader and Hardie (1996). Aca-demics also will need to develop clear definitions re-garding precisely what brand equity represents andhow it should be measured.

7. Product Assortment. Manufacturers (and re-tailers) could benefit from methods to determine thecosts and benefits of broader versus narrower productassortment (e.g., number of flavors and/or varieties).

8. Category Management. Manufacturers needsimple models and decision support systems to auto-mate elements of the category management function.

Even though research on scanner data has been ac-tively pursued by academics for well over a decade,the length of the above list shows that there are sig-nificant gaps in our knowledge and numerous oppor-tunities for research projects of high academic andcommercial impact.

Long-Term NeedsIn addition to the topics described above, we believethat there are also several important long-term needsfor research and action. These include the following:

1. From Tactics to Strategy. Most of the scannerdata analysis to date has focused on short-term tactical

issues (e.g., deciding whether the price should be $2.99or $2.79). We believe that for this type of analysis togain greater attention with senior management, it mustalso address strategic issues such as brand equity, cus-tomer equity, competitive reactions, category expan-sion, and everyday pricing (EDLP) versus promotionalpricing (Hi-Lo).

2. From Sales/Share to Profit/Industry Sur-plus. Both academic studies and industry practice fo-cus almost exclusively on sales and share. Academicshave not studied the impact of marketing activities onprofits partly because suitable data has not been avail-able.16 Interestingly, practitioners told us that they hadsimilar problems: marketing managers seldom evalu-ate profit impact because it is hard to assess and allo-cate promotion costs. We believe that it is important tomove beyond sales analysis and to carefully study theimpact of price, promotions, and other mix elementson short- and long-run profits, and the implications forhow total industry surplus is shared among manufac-turers and retailers.

3. From Descriptive to Prescriptive Mod-els. Most scanner data models are descriptive in na-ture. Although these models provide broad guidelines(e.g., if your price elasticity is low in absolute terms, aprice reduction may not be advisable), they still leavemanagers uncomfortable about making specific deci-sions. A typical complaint along these lines is, “Youare telling me that the price elasticity for my brand is�2, but what do I do with it?” To aid managers inmaking these decisions, we need to develop prescrip-tive models that will incorporate the key aspects ofdescriptive models as well as assess competitive reac-tions, impact on profits, and long-run health of thebusiness.

4. From Method Confusion to Industry Stan-dards. We found packaged goods companies to bebombarded with a variety of methods from third-party

16The Marketing Science Institute attempted to assemble a “brand-level data base” in the early 1990s that would enable researchers tostudy the impact of marketing tactics and strategy on profitability.It reportedly ran into resistance from packaged goods companiesregarding the release of proprietary information on costs and prof-itability and the project was dropped.



consultants, the details of which are often not disclosedto clients or outsiders. This creates methods confusionand makes it impossible to compare results and resolvecontroversies. As we discussed above, this problem ismost acute in the area of advertising. We believe thatit may be quite helpful for an industry council (or simi-lar forum) to actively promote open discussion and de-bate to help establish methods standards for scannerdata analysis.

Accelerating the Adoption of Scanner AnalyticsA series of barriers to the diffusion of new analyticalapproaches were noted by practitioners. First, man-agers must have confidence that scanner analysis canoffer tangible advantages over judgment and othermethods. At PepsiCo, one executive said that the pre-vailing perception in the organization was that thescanner analysis did not necessarily improve the “winrate” for marketing decisions. Second, many manufac-turers commented on what they perceived to be thehigh costs for scanner data and custom studies. AtProcter and Gamble, executives said they viewed scan-ner analysis as a cost and that cheaper solutions areconstantly sought out. At Kraft, however, executivessaid that they view scanner analysis as an investment.A third barrier, at least from the perspective of the datasuppliers, was a concern that introducing a new anal-ysis technique may be met with unrealistic client ex-pectations. An executive at ACNielsen summarizedthe problem: “As soon as you make something newavailable, all of the worst problems line up first.” Prac-titioners also noted that the quantitative orientation ofkey decision makers (e.g., senior vice-president of mar-keting) is a predictive factor in a company’s use ofscanner data analysis (i.e., going beyond a basic score-card function). Finally, large organizations are thoughtto be bigger users of more advanced analysis tech-niques. In the words of one IRI executive, “Size cor-relates with [analytical] sophistication.”

Academics can help to speed diffusion in manyways. By stepping back and arriving at fundamentalsubstantive findings, academics can offer practitionersthe power of good social science to build confidence inresults from scanner data analysis. They can also de-termine the generalizability of empirical results (e.g.,meta-analyses on price elasticity or advertising elastic-ity; see Tellis 1988 or Lodish et al. 1995) which helps

provide reasonable bounds when estimating models.Methodologically, academics can work to make mod-els more robust and more parsimonious. Though morecomplex models can sometimes offer technical advan-tages, these must be carefully weighed against the neg-ative effect that complexity has on commercial adop-tion. Academics can also show how managementjudgment can be incorporated into the modeling pro-cess (e.g., Blattberg and Hoch 1990) and they can pro-vide impartial assessments of different approaches,helping to resolve controversies and technical debates.Perhaps most importantly, in their teaching roles theycan provide learning experiences that give current andfuture managers confidence that using analyticalmethods indeed leads to better decision making.17

ReferencesAbraham, Magid M., Leonard M. Lodish. 1990. Getting the most out

of advertising and promotion. Harvard Bus. Rev. 68 (May–June)50–60.

——, ——. 1993. An implemented system for improving promotionproductivity using store scanner data. Marketing Sci. 12 (Sum-mer) 248–269.

Ailawadi, Kusum, Scott A. Neslin. 1998. The effect of promotion onconsumption: buying more and consuming it faster. J. Market-ing Res. 35 (August) 390–398.

Allenby, Greg M. 1990. Hypothesis testing with scanner data: theadvantage of Bayesian methods. J. Marketing Res. 27 (Novem-ber) 379–389.

Assmus, Gert, John U. Farley, Donald R. Lehmann. 1984. How ad-vertising affects sales: a meta analysis of econometric results. J.Marketing Res. 21 (February) 65–74.

Bell, David R., Jeongwen Chiang, V. Padmanaban. 1999. The decom-position of promotional response: an empirical generalization.Marketing Sci. 18 (4) forthcoming.

Blattberg, Robert C., Stephen J. Hoch. 1990. Database models andmanagerial intuition: 50% model, 50% manager. ManagementSci. 36 887–899.

——, Edward I. George. 1991. Shrinkage estimation of price and pro-motional elasticities: seemingly unrelated equations. J. Amer.Statist. Assoc. 86 (June) 304–315.

——, Scott Neslin. 1990. Sales Promotion: Concepts, Methods, Strategies.Prentice-Hall, Englewood Cliffs, NJ.

——, Kenneth J. Wisniewski. 1989. Price-induced patterns of com-petition. Marketing Sci. 8 (Fall) 291–309.

17The authors thank the Marketing Science Institute for funding andthe practitioners at participating firms for their time. We also thankthe editor, area editor, and two anonymous reviewers for their com-ments and suggestions.



Bolton, Ruth N. 1989. The relationship between market characteris-tics and promotional price elasticities. Marketing Sci. 8 (2) 153–169.

Bresnahan, Timothy. 1997. Testing and measurement in competitionmodels. David M. Kreps, Kenneth F. Halls, eds., Advances inEconomics and Econometrics: Theory and Applications. CambridgeUniversity Press, Great Britain.

Broniarczyk, Susan, Wayne D. Hoyer, Leigh McAlister. 1998. Con-sumers’ perceptions of the assortment offered in a grocery cate-gory: the impact of item reduction. J. Marketing Res. 35 (May)166–176.

Bucklin, Randolph E., Sunil Gupta, S. Siddarth. 1998. Modeling theeffect of purchase quantity on consumer choice of product as-sortment. J. Forecasting 17 (June/July) 281–301.

——, James M. Lattin. 1992. A model of product category competi-tion among grocery retailers. J. Retailing 68 (Fall) 271–293.

——, Gary J. Russell, V. Srinivasan. 1998. A relationship betweenmarket share elasticities and brand switching probabilities. J.Marketing Res. 35 (February) 99–113.

Carpenter, Gregory S., Lee G. Cooper, Dominique M. Hanssens,David F. Midgley. 1988. Modeling asymmetric competition.Marketing Sci. 7 (Fall) 393–412.

Cattin, Phillipe, Dick R. Wittink. 1982. Commercial use of conjointanalysis: a survey. J. Marketing 46 (Summer) 44–53.

Christen, Markus, Sachin Gupta, John C. Porter, Richard Staelin,Dick R. Wittink. 1997. Using market-level data to understandpromotion effects in a nonlinear model. J. Marketing Res. 34 (Au-gust) 322–334.

Clarke, D. G. 1976. Econometric measurement of the duration of ad-vertising effects on sales. J. Marketing Res. 10 250–261.

Cooper, Lee G. 1988. Competitive maps: the structure underlyingasymmetric cross elasticities. Management Sci. 34 (June) 707–723.

Cummings, Peter, David White, Stefan Wisniowski. 1990. Strategicsimplicity. McKinsey Quart. (3) 80–90.

Davidson, Timothy A., E. Craig Stacey. 1997. 1996 marketing anal-ysis and modeling survey. Summary Report, Prevision Cor-poration, Wellesley, MA.

DeKimpe, Marnik G., Dominique M. Hanssens. 1995. The persis-tence of marketing effects on sales. Marketing Sci. 14 (Winter)1–21.

DeKluyver, C. A., R. J. Brodie. 1987. “Advertising-versus-marketingmix carryover effects: an empirical investigation. J. Bus. Res. 15269–287.

Dillon, William R., Sunil Gupta. 1996. A segment-level model of cate-gory volume and brand choice. Marketing Sci. 15 (1) 38–59.

Dreze, Xavier, Stephen J. Hoch, Mary E. Purk. 1994. Shelf manage-ment and space elasticity. J. Retailing 70 (4) 301–326.

Ehrenberg, A. S. C., L. R. England. 1990. Generalizing a pricing ef-fect. J. Indust. Econom. 34 (September) 47–68.

Fader, Peter S., Bruce G. S. Hardie. 1996. Modeling consumer choiceamong SKUs. J. Marketing Res. 33 (November) 442–452.

Foekens, Eijte W., Peter S. H. Leeflang, Dick R. Wittink. 1994. A com-parison and an exploration of the forecasting accuracy of a log-linear model at different levels of aggregation. Internat. J. Fore-casting 10 (September) 245–261.

——, ——, ——. 1999. Varying parameter models to accommodatedynamic promotion effects. J. Econometrics 89 (March–April)249–268.

Gabor, Andrew, C. W. J. Granger. 1964. Price sensitivity of the con-sumer. J. Adv. Res. 4 (December) 40–44.

Givon, Moshe, Dan Horsky. 1990. Untangling the effects of purchasereinforcement and advertising carryover. Marketing Sci. 9 (2)171–187.

Grover, Rajiv, V. Srinivasan. 1987. A simultaneous approach to mar-ket segmentation and market structuring. J. Marketing Res. 24(May) 139–153.

Guadagni, Peter M., John D. C. Little. 1983. A logit model of brandchoice calibrated on scanner data. Marketing Sci. 2 (Summer)203–238.

Gupta, Sachin, Pradeep Chintagunta, Anil Kual, Dick R. Wittink.1996. Do household scanner data provide representative infer-ences from brand choices: a comparison with store data. J. Mar-keting Res. 33 (November) 383–398.

Gupta, Sunil. 1988. Impact of price promotions on when, what andhow much to buy. J. Marketing Res. 25 (November) 342–356.

Hoch, Stephen J., Xavier Dreze, Mary E. Purk. 1994. EDLP, Hi-Lo,and margin arithmetic. J. Marketing 58 (October) 16–27.

——, J., Byung-Do Kim, Alan M. Montgomery, Peter E. Rossi. 1995.Determinants of store-level price elasticity. J. Marketing Res. 32(February) 17–29.

Hollis, Nigel S. 1997. Looking longer-term: how best to identifylonger-term advertising effects. Presentation made at The Ad-vertising Research Foundation Conference, New York, June.

Inman, J. Jeffrey, Leigh McAlister, Wayne D. Hoyer. 1990. Promotionsignal: proxy for a price cut? J. Consumer Res. 17 (June) 74–81.

Jedidi, Kamel, Carl F. Mela, Sunil Gupta. 1999. Managing advertisingand promotion for long-run profitability. Marketing Sci. 18 (1)1–22.

Jones, John Philip. 1995. When Ads Work: New Proof that AdvertisingTriggers Sales. Simon and Schuster, New York.

——. 1998. Point of view: STAS and BehaviorScan—yet anotherview. J. Adv. Res. 38 (March/April) 51–53.

Kadiyali, Vrinda, Naufel Vilcassim, Pradeep Chintagunta. 1999.Product line extensions and competitive market interactions: anempirical analysis. J. Econometrics 89 (March–April) 339–364.

Kalwani, Manohar U., Donald G. Morrison. 1977. A parsimoniousdescription of the Hendry system. Management Sci. 23 (January)467–477.

——, Chi Kin Yim. 1992. Consumer price and promotion expecta-tions: an experimental study. J. Marketing Res. 29 (February) 90–100.

Kalyanam, Kirthi, Thomas S. Shively. 1998. “Estimating irregularpricing effects: a stochastic spline regression approach.” J. Mar-keting Res. 35 (February) 16–29.

Kalyanaram, Gurumurthy, John D. C. Little. 1994. An empirical anal-ysis of latitude of price acceptance in consumer package goods.J. Consumer Res. 21 (December) 408–418.

Kamakura, Wagner A., Gary J. Russell. 1989. A probabilistic choicemodel for market segmentation and elasticity structure. J. Mar-keting Res. 26 (November) 420–434.



——, ——. 1993. Measuring brand value with scanner data. Internat.J. Res. in Marketing 10 (March) 9–22.

Kaul, Anil, Dick R. Wittink. 1995. Empirical generalizations aboutthe impact of advertising on price sensitivity and price. Mar-keting Sci. 14 (3) G151–G160.

Krum, Franklin. 1994. Quantum leap. Progressive Grocer (January)41–43.

Kurt Salmon Associates, Inc. 1993. Efficient Consumer Response. FoodMarketing Institute, Washington, DC.

Lal, Rajiv, V. Padmanabhan. 1995. Competitive response and equi-libria. Marketing Sci. 14 (3) G101–G108.

Leeflang, Peter S. H., Dick R. Wittink. 1996. Competitive reactionversus consumer response: do managers overreact? Internat. J.Res. in Marketing 13 (April) 103–120.

Leone, Robert P., Srini S. Srinivasan. 1996. Coupon face value: itsimpact on coupon redemptions, brand sales, and brand prof-itability. J. Retailing 72 (Fall) 273–289.

Little, John D. C. 1994. Modeling market response in large customerpanels. Robert C. Blattberg, Rashi Glazer, John D. C. Little, eds.,The Marketing Information Revolution. Harvard Business SchoolPress, Boston, MA.

Lodish, Leonard M. 1997. Point of view: J. P. Jones and M. H. Blairon measuring advertising effects—another point of view. J.Adv. Res. 37 (Sept./Oct.) 75–79.

——. 1998. STAS and BehaviorScan—it’s not just that simple. J. Adv.Res. 38 (March/April) 54–56.

——, Magid Abraham, Stuart Kalmenson, Jeanne Livelsberger, BethLubetlein, Bruce Richardson, Mary Ellen Stevens. 1995. HowT.V. advertising works: a meta-analysis of 389 real world splitcable TV advertising experiments. J. Marketing Res. 32 (May)125–139.

Mela, Carl F., Sunil Gupta, Donald R. Lehmann. 1997. The long-termimpact of promotion and advertising on consumer brandchoice. J. Marketing Res. 34 (May) 248–261.

Monroe, Kent B. 1990. Pricing: Making Profitable Decisions. McGraw-Hill, New York.

Montgomery, Alan. 1997. Creating micro-marketing pricing strate-gies using supermarket scanner data. Marketing Sci. 16 (4) 315–337.

Narasimhan, Chakravarthi. 1984. A price discrimination theory ofcoupons. Marketing Sci. 3 (Spring) 128–147.

——, Scott A. Neslin, Subrata K. Sen. 1996. Promotional elasticitiesand category characteristics. J. Marketing 60 (2) 17–30.

Narisetti, Raju. 1996. P&G ad chief plots demise of the coupon. WallStreet J. April 17, B1.

Neslin, Scott A. 1990. A market response model for coupon pro-motions. Marketing Sci. 9 (2) 125–145.

——, Caroline Henderson, John Quelch. 1985. Consumer promo-tions and the acceleration of product purchases. Marketing Sci.4 (Spring) 147–165.

——, Greg Allenby, Andrew Ehremberg, Steve Hoch, Gilles Laurent,Robert Leone, John Little, Leonard Lodish, Robert Shoemaker,Dick Wittink. 1994. A research agenda for making scanner datamore useful to managers. Marketing Letters 5 (4) 395–412.

——, Linda G. Schneider Stone. 1996. Consumer inventory sensitiv-ity and the postpromotion dip. Marketing Letters 7 (January) 77–94.

Pedrick, James H., Fred S. Zufryden. 1991. Evaluating the impact ofadvertising media plans: a model of consumer purchase dy-namics using single-source data. Marketing Sci. 10 (Spring) 111–130.

Rao, Ram C., Ramesh V. Arjunji, B. P. S. Murthi. 1995. Game theoryand empirical generalizations concerning competitive promo-tions. Marketing Sci. 14 (3) G89–G100.

Rao, Vithala R., Darius J. Sabavala. 1981. Inference of hierarchicalchoice processes from panel data. J. Consumer Res. 8 (June) 85–96.

Raju, Jagmohan S., V. Srinivasan, Rajiv Lal. 1990. The effects of brandloyalty on competitive price promotional strategies. Manage-ment Sci. 36 (March) 276–304.

Richardson, Kevin, Simon Dratfield. 1997. Bringing new capabilitiesto marketing analysis. Presentation made at The Advertising Re-search Foundation Conference, New York, June.

Riskey, Dwight R. 1997. How T.V. advertising works: an industryresponse. J. Marketing Res. 34 (May) 292–293.

Scott, Doug R., Kirk Ward. 1997. But which half is working? Mea-suring the impact of marketing activity on business results. Pre-sentation made at The Advertising Research Foundation Confer-ence, New York, June.

Sethuraman, Raj. 1996. A model of how discounting high-pricedbrands affects the sales of low-priced brands. J. Marketing Res.33 (November) 399–409.

Silva-Risso, Jorge M., Randolph E. Bucklin, Donald G. Morrison.1999. A decision support system for planning manufacturers’sales promotion calendars. Marketing Sci. 18 (3) this issue.

Simonson, Itamar. 1990. The effect of purchase quantity and timingon variety-seeking behavior. J. Marketing Res. 27 (May) 150–162.

——, Russell S. Winer. 1992. The influence of purchase quantity anddisplay format on consumer preference for variety. J. ConsumerRes. 19 (June) 133–138.

Srinivasan, Srini S., Robert P. Leone, Francis J. Mulhern. 1995. Theadvertising exposure effect of free standing inserts. J. Adv. 24(Spring) 29–40.

Tellis, Gerard J. 1988. The price elasticity of selective demand. J. Mar-keting Res. 25 (November) 331–341.

Thaler, Richard. 1985. Mental accounting and consumer choice. Mar-keting Sci. 4 (Summer) 199–214.

Von Gonten, Michael F., James F. Donius. 1997. Advertising expo-sure and effects: new panel-based findings. Presentation madeat The Advertising Research Foundation Conference, New York,June.

Van Heerde, Harald, Peter S. H. Leeflang, Dick R. Wittink. 1999. Theestimation of pre- and postpromotion dips with store-levelscanner data. Working Paper, Yale School of Management,New Haven, CT.

Walsh, John W. 1995. Flexibility in consumer purchasing for uncer-tain future tastes. Marketing Sci. 14 (2) 148–165.

Wansink, Brian, Rohit Deshpande. 1994. Out of sight, out of mind:



pantry stockpiling and brand-usage frequency. Marketing Let-ters 5 (January) 91–100.

Wedel, Michel, Wagner A. Kamakura. 1998. Market Segmentation:Conceptual and Methodological Foundations. Kluwer AcademicPublishers, Boston, MA.

Wittink, Dick R., Michael Addona, William Hawkes, John Porter.

1988. SCAN*PRO: the estimation, validation and use of pro-motional effects based on scanner data. Working Paper. ACNielsen, Schaumburg, IL.

——, Phillipe Cattin. 1989. Commercial use of conjoint analysis: anupdate. J. Marketing 53 (July) 91–96.

——, Marco Vriens, Wim Burhenne. 1994. Commercial use of con-joint analysis in Europe: results and critical reflections. Internat.J. Research in Marketing 11 (January) 41–52.

This paper was received December 17, 1997, and has been with the authors 10 months for 1 revision; processed by Dick Wittink.

Documents

Commercial Use of UPC Scanner Data: Industry and Academic