Intern. J. of Research in Marketingportal.idc.ac.il/en/main/research/ijrm/documents/pdf of...

Competitive information, trust, brand consideration and sales: Two field experiments☆

Guilherme Liberali a,b,⁎, Glen L. Urban b,1, John R. Hauser c,1

a Erasmus School of Economics, Erasmus University Rotterdam, 3000 DR, Rotterdam, The Netherlandsb MIT Sloan School of Management, Massachusetts Institute of Technology, E52-536, 77 Massachusetts Avenue, Cambridge, MA 02139, United Statesc MIT Sloan School of Management, Massachusetts Institute of Technology, E52-538, 77 Massachusetts Avenue, Cambridge, MA 02139, United States

a b s t r a c ta r t i c l e i n f o

Article history:First received in 3, April 2012 and was underreview for 2 monthsAvailable online 17 October 2012

Area Editor: Gary L. Lilien

Keywords:Competitive informationBrand considerationElectronic marketingInformation searchSignalingTrust

Two field experiments examine whether providing information to consumers regarding competitive prod-ucts builds trust. Established theory suggests that (1) competitive information leads to trust because it dem-onstrates the firm is altruistic, and (2) trust leads to brand consideration and sales. In year 1, an Americanautomaker provided experiential, product-feature, word-of-mouth, and advisor information to consumersin a 2!2!2!2 random-assignment field experiment that lasted six months. Main-effect analyses,conditional-logit models, and continuous-time Markov models suggest that competitive information en-hances brand consideration and possibly sales and that the effects are mediated through trust. However, ina modification to extant theory, effects are significant only for positively valenced information. The year-2 ex-periment tested whether a signal that the firm was willing to share competitive information would engendertrust, brand consideration, and sales. Contrary to many theories, the signal did not achieve these predictedoutcomes because, in the year-2 experiment, consumers who already trusted the automaker were more like-ly to opt in to competitive information. In addition to interpreting the field experiments in light of extant the-ory, we examine cost effectiveness and describe the automaker's successful implementation of revisedcompetitive-information strategies.

© 2012 Elsevier B.V. All rights reserved.

1. Introduction and motivation

In 2003–2005, an American automaker (“AAM”) faced a situationcommon among firms that had once dominated their industries. Newcompetitors had entered with products perceived to be higher in qual-ity and better matched to consumer needs. The automaker's brands nolonger connoted quality or innovation; brand strength had declined. By2003, over half of the consumers in the U.S. (and almost two-thirds inCalifornia) would not even consider AAM brands; its brands wereeffectively competing for only a fraction of the market. Ingrassia (210,p. 163) cites the lack of brand consideration as a cause of the declineof American brands. Although the automaker had recently investedheavily in design and engineering, the automaker would never again

regain strength nor market share unless its brands could earn con-sumers' brand consideration.

The situation of AAM was similar to that faced by manyonce-dominant brands. After the financial crises of 2008, consumerswere hesitant to consider and purchase financial services fromestablished firms such as New York Life (Green, 2010). Research inMotion (Blackberry), Motorola, and Nokia once dominated thesmart phone market, but lost market share and brand image toApple's iPhone. Many consumers reject these once-dominant firmsas no longer relevant. Firms entering dominated markets also facethe challenge of earning consumers' brand consideration. Examplesinclude Barnes & Noble's Nook (after Amazon's Kindle) and the Kin-dle Fire (after Apple's iPad). In other situations, a firm might stumble,perhaps through no fault of its own, and face an uphill battle to beconsidered as a viable alternative again. Examples include Tylenolafter the 1982 poisonings, the Audi Quattro after the 1986 publicityon sudden acceleration, Toyota after widespread media reports in2010 that software problems led to sudden acceleration, and Genentechafter production problems in critical drugs. In B2Bmarkets, the Brazilianaircraft manufacturer Embraer launched a new line of well-designedand high-quality business jets, but found it hard to enter buyers'consideration sets when competing with such established firms asBombardier, Dassault, Cessna, Hawker Beechcraft, and Gulfstream(Aboulafia, 2007).

Many authors recommend that firms earn brand consideration byfirst earning trust from consumers. If consumers trust a brand, then

Intern. J. of Research in Marketing 30 (2013) 101–113

☆ This research was supported by the MIT Sloan School of Management, the Center forDigital Business at MIT (http://ebusiness.mit.edu), and an anonymous U.S. automobilemanufacturer. We gratefully acknowledge the contributions of our industrial collabora-tors, research assistants, and faculty colleagues: Hunt Allcott, Eric Bradlow, MichaelBraun, Luis Felipe Camargo, Daria Dzyabura, Patricia Hawkins, Xitong Li, Natalie Mizik,Nick Pudar, Daniel Roesch, Dmitriy A. Rogozhnikov, Joyce Salisbury, Duncan Simester,Catherine Tucker, and Juanjuan Zhang.⁎ Corresponding author at: Tel.: +31 6 2271 2881; fax: +31 10 408 9169.

E-mail addresses: [email protected] (G. Liberali), [email protected] (G.L. Urban),[email protected] (J.R. Hauser).

1 Tel.: +1 617 253 6615.

0167-8116/$ – see front matter © 2012 Elsevier B.V. All rights reserved.http://dx.doi.org/10.1016/j.ijresmar.2012.07.002

Contents lists available at SciVerse ScienceDirect

Intern. J. of Research in Marketing

j ourna l homepage: www.e lsev ie r .com/ locate / i j resmar

consumers are motivated to invest the time and effort to learn moreabout the features of the brands (see, e.g., Urban, 2005 and citationstherein). Research suggests that to earn trust a firm should makeitself vulnerable by acting altruistically and putting consumers'needs above its own (Kirmani & Rao, 2000; Moorman, Deshpandé,& Zaltman, 1993; Rousseau, Sitkin, Burt, & Camerer, 1998). For exam-ple, New York Life stated publicly that “the guarantees we make last alifetime” and that “our promises have no expiration date” (Green,2010). Embraer, primarily a manufacturer of commercial regionaljets, took a risky strategy of announcing a new line of jets specificallydesigned for business travel (Aboulafia, 2007). Even Procter &Gamble's Global Marketing Officer began new initiatives because“market share is trust materialized” (Bloom, 2007).

The theory ofmaking oneself vulnerable is persuasive, but for amajorautomaker, such a strategy puts billions of dollars at risk. Most evidenceto date is based on cross-sectional surveys (structural equation modelsand meta-analyses) or laboratory experiments. For example, in a reviewof the literature, Geyskens, Steenkamp, andKumar (1998) cite 20 studiesbased on surveys, seven based on laboratory experiments, and nonebased on field experiments. Even today, we find few field experimentsand none in which a once-dominant firm made itself vulnerable togain trust. Before committing to a trust-based strategy, AAM decidedthat the strategy had to be proven in rigorous field experiments. Wewere interested in partnering because we felt the need to field-test the-ories that were developed from surveys and laboratory experiments.

Specifically, AAMwould signal that it was acting in the consumers'best interests by providing unbiased competitive information to itscustomers. If the theories were correct, consumers would come totrust AAM and consider its brands. The vulnerability signal wasrisky because AAM would not win all head-to-head comparisonswith its competitors and because goodwill was lacking among con-sumers familiar with pre-2003 products. However, if the theorieswere correct, competitive information would signal trust and have apositive impact. AAM's post-2003 products would win on sufficientlymany comparisons that brand consideration would overcome lossesdue to adverse short-term comparisons.

In this paper, we describe the field experiments and general les-sons. The year-1 data suggest that competitive information leads totrust, brand consideration, and sales but that the effect of competitiveinformation is mediated through trust. However, contrary to extanttheory, these effects are significant only for positively valenced infor-mation. The year-2 data suggest that contrary to many theories, asignal alone, indicating that the firm is willing to share competitiveinformation, does not engender trust, brand consideration, andsales. The signal was not effective because consumers who alreadytrusted the automaker were more likely to opt in to competitive in-formation. Managerial analyses suggest that competitive informationis effective when targeted cost-effectively to consumers who are oth-erwise skeptical of the brand. We begin with the theory.

2. Underlying theory

There is considerable precedent in the trust and signaling litera-tures to support the theory that providing competitive informationwould enhance trust, leading to brand consideration and sales.

2.1. Competitive information, vulnerability, and trust

Morgan and Hunt (1994) provide evidence that commitment leadsto trust and that the sharing of meaningful and timely information isan antecedent of trust. This classic article was of particular interest be-cause the Morgan–Hunt data were based on independent automobiletire retailers. However, the theories cut across industries. In a study ofmarket research suppliers,Moorman et al. (1993) suggest that sincerity,timeliness, and the willingness to reduce uncertainty (presumablyby sharing information) all lead to trust. Indeed, vulnerability to gain

trust is a common theme throughout the literature (Kirmani & Rao,2000; Moorman et al., 1993; Rousseau et al., 1998). Other researchersdefine trustwith the related concepts of credible information and actingbenevolently toward the other party (Doney & Cannon, 1997; Ganesan,1994; Ganesan & Hess, 1997; Geyskens et al., 1998; Gurviez & Korchia,2003; Maister, Green, & Galford, 2000; Sirdeshmukn, Singh, & Sabol,2002; Urban, 2004).

Providing competitive information to achieve vulnerability, credi-bility, and an image of altruism is a common recommendation (Trifts& Häubl, 2003; Urban, 2004; Urban, Amyx, & Lorenzon, 2009; Urban,Sultan, & Qualls, 2000). Bart, Shankar, Sultan, and Urban (2005) pro-vide evidence in online settings that clear, unbiased navigation andpresentation enhance trust. Shankar, Urban, and Sultan (2002)argue that quality and timeliness of information enhances trust. Inlaboratory experiments, Trifts and Häubl (2003) show that competi-tive price information enhances an online retailer's perceived trust-worthiness. These authors go on to state, citing other authors, that“the sharing of relevant and potentially self-damaging informationcan be viewed as a type of communication openness, which is an im-portant form of trust-building behavior and has been found to have adirect positive relationship with the level of trust” (p. 151).

However, gaining trust is only useful if it leads to sales. Fortunately,there is ample evidence supporting this conclusion. Bart et al. (2005)show that trust mediates relationships between website design charac-teristics and purchase intensions. Crosby, Evans, and Cowles (1990)show that enhanced trust leads to continued exchanges between partiesand to sales. Doney and Cannon (1997) demonstrate that trust influencesbuyer's anticipated future interactions. Ganesan (1994) shows that trustand interdependence determine the long-term orientation of both retailbuyers and their vendors. Geyskens et al. (1998), in a meta-analysis of24 papers (27 studies), suggest that trust mediates roughly 49% oflong-termoutcomes. Trifts andHäubl (2003) show that the effect of com-petitive price information is mediated through trust. Hoffman, Novak,and Peralta (1999, p. 85) posit that “the most effective way for commer-cial web providers to develop profitable exchange relationships with on-line customers is to earn their trust.” Sirdeshmukn et al. (2002) provideevidence that benevolent management policies and procedures enhancetrust and that trust enhances long-term loyalty. Büttner and Göritz(2008) provide evidence that trust enhances intentions to buy.

Based on this literature, we posit that sharing competitive infor-mation with consumers will demonstrate that AAM is both altruisticand vulnerable. Altruism and vulnerability will cause consumers totrust the automaker, which, in turn, will lead to brand considerationand sales. We also posit that the effect of competitive informationwill be mediated through trust. The following hypotheses are testedin the year-1 field experiment.

H1. Competitive information provided to the test group will increaseconsumers' trust of the automaker relative to the control group.

H2. Competitive information provided to the test group will increasebrand consideration relative to the control group.

H3. Conditional on brand consideration, competitive information pro-vided to the test group will increase sales relative to the control group.

H4. The increase in brand consideration and sales in the test group ismediated through trust.

The literature does not distinguish between positively valenced andnegatively valenced competitive information: both are assumed to com-municate altruism and vulnerability and hence lead to trust, brand con-sideration, and sales. However, we know that negative information perse decreases brand consideration and sales relative to positive informa-tion (e.g., automotive experiments by Urban, Hauser, & Roberts, 1990,p. 407). Thus, we expect that H1 through H3 are more likely to besupported for positively valenced competitive information than for

102 G. Liberali et al. / Intern. J. of Research in Marketing 30 (2013) 101–113

negatively valenced competitive information. A priori, we do not knowthe relative strengths of the negative information and the altruism vul-nerability effects; thus, it becomes an empirical question whether H1through H3 are supported for negatively valenced information.

2.2. Trust as a signal

Automotive competitive information comes in many forms, varyingfrom a simple list of competitive features to community forums oronline advisors, all the way to providing consumers the ability to testdrive competitive vehicles. Some data are inexpensive (simple lists)and some are quite costly (competitive test drives).Wewould thereforelike to disentangle the effect of actually providing competitive informa-tion from the act of sending a vulnerability (trust) signal by offering toprovide competitive information.

Kirmani and Rao (2000) provide a comprehensive review of theliterature on signaling unobserved product quality. These researchersargue that revenue-risking signals work when consumers can inferthat it is in the long-term revenue interests of high-quality firms toprovide the signal but not in the long-term revenue interests oflow-quality firms to do so. The theories require that (1) informationis hard to obtain pre-purchase, (2) quality can be assessed post-purchase, and (3) that the “bond” of vulnerability (supplied in equi-librium only by high-quality firms) is credible. Condition 2 is clearlymet in automotive markets. Condition 1 is met for expensive anddifficult-to-obtain information such as competitive test drives, but itmay not be met for simple lists of features. Condition 3 requiresthat the competitive information is expensive to provide so that con-sumers do not interpret the information as “cheap talk.” This condi-tion is met for the competitive information provided in AAM'sexperiments. Only an automaker with post-evaluation high-qualitybrands would want to risk providing competitive information thatwas expensive and otherwise hard to obtain.

Signaling theory is extensive, beginningwith Spence's (1973) classicarticle. See also Milgrom and Roberts (1986). In marketing, Biswas andBiswas (2004, p. 43) provide experiments that “signals are stronger re-lievers of risk in the online setting, especially for products with highnon-digital attributes.” Erdem and Swait (1998, p. 137) argue that thebrand itself can be a signal, especially if “the information about a brand'sposition that is communicated to the consumer by a firm [is] perceivedas truthful and dependable.” Erdem and Swait (2004) argue explicitlythat brand credibility, operationalized as trustworthiness and expertise,is a signal that increases brand consideration.

Given the strong support for a signaling theory of trust and thefact that the willingness to accept vulnerability by providing compet-itive information could signal that an automaker is a high-quality,trusted brand, we state the following hypotheses that we sought totest in the year-2 field experiment:

H5. A signal that competitive information is readily available in-creases trust in the test group relative to the control group.

H6. A signal that competitive information is readily available increasesbrand consideration in the test group relative to the control group.

H7. Conditional on brand consideration, a signal that competitive in-formation is readily available increases sales in the test group relativeto the control group.

H8. If H5 is true and if either H6 or H7 is true, then the increase in brandconsideration and/or sales due to the signal are mediated through trust(H8 is conditional on whether H5 through H7 are supported).

2.3. Types of competitive information

Urban et al. (2000) suggest that virtual advisors and complete-and-unbiased information on competitive products are important to

building trust. Urban et al. (2009) suggest that firms provide honestopen advice and information on competitive offerings. Häubl andTrifts (2000) provide evidence that virtual recommendation agents(advisors) and comparative product information assist consumers inbrand consideration. Arora et al. (2008) suggest further that informa-tion sources be personalized by the firm to the consumer (as in custom-ized brochures) or capable of being customized by the consumer (as inthe availability of drive experiences for a wide range of vehicles). We(and AAM) wanted to test different types of competitive information.The generic types were chosen to span the range of information avail-able in automotive markets (for managerial decisions made at AAM)and as representing generic types of information that would be avail-able in other categories (for generality). To avoid cheap talk, the infor-mation sources (in year 1) were all expensive for AAM to provide andincluded the following:

• Direct product experience. In automotive markets, this means testdrives or their equivalent, such as renting a car or truck. Unlike otherinformation sources, no automaker (at the time) provided competitivetest drives (auto showswere a poor substitute). Direct product experi-ence is typical in other high involvement (but less expensive) catego-ries. For example, Proctor & Gamble routinely sends out productsamples and encourages consumers to dodirect comparisonswith cur-rently used brands, and exercise equipment manufacturers place theirproducts in fitness centers to allow consumers to try them.

• Print and online information. In automotive markets, this takes theform of competitive brochures or information abstracted from thosebrochures. In othermarkets, such information is available by searchingproduct catalogs or from information aggregation websites, such asCNET in electronic goods.

• Word-of-mouth. In automotive markets (in 2003–2005), word-of-mouth information on competitors was available through online auto-motive communities. This type of information is a surrogate for infor-mation now available in Angie's List, Cyworld, Facebook, Google+,Yelp, and other social networks.

• Trusted advisors. Many websites provide unbiased online advisors(and some biased advisors). Such advisors are available in many prod-uct categories.

3. Year 1: randomized experiments for competitive information

The year-1 field experiment tested Hypotheses H1 through H4,whether competitive information leads to trust and trust leads tobrand consideration and sales. To ensure that the year-1 field exper-iment tests competitive information and not just the signal that com-petitive information is available, consumers in year 1 are givenincentives to experience the competitive information. The year-2field experiment tested Hypotheses H5 through H8, whether thesheer act of offering competitive information signals vulnerabilityand altruism which in turn leads to trust, brand consideration, andsales. In year 2, competitive information was made available, but con-sumers experienced the information only if they opted in.

3.1. Consumer panel observed over six months

The year-1 panel ran monthly from October 2003 to April 2004(this was five years prior to the bankruptcies of two American auto-makers, both of which are now profitable). Members of HarrisInteractive's panel were screened to be in the market for a new vehi-cle in the next year, on average within the next 6.6 months, and wereinvited to participate and complete six monthly questionnaires. Intotal, Harris Interactive enrolled 615 Los Angeles consumers ofwhom 317 completed all six questionnaires for an average comple-tion/retention rate of 51.5%. We were unable to obtain exact recruit-ment rate statistics for year 1, but Harris Interactive estimates aninitial recruitment rate of approximately 40%.

103G. Liberali et al. / Intern. J. of Research in Marketing 30 (2013) 101–113

Consumers were assigned randomly to four experimental treat-ments, each representing one of the four generic forms of competitiveinformation. In year 1 (but not year 2), consumers assigned to a treat-ment were given sufficient incentives to experience that treatment.Treatments were assigned in a fully crossed orthogonal design givingus a 2!2!2!2 full-factorial field experiment such that variousrespondents received 0, 1, 2, 3, or 4 treatments. Although assign-ments were random with a goal of 50–50 assignment, the logisticswere such that only approximately 40% of the panel members wererandomly assigned to competitive test drives. The other treatmentswere close to 50–50. By design, the competitive online advisor wasavailable in all months, the competitive community ran for all butthe last month, the customized brochures were mailed in months 2and 3, and the competitive test drive took place in month 4 (this dif-ferential assignment is analyzed with models that take account of thedifferential timing of assignments). The exact numbers of consumersassigned to each treatment in year 1 is summarized in Table 1.

3.2. Competitive information experimental treatments

All experimental treatments were produced and managed profes-sionally and required substantial investments ranging from approxi-mately $150,000 (brochures) to $1 million (competitive test drives).To avoid cheap talk, all treatments would be quite expensive on a na-tional basis. Direct product experience was represented by a testdrive experience at a California test track in which consumers coulddrive vehicles from Acura, BMW, Buick, Cadillac, Chevrolet, Chrysler,Dodge, Ford, Honda, Lexus, Lincoln, Mercedes, Pontiac, Saab, Saturn,Toyota, Volkswagen, and Volvo without any sales pressure (Fig. 1a).

Print and online information was represented by customized bro-chures (year 1) and competitive brochures (year 2). Customized bro-chures were glossy brochures tailored to the needs of individualconsumers and mailed directly to the consumers (Fig. 1b). Competitivebrochures were less customized, web-based, and included brochuresfrom all competitors. While the year-1 print information was not com-petitive information, AAM believed it would engender trust and signalaltruism. All other information is competitive.

Word-of-mouth information was represented by an unbiased on-line CommuniSpace™ forum in which consumers could participate inover 30 discussions about both AAM and competitive vehicles(Fig. 1c). Commensurate with concerns that such information mightmake the automaker vulnerable, approximately 20% of the commentsabout AAM brands were negative.

Trusted advisors were represented by an online advisor that wasco-branded with Kelley Blue Book and similar to that developed byUrban and Hauser (2004) (see Fig. 1d). In year 1, the online advisorrecommended competitors approximately 83% of the time.

3.3. Dependent measures: brand consideration and purchase

Brand considerationwasmeasuredwith drop-downmenus inwhichthe consumer indicated the brands that he or she would consider. Brand

consideration is a quantal measure (consider vs. not consider) where aconsumer is coded as considering AAM if the consumer indicated thathe or she would consider one of AAM's brands. To avoid demandartifacts, AAM was not identified as the sponsor of the study and AAMvehicles occurred throughout the list of potential vehicles. Because con-sumersmight evaluate and reject a brand, they can report no brand con-sideration in month t, even if they considered a brand in month t!1.Purchase (sales) is also a quantal measure (purchase or not purchase)measured with a standard stated purchase measure. Once consumerspurchase a vehicle, we assume they cannot un-consider and un-purchase that vehicle during the six-month observation period.

3.4. Trust

We hypothesize that trust is a dependent measure and a mediatorfor brand consideration and purchase (trust is a firm-specific propertyin our experiments). While the definition of trust varies widely in theliterature, a commondefinition is “a single, global, unidimensionalmea-sure of trust” (Geyskens et al., 1998, p. 225). We include a global mea-sure: “Overall, this company is trustworthy.” Other authors includebenevolence and credibility (e.g., Doney & Cannon, 1997; Ganesan,1994; Ganesan & Hess, 1997; Geyskens et al., 1998). We implementedbenevolence with “I believe that this company is willing to assist andsupportme,” andwe implemented crediblewith “Overall, this companyhas the ability to meet customer needs.” Sirdeshmukn et al. (2002)suggest that trust includes operational competence, which weimplemented with “The company is very competent in its dealingswith its customers.” Finally, Anderson and Narus (1990, p. 45) suggestthat the trustworthy firm “will perform actions that will result in posi-tive outcomes for the [consumer]” and Anderson andWeitz (1989) de-fine trust as a [consumer's] belief that “its needs will be fulfilled in thefuture by actions undertaken by the other party.”We implemented pos-itive outcomes with “This company makes excellent vehicles.” All fiveitems were measured with 7-point Likert questions. Consistency withone's products, listening to customers and placing customers ahead ofprofit are becoming increasingly important to consumers, as indicatedby the Edelman Trust Barometer (http://trust.edelman.com/what-the-2012-edelman-trust-barometer-means-for-purpose/).

Together, the five items reflected the goals of AAM on how theymight engender trust among consumers. Before we use the fiveitems to evaluate the impact of competitive information, we establishthat the five items form a unidimensional construct and that the con-struct is related to overall trust. Using methods recommended inChurchill (1979), we purify the scale. The scale is unidimensionalwith high construct reliability and maximized with all five items:Cronbach's !=0.95. As a further test, we compare a four-item scalewithout the global item; the reduced scale is highly correlated withthe global item ("=0.88). We therefore conclude that the fiveitems represent a single scale and that it is reasonable to call thescale “trust” for the purpose of evaluating the implications of compet-itive information. In the following analyses, trust is measured by thesum!score (divided by the number of items such that the trustscale ranges from 1 to 7).

3.5. Randomization tests in year 1

All models of consideration and purchase use treatment-assignmentdummies. Following standard experimental procedures, the impact of atreatment is measured for all respondents for whom the treatment wasavailable whether they experienced the treatment. This approach is con-servative and avoids effects that might be due to differential take-up ofthe treatments. For completeness, we compared treatment-assignmentanalyses to self-reported-treatment analyses. The pattern of coefficientsand their significance was similar for both analyses.

Qualitative data are consistentwith the hypothesis that take-upwasrandom in year 1. For example, a portion of consumers experienced

Table 1Consumers randomly assigned to treatments in year 1.

Number of respondents whowere assigned to the indicated treatment in thatmonth.

Treatment Month2

Month3

Month4

Month5

Month6

Treatmentcell

Competitivetest drives

Yes 0 0 124 0 0 124No 317 317 193 317 317 193

Customizedbrochures

Yes 164 164 0 0 0 164No 153 153 317 317 317 153

Competitiveforum

Yes 151 151 151 151 0 151No 166 166 166 166 317 166

Competitiveadvisor

Yes 156 156 156 156 156 156No 161 161 161 161 161 161


technical difficulties with the competitive online advisor, and a fewcould not come to the competitive test drive due to last-minute sched-uling issues. Take-up rateswere 91.1% for competitive test drives, 99.4%for brochures, 97.4% for the community forum, and 82.1% for the onlineadvisor. The high take-up rates and the similarity of coefficients suggestthat treatment take-up (given it was offered) was sufficiently randomthat take-up selection had little or no effect in year 1. We thus leaveanalyses of self-selected take-up to year 2 when the trust signal ismore cleanly implemented.

Although we use treatment assignments as independent mea-sures, it is useful to examine further whether take-up was random.Specifically, we examine consumers who (1) were not assigned to atreatment, (2) were assigned to a treatment but did not report partic-ipation, and (3) were assigned and reported participation. If therewere non-random take-up, then consumers in group (3) would be anon-random draw from (2 and 3). However, if that were true,non-random take-up from (2 and 3) to (3) would leave a non-random set of consumer behaviors in (2). We find no differences inthe dependent measures between groups (1) and (2), thus providingfurther evidence that take-up was sufficiently random. For example,measured brand consideration does not vary between groups (1)and (2) for competitive test drives (t=.05, p=.96), customized bro-chures (t=.60, p=.56), competitive forums (t=.90, p=.37), orcompetitive advisors (t=1.14, p=.26). All t-tests in this paper aretwo-tailed.

4. Main effects of treatments in year 1

We begin with treatment main effects from the fully crossed2!2!2!2 experiment. We explore interactions, dynamics, hetero-geneity, and other issues in Section 5.

Themain effects are summarized in Table 2. The first column is thepercent increase in consumers who consider an AAM vehicle at theend of the experiment. For example, among consumers assigned tocompetitive test drives, brand consideration increased by 20.5% rela-tive to consumers who were not assigned to competitive test drives.This difference is significant (t=3.6, pb .01). The treatment vs. con-trol difference is not significant for customized brochures, thecompetitive forum, and the competitive advisor. The increase in cu-mulative purchase follows the same pattern with an 11.1% differencefor competitive test drives (t=2.4, p=.02) and insignificant effectsfor the other treatments. The impact on trust is similar but not iden-tical. Competitive test drives increase trust (11.6%), but the competi-tive advisor has a negative effect (!8.9%), and both are significant atpb .01. Neither customized brochures nor the competitive forumchange trust significantly (we examine more-complete models andtrust mediation in Section 5).

We might posit the effect of competitive information to be eitherlarger or smaller among consumers who own AAM vehicles. This ef-fect might be larger because current vehicles are improved relativeto prior vehicles, but it might be smaller because consumers who do

Fig. 1. Year-1 (Random Assignment) Competitive-Information Treatments.


not own AAM vehicles have less experience with older, less well-received vehicles. Empirically, the data suggest that there is no interac-tion effect due to prior AAMownership, implying either that the two ef-fects cancel or that neither is strong. Competitive test drives is the onlytreatment with a significant impact among non-AAM owners, and themagnitude of that impact is virtually identical to the magnitudeamong all consumers (lower left of Table 2). We explore interactionsmore formally in Section 5. We find no differential impact due to ageor sex. The age–sex comparisons are not shown to simplify the tables.

Analyses of the main effects are the simplest and cleanest set ofanalyses. Heterogeneity is mitigated because of randomization intreatment assignment. However, we can improve insight by account-ing for dynamics, interactions among treatments, the conditioning ofpurchase on brand consideration, the potential for trust mediation,and heterogeneity.

5. Dynamics of trust, brand consideration, and purchase

5.1. Average trust, brand consideration, and purchase by month

Table 3 summarizes average trust, brand consideration, and pur-chase by treatment for months 2 to 6. The response to competitive in-formation is more complex. For example, trust increases substantiallyin month 4 among consumers who experience test drives anddeclines in months 5 and 6, but it does not decline to pre-month 4levels. A large change in brand consideration occurs in month 4when consumers experience competitive test drives, but the effectendured through subsequent months. These observations might bedue to an effect where competitive test drives increase trust whichthen decays slowly rather than immediately. In addition, in automo-tive markets, a consumer might come to a competitive test drive inone month, consider and seriously evaluate vehicles in the nextmonth, and purchase in a third month. To untangle these effects, weneed analyses beyond mere inspection of Table 3. We need analysesbased on a theory of automotive purchasing dynamics.

Table 3 reports themain effects, but the experiment is a fully crossed2!2!2!2 design with some consumers experiencing other forms ofcompetitive information in earlier months, some in later months, andsome not at all. More sophisticated analyses are necessary to accountfor the fully crossed design, potential interactions among treatments,and the time-varying nature of the treatments. Finally, the measures inTable 3 are averages and do not account for individual differences. Tountangle the true effect of competitive information we now model the

complexities of the dynamics of the automotive market, the complexityof the experimental design, and potential individual differences.

5.2. Trust dynamics

Trust builds or declines over time. A consumer might experience atreatment in month t and, as a result, increase his or her trust in abrand. However, the consumer may not trust the brand enough toconsider it. Another treatment in month t+1 might increase trustfurther and be enough to encourage brand consideration. To capturethis phenomenon, we model trust as a variable that can increase ordecrease over time as a result of treatments. Trust might also decay.Specifically, we model trust in month t as a sum of # times trust inmonth t!1 plus effects due to the treatments in month t, where #and the coefficients of the treatment effects are to be estimated(#"1). We provide detailed equations in Section 6.

5.3. Interactions among the treatments

Hauser, Urban, and Weinberg (1993) examine consumers' searchfor information concerning automobiles and find that the value ofan information source sometimes depends upon whether consumershad previously experienced another information source. Theseresearchers' data suggest two-level interactions but no three-levelinteractions. In light of this prior research, we examine models thatallow interactions among the treatments.

5.4. Individual differences among consumers

Although the treatments were randomized, consumers with dif-ferent purchase histories and different demographics (age and sex)might react differently. For example, consumers who now own AAMvehicles may base their trust, brand consideration, or purchase ontheir prior ownership experience. To capture purchase history,

Table 2Main-effect analyses in year 1 randomly assignment field experiment.

Brand consideration(difference inpercent in last month)

Purchase(cumulative differencein percent)

Trust(% lift in lastmonth)

TreatmentCompetitivetest drives

20.5%a 11.1% a 11.6%a

Customizedbrochures

3.2% 4.8% 5.6%

Competitiveforum

0.5% 3.3% 2.9%

Competitiveadvisor

!2.4% !4.4% !8.9%a

Treatment among non-AAM ownersCompetitivetest drives

20.0%a 7.3% 12.3%a

Customizedbrochures

2.2% 5.0% 3.9%

Competitiveforum

2.0% 6.1% 2.4%

Competitiveadvisor

1.1% !0.9% !5.7%

a Significant at the 0.05 level.

Table 3Average brand trust, brand consideration (%), and purchase (%) by month in year 1 ran-domly assignment field experiment (cells are in bold if the treatment was available inthat month).

Treatment Month 2 Month 3 Month 4 Month 5 Month 6

Average trust among respondents assigned to the indicated treatment in that month(five-item scale)

Competitive test drives Yes 4.97 4.94 5.17 5.14 5.07No 4.66 4.66 4.55 4.65 4.54

Customized brochures Yes 5.01 4.94 4.93 4.97 4.87No 4.53 4.58 4.65 4.69 4.61

Competitive forum Yes 4.81 4.87 4.78 4.82 4.82No 4.75 4.67 4.81 4.85 4.68

Competitive advisor Yes 4.62 4.55 4.54 4.60 4.52No 4.93 4.97 5.04 5.07 4.97

Brand consideration among respondents assigned to the indicated treatment in thatmonth (%)

Competitive test drives Yes 49.2 55.6 61.3 59.7 57.3No 37.3 38.3 41.5 36.8 36.8

Customized brochures Yes 43.9 43.9 52.4 47.6 46.3No 39.9 46.4 45.8 43.8 43.1

Competitive forum Yes 45.0 47.0 53.0 45.0 45.0No 39.2 43.4 45.8 46.4 44.6

Competitive advisor Yes 37.2 46.2 47.4 43.6 43.6No 46.6 44.1 50.9 47.8 46.0

Purchase among respondents assigned to the indicated treatment in that month (%)Competitive test drives Yes 5.6 2.4 12.1 3.2 4.8

No 2.6 6.2 3.6 2.1 2.6Customized brochures Yes 4.9 5.5 7.3 1.2 4.9

No 2.6 3.9 6.5 3.9 2.0Competitive forum Yes 4.0 5.3 7.9 2.6 3.3

No 3.6 4.2 6.0 2.4 3.6Competitive advisor Yes 3.2 5.1 6.4 2.6 1.9

No 4.3 4.3 7.4 2.5 5.0


we include dummy variables for “own other American” and “ownJapanese.” There is no dummy variable for “own European.”

Heterogeneity might also be unobservable. One way to correct forunobserved heterogeneity in the propensity to trust, consider, orpurchase anAAMvehiclewould be to computemonth-to-month differ-ences in trust, brand consideration, and purchase. However, both brandconsideration and purchase are quantal (0 vs. 1) measures andmonth-to-month differences would implicitly assume (a) no decayand (b) perfect reliability of repeated measures. With repeated noisymeasures, the best estimate of the true score at t is not the score att!1 but rather a function of reliability times the lagged score(Nunnally & Bernstein, 1994, 222). While we can never rule outunobserved heterogeneity completely, we can examine whetherunobserved heterogeneity is likely to be a major effect or a second-order effect. Specifically, (a) a coefficient close to one in a trust regres-sion (#=0.833, yet to be shown), (b) explicit controls for observableheterogeneity, (c) consistency with the main-effect analyses (yet tobe shown), and (d) continuous-time Markov analyses based ondifferences (yet to be shown) all suggest that effects due to unobservedheterogeneity are negligible and unlikely to reverse primary insights.

6. Modeling dynamics in year 1

Ignoring for a moment dynamics, persistence, more complete prior-ownership effects, interactions among treatments, and unobservedexternal shocks, we see that trust is correlatedwith both brand consider-ation ("=0.22) and purchase ("=0.17), both significant at the 0.01level. The correlation with lagged trust is higher for brand consideration("=0.61) but lower for purchase ("=.05). This result is consistent witha dynamic interpretation that trust at the beginning of a month is thedriver of brand consideration during that month. However, these simplecorrelations do not capture all of the dynamics and, technically, misuse acorrelation coefficient for two quantal outcomes (brand considerationand purchase).

To account for theoretical dynamics and to respect the quantal na-ture of the outcome variables, we use conditional-logit analyses ofbrand consideration and purchase (see Fig. 2). Specifically, we askwhether the treatments increase brand consideration and, amongthose consumers who consider AAM vehicles, whether the treatmentsalso affect purchase.

In the conditional-logit analyses, we include lagged brand consid-eration as an explanatory variable to focus on changes in brand con-sideration. We include dummy variables for observation months to

account for unobserved marketing actions and environmental shocks(month 1 is a pre-measure and the month-2 dummy variable is set tozero for identification). The month dummy variables also account forany measurement artifact that might boost brand consideration(e.g., “Hawthorne” effect). To account for observed heterogeneity inpast purchases, we include prior ownership of AAM, other American,and Japanese (relative to European) vehicles. Age and sex effectswere examined but suppressed to simplify Table 4, as they were notsignificant.

Let Rit be a measure of consumer i's trust in month t and let xijt=1 ifconsumer iwas assigned to treatment j inmonth t, and xijt=0 otherwise.Let yik=1 if consumer i has characteristic k, and let yik=0 otherwise. Let$t=1 in month t and $t=0 otherwise. Trust dynamics are modeled withEq. (1) where we estimate the #R, wj

R, vkR, utR, and bR:

Rit ! #RRi;t!1 "#4

j!1wR

j xijt "#K

k!1vRkyik "#

6

t!3uRt $t " bR: #1$

Let Cit=1 if consumer i considers an AAM brand in month t andCit=0 otherwise. Let Pit=1 if consumer i purchases an AAM vehiclein month t and Pit=0 otherwise. The conditional-logit models arespecified by Eq. (2):

Prob Cit ! 1# $ ! ef it

1" ef itwhere f it ! #CCi;t!1 "#

4

j!1wC

j xijt "#K

k!1vCk yik "#

6

t!3uCt $t " bC :

Prob!Pit ! 1 Cit ! 1j $ ! egit

1" egitwhere git ! #

4

j!1wP

j xijt "#K

k!1vPkyik "#

6

t!3uPt $t " bP :

#2$

6.1. Direct effects of treatments

We begin with main effects of the treatments, as shown in the firstand second columns of parameters in Table 4. The brand considerationanalysis includes all respondents. The purchase analysis is conditionalon brand consideration: only those respondents who consider AAM inthatmonth are includedwhen estimating the purchase logit, and the ef-fect on purchase is incremental above and beyond the effect on brandconsideration (standard errors available upon request).

Brand-consideration analyses explain substantial informationwith aU2 of 24.8% (U2, sometimes called a pseudo-R2, measures the percent ofuncertainty explained, Hauser, 1978). Brand consideration is increasedif consumers own AAM or other American vehicles and decreased ifthey own Japanese vehicles. Brand consideration is also higher inmonths 3 to 6 relative to month 2. The only significant direct treatmenteffect is due to competitive test drives. Purchase, conditional on brandconsideration, also increases with competitive test drives (marginallysignificant), but there are no direct effects of prior ownership ormonth of measurement on purchase. The purchase model explainsless uncertainty, suggesting that the treatments affected brand consid-eration more strongly than purchase.

6.2. Trust as a mediator

There is ample precedent in the literature for trust as a mediator ofpurchase or purchase intentions (e.g., Bart et al., 2005; Büttner &Göritz, 2008; Erdem & Swait, 2004; Morgan & Hunt, 1994; Porter &Donthu, 2008; Urban et al., 2009; Yoon, 2002). In a series of experiments,Trifts and Häubl (2003) demonstrate that competitive price informationaffects preference, but the effect on preference ismediated through trust.

We use the methods of Baron and Kenney (1986) to test whethercompetitive information treatments were mediated through trust. Spe-cifically, if the treatments affect trust and if the treatments affect brandconsideration (or purchase), we estimate a third model. We add an in-dicator of trust as an explanatory variable in the conditional-logitmodels. If the treatments are mediated through trust, then (1) the

Not Consider AAM Consider AAM

Consider and Purchase AAM

Fig. 2. Consideration and Purchase Dynamics: Conditional-Logit Analyses.


indicator of trust should be significant in the new models, and (2) thedirect effect of treatments on consideration (or purchase) should nowbe insignificant. Partial mediation includes (1) but requires only thatthe direct effect decrease in magnitude.

Wemust be careful when we add trust to the model. We use laggedtrust to be consistent with the dynamics of measurement and causality.Lagged trust has the added benefit that joint causality in measurementerrors is reduced because the trust measures occur in different monthsthan the brand consideration and purchase measures. Nonetheless, toaccount for unobserved shocks that affect trust in month t!1 andbrand consideration (purchase) in month t, we use estimated laggedtrust in an equation that predicts brand consideration (purchase) withreported treatments (see Online Appendix for other trust regressions).That is, we add R̂i;t!1 as an explanatory variable on the right-handsides of Eq. (2) where R̂i;t!1 is estimated with Eq. (1). Traditional medi-ation analyses use lagged trust directly. In our data, these tests also in-dicate mediation and have similar implications.

Wefirst examine the trust regression. Competitive test drives clearlyincrease trust, and there is evidence that customized brochures increasetrust. The impact of customized brochures is consistent with publishedstudies of customization (e.g., Ansari &Mela, 2003; Hauser et al., 2010).The effect of customized brochures is less apparent in the main-effectanalyses because, although the effect is strong in earlier months, itbecomes insignificant in the last month. Review Table 3. Consistentwith the main-effect analyses, the conditional-logit analyses and thetrust regression identify no impact on brand consideration and pur-chase for the community forum and the competitive advisor.

We now add lagged estimated trust to the conditional-logit analyses.Such two-stage estimates are limited-information maximum-likelihoodestimates. The two-stage estimates are consistent but require bootstrapmethods to estimate the standard errors for the coefficients (Berndt,Hall, Hall, & Hausman, 1974; Efron & Tibshirani, 1994; Wooldridge,2002, p. 354, 414). The parameter estimates and standard errors arebased on 1000 bootstrap replications (Table 4 reports significance; stan-dard errors are available upon request).

Following Baron and Kenney (1986), the treatments are mediatedthrough trust if (a) including lagged trust in the model increases fit

significantly (and the lagged trust variable is significant) and (b) thetreatments are no longer significant when estimated lagged trust is inthe model. The increase is significant for brand consideration and mar-ginally significant for purchase (%1

2=86.7, pb .001 and %12=3.1, p=

0.08, respectively). Once we partial out lagged trust, there remain nosignificant direct effects due to the treatments. This result suggeststrust mediation.

6.3. Testing interaction effects for prior ownership and for multipletreatments

Prior ownership might influence the impact of the treatments andtheremight be interactions due tomultiple treatments. To test whetherprior ownership affects the impact of competitive information, wecrossed prior ownership of an AAM vehicle with the treatment-assignment dummies. For trust, brand consideration, and purchase,the interactions are not significant (F=1.91, p=.11; %4

2=4.3, p=.37,and %4

2=7.0, p=.13, respectively).We also tested interactions among the treatments. Treatment

interaction-effects do not add significantly to a trust regression usinga specification that allows all interactions (F=.85, p=.59). We contin-ue to use estimated lagged trust (without interactions) and estimate aconditional-logit model that allows interactions. The fully saturatedbrand consideration model that allows all possible interactions is notsignificant relative to a main-effects model and provides no additionalinsight (%11

2 =12.0, p=.10+). A few coefficients are significant, butall include competitive test drives with slight variations in parametermagnitudes that depend upon the other combinations of treatments.To avoid over-fitting with a fully saturated model, we examined amore parsimonious model in which we add a variable for two or moretreatments. This parsimoniousmodel is consistent with earlier automo-tive studies (see, e.g., Hauser et al., 1993). The “two ormore treatments”variable is not significant for brand consideration or purchase. Neitherthe fully saturated nor the parsimonious analysis highlights any inter-pretable interactions suggesting that the fully saturated model wasover parameterized. The fourth (brand consideration) and sixth (condi-tional purchase) data columns of Table 4 display models with

Table 4Conditional-logit analyses and trust regression with time-specific stimuli — year 1 random assignments.

Dependent measure Conditional-logit analyses (five months, 317 respondents for brand consideration model, only thosewho consider for conditional-purchase model)

Trust regression(lagged trust is used in this regression)

Direct effects not mediated Mediated by trust (bootstrap estimates)

Consider brand Purchase if consider Consider brand Purchase if consider

Constant !1.492a !2.567a !3.945a !4.348a !3.896a !3.977a .640b

Lagged brand consider 2.537a 2.394a 2.405a

Lagged trust hat .531a .525a .245b .249 .860a

Competitive test drives .579a .938b .392 .346 .879 .498 .380a

Customized brochures .079 .477 ! .059 ! .202 .405 .575 .171a

Competitive forum .144 .122 .133 .124 .182 .522 .045Competitive advisor ! .023 ! .103 .136 .133 ! .059 .025 ! .057Prior ownership AAM .399a .137 .327b .734a .079 .219 .000Prior own other American .304a ! .005 .253b .238b .037 .009 .021Prior ownership Japanese ! .577a ! .188 ! .464a ! .478a ! .126 ! .105 ! .019Month 3 .313 .200 .461a .461a .269 .275 ! .220a

Month 4 .419b .264 .433b .423 .270 .281 ! .295a

Month 5 .523a ! .238 .390b .293b ! .276 ! .253 ! .127a

Month 6 .722a .185 .654a .644a .194 .253 ! .251a

Prior ownership of AAM crossed withCompetitive test drives ! .211 1.61Customized brochures .109 ! .116Competitive forum ! .495 ! .966Competitive advisor ! .444 ! .280

Two or more treatments .208 ! .169Log likelihood !820.6 !218.2 !777.2 !774.8 !216.6 !213.1 Adjusted-R2

U2 (aka pseudo-R2) 24.8% 3.1% 28.8% 29.0% 3.8% 5.3% .749

Sex and age coefficients are not shown (not significant). Trust regression interactions are not significant.a Significant at the 0.05 level.b Significant at the 0.10 level.


interactions due to prior ownership and due to two ormore treatments.The addition of prior ownership and interactions among the treatmentsare not significant (%5

2=4.9, p=.42 and%52=7.1, p=.21 for brand con-

sideration and purchase, respectively).

6.4. Interpretation of year-1 results relative to year-1 hypotheses

H4 is clearly supported. Enhanced trust at the end month t!1 issignificantly correlated with brand consideration in month t. Whenthere is an effect due to competitive information, it is mediatedthrough trust. Enhanced trust at the end of month t!1 has a signif-icant effect on consideration in month t and a (marginally) signifi-cant effect on purchase in month t. These field-experiment resultsare consistent with results from the structural-equation analysis ofquestionnaires and laboratory experiments found in the literature.More importantly, given the in vivo nature of the field experimentand the modeling of dynamics, causality and external validity arelikely.

H1 through H3 are more complicated. Contrary to predictions andrecommendations in the literature, all forms of competitive informationdo not enhance trust. Neither the competitive forum (word of mouth)nor the competitive advisor enhanced trust, brand consideration, orpurchase. Only competitive test drives and possibly customizedbrochures enhanced trust, brand consideration, and purchase. Notonly was this complexity not predicted by theory, but the results alsosurprised AAM. All four generic forms of competitive informationwere provided altruistically and made AAM vulnerable, but vulnerabil-ity and altruism were not sufficient to affect trust, and through trust,brand consideration and purchase.

The experiments alone do not explain why some forms of competi-tive information were better than others, but our belief and the judg-ment of managers at AAM suggest that the valence of the informationmade a difference. AAM does well relative to competition in competitivetest drives; the customized brochures highlighted the benefits of AAM.However, there was substantial negative information in both the com-petitive forum and the competitive advisor. Empirically, it appears thatthe effect of negative information countered the effect of trust from com-petitive information.We believe that this possibility is themost likely in-terpretation, but we cannot rule out other interpretations, such as thestrength of the signal. We can rule out cheap talk because both themost costly and least costly forms of competitive informationwere effec-tive. This moderation of the vulnerability/altruism-to-trust link by thevalence of information is worthy of further analyses.

Although it may seem intuitive a posteriori, an interpretation thatnegatively valenced information does not enhance trust refines the-ories of trust signaling. Many authors posit only that the firm shouldact altruistically and provide information that is truthful and de-pendable (e.g., Erdem & Swait, 2004; Urban et al., 2009; Urban etal., 2000). Prior to the field experiments, the managers at AAM be-lieved that vulnerability and altruism would signal trust and thatthe enhanced trust would lead to brand consideration. The insightabout the valence of the competitive information was manageriallyimportant.

6.5. Continuous-time analyses

As a final check, we estimate continuous-timeMarkov models thatrelax three restrictions. These Markov models allow flows to happenin continuous time (even though observations of the results ofthose flows are at discrete time intervals). In addition, because thedependent variables are differences in observed states (not consider,consider but not purchase, and consider and purchase), the Markovmodels are analogous to difference-in-differences models and thusaccount for unobserved heterogeneity in the propensity to considerAAM. Finally, all transitions are estimated simultaneously with a sin-gle likelihood function (details are in Appendix A). The results

reinforce the implications of the conditional-logit analyses: competi-tive test drives have a significant effect on brand consideration, thateffect is mediated through trust, and neither competitive forumsnor competitive advisors have significant effects.

7. Year 2: field test of trust signaling

Year 1 established that positively valenced competitive information(competitive test drives and, possibly, customized brochures) enhancestrust, which, in turn, precedes brand consideration. However, the mosteffective competitive information (competitive test drives) is extremelyexpensive to provide and may not be cost-effective. Conversely, com-petitive information might be cost-effective if a firm could signal trust-worthiness by simply offering to make competitive informationavailable. If the signal alone were sufficient, a national launch wouldbe feasible. Trust signaling (H5 through H8) suggests that the offer ofcompetitive information itself engenders trust. However, there is a dan-ger that a trust signal would attract only those consumers who alreadytrust and/or feel favorably toward the brand that is signaling trust. If thispossibility is supported, H5 through H8 might be rejected.

To test trust signaling, year-2 consumers were assigned randomly toone of two groups. Consumers in the control group received no treat-ments whatsoever. Consumers in the test group received an advertise-ment inviting them visit a “My Auto Advocate” website (screenshotsavailable in an Online Appendix). We call consumers who visited thewebsite based on the advertisement alone the “website-visited” testgroup. Consumers in the test groupwhodid not visit the “MyAuto Advo-cate” website in response to advertising were invited to an “Internetstudy” that included a visit to the “My Auto Advocate” website. We callthese consumers the “website-forced-exposure” test group. At the “MyAuto Advocate” website, both the website-visited test group and thewebsite-forced-exposure test group could select (opt in to) any combi-nation of five treatments. Taken together, these two sub-groups makeup the test group that tests trust signaling. The website-forced-exposure group represents a stronger signal andmore intensive nationaladvertising.

The panel ran monthly from January to June 2005. In year 2, mem-bers of Harris Interactive's panel were screened to be in the marketfor a new vehicle, on average within the next 2.2 years, and invited toparticipate and complete six monthly questionnaires. This 2.2-year av-erage was designed to draw in more consumers (relative to the year-1twelve-month intenders). Once consumers visited the “My Auto Advo-cate” website, they were given incentives to opt in to the treatments,but unlike in year 1, theywere not assigned to treatments. For example,consumers received 20 reward certificates (worth $1 each) for partici-pating in the competitive test drives. Incentives for the other treatmentswere the order of 5 reward certificates.

Harris Interactive invited 6092 Los Angeles consumers of which1720 completed all six questionnaires for an average response/completion/retention rate of 21.7%. This rate was not significantlydifferent across the three groups (control vs. website-visited vs.website-forced-exposure, p=.25). Brand consideration, purchase, andtrust were measured as in year 1.

7.1. Treatments in year 2

The treatments in year 2 were similar to year 1. The competitivetest drive treatment and the word-of-mouth treatment were virtuallythe same with minor updates. The competitive online advisor wasimproved slightly with a better interface and a “garage” at which con-sumers could store vehicle descriptions. The online advisor still favoredother manufacturers' vehicles in year 2, although somewhat less sothan in year 1. The major change was the brochures. Year 2 used elec-tronic brochures for AAM vehicles (called eBooklets). These brochureswere online or downloadable, not mailed, and were less customized.An additional treatment, eBrochures, allowed consumers to download


competitive brochures. Although many competitive brochures wereavailable on automakers' websites, the single-source webpage madeit more convenient for consumers to compare vehicles. Screenshotsfor “My Auto Advocate” and the treatments are available in an OnlineAppendix. Table 5 summarizes the numbers of consumers who optedin to treatments in year 2. All treatments except competitive test driveswere reasonably popular and available in all months.

7.2. Testing whether the trust signal was effective

We first examine brand consideration and purchase in the test(website-visited and website-forced-exposure) vs. the control groups.There were no significant differences in trust, brand consideration, orpurchase intentions (t=.9, p=.37, t=.14, p=.88, and t=.18, p=.86,respectively). Similarly, the differences between the website-visitedsub-group and control group were not significant (t=1.4, p=.15,t=!1.5, p=.14, and t=! .05, p=.96, respectively). These results sug-gest that a trust signal provided little or no lift in trust, brand consider-ation, and purchase relative to the control. Contrary to extant theory,we reject H5, H6, andH7. Because these hypotheses are rejected,we can-not test the conditional hypothesis, H8.

There are at least two complementary explanations for the null ef-fect in year 2. First, the null effect may be because offering competi-tive information does not increase trust. In this case, it might havebeen that only those consumers who already trusted AAMwere likelyto opt in to the treatments. A second (and complementary) explana-tion is that fewer consumers opted in to the effective treatments (testdrives and possibly brochures) than potentially negative treatments(word-of-mouth and online advisors).

7.3. Examining potential explanations

If the signal did not enhance trust, brand consideration, and purchasein the test group relative to the control group, then we should see evi-dence that consumers who were more favorable to AAM opted in to thetreatments (to be most effective, the signal needed to reach consumerswho do not trust the automaker, not those who already trust the auto-maker). To test whether the trust signal reached targeted consumers wecompare consumers in the control group (who received no treatments)to those in the test group who did not opt in to any treatments.

Among no-treatment consumers, the non-treated members of thetest group had significantly lower trust, brand consideration, and pur-chase intentions than the control group (t=6.1, p=.0, t=6.1, p=.0,and t=2.0, p=.05, respectively). By implication, consumers who

opted in to competitive-information treatments were consumers whowere more trusting of AAM (or, at least, more favorable toward theautomaker). Comparing the control group to non-treated members ofthe website-forced-exposure test group gives similar results. The trustsignal ultimately targeted consumers more likely to be favorabletoward AAM. The signal alone did not engender trust, brand consider-ation, and purchase.

We gain further insight by redoing for year 2 the main-effects analy-ses (as in Table 2) and conditional-logit analyses (as in Table 4). Details ofthe year-2 analyses are available in an Online Appendix. The opt-inmaineffects (year 2), relative to random-assignmentmain effects (year 1), areconsistent with a hypothesis that for each treatment, consumers whoopted in to that treatment were a priori more favorable to AAM. Whenwe account for prior ownership, prior propensities, dynamics, andmedi-ationwith conditional-logit analyses, we find that the effects of the treat-ments are consistent with those observed in year 1. Thus, the null effectof the trust signal is probably observed because the signal did notencourage opt in among those to whom the signal was targeted.

8. Hypotheses revisited

Table 6 summarizes the implications of the two field experiments.H1 through H4 suggest that if firms build trust with consumers, trustwill lead to brand consideration and purchase. H5 through H8 suggestfurther that trust signals alone should achieve these outcomes. Thesituation in the automotive industry in 2003–2005 provided an excel-lent test of these theories. Because of past experiences with AAM's ve-hicles,many consumerswould not even consider those vehicles in 2003–2005. Because the vehicles from AAM had improved relative to prioryears, the automaker had an opportunity to build trust by providingcompetitive information (year 1) or by signaling trustworthiness (year2). Current theories suggest altruism and vulnerability will build trustbut do not distinguishwhether the information provided to the consum-er should favor the brand. In fact, altruism and vulnerability are greater ifthe competitive information does not always favor the firm's brands.

The year-1 experimentswere consistentwith H4. Trust in onemonthwas correlated with brand consideration in the next month, and the ef-fect of competitive information was mediated through trust. However,the year-1 experiments also identified certain competitive informationtypes as more effective than other competitive information types. Expe-riential information (competitive test drives) was the most effectivecommunications strategy. Tangible experience convinced consumersthat the automaker's products had improved relative to the competition.

Table 5Consumers who selected treatments in year 2 signaling experiment (test of signaling trust through advertising then website opt in to treatments).

Number of respondents who selected the indicated treatment in that month.

Treatment Month 2 Month 3 Month 4 Month 5 Month 6 Treatment “cell”

Competitive test drives Website-visiteda 70 0 0 0 0 70Website-forcedb 140 0 0 0 0 140Not treatedc 1182 1322 1322 1322 1322 1182

Competitive eBrochures Website-visiteda 88 178 199 202 211 289Website-forcedb 149 361 411 425 432 621Not treatedc 1173 961 911 897 890 701

AAM eBooklets Website-visiteda 49 158 184 194 205 252Website-forcedb 114 355 411 417 438 549Not treatedc 1208 967 911 905 884 773

Competitive forum Website-visiteda 71 139 168 194 208 249Website-forcedb 114 294 352 409 420 538Not treatedc 1208 1028 970 913 902 784

Competitive advisor Website-visiteda 97 180 206 226 246 290Website-forcedb 199 378 441 493 535 645Not treatedc 1123 944 881 829 787 677

a Website-visited = consumers visited the “My Auto Advocate” website on their own and opted in to the treatment.b Website-forced = consumers were forced to visit the “My Auto Advocate” website, but opted in to the treatment.c Not treated = consumers in control group plus consumers in test group who did not opt in to the treatment.


Neither word-of-mouth (community forums) nor competitive ad-visors increased trust in the test group relative to the control group.From the year-1 experiments alone, we cannot determine whetherthese forms of competitive information are not effective or whetherAAM's implementation of these forms of competitive informationwas not effective. Community forums relied on other consumers'opinions— opinions contaminated with past experience. Online advi-sors relied in part on past consumer experience and may have laggedimprovement in vehicles. We posit that positively valenced informa-tion enhanced and negatively valenced information diminished theeffectiveness of competitive-information implementations.

Hypothesis. Unbiased competitive information can build trust, and trustenhances brand consideration and purchase. The firm builds trust if com-petitive information enables the firm to communicate to consumers thatit is acting altruistically and that it has products that meet their needs.However, the availability of competitive information alone does not en-hance trust; consumers must process the competitive information.

We believe that this hypothesis applies across a variety of situationsand product categories as discussed in the introduction to this paper.Naturally, this hypothesis is subject to tests in different categorieswith different implementations of generic competitive-informationtreatments and with different signals of trust.

9. Cost effectiveness and managerial implications

We nowmove from theory back to practice. Because the experimentsestablished that some forms of competitive information engender trustand that trust leads to brand consideration, AAMultimately implementedcompetitive-information strategies. The automaker first used the moregeneral insights to refine competitive test drives. The following calcula-tions illustrate the motivation behind managers' decisions at the time.We disguise the proprietary datawith comparable publicly available data.

For illustration, we assume a 15% market share. Based on this share,competitive test drives provide an 11.1% sales lift (year-1 data) with anapproximate cost of $120 per participating consumer (with incentives).These calculations suggest that the cost of an incremental vehicle sale isapproximately $7200. [$7207=$120/(0.15$0.111).] Typical marginsfor the automotive industry are approximately 9%, and the average priceof a new car is about $28,400 (Thomas & Cremer, 2010, http://www.ehow.com/facts_5977729_average-cost-new-car.html, visited July2012). These public numbers suggest an incremental profit of approxi-mately $2500 per vehicle sold, much less than the $7200 cost per vehiclesold based on a competitive test drive. Post-sales parts and other consid-erations are unlikely tomake up the $4700 difference. As implemented in

the year-1 randomized experiments, competitive test drives are notprofitable.

Large-scale competitive information strategies were put on holdduring the distractions of the automotive and financial crises of2005–2009. A multi-city competitive test-drive format was neitherfeasible nor cost-effective. Meanwhile, the concept of competitivetest drives gained traction in situations where test drives could beimplemented cost efficiently and targeted at skeptical consumers.When the financial situation improved, AAM tested competitive testdrives for SUVs with a dealer in Arizona (automotive consumers oftenuse agendas to screen on body-type; Hauser, 1986; Urban, Liberali,MacDonald, Bordley, & Hauser, 2012). These competitive test drivesproved to be cost-effective — approximately $100–200 per incrementalsale. Costs were substantially lower because the test drives were fromthe dealer's lot (no need to rent a test track), because fewer vehicleswere necessary (only SUVs), and because the dealer could borrow orrent vehicles from competitive dealers. On the benefit side, grossmargins were higher than average for SUVs. AAM continued to experi-ment with competitive test drives in key local markets when high-valueskeptical consumers could be targeted cost-effectively. In late 2010, thehead of U.S. marketing for AAM launched a yearlong series of weekendcompetitive test drives at dealerships. Each event invited a few thousandpotential buyers to compare AAM vehicles with competitive vehicles.

Year 2 taught managers that they needed to do more than signal thatcompetitive information was available. Communication strategies shouldbe targeted at consumers who do not already trust the automaker. Man-agers now place a premiumon targeting competitive information towardskeptical consumers. Newmethods include interactive screening to iden-tify consumers who answer questions that indicate they do not trustAAM. Targeted consumers would receive substantial incentives to partic-ipate in competitive-information treatments.

Based on the 2003–2005 data and managerial judgment, managersbelieve that providing competitive information is effective when thereis good news and when it is cost effective. As of this writing, AAM in-cludes key competitive comparisons on its website using standardizedPolk data on prices, specifications, and equipment for preselected com-petitive and consumer-specified vehicles. In 2009, AAM used a nationaladvertising campaign that encouraged consumers to compareAAMvehi-cleswith those of competitors regarding fuel economy and styling.ManyAAMdealers offer unsolicited extendedweekend test drives and encour-age competitive comparisons. AAM believes that competitive informa-tion builds trust, brand consideration, and sales and is profitable only ifimplemented cost-effectively to skeptical consumers for categories inwhich AAM has good vehicles relative to competitors. This more nu-anced trust-based strategy is believed to be more profitable than a gen-eral strategy of trust signaling.

Table 6Summary of hypotheses, support, tests, and implications.

Hypotheses Supported or not supported Evidence Implications

H1. Trust enhanced by competitiveinformation (in test vs. control).

Supported for positively valencedinformation but not for information withsubstantial negative content (positive vs.negative remains a hypothesis).

Competitive test drives and customizedbrochures have significant effects intrust regressions. Other treatmentshave no significant effect.

There exist situations where competitiveinformation leads to trust, especially positivelyvalenced information. Negatively valencedinformation may not engender trust.

H2. Brand consideration enhancedby competitive information(in test vs. control).

Strongly supported for positively valencedinformation.

Trust-to-brand consideration issignificant in conditional-logit andcontinuous-Markov analyses.

If a firm can engender trust it can enhance brandconsideration.

H3. Sales enhanced by competitiveinformation (in test vs. control).

Marginally supported for positivelyvalenced information.

Trust-to-sales is marginally significantin conditional-logit and significant incontinuous-Markov analyses.

If a firm can engender trust it might be able toobtain sales.

H4. Mediation by trust. Strongly supported. Baron–Kenney tests support trustmediation in both conditional-logit andcontinuous-Markov analyses.

Other means to gain trust might gain brandconsideration and sales.

H5–H7. Offering to providecompetitive information enhancestrust, brand consideration, and sales.

Rejected. Opt-in communications attractedconsumers who are already favorable.

Year-2 null effects. Also comparisons ofcontrol consumers to consumers in thetest groups who did not opt in totreatments.

Signaling altruism (trust) by providingcompetitive information did not enhance trust,consideration, or sales. Consumers mustexperience the information.

H6. Mediation of signal by trust. Not testable in the data. H8 is conditioned on support for H5–H7. Trust was not enhanced by offer of information.


Online Appendices available from the authors

OA1. Screenshots of year-2 treatments.OA2. Alternative specifications of trust regressions in year 1.OA3. Main effects, trust regression, and conditional-logit analyses

in year 2.Year 1 and year 2 data as used in tables are available from the

authors.

Appendix A. Continuous-time Markov analysis in year 1

The analyses in the text model month-to-month dynamics but donot allow flows to happen in continuous time nor to do they allowreverse flows from “consider” to “not consider.” We address theseissues with continuous-time Markov analyses (Cox & Miller, 1965;Hauser & Wisniewski, 1982, hereafter “Markov” analyses). There aretwo added advantages of Markov analyses: (a) a single likelihood func-tion estimates all parameters for all defined flows simultaneously, and(b) treatments affect differences in behavioral states directly. By theMarkov property, observing a customer in “consider” in month t istreated differently if the customer was in “not consider” versus in

“consider” at month t!1. By focusing on differences in behavioralstates, the Markov analyses are less sensitive to unobservedheterogeneity.

The Markov analyses complement the conditional-logit analyses inFig. 2; the concepts are similar, but we model a more complete set offlows and allow the flows to occur in continuous time (Fig. A1). Con-sumers “flow” among states with instantaneous flow rates dependentupon the treatments and other variables. Mathematically, for j% i, aijnΔtis the probability that the consumer flows from state i to state j in thetime period between t and t+Δt for very small Δt during the nthmonth. We specify as relevant the flow rate as a log-linear function ofthe treatment assignments, prior ownership, age, sex, month dummies,and interactions as relevant— the same types of specifications as in theconditional-logit analyses. Although we model instantaneous flowrates, we only observe the state that describes each consumer at theend of each month. Fortunately, using the aijn's, we can calculate theprobability, pijn, that the consumer was in state i at the beginning ofthe nth month and in state j at the end of the month. Specifically,

Pn ! eAn Tn!Tn!1# $&#!

r!0

Arn Tn!Tn!1# $r

r!&Vn expΛn% &V!1

n ; #A1$

where Pn is the matrix of the pijn's, An is the matrix of the aijn's, Tn is thetime at the end of the nth month, Vn is the matrix of eigenvectors ofAn(Tn!Tn!1), and [expΛn] is the matrix with the exponentiation ofthe eigenvalues on the diagonal.

Prior applications in marketing used regression approximations ofEq. (A1) (Hauser & Wisniewski, 1982). With today's computers, weuse maximum-likelihood methods with all flows estimated simulta-neously. See Kulkarni (1995) for a review of computational methodsto deal with matrix exponentiation. While we would like to repeat theMarkov analyses for all of the specifications tested by conditional-logitanalyses, the convergence of the Markov estimates and the computa-tion times appear to be most appropriate for more parsimoniousmodels. Thus, we use the Markov analyses as a confirmation of theconditional-logit analyses by carefully selecting the explanatory vari-ables based on the conditional-logit analyses (we do not need laggedconsideration in the Markov analyses because the analyses are basedon transitions from “not consider”, rather than based on estimating con-sideration as a function of lagged consideration and other variables). Forsimplicity of exposition, we report key analyses in Table A1. Other anal-yses and R-code are available from the authors.Fig. A1. Continuous-Time Markov Flow Dynamics In Each Period.

Table A1Continuous time Markov process analysis — year 1 random assignments.

Dependent measure Continuous time Markov estimation not mediated Continuous time Markov estimation mediated by trust

Consider brand Purchase Consider brand Purchase

Not consider toconsider(1!2)

Consider to notconsider(2!1)

Consider topurchase(2!3)

Not consider toconsider(1!2)

Consider to notconsider(2!1)

Consider topurchase(2!3)

Constant .139 .231 .120a .146a .249a .102a

Lagged trust hat .221a ! .230b .313a

Competitive test drives 1.060b ! .408 ! .003 1.001 ! .262 ! .124Customized brochures .130 .140 .252 .107 .217 .169Competitive forum ! .227 ! .236 .124 ! .273 ! .300 .241Competitive advisor ! .003 ! .342 ! .144 ! .088 ! .359 ! .089Prior ownership of AAM ! .407 ! .773 ! .403 ! .624Prior own other American .599a .070 .525 a .039Prior ownership of Japanese ! .305 .292 ! .249 .197Month 3 .012 ! .172 .445 .032 ! .289 .565Month 4 ! .971a .544 ! .394 !1.004a .423 ! .366Month 5 ! .698a .122 .224 ! .760a .083 .194Log likelihood !616.46 !608.12

All flows are estimated simultaneously.a Significant at the 0.05 level.b Significant at the 0.10 level.


The Markov analyses reinforce the conditional-logit and main-effectanalyses. Competitive test drives have a significant effect on consider-ation, but that effect is likely mediated through lagged trust. Laggedtrust has a significant effect on key flows. We also modeled potentialmisclassification of “consider” vs. “not consider” as in Jackson, Sharples,Thompson, Duffy, and Couto (2003). The misclassification analyses im-proved fit but provided no additional managerial insights. Estimatedmisclassification was moderate.

References

Aboulafia, R. (2007). Embraer eyes options for growth. Aerospace America, 18–20.Anderson, J. C., & Narus, J. A. (1990). A model of distributor firm and manufacturer firm

working partnerships. Journal of Marketing, 54(January), 42–58.Anderson, E., & Weitz, B. (Fall). Determinants of continuity in conventional industrial

channel dyads. Marketing Science, 9(4), 310–323.Ansari, A., & Mela, C. F. (2003). E-customization. Journal of Marketing Research, 40(May),

131–145.Arora, N., Dreze, X., Ghose, A., Hess, J., Iyengar, R., Jing, B., et al. (2008). Putting one-to-one

marketing to work: Personalization, customization, and choice. Marketing Letters, 19,305–321.

Baron, R. M., & Kenney, D. A. (1986). The moderator–mediator variable distinction insocial psychological research: Conceptual, strategic, and statistical considerations.Journal of Personality and Social Psychology, 51(6), 1173–1182.

Bart, Y., Shankar, V., Sultan, F., & Urban, G. L. (2005). Are the drivers and role of onlinetrust the same for all web sites and consumers? A large-scale exploratory empiricalstudy. Journal of Marketing, 69(October), 133–152.

Berndt, E. K., Hall, B. H., Hall, R. E., & Hausman, J. (1974). Estimation and inference innonlinear structural models. Annals of Social Measurement, 3, 653–665.

Bloom, J. (March 1). Jim Stengel exhorts 4A's: It's not about telling and selling. AdvertisingAge, http://adage.com/article/special-report-4asmedia07/jim-stengel-exhorts-4a-s-telling-selling/115259/.

Biswas, D., & Biswas, A. (Summer). The diagnostic role of signals in the context of per-ceived risks in online shopping: Do signals matter more on the web? Journal of In-teractive Marketing, 18(3), 30–45.

Büttner, O. B., & Göritz, A. S. (2008). Perceived trustworthiness of online shops. Journalof Consumer Behavior, 7(1), 35–50.

Churchill, G. A., Jr. (1979). A paradigm for developing better measures of marketingconstructs. Journal of Marketing Research, 16(February), 64–73.

Cox, D. R., & Miller, H. D. (1965). The theory of stochastic processes. London, UK: Chapman& Hall.

Crosby, L. A., Evans, K. R., & Cowles, D. (1990). Relationship quality in services selling:An interpersonal influence perspective. Journal of Marketing, 54(July), 68–81.

Doney, P. M., & Cannon, J. P. (1997). An examination of the nature of trust in buyer–seller relationships. Journal of Marketing, 61(April), 35–51.

Efron, B., & Tibshirani, R. J. (1994). An introduction to the bootstrap. New York, NY:Chapman & Hall/CRC.

Erdem, T., & Swait, J. (2004). Brand credibility, brand consideration, and choice. Journalof Consumer Research, 31(June), 191–198.

Erdem, T., & Swait, J. (1998). Brand equity as a signaling phenomenon. Journal of Con-sumer Psychology, 7(2), 131–157.

Ganesan, S. (1994). Determinants of long-term orientation in buyer–seller relation-ships. Journal of Marketing, 58(April), 1–19.

Ganesan, S., & Hess, R. (1997). Dimensions and levels of trust: Implications for commit-ment to a relationship. Marketing Letters, 8(4), 439–448.

Geyskens, I., Steenkamp, J. -B. E. M., & Kumar, N. (1998). Generalizations about trust inmarketing channel relationships using meta-analysis. International Journal ofResearch in Marketing, 15, 223–248.

Green, C. H. (2010). NY Life ex-Chairman Sy Sternberg on trust in the life insurance business.Trusted Advisor Associates, Ltd. Blog , http://trustedadvisor.com/trustmatters/ny-life-ex-chairman-sy-sternberg-on-trust-in-the-life-insurance-business-trust-quotes-16.Accessed February 15, 2012

Gurviez, P., & Korchia, M. (2003). Proposal for a multidimensional brand trust scale.32nd EMAC Conference, Marketing: Responsible and Relevant? May 20–23. Glasgow.

Häubl, G., & Trifts, V. (2000). Consumer decision making in online shopping environ-ments. Marketing Science, 19(1), 4–21.

Hauser, J. R. (1978). Testing the accuracy, usefulness and significance of probabilisticmodels: An information theoretic approach. Operations Research, 26(3), 406–421.

Hauser, J. R. (1986). Agendas and consumer choice. Journal of Marketing Research,23(3), 199–212.

Hauser, J. R., Urban, G. L., & Weinberg, B. (1993). How consumers allocate their timewhen searching for information. Journal of Marketing Research, 30(4), 452–466.

Hauser, J. R., & Wisniewski, K. J. (1982). Dynamic analysis of consumer response tomarketing strategies. Management Science, 28(5), 455–484.

Hauser, J. R., Olivier, T., Theodorosm, E., Daria, D., & Rene, B. (2010). Cognitive Simplic-ity and Consideration Sets. Journal of Marketing Research, 47, 485–496 June.

Hoffman, D. L., Novak, T. P., & Peralta, M. (1999). Building consumer trust online. Com-munications of the ACM, 42(4), 80–85.

Ingrassia, P. (2010). Crash course: The American automobile industry's road from glory todisaster. New York, NY: Random House.

Jackson, C. H., Sharples, L. D., Thompson, S. G., Duffy, S. W., & Couto, E. (2003). Multi-state Markov models for disease progression with classification error. The Statisti-cian, 52, 193–209.

Kirmani, A., & Rao, A. R. (2000). No pain, no gain: A critical review of the literature onsignaling unobservable product quality. Journal of Marketing, 64(April), 66–79.

Kulkarni, V. (1995). Modeling and analysis of stochastic systems. London, UK: Chapman& Hall/CRC.

Maister, D. H., Green, C. H., & Galford, R. M. (2000). The trusted advisor. New York, NY:Simon & Schuster.

Milgrom, P., & Roberts, J. (1986). Price and Advertising Signals of Product Quality.Journal of Political Economy, 94(4), 796–821.

Moorman, C., Deshpandé, R., & Zaltman, G. (1993). Factors affecting trust in market re-search relationships. Journal of Marketing, 57(January), 81–101.

Morgan, R. M., & Hunt, S. D. (1994). The commitment–trust theory of relationship mar-keting. Journal of Marketing, 58(3), 20–38.

Nunnally, J. C., & Bernstein, I. H. (1994). Psychometric theory, 3E. New York, NY:McGraw Hill, Inc.

Rousseau, D. M., Sitkin, S. B., Burt, R. S., & Camerer, C. (1998). Not so different after all: Across-discipline view of trust. Academy of Management Review, 23(3), 393–404.

Porter, C. E., & Donthu, N. (2008). Cultivating trust and harvesting value in virtual com-munities. Management Science, 54(1), 113–128.

Shankar, V., Urban, G. L., & Sultan, F. (2002). Online trust: A stakeholder perspective,concepts, implications and future directions. The Journal of Strategic InformationSystems, 11(December), 325–344.

Sirdeshmukn, D., Singh, J., & Sabol, B. (2002). Consumer trust, value, and loyalty in re-lationship exchanges. Journal of Marketing, 86(January), 15–37.

Spence, M. (1973). Job market signaling. Quarterly Journal of Economics, 87(3),355–374.

Thomas, C., & Cremer, A. (2010). Audi: Taking a new route in the U.S. Bloomberg Busi-ness Week. November, 22–28, 32–33.

Trifts, V., & Häubl, G. (2003). Information availability and consumer preference: Canonline retailers benefit from providing access to competitor price information?Journal of Consumer Psychology, 13(1 & 2), 149–159.

Urban, G. L. (Winter). The emerging era of customer advocacy. MIT Sloan ManagementReview, 45(2), 77–82.

Urban, G. L. (2005). Don't just relate— Advocate. Pearson Education, Upper Saddle River,N.J.: Wharton School Publishing.

Urban, G. L., Amyx, C., & Lorenzon, A. (2009). Online trust: State of the art, new fron-tiers, and research potential. Journal of Interactive Marketing, 23, 179–190.

Urban, G. L., & Hauser, J. R. (2004). ‘Listening-in’ to find and explore new combinationsof customer needs. Journal of Marketing, 68(April), 72–87.

Urban, G. L., Hauser, J. R., & Roberts, J. H. (1990). Prelaunch forecasting of new automo-biles: Models and implementation. Management Science, 36(4), 401–421.

Urban, G. L., Liberali, G., MacDonald, E., Bordley, R., & Hauser, J. R. (2012). Morphingbanner advertisements. Cambridge, MA: MIT Sloan School of Management.

Urban, G. L., Sultan, F., & Qualls, W. J. (Fall). Placing trust at the center of your internetstrategy. MIT Sloan Management Review, 42(1), 39–48.

Yoon, S. -J. (2002). The antecedents and consequences of trust in online purchase de-cisions. Journal of Interactive Marketing, 16(2), 47–63.

Wooldridge, J. M. (2002). Econometric analysis of cross section and panel data.Cambridge, MA: MIT Press.


Performance implications of deploying marketing analytics

Frank Germann b,1, Gary L. Lilien a,⁎, Arvind Rangaswamy c,2

a Smeal College of Business, Pennsylvania State University, 484 Business Building, University Park, PA 16802, USAb University of Notre Dame, 395 Mendoza College of Business, Notre Dame, IN 46556, USAc Smeal College of Business, Pennsylvania State University, 210E Business Building, University Park, PA 16802, USA


Article history:First received in 29, February 2012 and wasunder review for 3! monthsAvailable online 27 November 2012

Area Editor: Dominique M. Hanssens

Keywords:Marketing analyticsMarketing modelsMarketing ROI

A fewwell-documented cases describe how thedeployment ofmarketing analytics produces positive organizationaloutcomes. However, the deployment ofmarketing analytics varieswidely acrossfirms, andmany C-level executivesremain skeptical regarding the benefits that they could gain from their marketing analytics efforts. We draw onupper echelons theory and the resource-based view of the firm to develop a conceptual framework that relatesthe organizational deployment of marketing analytics to firm performance and that also identifies the key ante-cedents of that deployment. The analysis of a survey of 212 senior executives of Fortune 1000firms demonstratesthat firms attain favorable and apparently sustainable performance outcomes through greater use of marketinganalytics. The analysis also reveals important moderators: more intense industry competition and more rapidlychanging customer preferences increase the positive impact of the deployment of marketing analytics on firmperformance. The results are robust to the choice of performancemeasures, and, on average, a one-unit increasein the degree of deployment (moving a firm at themedian or the 50th percentile of deployment to the 65th per-centile) on a 1–7 scale is associated with an 8% increase in return on assets. The analysis also demonstrates thatsupport from the topmanagement team, a supportive analytics culture, appropriate data, information technolo-gy support, and analytics skills are all necessary for the effective deployment of marketing analytics.


1. Introduction

A recent Google search for “marketing analytics” returned morethan 500,000 hits. Marketing analytics, a “technology-enabled andmodel-supported approach to harness customer and market data toenhance marketing decision making” (Lilien, 2011, p. 5) consists oftwo types of applications: those that involve their users in a decisionsupport framework and those that do not (i.e., automated marketinganalytics). During the past half century, the marketing literature hasdocumented numerous benefits of the use of marketing analytics, in-cluding improved decision consistency (e.g., Natter, Mild, Wagner, &Taudes, 2008), explorations of broader decision options (e.g., Sinha& Zoltners, 2001), and an ability to assess the relative impact of deci-sion variables (e.g., Silk & Urban, 1978). The common theme in thisliterature is the improvement in the overall decision-making process(e.g., Russo & Schoemaker, 1989, p. 137).

Rapid technological and environmental changes have transformedthe structure and content of marketing managers' jobs. These changesinclude (1) pervasive, networked, high-powered information tech-nology (IT) infrastructures, (2) exploding volumes of data, (3) more

sophisticated customers, (4) an increase in management's demandsfor the demonstration of positive returns on marketing investments,and (5) a global, hypercompetitive business environment. In thischanging environment, opportunities for the deployment of market-ing analytics to increase profitability seemingly should abound. In-deed, an entire stream of research in marketing documents thepositive performance implications of deploying marketing analytics(e.g., Hoch & Schkade, 1996; Kannan, Kline Pope, & Jain, 2009;Lodish, Curtis, Ness, & Simpson, 1988; McIntyre, 1982; Natter et al.,2008; Silva-Risso, Bucklin, & Morrison, 1999; Zoltners & Sinha, 2005).

However, there continue to be many skeptics with regard to the“rational analytics approach” to marketing. For example, in a recentinterview with one of the authors, a (former) senior executive at oneof the world's leading car manufacturers claimed that “…marketinganalytics-based results usually raise more questions than theyanswer,” and he asserted that “the use of marketing analytics oftenslows you down.” He also claimed that the “…performance implica-tions of marketing analytics are at best marginal.” When we inquiredabout documentation for his views, he referred us to Peters andWaterman's (1982) highly influential book, In Search of Excellence, inwhich the authors denounce formal analysis because of its abstractionfrom reality and its tendency to produce “paralysis through analysis”(p. 31). More recently, a study of 587 C-level executives of large inter-national companies revealed that only approximately 10% of the firmsregularly employ marketing analytics (McKinsey & Co., 2009). AndKucera and White (2012) note that only 16% of the 160 business


⁎ Corresponding author. Tel.: +1 814 863 2782; fax: +1 814 863 0413.E-mail addresses: [email protected] (F. Germann), [email protected] (G.L. Lilien),

[email protected] (A. Rangaswamy).1 Tel.: +1 574 631 4858; fax: +1 574 631 5255.2 Tel.: +1 814 865 1907; fax: +1 814 865 7064.





leaders who responded to their survey reported using predictive ana-lytics, although those users “significantly outpace those that do not intwo important marketing performance metrics3” (p. 1).

John Little diagnosed the issue more than 40 years ago as follows:“The big problem with … models is that managers practically neveruse them. There have been a few applications, of course, but the practiceis a pallid picture of the promise” (Little, 1970, p. B-466). Revisiting theissue, Little (2004, p. 1858) reports that “The good news is that moremanagers than ever are using models…what has not changed is orga-nizational inertia.” Winer (2000, p. 143) concurs: “My contacts in con-sumer products firms, banks, advertising agencies and other largefirms say that [model builders] are a rare find and that models are notused much internally. Personal experience with member firms of MSIindicates the same.”

The low prevalence of marketing analytics use implies that manymanagers remain unconvinced about the benefits that accrue fromthat use. In addition, most research studies that document these bene-fits have focused on isolated firm or business unit “success stories”without systematically exploring performance implications at the firmlevel. Given the lack of compelling evidence about the performance im-plications of marketing analytics, the objective of this research is to ad-dress two questions: (1) Does widespread deployment4 of marketinganalytics within a firm lead to improved firm performance? and (2) Ifthe answer to (1) is “yes,” what leads to the widespread deploymentof marketing analytics within firms? With the usual caveats and cau-tions, particularly with regard to making causal inferences usingnon-experimental data, we find that the answer to question 1 appearsto be “yes” and, hence, the answer to question 2 has high managerialrelevance, as well as academic importance.

To address our research questions, we propose a conceptual frame-work that relies on both the resource-based view (RBV) of the firm(Barney, 1991; Wernerfelt, 1984) and upper echelons theory (Hambrick& Mason, 1984) to model the factors that link marketing analyticsdeployment to firm performance, as well as the factors that drive thedeployment of marketing analytics. We assess the validity and value ofthat framework with data drawn from a survey of 212 senior executivesat Fortune 1000 firms, supplemented by secondary source objective per-formance data for those firms.We find that the deployment of marketinganalytics has a greater impact on firm performance when the industry ischaracterized by strong competition and when customer preferenceschange frequently in the industry. We also find that top managementteam (TMT) advocacy and a culture that is supportive of marketing ana-lytics are the keys to enabling a firm to benefit from the use of marketinganalytics, and our analyses suggest that the benefits realized bymarketinganalytics deployment may be sustainable.

We proceed as follows: We first present our conceptual frameworkand hypotheses and, then, describe our data and our methodology. Wethen present our findings and discuss their theoretical and managerialimplications, as well as the limitations of our research.

2. Conceptual framework

The conceptual framework in Fig. 1 depicts what we refer toas the marketing analytics chain of effects. The framework articu-lates our predicted relationships, including the hypothesized rela-tionship between the deployment of marketing analytics and firmperformance.

We propose that marketing analytics deployment, which we defineas the extent to which insights gained from marketing analytics guideand support marketing decision making within the firm, has a positiveimpact on firm performance. However, this positive impact on firmperformance is likely to be moderated by three industry-specificfactors: (1) the degree of competition faced by the firm, (2) the rate ofchange in customer preferences, and (3) the prevalence of marketinganalytics use within the industry. Furthermore, we identify TMT advo-cacy of marketing analytics as a vital antecedent of the deployment ofmarketing analytics. We suggest that a firm's TMT must not only com-mit adequate resources in the form of employee analytic skills, data,and IT but also nurture a culture that supports the use of marketing an-alytics. Such a culture can ensure that the insights gained frommarket-ing analytics are deployed effectively.

In the following section, we first elaborate on the link between thedeployment of marketing analytics and firm performance. Next, weconsider the antecedents of the deployment of marketing analytics;i.e., the resources and organizational elements that we posit mustbe in place for marketing analytics to be deployed effectively.

2.1. The performance implications of deploying marketing analytics

A few authors (primarily authors writing for non-academic journals)suggest that the use of marketing analytics can slow firms down,leading to missed market opportunities that are seized by moreagile and non-analytics-oriented competition. For example, citingGeneral Colin Powell's leadership primer, Harari (1996, p. 37) sug-gests that “excessive delays in the name of information-gatheringbreeds analysis paralysis,” which leads to missed opportunities and,hence, subpar firm performance. Peters and Waterman (1982) predictan analogous effect. Additionally, based on our discussions with execu-tives, we conclude that many top managers share similar notionsregarding the performance outcomes of marketing analytics use.

However, there are many firm-specific case studies that describethe positive performance impact of marketing analytics use. Forexample, Elsner, Krafft, and Huchzermeier (2004) demonstrate howRhenania, a medium-sized German mail order company, used a dy-namic, multilevel response modeling system to answer its most im-portant direct marketing questions: When, how often, and to whomshould the company mail its catalogs? The model allowed the compa-ny to increase its customer base by more than 55% and quadrupled itsprofitability during the first few years following implementation, andthe firm's president asserted that the firm was saved by deployingthis model.

Marketing analytics can also significantly improve a firm's abilityto identify and assess alternative courses of action. For example, inthe 1980s, Marriott Corporation was running out of adequate down-town locations for its new full-service hotels. To maintain growth,Marriott's management planned to locate hotels outside downtownareas to appeal to both business and leisure travelers. A marketinganalytics approach called conjoint analysis facilitated the company'sdesign and launch of its highly successful Courtyard by Marriottchain, establish amultibillion dollar business, and create a new prod-uct category (Wind, Green, Shifflet, & Scarbrough, 1989).

In another example, Kannan, Kline Pope, and Jain (2009) reporthow marketing analytics at the National Academies Press (NAP) ledto a better understanding of customers and to a better manner ofreaching the customers. The NAP was concerned about the best wayto price and distribute its books in print and in pdf format via the In-ternet. It built a pricing model that allowed for both substitution andcomplementarity effects among the two formats and calibrated themodel using a choice modeling experiment. The results permittedthe NAP to launch its entire range of digital products with a variablepricing scheme, thereby maximizing the reach of its authors' work.

The common theme of the above firm-specific examples is that thedeployment of marketing analytics allows firms to develop and offer

3 The metrics are “incremental lift from a sales campaign” and “click through rate(for mass campaigns).” Those firms that use customer analytics also report a signifi-cantly greater ability to measure customer profitability and lifetime value and are alsomore likely to have staff dedicated to data mining.

4 We use the term “deployment” or “to deploy” to mean “to put into use, utilize orarrange for a deliberate purpose,”without reference to the financial, human, or techni-cal investment that might be necessary for the enablement of such deployment.

115F. Germann et al. / Intern. J. of Research in Marketing 30 (2013) 114–128

products and services that are better aligned with customer needsand wants, which, in turn, leads to improved firm performance.Thus, we propose the following main effect:

H1. The greater the deployment of marketing analytics, the better thefirm's performance.

2.1.1. Competitive industry structureMost firms compete with a number of rivals (Debruyne & Reibstein,

2005), although the degree of rivalry varies considerably across indus-tries (DeSarbo, Grewal, & Wind, 2006). The level of competition that afirm faces also has many concomitant effects, including the degree ofcustomer satisfaction that the firm must attain to operate successfully.For example, Anderson and Sullivan (1993)find thatfirmswith less sat-isfied customers that face less competition perform approximately thesame or even better than firms with more satisfied customers that op-erate in more competitive environments. Thus, firms that confrontmore competition must strive for higher levels of customer satisfactionto perform well.

Assuming that marketing analytics provide better insights aboutcustomer needs, firms in industries with greater competition shouldearn higher returns (because of more clearly targeted offerings, forexample, which result in greater customer satisfaction) than firmsin less competitive industries. Thus, we propose:

H2. Themore intense the level of competition among industry partic-ipants, the greater the positive impact of marketing analytics deploy-ment on firm performance.

We note that if “analysis-paralysis” is a serious concern associatedwith the deployment of marketing analytics, then the correspondingnegative performance implications should be even greater in compet-itive environments because competitors move more swiftly in suchenvironments (e.g., DeSarbo et al., 2006). Under these circumstances,we should observe a negative interaction between marketing analyt-ics deployment and level of competition (as opposed to our predictedpositive interaction).

2.1.2. Customer preference changesCustomer preferences regarding product features, price points, dis-

tribution channels, media outlets, and other elements of the marketingmix change over time (e.g., Kotler & Keller, 2006, p. 34). The rate of suchchange varies: fashions change seasonally, whereas preferences for

consumer electronics appear to change almost monthly (e.g., Lamb,Hair, & McDaniel, 2009, p. 58), but preferences regarding constructionequipment, hand tools, and agricultural products appear to be muchmore stable over time.

The more customers' needs fluctuate, the greater is the uncertaintythat firms face in making decisions and the more critical scanning andinterpreting the changing environment becomes (Daft & Weick, 1984).Marketing analytics offer various means to assist firms in monitoringthe pulse of the market and providing early warning of preferencechanges. Additionally, a stable, predictable environment reduces theneed for marketing analytics because such an environment requires alimited number of decision variables to manage for organizational suc-cess (Smart & Vertinsky, 1984). Therefore, we propose:

H3. The more rapidly customer preferences change in an industry,the greater the positive impact of the deployment of marketing ana-lytics on firm performance.

2.1.3. Prevalence of marketing analytics useThe prevalence of the use of marketing analytics within an indus-

try may attenuate their positive performance implications. Porter(1996, p. 63) notes that as firms evolve, “staying ahead of rivalsgets harder,” partially because of the diffusion of best practices, facil-itated, for example, by inputs from strategy consultants. Competitorsare quick to imitate successful management techniques, particularlyif they promise superior methods of understanding andmeeting cus-tomers' needs. Such imitation eventually raises the bar for everyone(e.g., Chen, Su, & Tsai, 2007; D'Aveni, 1994; MacMillan, McCaffery, &Van Wijk, 1985). Thus, the greater the overall use of marketing ana-lytics in an industry, the lower is the upside potential for a firm to in-crease its use. Hence, we propose:

H4. The more prevalent the use of marketing analytics in an industry,the lower is the positive impact of the deployment of marketing ana-lytics on the performance of individual firms in that industry.

To summarize our hypotheses regarding research question #1, wepredict that the deployment of marketing analytics has positive per-formance implications in general5 and that this effect is even stronger

5 A concave (downward sloping) response function would admit diminishingreturns to deployment and would model a “paralysis of analysis effect”. We report atest for such an effect in Section 4.3.4 and do not find that effect.

Top MgmtTeam

Advocacy

AnalyticsCulture

AnalyticsSkills

Data and IT

Deploymentof Analytics

FirmPerformance

Competition (H2) (+)Needs & Wants Change (H3) (+)

Analytics Prevalence (H4) (-)

+

+

+

+

+

+ (H1)

+

+

The Deployment ofMarketing Analytics

The Performance Implications ofDeploying Marketing Analytics

Fig. 1. Conceptual framework.

116 F. Germann et al. / Intern. J. of Research in Marketing 30 (2013) 114–128

in industries characterized by strong competition and inwhich custom-er preferences change frequently and weaker in industries in which thedeployment of marketing analytics is commonplace.

We next discuss the factors that lead to the deployment of mar-keting analytics.

2.2. Antecedents of the deployment of marketing analytics

Adapting a resource-based view (RBV—Barney, 1991; Wernerfelt,1984), Amit and Schoemaker (1993) suggest that firms create com-petitive advantage by assembling, integrating, and deploying theirresources in a manner that allows them to work together to createfirm capabilities. Firm capabilities can provide a sustainable compet-itive advantage when they are protected by isolating mechanismsthat thwart competitive imitation (Rumelt, 1984).

Building on the RBV literature, we suggest that marketing analyticsmust be appropriately assembled and embedded within the fabric ofthe firm to be deployed effectively, which potentially results in a sus-tainable competitive advantage. Furthermore, we single out TMT advo-cacy of marketing analytics as a key driver of that process.

2.2.1. TMT advocacy, analytics culture, and sustainable competitiveadvantage

According to upper echelons theory (Hambrick & Mason, 1984),organizations are a reflection of their TMT; thus, for marketing ana-lytics to become an integral part of a firm's business routines and,ultimately, its culture, it must be strongly supported by the firm'sTMT (Hambrick, 2005).

We posit that a culture that is supportive of marketing analyticsis critical for its effective deployment because that culture carriesthe logic of how and why “things happen” (Deshpande & Webster,1989, p. 4). These norms are especially important because the per-son (or organizational unit) that carries out the marketing analytics(e.g., marketing analyst or researcher) frequently is not responsiblefor implementing the insights gained, namely, executives in marketingand other functions (Carlsson & Turban, 2002; Hoekstra & Verhoef,2011; Van Bruggen & Wierenga, 2010; Wierenga & van Bruggen,1997). An analytics culture provides decision makers with a pattern ofshared values and beliefs (Deshpande, Farley, & Webster, 1993; Ouchi,1981), which in turn, should positively influence the degree to whichthey incorporate the insights gained from marketing analytics in theirdecisions. Furthermore, culture is sticky, difficult to create, and evenmore difficult to change (e.g., Schein, 2004), suggesting that it mayprotect against competitive imitation of a firm's analytics investments,thus delivering sustainable rewards from a firm's marketing analyticsinvestments.

2.2.2. Analytics skillsTo deploymarketing analyticswithin a firm, the firmmust also have

access to people (either internally or among its partners) who have theknowledge to execute marketing analytics. Thus, the TMT must ensurethat people with the requisite marketing analytics skills are presentwithin the company or available outside the firm. We distinguish be-tween technical marketing analytics skills and other individual-level,analytics-based knowledge structures that are tacit (Grant, 1991). Tech-nicalmarketing analytics skills likely derive primarily from classroomorother structured learning situations and consist of the range of market-ing models and related concepts that the analyst could deploy. In con-trast, tacit knowledge of marketing analytics includes skills acquiredprimarily through real-world learning.

We anticipate that higher levels of marketing analytics skills willincrease the extent of marketing analytics deployment because peo-ple use the tools and skills they understand and with which theyare comfortable (Lounsbury, 2001; Westphal, Gulati, & Shortell,1997). Additionally, better skills should lead to more useful resultsfrom using those skills, thus facilitating the organization-wide

marketing analytics adoption process. Therefore, a firm's employees'analytics skills should have both a direct, positive impact on the orga-nizational deployment of analytics and an indirect effect on organiza-tional deployment through the positive impact on analytics culture.

2.2.3. Data and IT resourcesA firm's physical IT infrastructure and data resources are two other

critical tangible assets that the TMT must implement to allow for theeffective deployment of marketing analytics. Physical IT resourcesform the core of a firm's overall IT infrastructure and include computerand communication technologies and shared technical platforms anddatabases (Ross, Beath, & Goodhue, 1996). Data result from measure-ments and provide the basis for deriving information and insightsfrommarketing analytics (Lilien & Rangaswamy, 2008). Marketing ana-lytics are often based on vast amounts of customer data (Roberts,Morrison, & Nelson, 2004), which require sophisticated IT resources toeffectively obtain, store, manipulate, analyze, and distribute across thefirm. Therefore, IT and data are closely related tangible resources, suchthat onewould be significantly less valuable without the other. Buildingon thismutual dependence, we posit that both IT and data resources areimportant prerequisites for marketing analytics use.

To summarize our hypotheses regarding research question #2,we propose that TMT advocacy of marketing analytics is an impor-tant precursor to the effective deployment of marketing analytics.We further propose that a firm's TMT must not only ensure that em-ployees with the requisite analytics skills and an adequate data andIT infrastructure are in place but also nurture a culture that supportsthe use of marketing analytics. Such a culture can ensure that the in-sights gained from marketing analytics are deployed effectively.

3. Data and methods

3.1. Scale development

We adapted existing scales when they were available. However, ourstudy is among the first to empirically explore the performance implica-tions ofmarketing analytics, and scales for several of our constructswerenot available.Wedeveloped themissing scales, following a four-phase it-erative procedure, as recommended in the literature (Churchill, 1979):First, we independently generated a large pool of items for each of theconstructs from an extensive literature review. Second, we engaged fif-teen senior-level, highly regarded marketing academics to expand ourlist of items and evaluate the clarity and appropriateness of each item.Third, we personally administered pretests to six topmanagers to assessany ambiguity or difficulty that they experienced when responding.Fourth, we conducted a formal pretest with 31 seniormanagers. Becausethe fourth stage/pretest revealed no additional concerns, we finalizedthe scale items, which are listed in Appendix A.6

3.2. Data collection procedure

We conducted a mail survey among executives of Fortune 1000firms. We first randomly selected 500 entries from the Fortune 1000

6 We note that we employed single-item measures for some of our constructs. Sev-eral researchers have demonstrated that in certain contexts, measures that compriseone item generate excellent psychometric properties (e.g., Bergkvist & Rossiter,2007; Drolet & Morrison, 2001; Robins, Hendin, & Trzesniewski, 2001; Schimmack &Oishi, 2005). In particular, single-item measures have been found to be very usefulwhen the construct is unambiguous (Wanous, Reichers, & Hudy, 1997). Furthermore,single-item measures are also useful when participants are busy (which certainly ap-plies to top executives) and perhaps dismissive of and/or aggravated by multiple-item measures that, in their view, measure exactly the same construct (Wanous etal., 1997). Such respondent behavior has been found to inflate across-item error termcorrelation (Drolet & Morrison, 2001). Our pretests revealed that three of our con-structs (i.e., competition, needs and wants change, and marketing analytics preva-lence) are unambiguous in nature, leading us to employ single-item measures forthem.


list and then leveraged the corporate connections of two major U.S.universities to obtain the names of 968 senior executives (primarilyalumni) working at these firms.

We addressed these respondents using personalized letters, inwhich we asked them to complete the survey in reference to eithertheir strategic business unit (SBU) or their company, whicheverthey felt was more appropriate. We also provided a nominal incentive(1 USD, called a token of thanks, which emerged as the most effectiveincentive in a pretest). Of the 968 executives contacted, 36 returnedthe surveys and indicated they were not qualified to respond and20 surveys were returned because of incorrect addresses. Weobtained 212 completed surveys (of the 912 remaining surveys),which yielded an effective response rate of 23.25%. We controlledfor possible nonresponse bias by comparing the construct means forearly and late respondents (Armstrong & Overton, 1977) but foundno significant differences. As we show in Table 1, most (71%) of therespondents in our sample had titles of director or higher, which sug-gests that they should be knowledgeable about their firms' capabili-ties and actions.

We also asked the respondents to report their confidence levelswith regard to the information they provided (Kumar, Stern, &Anderson, 1993). The sample mean score was 5.59 (out of 7 [SD=.81]), indicating a high level of confidence. Additionally, we receivedmultiple (either two or three) responses from 35 firms/SBUs in oursample, allowing us to cross-check the responses when we receivedmore than one response from a firm.7,8

3.3. Scale assessment

We assessed the reliability and validity of our constructs usingconfirmatory factor analysis (Bagozzi, Yi, & Phillips, 1991; Gerbing& Anderson, 1988). We included all independent and dependent la-tent variables in one confirmatory factor analysis model, which pro-vided satisfactory fit to the data (comparative fit index [CFI]=.97;root mean square error of approximation [RMSEA]=.05; 90% confi-dence interval [CI] of RMSEA=[.033; .068]). On the basis of the esti-mates from this model, we examined the composite reliability anddiscriminant validity of our constructs (Fornell & Larcker, 1981). Allcomposite reliabilities exceed the recommended threshold value of.6 (Bagozzi & Yi, 1988); the lowest reliability is .75. The coefficient al-phas of our constructs are all greater than .7. We also assessed dis-criminant validity using the criteria proposed by Fornell andLarcker (1981). The results demonstrate that the squared correlationbetween any two constructs is always lower than the average vari-ance extracted (AVE) for the respective constructs, providing sup-port for discriminant validity. Finally, the correlations between therespective constructs are all significantly different from unity(Gerbing & Anderson, 1988). Overall, the results indicate that our la-tent constructs demonstrate satisfactory levels of composite reliabil-ity and discriminant validity. We present the correlations among theconstructs in Table 2 and the AVE and coefficient alphas in theAppendix A along with the scale items.

Although we were able to establish discriminant validity, some ofour constructs are highly correlated. For example, the correlation be-tween analytics skills and analytics culture is 0.825. As per our mea-sures, “analytics skills” refer to the type of analytics skills that the

employees possess, whereas “analytics culture” indicates sharedbeliefs with regard to how analytics will influence the company. Al-though one would expect these two constructs to be highly correlated,we assert that they do not measure the same thing, much in the samemanner that a physician who measures a patient's height and weight,two highly correlated items, might argue that height and weight mea-sure different important things and thus both should be measured.

3.3.1. Descriptive statisticsTable 3 contains descriptive statistics for our sample firms and indi-

cates that the sample represents a broad range of firms. Table 4 lists thenames of some sample firms. In Table 5,we provide the summary statis-tics and correlations for our variables and, in Table 6, we present histo-grams for our focal variables. As the histograms show, the sampledfirms display awide range of values for our focal variables. For example,on the seven-point scale measuring TMT advocacy of marketing analyt-ics, approximately 18% of the sample firms fall within the 6–7 range and16% within the 1–3 range (M=4.5; SD=1.7). Furthermore, with re-gard to analytics culture, approximately 25% of the sample firms fallwithin the 6–7 range, and approximately 14% score within the 1–3range (M=4.6; SD=1.6). We also asked the respondents (1) whethertheir marketing analytics applications are designed primarily in-houseor by outside experts and (2) whether the primary day-to-day opera-tions of marketing analytics are managed in-house or outsourced.Table 7 presents the responses to these questions and demonstratesthat the majority of the Fortune 1000 firms design and manage theirmarketing analytics (applications) in-house. We also make note of thelow percentage of respondents who did not know the answer to thesequestions, another sign that our respondents are quite knowledgeableabout the domain under study.

3.4. Conceptual model testing procedures

Our conceptual model proposes both direct and moderating effects(Fig. 1). To model and test these effects simultaneously, we used struc-tural equation modeling (SEM); recent methodological advances havemade it feasible to include multiple interactions in a path model(Klein & Moosbrugger, 2000; Marsh, Wen, & Hau, 2004; Muthén &Asparouhov, 2003). We used Mplus Version 6.11 and estimated ourmodel using the full-information maximum likelihood approach(Klein & Moosbrugger, 2000; Muthén & Muthén, 2010, p. 71).

4. Results

4.1. SEM model fit

Fig. 2 summarizes the results of our SEM, depicting two of the threeinteractions (i.e., competition and needs and wants change) as statisti-cally significant. Becausemeans, variances, and covariances are not suf-ficient statistics for our SEM estimation approach, our model does notprovide the commonly used fit statistics, such as RMSEA and CFI. In-stead, in accordance with Muthén (2010), we assessed fit in twosteps. First, we re-estimated our SEM without the interaction termsand compared that model with our original model via a chi-square dif-ference test using the associated loglikelihoods (Muthén & Muthén,2011; Satorra & Bentler, 1999). This test yielded a !2 (3) difference of

Table 1Profile of Fortune 1000 firm respondents.

Position Number of participants Percentage

President, CEO 7 3EVP, (Sr.) VP, CMO, CFO, COO 78 37(Sr.) Director, Executive Director 65 31(Sr.) Marketing Manager 47 22Other (e.g., Marketing Strategist) 15 7Total 212 100

7 We received two responses from 33 firms/SBUs and three responses from 2 firms/SBUs. Because we had contacted 968 executives who worked for 500 randomly select-ed Fortune 1,000 firms, we evidently contacted multiple executives working for thesame firms/SBUs, which accounts for most of these multiple responses. In a few in-stances (n=5), executives also invited their coworkers to participate in the survey.

8 Although this multiple-response sample is too small for a formal multitrait,multimethod assessment, it enabled us to assess whether the respective respondentgroups’means for the key constructs were statistically different (e.g., Srinivasan, Lilien,& Rangaswamy, 2002). T-tests indicated that none of the means were statistically sig-nificantly different from each other.


28.124, which is highly significant (pb .0001) which clearly favors themodel with interactions. Second, we (re)estimated the model withoutinteractions with the conventional SEM estimation approach to derivethe usual model fit statistics (e.g., RMSEA and CFI). This conventionalmodel (without interactions) fits the data quite well (!2 (175)=243;

CFI=.97; RMSEA=.04; 90% C.I.=[.03; .06]), and the paths are verysimilar to those of the moderated model. Based on these results, weconclude that the “un-moderated” model fits the data well and thatthe moderated model enhances the model fit.

4.2. Specific model paths and hypothesis test results

All of the paths from TMT advocacy to the respective subsequentlatent constructs are positive and significant, suggesting that the TMTplays a key role in establishing an organizational setting in which mar-keting analytics can be deployed effectively. Additionally, as predicted,an analytics-oriented culture has a positive and significant effect onthe deployment of analytics ("=.317, pb .01), in line with our proposi-tion that strengthening a firm's analytics-oriented culture leads to anactual increase in the deployment of marketing analytics. In addition,we find that enhancements to a firm's marketing analytics skills haveboth a direct and positive impact on the deployment of analytics ("=.427, pb .001) and a positive, indirect effect through analytics culture("=.120, pb .05). That is, employees' marketing analytics skills directlyinfluence the degree to which the firm uses analytics-based findings inmarketing decisionmaking; they also exert an indirect influence by en-hancing the organization's analytics-oriented culture. We also find thatthe presence of a strong data and IT infrastructure promotes marketinganalytics skills within the firm ("=.621, pb .001).9

As hypothesized in H1, higher levels of deployment of marketing an-alytics leads to an increase in firm performance ("=.106, pb .01).Moreover, as hypothesized in H2, we find a positive and significant de-ployment of analytics " competition interaction ("=.081, pb .05),which shows that the use of analytics is more effective inmore compet-itive environments than in less competitive environments.10 Similarly,in support of H3, the use of analytics is more effective in environmentsin which customers' needs and wants change frequently ("=.060,pb .01). However, we do not find support for H4 concerning the analyt-ics " prevalence interaction ("=! .034, ns).

4.3. Robustness checks

4.3.1. Validity of the performance measure/monomethod biasBecause our independent and dependent measures originate from

the same respondents, leading to the possibility of monomethod bias(Podsakoff, MacKenzie, Lee, & Podsakoff, 2003), we collected perfor-mance data from independent sources to validate our performancemeasure. We obtained information on firm-specific net income andtotal assets for as many firms as possible by retrieving their 10 Kand other filings with the U.S. Securities and Exchange Commission

9 Because data and IT go hand in hand, this may imply an interaction effect betweenthe two in our model. As a robustness check, we added a fourth item to the “Data andIT” construct that captured the interaction between the data and IT items and then re-ran our model. The results did not change in any substantive way.10 The competition variable was skewed to the left. As a robustness check, we reranour analysis, substituting the competition variable with a dummy variable (1=highcompetition [survey score of 6 or 7]; 0=low competition [survey score between 1and 5]). The results did not change in any substantive way.

Table 3Sample firm profiles.

Industry groups # %

Services 88 41.5Manufacturing 65 30.7Trade 22 10.4Construction and Mining 7 3.3Finance and Insurance 30 14.1

Total 212 100Sales # %b$1 Million 5 2.4$1 Million to $10 Million 14 6.6$10 Million to $100 Million 23 10.8$100 Million to $1 Billion 57 26.9$1 Billion to $5 Billion 74 34.9>$5 Billion 39 18.4

Total 212 100Number of employees # %0–100 20 9.4101–1000 37 17.51001–10,000 39 18.410,001–100,000 60 28.3100,001–200,000 32 15.1>200,000 24 11.3

Total 212 100

Note: The profiles pertain to either the strategic business unit (SBU) or the overallcompany associated with our respondents, depending on which UNIT the respondentsselected when completing the survey.

Table 2Construct correlations and variances.

Constructs Correlations

1 2 3 4 5 6

1. TMT Advocacy 1.257 0.649 0.570 0.188 0.476 0.0472. Analytics Culture 0.806 (0.03) 1.677 0.681 0.176 0.543 0.0333. Marketing Analytics Skills 0.755 (0.04) 0.825 (0.03) 2.777 0.318 0.608 0.0704. Data and IT 0.434 (0.07) 0.419 (0.07) 0.564 (0.06) 0.638 0.196 0.1075. Deployment of Analytics 0.690 (0.05) 0.737 (0.04) 0.780 (0.03) 0.443 (0.06) 1.788 0.0626. Firm Performance 0.216 (0.07) 0.181 (0.08) 0.265 (0.07) 0.327 (0.08) 0.248 (0.07) 0.373

Note: The correlations and their standard errors (provided in brackets underneath) are in bold, the squared correlations are in italics, and the variances are provided on the diagonal.

Table 4Sample firms (partial list).

• IBM • Kraft Foods• Honeywell • FedEx• American Express • Sears Holdings• Marriott International • JP Morgan Chase• Raytheon • UPS• Capital One • Deere & Company• DuPont • Alcoa• Hewlett-Packard • Aramark• Ford Motor Co • Citigroup• Pfizer • Baxter International• AT&T • General Mills• Xerox • 3 M• Johnson & Johnson • Motorola• Progressive • Starbucks• Boeing • Verizon• Amazon.com • Charles Schwab• ConAgra Foods • Dick's Sporting Goods• Apple • Harley-Davidson• Oracle • Hershey


from the EDGAR database. We also consulted COMPUSTAT, MergentOnline and the firms' websites. With these financial data, we comput-ed the respective firm's return on assets (ROA). These proceduresyielded financial performance data for 68 of the 212 responses.After matching the time horizon of the performance measures, wecomputed a 2-year average ROA for the 2 years preceding our prima-ry data collection (see, for example, Boulding, Lee, & Staelin, 1994).We also standardized the ROA measure with respect to each firm'scompetitors (from Mergent Online).

To address same-source bias, we used the objective performancedata (i.e., ROA) to reanalyze our conceptual framework. Given thesmall sample size and the consequent lack of statistical power (n=68), it was not feasible to simultaneously test all of the hypothesizedeffects of our framework in a single SEM model. Instead, weconducted two separate analyses: first, we used a SEM to estimatethe direct (un-moderated) effects in our conceptual framework. Sec-ond, we used an ordinary least squares (OLS) regression model to (re)examine the link between deployment of analytics and firm perfor-mance and to (re)test H1–H4. We substituted the ROA objective per-formance measure for the perceptual performance measure in bothanalyses.

The SEM results remain consistent regardless of the use of objec-tive or subjective data; in fact, the link from deployment to perfor-mance is even stronger with objective data than with subjectivedata. We provide the SEM results with objective data in Fig. 3.

We report the regression results with objective data in Table 8(model 1). We used a simple average of the items measuring deploy-ment of analytics as our deployment construct in that analysis. We re-peated the analyses using the factor scores from our SEM for ourdeployment construct. These two measures were highly correlated(correlation>.94), and none of our inferences were affected by thechoice of deployment construct. Overall, the regression model is sig-nificant, and our inferences did not change.

In summary, the signs of the SEM and regression model coeffi-cients using objective data are consistent with those obtainedusing the survey-based data. However, the deployment of analytics

" competition interaction did not reach significance in the regressionmodel (t=1.60), a result that could be due to the small sample sizefor the objective data (n=68).

4.3.2. Multiple respondents for some firmsAs noted, we obtained data frommultiple respondents from 35 or-

ganizational units. To address potential issues of non-independenceamong these observations in our data, we averaged the responses ofmultiple respondents11 from each firm (e.g., Homburg, Grozdanovic,& Klarmann, 2007) and then re-estimated the SEM using individualresponses as if we had only obtained single responses (i.e., the aver-age responses for those organizational units for which we obtainedmultiple responses). The results remain virtually the same, and ourinferences do not change.

4.3.3. Multigroup analysis — B2B vs. B2CThere are many differences between business-to-business (B2B)

and business-to-consumer (B2C) firms (see Grewal & Lilien, 2012)that might lead one to expect that there would be differences in therole and impact of marketing analytics within B2B and B2C firms. Toassess this possibility, we performed a multigroup confirmatory fac-tor analysis to compare the factor loadings of B2B with B2C firms.To test for partial measurement invariance across groups, we com-pared a model in which all parameters could be unequal across thetwo groups with one in which we constrained the factor loadings tobe equal. The model with all parameters freely estimated fit thedata well (!2 (252)=321.541; CFI=.97; RMSEA=.05), as did thepartial invariance model with factor loadings constrained to be equal(!2 (270)=336.227; CFI=.97; RMSEA=.05). Furthermore, the !2

difference test indicated that the two models were not statisticallysignificantly different (!2 (18)=14.7, p=.68), thereby suggestingthat our findings hold across different types of firms.

11 The t-tests of the key variables across these respondents' reports indicated that therespective means were not statistically different.

Table 5Correlations and summary statistics.

Variables Correlations

1 2 3 4 5 6 7 8 9 10

1. TMT attitude toward marketing analytics 1.0002. Annual reports highlight use of marketing analytics 0.579 1.0003. TMT expects quantitative analyses 0.578 0.778 1.0004. If we reduce marketing analytics use, profits will suffer 0.379 0.558 0.601 1.0005. Confident that use of marketing analytics improves customer satisfaction 0.492 0.641 0.635 0.697 1.0006. Most people are skeptical of any kind of analytics-based results (R) 0.401 0.552 0.581 0.676 0.713 1.0007. Appropriate marketing analytics tool use 0.497 0.546 0.600 0.630 0.637 0.561 1.0008. Master many different marketing analysis tools and techniques 0.466 0.569 0.591 0.648 0.615 0.558 0.837 1.0009. Our people can be considered experts in marketing analytics 0.572 0.599 0.642 0.621 0.648 0.637 0.736 0.738 1.00010. We have a state-of-art IT infrastructure 0.283 0.281 0.264 0.300 0.331 0.283 0.431 0.343 0.361 1.00011. We use IT to gain a competitive advantage 0.190 0.230 0.285 0.220 0.180 0.103 0.315 0.312 0.320 0.34412. In general, we collect more data than our primary competitors 0.269 0.268 0.349 0.326 0.319 0.253 0.432 0.422 0.420 0.39213. Everyone in our UNIT uses analytics insights to support decisions 0.459 0.537 0.586 0.577 0.624 0.560 0.649 0.639 0.645 0.31214. We back arguments with analytics based facts 0.404 0.436 0.516 0.460 0.498 0.444 0.586 0.598 0.562 0.23415. We regularly use analytics in the following areas 0.335 0.550 0.502 0.442 0.598 0.460 0.489 0.517 0.509 0.31016. Firm performance - total sales growth 0.077 0.007 0.006 0.089 0.124 0.062 0.100 0.149 0.147 0.20717. Firm performance - profits 0.293 0.150 0.186 0.106 0.113 0.167 0.193 0.203 0.230 0.34318. Firm performance - return on investment 0.276 0.158 0.182 0.134 0.136 0.216 0.204 0.197 0.236 0.31319. We face intense competition !0.060 !0.115 !0.060 !0.058 !0.078 !0.082 !0.050 !0.058 !0.113 !0.04220. Our customers' needs and wants change frequently !0.090 !0.083 !0.103 !0.130 !0.172 !0.084 !0.040 !0.057 !0.042 0.05121. Marketing analytics are used extensively in our industry !0.052 0.126 0.101 0.069 0.061 0.069 0.014 0.032 0.049 !0.06322. Size !0.005 0.017 0.057 0.081 0.084 0.043 0.069 0.067 0.091 0.04423. Objective ROA (Time 1) 0.278 0.276 0.334 0.168 0.288 0.287 0.320 0.283 0.294 0.05624. Objective ROA (Time 2) 0.276 0.300 0.270 0.151 0.229 0.244 0.219 0.187 0.258 0.060

Summary statisticsMean 3.571 5.029 5.014 4.699 4.714 4.455 3.596 3.790 3.720 4.696Standard Deviation 1.705 1.506 1.419 1.589 1.511 1.618 1.860 1.704 1.771 1.576


4.3.4. Robustness of the deployment to performance linkOur study reveals a statistically significant positive relationship

between the deployment of marketing analytics and firm perfor-mance (both subjective and objective). This result is of great manage-rial importance, and, therefore, we subjected this relationship toadditional scrutiny via (1) testing for the linearity of this relationship,(2) assessing the effects of various controls, (3) subjecting it to areverse-causality test, (4) assessing the contemporary vs. carryovereffects of deployment on performance, (5) testing for the effects ofunobserved heterogeneity, and (6) assessing the unidimensionalityof our performance construct. We elaborate on these robustnesstests below.

First, we ran an OLS regression model similar to that reported inTable 8 and included a quadratic term to check for curvilinear effectsof the deployment of analytics. The squared term was not statisticallysignificant, suggesting the absence of a curvilinear effect, at leastwithin the range of our data.

Second, we included organization size (number of employees) andindustry dummy variables as controls in the regression model. Firmsize would account for the fact that larger firms could benefit fromeconomies of scale and scope, rendering their use of analytics moreeffective. Industry dummies would account for differences in industrysegments. We used standard industrial classifications to group thesample firms into five categories (see Table 3): services, manufactur-ing, finance/insurance, trade, and construction/mining. The size andindustry dummy variables had neither a main nor a moderating effecton the relationship between deployment of analytics and firm perfor-mance, and our inferences did not change. Thus, our results appearrobust to firm size and industry segments.

Third, it might be that firms that perform well have more leewayand, hence, more resources to deploy marketing analytics than dothose that perform poorly, implying that firm performance may affectthe deployment of marketing analytics, and not vice versa. To (at leastpartially) assess this potential reverse-causality issue, we collectedadditional objective performance data for the year following our sur-vey. We followed the same procedure as outlined earlier to collect the

objective performance data and then calculated the 2-year averageROA using the newly collected data, as well as the data for the yearpreceding our primary data collection. We then used this new objec-tive performance data to reanalyze our conceptual model. As before,we relied on SEM to estimate the direct (un-moderated) effects inour conceptual model and used OLS regression to examine the linkbetween deployment of marketing analytics and firm performance.We report the SEM results in Fig. 4 and include the regression resultsin Table 8 (model 2). As the results show, the outcomes did notchange in any substantive manner, providing support for the notionthat the deployment of marketing analytics is an antecedent of firmperformance, not vice versa.

Fourth, to assess the timing of the performance effects of deploy-ment of marketing analytics, we combined the objective performancemeasures as follows:

#! PerformanceTime 1" # $ 1–#% & ! PerformanceTime 2" #;

where # can range from 0 to 1, PerformanceTime 1 is our initial objec-tive ROA measure and PerformanceTime 2 is the ROA measure with a1-year lag. We then re-estimated our OLS regression model, withthe resulting linear combination values as the dependent variable(with # varying in increments of 0.1 from 0 to 1), and assessedwhich linear combination yields the best fitting model as determinedby Adj. R2. Fig. 5 provides the results of our analyses.

The results reveal that the highest Adj. R2 occurs when "=.4 (this isthe maximum likelihood estimate for # assuming Normal distribution ofthe error terms of the OLS regression), suggesting that the performanceeffects of the deployment of analytics appear to be observed both imme-diately and with a slightly stronger carryover. This finding further dis-counts the possibility of a reverse-causality effect, with the effects beingslightly stronger in Time 2 than in Time 1 (A value of #=.5would indicatethat the short-term and longer-term effects are the same).

Fifth, we estimated a mixture regression model (DeSarbo & Cron,1988) to explore the possibility of unobserved heterogeneity amongfirms. The lowest Bayesian Information Criterion (BIC) emerged for

Table 5Correlations and summary statistics.

Correlations

11 12 13 14 15 16 17 18 19 20 21 22 23 24

1.0000.637 1.0000.248 0.352 1.0000.205 0.314 0.813 1.0000.205 0.354 0.542 0.479 1.0000.156 0.206 0.172 0.148 0.174 1.0000.218 0.242 0.197 0.172 0.222 0.451 1.0000.204 0.193 0.208 0.188 0.181 0.496 0.832 1.000

!0.034 !0.017 !0.118 !0.154 !0.117 0.018 !0.068 !0.097 1.0000.013 0.005 !0.092 !0.060 0.022 0.031 0.020 0.005 0.167 1.0000.033 0.117 0.079 0.097 0.115 !0.032 !0.023 !0.011 0.007 0.052 1.0000.012 0.007 !0.006 !0.030 !0.024 !0.012 !0.157 !0.162 0.146 0.177 !0.070 1.0000.291 0.279 0.342 0.397 0.444 0.275 0.318 0.375 0.061 !0.037 0.177 0.048 1.0000.204 0.145 0.347 0.323 0.362 0.217 0.341 0.371 0.082 !0.003 0.230 0.001 0.508 1.000

4.219 4.505 5.241 4.580 5.189 4.839 5.196 5.006 5.422 3.743 3.408 3.561 4.962 4.6741.755 1.744 1.422 1.383 1.435 1.208 1.268 1.262 1.635 1.966 1.638 1.467 1.541 1.234


a one-class model (consistent with our multi-group analysis above),which suggests that unobserved heterogeneity was not relevant forour model. Thus, our findings appear to be generalizable to all typesof Fortune 1000 firms.

Sixth, the correlations among the subjective performance measures(items 16–18 in Table 5) suggest that our performance construct maynot be unidimensional: the correlation between profits and return oninvestment (ROI) is quite high (r=.832), whereas the correlations be-tween sales growth and profits (r=.451) and sales growth and ROI(r=.496) are significantly lower. Therefore, we analyzed the effect ofthe deployment of analytics on performance with regard to salesgrowth and profits/ROI separately. In the SEM model, the main effectof the deployment of analytics on performance increased in both in-stances, i.e., when using only the single-item sales growth measure("=.171 vs. .106) and when using the construct comprised of theprofits and ROI items ("=.198 vs. .106). Furthermore, when employingsales growth as the outcome measure, competition no longer emergesas a significant moderator of analytics deployments' effect on perfor-mance ("deployment " competition=.063 vs. .081; the interaction betweenneeds and wants change and deployment of analytics remains mar-ginally significant: "deployment " needs and wants change=.076 vs. .06).In contrast, both interactions, i.e., competition"deployment of ana-lytics and needs andwants change"deployment of analytics becomestronger when including the profits/ROI performance variables inthe SEM ("deployment x competition=.149 vs. .081 and "deployment x

needs and wants change=.113 vs. .06). All other paths remain virtuallythe same in the respective models.

Thus, although the use of marketing analytics appears to positivelyaffect sales growth, profits and ROI, our analysis suggests that the de-ployment of analytics may have a somewhat stronger effect onprofits/ROI than on sales growth.We offer the following possible expla-nations for this finding: First, manymarketing analytics applications aregeared toward identifying the most profitable customer segment(s)(e.g., Reinartz & Kumar, 2000), applications designed to improve profitsand ROI, as opposed to sales. Second, our sample is drawn from Fortune1000 firms — all large firms — and their scale may prevent them fromgrowing as quickly as smaller firms. Thus, this finding may be specificto our sample and should be explored more broadly.

Table 9 summarizes our robustness checks of the deployment toperformance link.

4.3.5. Deployment of analytics as mediatorOur conceptual model assumes that the deployment of analytics

mediates the effect of analytics culture and analytics skills on firmperformance. To test this assumption, we conducted a formal test ofmediation, following the procedure recommended by Baron andKenny (1986). We used both of the objective performance measuresas the respective dependent variables, deployment of analytics asthe mediator, and analytics skills or analytics culture as the respectiveindependent variables. Deployment of analytics emerges as a media-tor for both independent variables irrespective of the objective per-formance measure used.

5. Discussion and conclusions

Our research objective was to determine whether the deploymentof marketing analytics leads to improved firm performance and toidentify the factors that lead firms to deploy marketing analytics.Our findings address these two research objectives and provide in-sights of value for both marketing theory and practice.

5.1. Theoretical implications

Our study helps explain what drives the adoption of marketing an-alytics by firms and why that adoption leads to improved firmperformance.

We find support for our hypotheses that the positive effect of mar-keting analytics deployment on firm performance is moderated bythe level of competition that a firm faces, as well as by the degree towhich the needs and wants of its customers change over time. How-ever, contrary to our hypothesis, the prevalence of marketing analyt-ics use in a given industry does not moderate the effect of marketinganalytics on firm performance. We suggest a possible explanation forthis (non)result: consistent with McKinsey & Co.'s (2009) findings,the prevalence of marketing analytics use in the industries that weexamined is relatively low. That is, the average response of executiveswho participated in our survey to the statement “marketing analyticsare used extensively in our industry” was a 3.4 on a 7-point scale(SD=1.6). Perhaps the moderating effect of marketing analytics'prevalence does not emerge until the industry-wide use of marketinganalytics reaches a higher level than evidenced in our sample. Ourdata simply may not provide the necessary range to manifest suchan effect,12 an issue we plan to examine in more detail in the future.An alternative explanation for the non-significant interaction couldbe that competitors cannot compete away a firm's marketing analyt-ics capability that is implemented properly.

We posit and show empirically that a firm's TMT must ensure thatthe firm (1) employs people with requisite analytics skills, (2) de-ploys sophisticated IT infrastructure and data, and (3) develops a cul-ture that supports marketing analytics so that the insights gainedfrom marketing analytics can be deployed effectively within the firm.

The people who performmarketing analytics (e.g., marketing ana-lysts) are frequently not those who implement the insights gainedfrom marketing analytics (e.g., marketing executives), but bothgroups should support the use of marketing analytics if the firm is topossess a strong marketing analytics-oriented culture (Deshpande etal., 1993). Therefore, a suitable analytics culture that promotes theuse of marketing analytics is a critical component of our framework.Additionally, the centrality of an analytics culture, which is stickyand difficult to change or replicate, suggests that the deployment ofmarketing analyticsmay provide the necessary firm capability proper-ties that can lead to a sustainable competitive advantage (Barney,1991).

5.2. Managerial implications

Our findings offer several useful implications for managerial prac-tice. First, the low prevalence of marketing analytics use indicates thatfew managers are convinced of the benefits of marketing analytics.However, our results suggest that most firms can expect favorableperformance outcomes from deploying marketing analytics. More-over, these favorable performance outcomes should be even greaterin industries in which competition is high and in which customerschange their needs and wants frequently.

The use of objective performance data as the dependent variablein our regressionmodel enables us to quantify the actual performanceimplications of, for instance, a one-unit increase (on a scale of 1 to 7)in marketing analytics deployment. Consider Firm A in our sample,which is at the median (50th percentile) in deployment of marketinganalytics and operates in an industry characterized by average com-petition and average changes in customer needs and wants. For FirmA, a one-unit increase in the deployment of marketing analytics is as-sociated with an 8% increase in ROA. Now, consider Firm B in oursample, which is also at the median (50th percentile) deploymentof marketing analytics but which operates in highly competitive in-dustries with frequently changing customer needs and wants. ForFirm B, a one-unit increase is associated with a 21% average increase

12 We also examined potential curvilinear effects of marketing analytics prevalencebut did not find any such effects.


Table 6Histograms of focal variables.

Note: (Combined) signifies that the graph reports the average scores of the variables that form the respective latent variables. As the histograms illustrate, the firms in the sampledisplay a wide range of values for our focal variables.

Table 7Locus of marketing analytics development and execution.

1=Primarily in-house; 2=Primarily external; 3=Combination of in-house and external; 4=Don't know.


in ROA.13 The 8% increase in ROA translates to an expected increase ofapproximately $70 million in net income for the firms in our sample;the 21% increase indicates an increase of $180 million in net income.14

Second, if implemented properly, the use of marketing analyticscould be a source of a sustainable competitive advantage for a firm.Our study should aid managers in avoiding what appears to be a com-mon misconception, i.e., that simply hiring marketing analysts whoknow how to perform marketing analytics will be sufficient for a firmto benefit from marketing analytics. Instead, we find that TMT involve-ment and a suitable analytics culture that supports the use of marketinganalytics (along with the appropriate IT and data infrastructure) arenecessary for the firm to see the benefits of greater deployment.

AnalyticsSkills

Analytics Culture

y6y5y4

Data and IT

y10

y11

y12

FirmPerformance

y16

y17

y18

.514***

.947***

.294***

.621***

.317**

.427*** .106**

.378***

Deployment of Analytics

y13

y14

y15

Top MgmtTeam

Advocacy

y1 y2 y3Competition Needs & Wants

ChangeAnalytics

Prevalence

y19

y20

y21.037-.076† -.036

.081* .060** -.034

y9y8y7

Fig. 2. Structural equation model results. We used full information maximum likelihood to estimate the model; ***t"3.291, pb .001; **t"2.576, pb .01; *t"1.96, pb .05; †t"1.645,pb .10.

AnalyticsSkills

AnalyticsCulture

y6y5y4

Data and IT

y10

y11

y12

FirmPerformance

y16

.988***

.915***

.233***

.550*

.343**

.453*** .474***

.137 †


y13

y14

y15

Top MgmtTeam

Advocacy

y1 y2 y3

y9y8y7

Fig. 3. Structural equation model results using objective ROA (Time 1) as performance measure. Overall, the model fits the data reasonably well; !2=158.153; CFI=.922; RMSEA=.096, 90% confidence interval of RMSEA=[.068; .123]. ***t"3.291, pb .001; **t"2.576, pb .01; *t"1.96, pb .05; †t"1.645, pb .10.

13 Assuming Firm B's ROA is 0.05, a one-unit increase in deployment of analytics should, onaverage, be associated with an increase in ROA of about 0.01 (i.e., 0.05"1.21#0.06).14 We used our first objective performance measure in this analysis (i.e., the performancemeasure used in regression 1 in Table 8). The average net income of the firms in our samplewas $922million. We note that we repeated the analysis using our second objective perfor-mance measure, and our conclusions did not change in any significant way.


5.3. Limitations and further research

Although we believe that we have broken new ground with thiswork, there are clear limitations, several of which provide avenuesfor future research. First, while our robustness analysis shows thatthe effects that we report are associated with financial returns, ourmain measures are attitudinal, not objective. In addition, we do notexamine the actual return that a firm could expect from its invest-ments in marketing analytics. Thus, obtaining objective data on the

costs and benefits that we measure subjectively in this researchwould be useful.

Second, our findings are correlational, not causal. For example,we find that a higher level of analytics skills and culture ceterisparibus is associated with the deployment of analytics, which inturn, is associated with higher firm performance. However, wecannot make causal claims regarding these relationships. Future re-search could be based on longitudinal data for a sample of firms totrack changes in the precursors of the deployment of marketing

Table 8The effect of analytics deployment on (objective) firm performance (=DV).

Predictor Variable Model 1: ObjectiveROA(Time 1)

Model 2: ObjectiveROA(Time 2)

Parameterestimate

t-Value Parameterestimate

t-Value

Main EffectsDeployment of Analytics .45** 3.06 .24* 2.08Needs & Wants Change .04 .46 .06 .83Competition .11 1.09 .10 1.26Analytics Prevalence .08 .87 .11 1.43

InteractionsDepl"Competition .12 1.60 .11† 1.79Depl"Needs & Wants Change .13* 2.15 .13** 2.68Depl"Prevalence .03 .46 !0.04 !0.63

OtherConstant 5.00 29.14 4.75 35.58R2 32.5% 36.3%Adjusted R2 24.7% 28.9%F-value (7,60) 4.14 4.89F-probability b .001 b .001

Note: For ease of interpretation, we mean-centered the focal variables (i.e., deploy-ment of analytics, needs and wants change, competition, and analytics prevalence) be-fore creating the interaction terms (Echambadi & Hess, 2007). **t"2.576, pb .01;*t"1.96, pb .05; †t"1.645, pb .10.

AnalyticsSkills

AnalyticsCulture

y6y5y4

Data and IT

y10

y11

y12

FirmPerformance

y16

.987***

.916***

.234***

.548*

.354**

.447*** .328***

.137 †


y13

y14

y15

Top MgmtTeam

Advocacy

y1 y2 y3

y9y8y7

Fig. 4. Structural equation model results using objective ROA (Time 2) as performance measure. Overall, the model fits the data reasonably well; !2=149.744; CFI=.932; RMSEA=.089, 90% confidence interval of RMSEA=[.060; .117]. ***t"3.291, pb .001; **t"2.576, pb .01; *t"1.96, pb .05; †t"1.645, pb .10.

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

! =

0.0

! =

0.1

! =

0.2

! =

0.3

! =

0.4

! =

0.5

! =

0.6

! =

0.7

! =

0.8

! =

0.9

! =

1.0

Adj. R-Square

Fig. 5. Contemporary vs. carryover effects on firm performance. This linear combina-tion analysis shows that the highest Adj. R2 occurs for #=.4. This result suggests thatthe deployment to performance link is strongest with an objective performance vari-able that gives 40% of the weight (#=.4) to contemporary effects on firm performanceand 60% to carryover effects.


Table 9Robustness of the deployment to performance link.

Model Parameter estimates and significance levels (two-sided) Conclusion

H1: Deployment of analyticsis positively correlated withperformance measure (pb .05)

H2: The interaction betweendeployment of analytics andcompetition is significant(pb .05)

H3: The interaction betweendeployment of analytics andneeds and wants change issignificant (pb .05)

H4: The interaction betweendeployment of analytics andanalytics prevalence issignificant (pb .05)

OLS regression using objectiveperformance measure (ROA)at Time 1

#Depl. of analytics=.45; p=.003#Depl. of analytics"competition=.12; p=.114#Depl. of analytics"needs and wants change=.13; p=.036#Depl. of analytics"prevalence=.03; p=.645

✓ ✓

SEM in which we averaged theresponses of multiplerespondents of each firm

#Depl. of analytics = .093.; p = .018.#Depl. of analytics"competition=.07; p=.033#Depl. of analytics"needs and wants change=.06; p=.012#Depl. of analytics"prevalance=! .02; p=.352

✓ ✓ ✓

OLS regression using objectiveperformance measure (ROA)at Time 1 and includingquadratic term of deploymentof analytics

#Depl. of analytics=.38.; p=.016#Depl. of analytics2=! .16; p=.149#Depl. of analytics"competition=.08; p=.308#Depl. of analytics"needs and wants change=.14; p=.027#Depl. of analytics"prevalence=.07; p=.404

✓ ✓

OLS regression using objectiveperformance measure (ROA)at Time 1 and includingcontrol variables

#Depl. of analytics=.42; p=.008#Depl. of analytics"competition=.13; p=.113#Depl. of analytics"needs and wants change=.14; p=.040#Depl. of analytics"prevalence=.02; p=.835

✓ ✓

OLS regression using objectiveperformance measure (ROA)at Time 2

#Depl. of analytics=.24; p=.042#Depl. of analytics"competition=.11; p=.078#Depl. of analytics"needs and wants change=.13; p=.009#Depl. of analytics"prevalence=! .04; p=.534

✓ ✓

Mixture regression model(one-class model)


✓ ✓ ✓

SEM using single-item salesgrowth measure from surveyinstrument


✓

SEM using profit and ROImeasures from surveyinstrument

#Depl. of analytics=.198; p=.004#Depl. of analytics"competition=.149; p=.013#Depl. of analytics"needs and wants change=.113; p=.008#Depl. of analytics"prevalence=! .060.; p=.296

✓ ✓ ✓

126F.G

ermann

etal./Intern.J.ofResearch

inMarketing

30(2013)

114–128

analytics to determine how they affect deployment and howchanges in deployment affect firm performance. Such researchshould be feasible because many firms are still in the early stagesof deploying marketing analytics.

Third, our results are based on the overall deployment and impactof marketing analytics. Additional research is needed to understandthe performance implications associated with different types of ana-lytics (e.g., embedded automated models vs. interactive decisionsupport), as well as from various aspects of analytics implementa-tion, such as the nature of the decisions/actions supported by analyt-ics (e.g., segmentation, targeting, forecasting, pricing, sales), and thepenetration ofmarketing analytics into non-marketing decisions andactions.

Fourth, our results are based on and limited to very large U.S.firms. Extending this work to other geographies and to the much larg-er universe of medium-sized and small firms would be useful.

Despite these limitations, we believe that beyond their theoreticalinterest, our framework and findings should prove useful for managerswho are seeking a framework thatwill aid them in deploying their mar-keting analytics investments most effectively. Our results also provide abit of a cautionary tale: Without TMT advocacy and support, the neces-sary investments in data, analytic skills, and a supportive analytics cul-ture are unlikely to occur. We hope that themodest stepwe have takenhere to address the performance implications of marketing analyticswill prove provocative and spawn additional research in this importantarea.

Measure Items

Top management team advocacy$=.84Average variance extracted (AVE)=0.659

1. Our top management has a favorable attitude towards marketing analytics.2. Our annual reports and other publications highlight our use of analytics as a core competitive advantage.3. Our top management expects quantitative analysis to support important marketing decisions.

Analytics culture$=.87AVE=0.692

4. If we reduce our marketing analytics activities, our UNIT's profits will suffer.5. We are confident that the use of marketing analytics improves our ability to satisfy our customers.6. Most people in my unit are skeptical of any kind of analytics-based results (R).

Marketing analytics skills$=.90AVE=0.777

7. Our people are very good at identifying and employing the appropriate marketing analysis tool given theproblem at hand.8. Our people master many different quantitative marketing analysis tools and techniques.9. Our people can be considered as experts in marketing analytics.

Data and IT$=0.72AVE=0.503

10. We have a state-of-art IT infrastructure.11. We use IT to gain a competitive advantage.12. In general, we collect more data than our primary competitors.

Deployment of analytics$=.82AVE=0.657

13. Virtually everyone in our UNIT uses analytics based insights to support decisions.14. In our strategy meetings, we back arguments with analytics based facts.15. We regularly use analytics to support decisions in the following areas (average score across 12 areas tochoose from [pricing, promotion and discount management, sales-force planning, segmentation, targeting,product positioning, developing annual budgets, advertising, marketing mix allocation, new productdevelopment, long-term strategic planning, sales forecasting]+2 open ended areas).

Firm performance$=.81AVE=0.639

Please circle the number that most accurately describes the performance of your UNIT in the following areasrelative to your average competitor (1=well below our competition; 7=well above our competition) Pleaseconsider the immediate past year in responding to these items.16. Total Sales Growth.17. Profit.18. Return on Investment.

Competition 19. We face intense competition.Needs and wants change 20. Our customers are fickle—their needs and wants change frequentlyIndustry prevalence 21. Marketing analytics are used extensively in our industry.

Appendix A. Scale Items

References

Amit, R., & Schoemaker, P. (1993). Strategic assets and organizational rent. StrategicManagement Journal, 14(1), 33–46.

Anderson, E. W., & Sullivan, M. W. (1993). The antecedents and consequences ofcustomer satisfaction for firms. Marketing Science, 12(2), 125–143.

Armstrong, J. S., & Overton, T. S. (1977). Estimating non-response bias in mail surveys.Journal of Marketing Research, 14(3), 396–402.

Bagozzi, R. P., & Yi, Y. (1988). On the evaluation of structural equation models. Journalof the Academy of Marketing Science, 16(1), 74–94.

Bagozzi, R. P., Yi, Y., & Phillips, L. W. (1991). Assessing construct validity in organiza-tional research. Administrative Science Quarterly, 36(3), 421–458.

Barney, J. B. (1991). Firm resources and sustained competitive advantage. Journal ofManagement, 17(1), 99–120.

Baron, R. M., & Kenny, D. A. (1986). Moderator–mediator variables distinction in socialpsychological research: conceptual, strategic, and statistical considerations. Journalof Personality and Social Psychology, 51(6), 1173–1182.

Bergkvist, L., & Rossiter, J. R. (2007). The predictive validity of multiple-item versussingle-item measures of the same constructs. Journal of Marketing Research,44(3), 175–184.

Boulding, W., Lee, E., & Staelin, R. (1994). Mastering the mix: Do advertising promo-tion, and sales force activities lead to differentiation? Journal of Marketing Research,31(2), 159–172.

Carlsson, C., & Turban, E. (2002). DSS: Directions for the next decade. Decision SupportSystems, 33(2), 105–110.

Chen,M. J., Su,K.H.,&Tsai,W. (2007). Competitive tension:Theawareness-motivation-capabilityperspective. Academy of Management Journal, 50(1), 101–118.

Churchill, G. A., Jr. (1979). A paradigm for developing better measures of marketingconstructs. Journal of Marketing Research, 16(1), 64–73.

D'Aveni, R. (1994). Hypercompetition: Managing the dynamics of strategic maneuvering.New York: The Free Press.

Daft, R. L., & Weick, K. E. (1984). Toward a model of organizations as interpretationsystems. Academy of Management Review, 9(2), 284–295.

Debruyne, M., & Reibstein, D. J. (2005). Competitor see, competitor do: Incumbententry in new market niches. Marketing Science, 24(1), 55–66.

DeSarbo, W., & Cron, W. L. (1988). A maximum likelihood methodology for clusterwiselinear regression. Journal of Classification, 5(2), 249–282.

DeSarbo, W., Grewal, R., & Wind, J. (2006). Who competes with whom? A demand-basedperspective for identifying and representing asymmetric competition. StrategicManagement Journal, 27(2), 101–129.

Deshpande, R., Farley, J. U., &Webster, F. E. (1993). Corporate culture, customerorientation, andinnovativeness in Japanese firms: A quadrad analysis. Journal of Marketing, 57(1), 23–37.

Deshpande, R., &Webster, F. E., Jr. (1989). Organizational culture andmarketing: Definingthe research agenda. Journal of Marketing, 53(1), 3–15.

Drolet, A. L., & Morrison, D. G. (2001). Do we really need multiple-item measures inservice research? Journal of Service Research, 3(3), 196–204.


Echambadi, R., & Hess, J. D. (2007). Mean-centering does not alleviate collinearityproblems in moderated multiple regression models. Marketing Science, 26(3),438–445.

Elsner, R., Krafft, M., & Huchzermeier, A. (2004). Optimizing Rhenania's direct marketingbusiness through dynamic multilevel modeling (DMLM) in a multicatalog-brandenvironment. Marketing Science, 23(2), 192–206.

Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models withunobservable variables and measurement error. Journal of Marketing Research,18(1), 39–50.

Gerbing, D. W., & Anderson, J. C. (1988). An updated paradigm for scale developmentincorporating unidimensionality and its assessment. Journal of Marketing Research,25(2), 186–192.

Grant, R. M. (1991). The resource-based theory of competitive advantage. CaliforniaManagement Review, 33(3), 114–135.

Grewal, R., & Lilien, G. L. (2012). Business-to-business marketing: Looking back, lookingforward. In G. L. Lilien, & R. Grewal (Eds.), Handbook of business-to-business marketing(pp. 3–14). Northhampton: Edward Elgar Press.

Hambrick, D. C. (2005). Upper echelons theory: Origins, twists and turns, and lessonslearned. In K. G. Smith, & M. A. Hitt (Eds.), Great minds in management: The processof theory development. New York: Oxford University Press.

Hambrick, D. C., & Mason, P. A. (1984). Upper echelons: The organization as a reflectionof its top managers. Academy of Management Review, 9(2), 193–206.

Harari, O. (1996). Quotations from Chairman Powell: A leadership primer.ManagementReview, 84(12), 34–37.

Hoch, S. J., & Schkade, D. A. (1996). A psychological approach to decision supportsystems. Management Science, 42(1), 51–65.

Hoekstra, J. C., & Verhoef, P. C. (2011). The customer intelligence — Marketing interface:Its effect on firm performance. working draft.

Homburg, C., Grozdanovic, M., & Klarmann, M. (2007). Responsiveness to customersand competitors: The role of affective and cognitive organizational systems. Journalof Marketing, 71(3), 18–38.

Kannan, P. K., Kline Pope, B., & Jain, S. (2009). Pricing digital content product lines: Amodel and application for the National Academies Press. Marketing Science,28(4), 620–638.

Klein, A., & Moosbrugger, H. (2000). Maximum likelihood estimation of latent interac-tion effects with the LMS method. Psychometrika, 65(4), 457–474.

Kotler, P., & Keller, K. L. (2006).Marketing management (12th ed.). Upper Saddle River:Pearson Prentice Hall.

Kucera, T., &White, D. (2012). Predictive analytics for sales and marketing: Seeing aroundcorners. Aberdeen Group Research Brief (www.aberdeen.com)

Kumar, N., Stern, L. W., & Anderson, J. C. (1993). Conducting interorganizational researchusing key informants. Academy of Management Journal, 36(6), 1633–1651.

Lamb, C., Hair, J. F., Jr., &McDaniel, C. (2009).Marketing 3.0.Mason: Thomson/South-Western.Lilien, G. L. (2011). Bridging the academic–practitioner divide in marketing decision

models. Journal of Marketing, 75(4), 196–210.Lilien, G. L., & Rangaswamy, A. (2008). Marketing engineering: Connecting models

with practice. In Berend Wierenga (Ed.), Handbook of marketing decision models(pp. 527–560). New York: Elsevier.

Little, J. D. (1970). Models and managers: The concept of a decision calculus.ManagementScience, 16(8), B466–B486.

Little, J. D. (2004). Comments on models and managers: The concept of a decisioncalculus. Management Science, 50(12), 1841–1861.

Lodish, L. M., Curtis, E., Ness, M., & Simpson, M. K. (1988). Sales force sizing and deploy-ment using a decision calculus model at Syntex Laboratories. Interfaces, 18(1), 5–20.

Lounsbury, M. (2001). Institutional sources of practice variation: Staffing college anduniversity recycling programs. Administrative Science Quarterly, 46(1), 29–56.

MacMillan, I. C., McCaffery, M. L., & Van Wijk, G. (1985). Competitors' responses to easilyimitated new products—Exploring commercial banking product introductions.Strategic Management Journal, 6(1), 75–86.

Marsh, H.W.,Wen, Z., & Hau, K. -T. (2004). Structural equationmodels of latent interactions:Evaluation of alternative estimation strategies and indicator construction. PsychologicalMethods, 9(3), 275–300.

McIntyre, S. H. (1982). An experimental study of the impact of judgment-basedmarketingmodels. Management Science, 28(1), 17–33.

McKinsey & Co. (2009). McKinsey global survey results: Measuring marketing.McKinsey Quarterly, 1–8.

Muthén, B. O. (2010). Mplus discussion— Comment on how to evaluate the fit of a SEMwith interaction, posted October 26, 2010 (accessed July 2011). available at http://www.statmodel.com/discussion/messages/11/385.html?1309276920

Muthén, B. O., & Asparouhov, T. (2003).Modeling interactions between latent and observedcontinuous variables using maximum-likelihood estimation in Mplus.Mplus web notes:No. 6.

Muthén, L. K., & Muthén, B. O. (2010). Mplus user's guide (6th ed.). Los Angeles:Muthén & Muthén.

Muthén, L. K., &Muthén, B. O. (2011). Chi-square difference testing using the Satorra-Bentlerscaled chi-square. (Accessed July 2011). available at http://www.statmodel.com/chidiff.shtml

Natter, M., Mild, A., Wagner, U., & Taudes, A. (2008). Planning new tariffs at tele.ring:The application and impact of an integrated segmentation targeting, and positioningtool.Marketing Science, 27(4), 600–611.

Ouchi, W. G. (1981). Theory Z. Reading: Addison-Wesley Publishing Company.Peters, T. J., & Waterman, R. H. (1982). In Search of Excellence. New York: Harper & Row.Podsakoff, P. M., MacKenzie, S. B., Lee, J. -Y., & Podsakoff, N. P. (2003). Common method

biases in behavioral research: A critical review of the literature and recommendedremedies. Journal of Applied Psychology, 88(5), 879–903.

Porter, M. E. (1996). What Is Strategy? Harvard Business Review, 61–78.Reinartz, W., & Kumar, V. (2000). On the profitability of long lifetime customers: An

empirical investigation and implications for marketing. Journal of Marketing,64(3), 17–32.

Roberts, J., Morrison, P., & Nelson, C. (2004). Implementing a pre-launch diffusionmodel: Measurement and management challenges of the Telstra switching study.Marketing Science, 23(2), 180–191.

Robins, R. W., Hendin, H. M., & Trzesniewski, K. H. (2001). Measuring globalself-esteem: Construct validation of a single-item measure and the Rosenbergself-esteem scale. Personality and Social Psychology Bulletin, 27(2), 151–161.

Ross, J. W., Beath, C. M., & Goodhue, D. L. (1996). Develop long-term competitivenessthrough IT assets. Sloan Management Review, 38(1), 31–45.

Rumelt, R. P. (1984). Towards a strategic theory of the firm. In R. Lamb (Ed.), Competitivestrategic management. Englewood Cliffs: Prentice-Hall.

Russo, J. E., & Schoemaker, P. J. H. (1989). Decision traps. New York: Doubleday andCompany.

Satorra, A., & Bentler, P. M. (1999). A scaled difference chi-square test statistic formomentstructure analysis. Psychometrika, 66(4), 507–514.

Schein, E. H. (2004). Organizational culture and leadership (3rd ed.). San Francisco:Jossey-Bass.

Schimmack, U., & Oishi, S. (2005). The influence of chronically and temporarily acces-sible information on life satisfaction judgments. Journal of Personality and SocialPsychology, 89(3), 395–406.

Silk, A. J., & Urban, G. (1978). Pre-test-market evaluation of new packaged goods: Amodel and measurement methodology. Journal of Marketing Research, 15(2),171–191.

Silva-Risso, J. M., Bucklin, R. E., & Morrison, D. G. (1999). A decision support system forplanning manufacturers' sales promotion calendars.Marketing Science, 18(3), 274–300.

Sinha, P., & Zoltners, A. A. (2001). Sales-force decision models: Insights from 25 years ofimplementation. Interfaces, 31(3), S8–S44.

Smart, C., & Vertinsky, I. (1984). Strategy and environment: A study of corporateresponses to crises. Strategic Management Journal, 5(3), 199–213.

Srinivasan, R., Lilien, G. L., & Rangaswamy, A. (2002). Technological opportunism andradical technology adoption: An application to E-business. Journal of Marketing,66(3), 47–60.

Van Bruggen, G. H., & Wierenga, B. (2010). Marketing decision making and decisionsupport: Challenges and perspectives for successful marketing management supportsystems. Foundations and trends in marketing, Vol. 4, Boston: Now Publishing.

Wanous, J. P., Reichers, A. E., & Hudy, M. J. (1997). Overall job satisfaction: How goodare single-item measures? Journal of Applied Psychology, 82(2), 247–252.

Wernerfelt, B. (1984). A resource-based view of the firm. Strategic Management Journal,5(2), 171–180.

Westphal, J. D., Gulati, R., & Shortell, S. M. (1997). Customization or conformity? Aninstitutional and network perspective on the content and consequences of TQMadoption. Administrative Science Quarterly, 42(2), 366–394.

Wierenga, B., & van Bruggen, G. H. (1997). The integration of marketing-problem-solvingmodes and marketing management support systems. Journal of Marketing, 61(3),21–37.

Wind, J., Green, P. E., Shifflet, D., & Scarbrough, M. (1989). Courtyard by Marriott:Designing a hotel facility with consumer-based marketing models. Interfaces,19(1), 25–47.

Winer, R. S. (2000). Comment on Leeflang andWittink. International Journal of Researchin Marketing, 17(2–3), 141–145.

Zoltners, A. A., & Sinha, P. (2005). The 2004 ISMS practice prize winner: Sales territorydesign: Thirty years of modeling and implementation. Marketing Science, 24(3),313–332.


User-generated versus designer-generated products: A performance assessmentat Muji

Hidehiko Nishikawa a,1, Martin Schreier b,⁎, Susumu Ogawa c,2

a Hosei University, Faculty of Business Administration, 2-17-1 Fujimi, Chiyoda-ku, Tokyo, 102-8160 Japanb WU Vienna University of Economics and Business, Marketing Department, Augasse 2-6, Vienna, 1090 Austriac Kobe University, School of Business Administration, 2-1 Rokkodai-cho Nada-ku, Kobe Hyogo, 657-8501 Japan


Article history:First received in 6, June 2012 and was underreview for 5 monthsAvailable online 17 October 2012

Area Editor: John H. Roberts

Keywords:User designUser-generated productsUser innovationCo-creationCustomer integrationCrowdsourcingMarket performanceNew product developmentIdea generation

In recent years, more and more consumer goods firms have started to tap into the creative potential of theiruser communities to fuel their new product development pipelines. Although many have hailed this para-digm shift as a highly promising development for firms, hardly any research has systematically comparedthe actual market performance of user-generated products with designer-generated ones. We fill this voidby presenting a unique data set gathered from the Japanese consumer goods brand Muji, which has drawnon both sources of ideas in parallel in recent years. We demonstrate that user-generated products, whichare found to generally contain higher novelty, outperformed their designer-generated counterparts on keymarket performance metrics. Specifically, in the first year after introduction, sales revenues fromuser-generated products were three times higher and gross margins were four times greater than those ofdesigner-generated products. These effects also increased over time: after three years, the aggregate salesrevenues of user-generated products were, on average, 1.25 billion yen (approximately 16 million dollars)higher, or five times greater, than the sales of designer-generated products. The corresponding average mar-gin was an impressive 619 million yen (approximately 8 million dollars) higher, or six times greater, than themargin for designer-generated products. Finally, user-generated products were more likely to survive thethree-year observation period than designer-generated products (i.e., were still on the market three yearsafter introduction). These findings clearly favor the paradigm shift identified in marketing research and ap-peal to managers considering the integration of user ideas into the process of new product development.We discuss our study's limitations and identify important avenues for future research.


1. Introduction

For decades (if not for much longer), professional marketers,designers, and engineers rather than consumers or users have beenthe dominant agents in the development of new consumer products.Although it has always been imperative to listen to customers to dis-cern emerging market trends and unmet consumer needs, designprofessionals employed by firms (or their subcontractors) are usuallyresponsible for translating the resulting opportunities into new prod-uct ideas and market offerings (Cooper, 2001; Crawford & DiBenedetto, 2000; Ulrich & Eppinger, 1995; Urban & Hauser, 1993).Until very recently, firms did not consider user involvement in theprocess of generating ideas for new product development (NPD) to

be particularly promising. For example, as Schulze and Hoegl (2008,p. 1744) note, “relying on the method of asking buyers to describepotential future products, big leaps to novel product ideas are generallynot likely.” One of the key assumptions underlying this dominant ideageneration paradigm is that a firm's professionals, unlike customers orusers, have the requisite expertise to invent new and useful productconcepts (e.g., Amabile, 1998; Weisberg, 1993). A firm's professionals“have acquired skills and capabilities that allow them to perform mostdesign tasks more effectively and at a higher level of quality” (Ulrich,2007, Chapter 3, 5ff) and they “often have a significant advantage […]over consumers, in terms of their knowledge, training, and experience”(Moreau & Herd, 2010, p. 807).

However, research on the sources of innovation (Von Hippel, 1988)has long challenged this fundamental assumption and the related,widely held view that users are of little value to firms in providingideas to solve their unmet consumer problems (as opposed to merelypointing out such problems to firms in the first place). In particular,this line of research has found that many major and minor innovationsacross various industrieswere originally developed by users rather thanprofessionals in firms (cf. Von Hippel, 2005). In their seminal work,


⁎ Corresponding author. Tel.: +43 1 31336 4682; fax: +43 1 31336 732.E-mail addresses: [email protected] (H. Nishikawa), [email protected]

(M. Schreier), [email protected] (S. Ogawa).1 Tel./fax: +81 3 3264 9643.2 Tel.: +81 78 8 3 6932; fax: +81 78 803 6977.





Lilien,Morrison, Searls, Sonnack, and vonHippel (2002) provide furtherevidence that active user involvement in idea generation might benefita firm's NPD efforts, at least in industrial markets. Specifically, Lilien etal. (2002) compare the potential value of several new industrial productconcepts jointly developed by 3 M personnel and selected lead users tothose developed by more conventional means (i.e., internal developersonly). They found that the former outperformed the latter on key inno-vation indicators that were assessed by 3 Mmanagers (e.g., the productconcept's novelty compared to the competition or its potential to createan entire new product family). Furthermore, they found that the salesforecasts for concepts developed by lead users were eight times higherthan those of internally developed ideas.

Onemight plausibly argue that workingwith knowledgeable indus-trial users is one thing, while working with potential end consumers isanother. However, in various consumer goods domains as well, usersare frequently found to innovate for themselves, and many of theirinnovations are commercially attractive (Franke, von Hippel, & Schreier,2006). For example, a recent survey of a representative sample of UK con-sumers revealed that six percent, or nearly threemillion consumers, inno-vated in the domain of household products. Similar national userinnovation statistics have been reported for other countries, includingthe US (5%, or almost 12 million consumers) and Japan (4%, or almost4 million consumers) (Von Hippel, Ogawa, & de Jong, 2011).

Thus, user innovation has come of age as many consumer goodsfirms are challenging the traditional paradigm and experimentingwith new ways to more actively integrate users into the idea genera-tion process. In extreme cases, firms like Threadless no longer employdesigners but rely exclusively on their user communities to generatenew products (Ogawa & Piller, 2006). Many established firms andbrands, such as LEGO, Dell, and Muji, have followed suit and nowcomplement their in-house efforts with public idea contests knownas “crowdsourcing” initiatives, where idea generation is outsourcedto a potentially large, unknown population (the crowd) in the formof an open call (Bayus, 2010; Howe, 2008; Surowiecki, 2004). Thus,crowdsourcing relies on a self-selection process among users whoare willing and able to respond to widely broadcast idea generationcompetitions (Jeppesen & Lakhani, 2010; Piller & Walcher, 2006;Poetz & Schreier, 2012).

Conceptually, the key insight is that some (but, of course, not all)ideas generated by these self-selected users might outperform theones generated by firm designers (Poetz & Schreier, 2012; Schreier,Fuchs, & Dahl, 2012). Users may have a competitive edge in idea gen-eration over designers through their experience as consumers.Leading-edge users, in particular, may have tried to solve consump-tion problems themselves, and their ideas may be commerciallyattractive because they foreshadowwhat other consumerswill demandin the future (cf. Von Hippel, 2005). The respective user population thatmight be activated via crowdsourcing is also naturally much larger andmore diverse compared to the team of designers employed by a givenfirm; a firm's user community may comprise thousands of talentedusers from highly diverse backgrounds. It follows that the generationof more (and more diverse) ideas increases the odds of generating afew truly exceptional ones (e.g., Gross, 1972; Schreier et al., 2012;Surowiecki, 2004; Terwiesch & Ulrich, 2009).

Empirically, researchers have recently begun to address importantquestions like what motivates users to participate (e.g., Hertel,Niedner, & Herrmann, 2003), how idea contests should be organized(e.g., Boudreau, Lacetera, & Lakhani, 2011), how user ideas can bescreened effectively (e.g., Toubia & Flores, 2007), or how consumersperceive user-driven firms (e.g., Fuchs & Schreier, 2011). In a recentstudy, for example, Schreier et al. (2012) find that consumers associ-ate user-driven firms with higher innovation abilities, that is, theyperceive such firms as being better able to generate innovative newproducts. In a series of experimental studies, they further find thatthis innovation inference prompts consumers to demand one andthe same set of products more strongly. This finding is especially

interesting because objective product properties were kept constantbetween experimental conditions.

From an NPD perspective, however, the important question is notif the design mode affects consumer perceptions of innovation ability,but rather if it affects the objective quality of new product ideas actu-ally generated. This question has recently been addressed by Poetzand Schreier (2012), who compared the quality of ideas for newbaby products generated by a firm's designers to those generated byusers in a public idea contest. Ideas from both sources were judgedby company executives (who were blind to the source of ideas), andthe key finding was that user ideas were characterized by higher nov-elty and higher customer benefit, but also by lower feasibility.

Although these findings are promising, it has yet to be exploredwhether the best of these user ideas, if realized, would eventuallyperform better on the market. There are several reasons that makethis extrapolation non-obvious. First and foremost, it should benoted that in most cases firms still have to translate user ideas intomarketable products; a process that demands much effort andinvolves many decisions that may or may not lead to ultimate marketsuccess. More specifically, lower feasibility scores of user ideas pointto the possibility that the related promise might never be realized.Second, and on a related note, it could be that high innovativenessand low feasibility eventually boost development and productioncosts and thereby reduce any potential margin. Third, it is always anempirical question how well managers' perceptions of ideas matchconsumer reactions to new products once they reach the shelves.Even if this match is strikingly high (i.e., no judgment and translationerrors are involved), it remains unclear how a certain increment ininnovativeness (e.g., 10%)will translate into incrementalmarket perfor-mance (e.g., X% more sales). Finally, it is worth noting that firms canonly market a very limited number of new products at a given pointin time. As such, what matters is not the entire distribution and theaverage idea quality, but rather the extreme values: the very bestideas available (e.g., Fleming, 2007; Singh & Fleming, 2010). It remainsunclear whether the very best ideas from both sources would differ tosuch an extent that one could observe managerially meaningful differ-ences in the ultimate consumer reactions.

Against this backdrop, it is surprising that a systematic, empiricalmarket performance assessment of user- vs. designer-generated con-sumer products is still missing in the literature, despite the broad andgrowing interest in this topical phenomenon. This is most likely dueto the difficulty of obtaining reliable quantitative market data fromreal-world practice. Such research is important because the marketperformance of user-driven initiatives will guide marketers in decid-ing whether to follow the paradigm shift and involve users in gener-ating new product ideas. Drawing on a unique data set gathered fromthe Japanese consumer goods brand Muji, we address this importantgap in the literature. In short, our study reveals that user-generatedproducts systematically outperform designer-generated ones in termsof important outcome variables, including actual sales revenues. Mostimportantly, we see our study as “initial evidence” in favor of user-generated products and as a contribution to a more critical reflectionon the dominant NPD paradigm. As such, this paper makes a secondcontribution in its conceptual discussion of the generalizability of ourfindings.

2. Study method

2.1. New product development at Muji

Muji began experimenting with idea contests ten years ago; thus,it is a forerunner among firms that market user-generated productson a broad scale. Muji has also recently marketed a number of user-generated products alongside products generated by the firm'sdesigners, which allows us to empirically assess the performance ofthe two approaches. In addition to granting us access to sensitive

161H. Nishikawa et al. / Intern. J. of Research in Marketing 30 (2013) 160–167

data (e.g., sales data, gross margins), the firm provided us withsubstantial support to complement and validate performance data(e.g., interviews with executives, coding of ideas).

Muji is an established Japanesemanufacturer and retailer brand of abroad range of consumer goods of Ryohin Keikaku Co., Ltd, with a par-ticular focus on interior and household products. Muji's products aresold in almost 500 Muji stores in 22 countries including Japan, China,the US, the UK, France, Italy, and Germany. In 2010, Muji generatedapproximately 170 billion yen in sales revenue (approximately2.2 billion dollars), producing net profits of almost 8 billion yen(approximately 100 million dollars). Over the last few decades,Muji has become known for developing and marketing innovativeconsumer solutions, which is reflected in their receipt of more than90 design and innovation prizes (e.g., two international DesignOscars awarded by iF International Forum Design, two Good DesignAward Gold awards from the Japan Institute of Design Promotion,one Michael E. Porter Prize for innovation). Their ability to generatesuccessful innovations sets a high standard of performance foruser-generated products.

For the purposes of our research, we focus on Muji's furnituredivision. By focusing on one product type, we minimize the potentialof confounding effects that arise from comparing different productcategories. The furniture division is strategically important to Mujibecause it accounts for nearly 20 percent of their total sales. Wealso selected Muji's furniture division because it has experimentedmore with user-generated products in recent years than other divi-sions; in other words, data for the furniture division contained thelargest number of observations for user-generated products. Onemight argue that this method of sample selection may bias our resultsbecause the furniture division began to rely on user input due to theunderperformance of designers, thereby lowering the performancestandard that user-generated products must exceed. In our inter-views at Muji, however, we found no evidence for such selection bias.

In 2002, Muji launched the first user-driven piece of furniture, the“Floor Sofa,”which is a large cushion that sits on the floor and adjuststo the body. Because of the novelty and experimental nature of thisproject, Muji collected binding customer pre-orders before investingin final product development to minimize the risk of failure. TheFloor Sofa was a great success. Within the first year, it generatedalmost 500 million yen in sales, a record that was not exceeded byany of Muji's designer-generated products until the beginning ofour observation period (2005). Prior to 2005, the firm carried outanother crowdsourcing experiment that led to the production of anew piece of furniture in 2003 (a small, movable wall shelf) that gen-erated first-year sales of 71 million yen (ranking it sixth among the15 new furniture products introduced in 2003).

Based on the success of these products, the firm decided to usecrowdsourcing on a more systematic basis. During our observationperiod (February 2005 to July 2009), 43 new products were developed,produced, and introduced to the Japanese market. Thirty-seven prod-ucts were designer-generated, and six were user-generated. Weexcluded the two previous user-generated products from our analysisbecause they differed from the other products in terms of the idea selec-tion process (binding customer pre-orders compared to selection by thefirm). However, the results do not change if the observation period isextended to include these two early user-generated products, as wellas the respective designer-generated products (results available fromthe authors upon request). In what follows, we describe the develop-ment of designer- and user-generated products in greater detail.

2.1.1. Designer-generated new productsMuji's conventional approach to NPD adheres closely to the recom-

mendationsmade in the traditional literature onmarketing and innova-tion. The active agents in idea generation are the firm's designers,engineers, and marketers, whowork in teams organized around specif-ic, new product development projects. Based on established market

research (e.g., market trend studies, focus groups with target con-sumers), the team identifies and selects important consumption prob-lems to be addressed by a new product (referred to as new product“themes”). In one product development project, for example, thetheme was limited storage capacity in consumers' bedrooms due tosmall room sizes. Based on insight from market research, members ofthe project team generate new product ideas (e.g., a flat-panel shelfmatching the firms' line of beds). The best ideas enter the next phaseof product development, where designers propose more detaileddesign concepts and product specifications based on the product'sintended functionality (e.g., a specific design of a flat-panel shelf that isas thin as possible yet large enough to guarantee proper use and stabili-ty). The team then decideswhat concept tomove to the next stage. Final-ly, a new product is introduced to the market. The key characteristic forour study is that it is the internal employees who generate new productideas aimed to address the selected consumer problem (i.e., the newproduct theme).

2.1.2. User-generated productsIn a second development track, Muji invites users to generate ideas

for new products. Anyone who registers on their website can partici-pate. Similar to the internal development of designer-generated prod-ucts, idea generation follows a specific theme for which solutions canbe proposed. Users can also strongly influence the selection of thetheme by participating in online discussion boards (as is the case inthe designer-generated track). The theme “Suwaru Seikatsu” (“SitDown Life”), which aimed to generate ideas for better and more com-fortable sitting solutions in consumers' homes, was one of the firstthemes to be successfully outsourced to users. The themewas broadcastto users, who responded with many new product ideas, including theeventual winner: the Floor Sofa. After the first stage of idea generation,users can also comment, vote, and improve on each other's ideas. Thebest idea is adopted by a Muji project team, which defines the productcharacteristics in more detail (e.g., external fabrics, cushion filling),seeks further feedback from the user community, and finally movesthe new product to the full production stage. Thus, Muji employeesare also heavily involved in the development process for new productsin this second track. In contrast to the first track, however, new productideas are generated by users rather than the firm's designers.

Adopted user ideas in our observation period included the follow-ing: a sofa-bed (a sofa that can easily be changed to a bed and viceversa); a wooden cupboard with a movable wagon to increase flexi-bility of use; a small sofa chair with slim arms to fit in small rooms;a highly versatile stacking shelf; a bed with a thin, yet robust, woodenframe to provide more storage space under the bed; and a sofa witharm rests that are comfortable for both sitting (arms) and lyingdown (head and legs).

A series of in-depth interviews with senior managers and projectmembers revealed that the twodesign tracks did not differ substantiallyon other important dimensions. In particular, projects centered arounddesigner-generated vs. user-generated products did not differ in termsof the human resources invested by the firm (e.g., the quantity andquality of project staff). Furthermore, it is unlikely that employee moti-vation differed substantially between the two development processesbecause the firm's employees usually work on several projects at once(including designer- and user-generated products), and both individualand team incentives are identical across projects. Finally, the firm's ideaand concept selection criteria are identical in both tracks; a minimumthreshold of expected sales and gross margins must be met before aproject moves to full production. Interviews with managers furtherrefuted the argument that adopted user-generated ideas “survived atougher screening” process. Therefore, it is unlikely that any differencesin performance between designer- and user-generated products areattributable to these alternative factors.

Nonetheless, the number of ideas generated differed systematicallybetween the two processes, which is consistent with the theoretical

162 H. Nishikawa et al. / Intern. J. of Research in Marketing 30 (2013) 160–167

work on user involvement and crowdsourcing (Poetz & Schreier, 2012;Schreier et al., 2012). The average estimated number of ideas per themewas approximately ten in the designer condition (this number did notvary much across projects because teams are expected to come upwith at least ten good ideas before a choice ismade), whereas the lowestnumber of ideas per theme was over 400 in the user condition (theobserved maximum was over 2000). Any random user idea, therefore,has clearly lower odds of being selected, ceteris paribus. However, thisdoes not bias our results. If both distributions empirically ended at thesame right-hand point and these top ideas were selected, there shouldstill be no difference in market performance, as there were no differ-ences in the selection criteria employed. If, however, extreme valuesof the user distribution stretched farther to the right (e.g., becausehaving more ideas increases the likelihood of generating some extremevalues; e.g., Gross, 1972; Terwiesch & Ulrich, 2009), the best user-generated ideasmay verywell perform better. Such a scenario is consis-tent with the theoretical arguments that favor user involvement.

2.2. Measures

2.2.1. Independent variableThe independent variable is the source of idea generation: profes-

sional designers vs. users (i.e., user-generated vs. designer-generatedproducts).

2.2.2. Dependent variablesOur key outcome variable is the products' actual market perfor-

mance, measured as aggregate unit sales generated in the first yearafter market introduction. In addition, we capture profitability bymeasuring the products' monetary value of sales and gross margins.Because the firm does not allocate fixed costs to individual products,we cannot assess the first-year profits of products, but only theirgross margins. Data from the first year that a product is on the marketform the most important basis for a product's performance assess-ment by Muji because furniture products are characterized by shortproduct life cycles (i.e., the first year usually produces the highestsales). To check the robustness of our results, we also gathered aggre-gate sales data three years after a product was introduced to the mar-ket (the average product life cycle of Muji products is approximatelythree years), as well as product survival information (i.e., informationabout which products survived our observation period and were stillon the market three years after introduction).

In addition to performance data, we coded all new products (inrandom order) based on important innovation indicators to gain in-sight into the nature of the underlying product ideas (e.g., Lilien etal., 2002). First, we asked the general manager of the furniture divi-sion and the person in charge of each new product development pro-ject to rate the products' novelty relative to competing products andits novelty based on customer needs addressed (measured on a10-point scale where 1=low and 10=high). Because both measuresof novelty were highly correlated for both raters (rRater 1=.52,pb .001, rRater 2=.84, pb .001), we averaged these measures to createa novelty index. The high correlation between raters (rBetween raters=.51, pb .001) suggests that novelty is captured by this index in a validway (e.g., Franke et al., 2006). We also captured the products' strate-gic impact on the firm's product portfolio, which we measured usingtwo items: the strategic importance of the product to the businessunit's future and the potential of the product to generate a new prod-uct line (measured on a 10-point scale where 1=low and 10=high).Because bothmeasureswere highly correlated for both raters (rRater 1=.50, pb .01, rRater 2=.43, pb .01), we averaged them to create a strategicimpact index. The high correlation between raters (rBetween raters=.47,pb .01) suggests that this variable captures a product's strategic impactin a valid way.

2.2.3. Control variablesWe considered the products' price and their year of market intro-

duction as control variables. We re-ran our main analysis (reportedbelow) and controlled for these variables in OLS regressions. As wefind that neither price nor inter-year variations in sales, etc. accountfor our findings (all reported effects remain significant), we do notreport the resulting statistics due to space constraints (results areavailable upon request).

3. Findings

To compare the market performance of user-generated anddesigner-generated products, we compared the means of market per-formance measures for both types of products and tested their statis-tical significance using t-tests. Given the small sample size of the userdesign condition (n=6), we also performed Mann–Whitney U-teststo assess the robustness of our results. The results indicate that bothtests lead to the same conclusions (see Table 1A).

3.1. Unit sales

In thefirst year after introduction,wefind that user-generated prod-ucts sell roughly twice as much as designer-generated products(MUser=30,182.33, MDesigner=14,191.76, pt-testb .08). Similarly, wefind that the distribution of the products' first-year unit sales is signifi-cantly different between the two conditions (pM-W-Ub .05). Five out ofsix user-generated products are represented in the top third of thebest sellers among all 43 products (ranked 2, 4, 7, 10 and 11), andonly one user-generated product had lower unit sales than the averagedesigner-generated product (see Fig. 1). The results are evenmore pro-nounced using aggregate sales data across three years: units of user-generated products sold three times more frequently than designer-generated ones (pt-testb .12; pM-W-Ub .05).

3.2. Sales

First-year sales revenues of user-generated products are more thanthree times higher than those of designer-generated products (MUser=500.97 million yen, MDesigner=141.04 million yen, pt-testb .05). Similar-ly, the distribution of sales revenues differs significantly between thetwo types of products (pM-W-U=.001). Four of the top six products, interms of sales, are user-generated (ranked 1, 2, 3 and 6), and the tworemaining user-generated products generated higher sales revenuesthan the average designer-generated product (greater than 147.80 mil-lion yen, see Fig. 2). Aggregate sales data across three years suggeststhat the size of this effect increases over time: three years after intro-duction, user-generated products generated an average of 1250 millionyen (approximately 16 million dollars) more in sales than designer-generated products, which is equal to over five times the sales ofdesigner-generated products (pt-testb .05; pM-W-U=.001).

3.3. Gross margin

User-generated products yielded gross margins (MUser=240.47 mil-lion yen) that were approximately four times higher than designer-generated products (MDesigner=56.22 million yen, pt-testb .05). The dis-tribution of gross margins is also significantly different between thetwo conditions (pM-W-U=.001). In particular, four of the top five prod-ucts, in terms of gross margins, are user-generated products (ranked 1,2, 3 and 5) and the remaining two user-generated products yieldedhigher gross margins than the average designer-generated product(greater than 67.90 million yen; see Fig. 3). The results are more pro-nounced if we use three years of sales data: the average margin ofuser-generated products is 619 million yen (approximately 8 milliondollars) higher, or six times greater, than designer-generated products


(pt-testb .05; pM-W-U=.001). These findings suggest that user-generatedproducts demonstrate stronger market performance than theirdesigner-generated counterparts.

Because performance should be related to survival, we additionallyanalyzed how many designer- and user-generated products survivedour observation period (three years after introduction). Products thatsurvived performed substantially better than those which werediscontinued (e.g., first-year unit sales were approximately 60% higherfor products that survived the third year). Indeed, only 17 out of 37designer-generated products were still on the shelves after three years.In contrast, five out of six user-generated products were still on themar-ket (pchi-square-difference-testb .09). User-generated products, therefore,have a substantially higher survival rate at Muji.

3.4. Novelty and strategic impact

The results with regard to the nature of each new product providecorroborating evidence for the findings reported above. In particular,we find that user-generated products are characterized by substan-tially higher novelty compared to designer-generated products(MUser=7.29, MDesigner=4.78, pt-testb .001, pM-W-Ub .001). Four ofthe top five products, in terms of novelty, are user-generated (ranked1, 2, 3 and 4) and the remaining two user-generated products arerated among the top 50% of the sample on novelty and consistentlyexhibit higher novelty than the average designer-generated product(greater than 5.90, see Fig. 4). The results for the products' strategicimpact are similar: user-generated products have a significantlymore pronounced strategic impact on the firm's future product port-folio than designer-generated ones (MUser=6.08, MDesigner=4.25,pt-test=.001, pM-W-Ub .01). Three of the top five products, in termsof strategic impact, are user-generated (ranked 1, 3 and 4), and the“worst” user-generated product has a strategic impact score (4.00)close to the average score of designer-generated products (see Fig. 5).

3.5. Sensitivity analysis

Although we only analyzed products within a single category(furniture), the sample consists of different products (e.g., sofa,table, chair, bed, shelf). To assess whether different product typeswithin the furniture category might partially account for the findingsreported above (beyond price considerations), we re-analyzed thedata using a more homogeneous subsample of products. We matchedthe six user-generated products with similar designer-generated ones(n=15; sofa, bed, shelf). As Table 1B shows, the results remainrobust; user-generated products outperform designer-generated oneson all dimensions.

Table 1Summary of findings (mean statistics).

(A) Analysis of the full sample (furniture division)

Designer-generatedproducts (n=37)

User-generatedproducts (n=6)

Significance

M M

Unit sales (after 1styear)

14,191.76 30,182.33 pM-W-Ub .05;pt-testb .08

Unit sales (after 3rdyear)

29,959.65 91,120.50 pM-W-Ub .05;pt-testb .12

Sales in yen (after 1styear)

141.04 million 500.97 million pM-W-U=.001;pt-testb .05

Sales in yen (after 3rdyear)

276.46 million 1,525.63 million pM-W-U=.001;pt-testb .05

Gross margin in yen(after 1st year)


Gross margin in yen(after 3rd year)


Newnessa 4.78 7.29 pM-W-Ub .001;pt-testb .001

Strategic impacta 4.25 6.08 pM-W-Ub .01;pt-test=.001

(B) Sensitivity analysis (product-type matching of user- and designer-generatedproducts)



M M Significance


8905.93 30,182.33 pM-W-Ub .05;pt-testb .08


18,305.53 91,720.50 pM-W-Ub .05;pt-testb .08

Sales in yen (after1st year)

174.43 million 500.97 million pM-W-Ub .05;pt-testb .07

Sales in yen (after3rd year)

313.03 million 1,525.63 million pM-W-Ub .01;pt-testb .05





Newnessa 4.93 7.29 pM-W-Ub .001;pt-testb .001

Strategic impacta 4.47 6.08 pM-W-Ub .001;pt-testb .01

a Measured on a 10-point scale where 1=low and 10=high

0

20

40

60

80

100

120

User

User

User

UserUse

rUse

r

Uni

t Sal

es in

Firs

t Yea

r (in

thou

sand

s of

uni

ts)

Source of product design(User: user-generated product, No label: designer-generated product)

Fig. 1. Unit sales per product (furniture).


In a second robustness check, we aimed to broaden the scope ofour analysis. In particular, we sought market performance data fromanother category, namely health and beauty products. This categoryhas brought about four user-generated products in recent years (inthe area of natural and homeopathic remedies and skin care) wherethe respective product launch was far enough in the past in order totrack aggregate performance data across three years. Table 2 showsthat user-generated products once again demonstrate better marketperformance. Although we identified several other Muji categorieswith user-generated products (e.g., kitchenware, stationery, garden-ing products, and clothing), user-generated products in these catego-ries were either too low in number (e.g., one or two products) and/ormarket introduction had occurred too recently (e.g., in 2010).

4. Discussion

There is broad and growing interest in understanding the topicalphenomenon of more actively involving users in firms' idea generationprocesses. Yet and despite conceptual advancements (Poetz & Schreier,2012; Schreier et al., 2012) it is surprising that an empiricalmarket per-formance assessment of user-driven products (i.e., products that arebased on user- vs. designer-generated ideas) is still missing in the liter-ature. This is most likely due to the difficulty of obtaining reliable mar-ket data from real-world practice. As such, it remains unclearwhether it

will pay off for firms to depart from the classic NPD paradigm anchoredin the idea that firm professionals – and not users – should generate theideas for new products to be marketed to broader customer segments.By drawing on a unique data set gathered from the Japanese consumergoods brand Muji, which has drawn on both sources of ideas in parallelin recent years, we provide initial research to fill this gap in the litera-ture. We demonstrate that user-generated products systematicallyand substantially outperform their designer-generated counterparts interms of key market performance metrics.

Of course, we do not see our study as the “final battle” that wouldonce and for all relegate the “loser” to the history books. Rather, wesee our study as initial evidence and as a contribution to a more crit-ical reflection on the dominant NPD paradigm. In a nutshell, our find-ings suggest that it might pay for firms to draw on user ideas inparallel to their established in-house efforts — not unlike what Mujiseems to have identified as their best practice. Furthermore, userswere only the providers of ideas; the overall NPD process entailsmany more stages, and it goes without saying that decisive in-houseefforts and capabilities are needed to convert any promising ideainto a successful new product. Bearing this in mind, we now discussthe generalizability of our findings and identify promising directionsfor future research.

First, it is worth noting that the majority of winning user-designers in our sample were highly involved in the category, had

0

200

400

600

800

1000

1200

Sal

es in

Firs

t Yea

r (in

mill

ions

of y

en)

UserUse

rUse

rUse

rUse

rUse

r

Source of product design (User: user-generated product, No label: designer-generated product)

Fig. 2. Sales per product (furniture).

0

100

200

300

400

500

600

Gro

ss M

argi

n in

Firs

t Yea

r (in

mill

ions

of y

en)

UserUse

rUse

rUse

rUse

rUse

r


Fig. 3. Gross margin per product (furniture).


enduring past exposure to the problems for which ideas were sought,had relatively high levels of technical expertise, and thus generallymatched the conceptual description of lead users (cf. Von Hippel,2005). Clearly, if firms are unable to activate this type of user, it isunlikely that the effects found in this study can be replicated. Thenotion of users' leading-edge status is also conceptually important.Different types of users might account for the discrepancy betweenpredictions made in the classical marketing and NPD literatures, onthe one hand, and research on user innovation, on the other hand.Whereas the former might have “average” users in mind (whichwould justify skepticism about user involvement because theseusers lack both technical and usage expertise), the latter typicallyconstructs their arguments around leading-edge users. In otherwords, conventional marketing is concerned with the “average user”and research on user innovation targets only users with leading-edge status who represent “extreme values” on the right-hand sideof the distribution of users.

The consequence of involving the “right” vs. “wrong” users mightloom even larger for more complex products than the ones studied inthis paper (Poetz & Schreier, 2012; Schreier et al., 2012). For complexproducts (e.g., cars), the expertise required for successful ideationmight be too high, thereby preventing the vast majority of averageusers from contributing meaningful ideas. On a related note, all ofthe furniture themes studied were based on a clear “usage problem”(e.g., improving the sitting experience in the living room). We argue

that such problems are well-suited for users because users have acompetitive edge over designers in terms of the needs-based infor-mation necessary to solve the problem effectively. If, in contrast, thethemewas based on technical details regarding technologies or mate-rials, company designers would be more likely than users to have theskills and expertise to solve these problems. These speculations offerpromising avenues for future research.

Furthermore, the openness of contests may favor user-based ideas.The user population that is accessed through crowdsourcing is naturallymuch larger and more diverse compared with a team of designersemployed by a given firm. It follows that, if more (and more diverse)ideas are generated, the odds of generating a few truly exceptionalideas will increase (e.g., Gross, 1972; Schreier et al., 2012; Surowiecki,2004; Terwiesch & Ulrich, 2009). Our data from Muji support thisprediction: out of the many ideas submitted by users (ranging from400 to more than 2000 per theme), the best user ideas outperformedthose generated by designers. Yet the relationship between the open-ness of contests, how the contest announcement is targeted, the activa-tion of effective self-selectionmechanisms, and the final outcome is notwell understood; thus, it constitutes fruitful ground for future research.

Interviews with managers at Muji further revealed that they havesucceeded in building the specific capability to interact with userseffectively. Nonetheless, each crowdsourcing initiative took substan-tial time to implement, is generally less controllable and manageablein the early stages, and it cannot easily be implemented on a large

1

2

3

4

5

6

7

8

9

10

New

ness

(1

= lo

w, 1

0 =

high

)

UserUse

rUse

rUse

rUse

rUse

r


Fig. 4. Newness per product (furniture).

1

2

3

4

5

6

7

8

9

10

Str

ateg

ic Im

pact

(1

= lo

w, 1

0 =

high

)

User

UserUse

rUse

rUse

rUse

r


Fig. 5. Strategic Impact per Product (Furniture).


scale (as reflected in the relatively small number of user-generatedcompared with designer-generated new products). Thus, our findingsmight not be easily replicated with data from “old-school” firms withno prior experience in active user involvement. As a result, oneimportant, but unanswered question remains: what are the specificcapabilities that make user-driven firms successful?

Finally, it is worth noting that Muji did not actively market userproducts as “designed by users” to the general public. We, therefore,only assessed the “objective”market performance of user-driven prod-ucts. As indicated above, however, recent research suggests that firmsmight benefit from a “designed by users” label over and above anyobjective product characteristics (Schreier et al., 2012). To assess thefull potential of marketing user-generated products, future researchshould examine this psychological effect on performance.

In conclusion, we note that further research is required to fullyunderstand the newly emerging paradigm of an active user involve-ment in NPD. Our study provides important initial evidence thatsuch efforts might pay off.

Acknowledgments

The authors thank Muji, in particular Satoshi Shimizu, TsunemiKawana, Tetsu Takahashi, Tomisaburou Hagiwara, Yukie Saito and

Ikuo Takusari, for their support in realizing this project. They furtherthank Pradeep K. Chintagunta, BerendWierenga, and research seminarparticipants at the Universitat Pompeu Fabra Barcelona, Università diBologna, and VU University Amsterdam for their feedback on earlierdrafts of the manuscript as well as Marnik G. Dekimpe, the AE, andthe IJRM reviewers for their helpful suggestions and guidance duringthe review process. Financial support is greatly acknowledged fromJSPS KAKENHI (Grant Numbers 22330130, 24530535) and the DanishStrategic Research Council (Grant Number 09-067157).

References

Amabile, Teresa M. (1998). How to kill creativity. Harvard Business Review, 76(5),76–87.

Bayus, Barry L. (2013). Crowdsourcing new product ideas over time: An analysis of thedell ideastorm community. Management Science, 59(1), 226–244.

Boudreau, K. J., Lacetera, N., & Lakhani, K. R. (2011). Incentives and problem uncertainty ininnovation contests: An empirical analysis. Management Science, 57(5), 843–863.

Cooper, Robert G. (2001). Winning at new products: Accelerating the process from idea tolaunch (3rd edition). Cambridge: MA: Perseus Books.

Crawford, C. Merle, & Di Benedetto, C. Anthony (2000). New products management (6thedition). Burr Ridge: IL: Irwin-McGraw Hill.

Fleming, Lee (2007). Breakthroughs and the 'long tail' of innovation. Sloan ManagementReview, 49(1), 69–74.

Franke, Nikolaus, von Hippel, Eric, & Schreier, Martin (2006). Finding commerciallyattractive user innovations: A test of lead user theory. Journal of Product InnovationManagement, 23(4), 301–315.

Fuchs, Christoph, & Schreier, Martin (2011). Customer empowerment in new productdevelopment. Journal of Product Innovation Management, 28(1), 17–32.

Gross, Irwin (1972). The creative aspects of advertising. Sloan Management Review,14(1), 83–109.

Hertel, Guido, Niedner, Sven, & Herrmann, Stefanie (2003). Motivation of softwaredevelopers in open source projects: An internet-based survey of contributors tothe Linux kernel. Research Policy, 32(7), 1159–1177.

Howe, Jeff (2008). Crowdsourcing: Why the power of the crowd is driving the future ofbusiness. New York: Crown Business.

Jeppesen, Lars Bo, & Lakhani, Karim R. (2010). Marginality and problem-solvingeffectiveness in broad search. Organization Science, 21(5), 1016–1033.

Lilien, Gary L., Morrison, Pamela D., Searls, Kathleen K., Sonnack, Mary, & von Hippel,Eric (2002). Performance assessment of the lead user idea generation process.Management Science, 48(8), 1041–1059.

Moreau, C. Page, & Herd, Kelly B. (2010). To each his own? How comparisons withothers influence consumers' evaluations of their self-designed products. Journalof Consumer Research, 36(February), 806–819.

Ogawa, Susumu, & Piller, Frank T. (2006). Reducing the risks of new product development.MIT Sloan Management Review, 47(2), 65–71.

Piller, Frank T., & Walcher, Dominik (2006). Toolkits for idea competitions: A novelmethod to integrate users in new product development. R&D Management,36(3), 307–318.

Poetz, Marion K., & Schreier, Martin (2012). The value of crowdsourcing: Can usersreally compete with professionals in generating new product ideas? Journal ofProduct Innovation Management, 29(2), 245–256.

Schreier, Martin, Fuchs, Christoph, & Dahl, Darren (2012). The innovation effect of userdesign: Exploring consumers' innovation perceptions of firms selling productsdesigned by users. Journal of Marketing, 76(5), 18–32.

Schulze, Anja, & Hoegl, Martin (2008). Organizational knowledge creation and thegeneration of new product ideas: A behavioral approach. Research Policy,37(10), 1742–1750.

Singh, Jasjit, & Fleming, Lee (2010). Lone inventors as sources of technologicalbreakthroughs: Myth or reality? Management Science, 56(1), 41–56.

Surowiecki, James (2004). The wisdom of crowds — Why the many are smarter than thefew and how collective wisdom shapes business, economies, societies and nations.New York: Doubleday.

Terwiesch, Christian, & Ulrich, Karl T. (2009). Innovation tournaments: Creating andselecting exceptional opportunities. Boston: Harvard Business Press.

Toubia, Olivier, & Flores, Laurent (2007). Adaptive idea screening using consumers.Marketing Science, 26(3), 342–360.

Ulrich, Karl T. (2007). Design: creation of artefacts in society. : Pontifica Press.Ulrich, Karl T., & Eppinger, Steven D. (1995). Product design and development (1st edition).

New Jersey: McGraw-Hill.Urban, Glen L., & Hauser, John R. (1993). Design and marketing of new products (2nd

edition). Englewood Cliffs: NJ: Prentice Hall.Von Hippel, Eric (1988). The sources of innovation. New York: Oxford University Press.Von Hippel, Eric (2005). Democratizing innovation. Boston: MIT Press.Von Hippel, Eric, Ogawa, Susumu, & de Jong, Jeroen P. J. (2011). The age of the

consumer-innovator. MIT Sloan Management Review, 53(1), 27–35.Weisberg, Robert W. (1993). Creativity: Beyond the myth of genius. New York: Freeman.

Table 2Sensitivity analysis (health and beauty division).

(A) Analysis of the full sample



Significance

M M


45,527.76 74,787.50 pM-W-Ub .07;pt-testb .25


106,452.61 170,995.25 pM-W-Ub .13;pt-testb .38



Sales in yen (after3rd year)






(B) Sensitivity analysis (product-type matching of user- and designer-generatedproducts)



M M Significance


38,428.50 74,787.50 pM-W-Ub .09;pt-testb .15


74,878.50 170,995.25 pM-W-Ub .06;pt-testb .13



Sales in yen (after 3rdyear)






Note that we also analyzed survival in this category. Importantly, we again find thatonly 21 out of 49 designer-generated products were still on the shelves after threeyears. At the same time, however, all four user-generated products were still on themarket (pchi-square-difference-testb .05). User-generated products thus demonstrate a sub-stantially higher survival rate. (Products that survived the three-year observation peri-od also performed substantially better than those that were discontinued, whichconfirms that lower-performing products were eliminated; e.g., first-year unit saleswere approximately 140% higher for products that survived the third year.)


Patterns in consumption-based learning about brand quality for consumerpackaged goods

Maciej Szymanowski a,⁎, Els Gijsbrechts b,1

a Rotterdam School of Management, Erasmus University, P.O. Box 1738, 3000 DR Rotterdam, The Netherlandsb Quantitative Marketing at Tilburg University, P.O. Box 90153, 5000 LE Tilburg, The Netherlands


Article history:First received in 17, September 2010 and wasunder review for 5 monthsAvailable online 16 May 2013

Area Editor: Luk Warlop

Keywords:Consumption-based learningConsumer packaged goodsCategory characteristicsHousehold characteristics

In this paper, we explore the patterns of consumption-based learning about brand quality in mature consumerpackaged goods (CPG) categories as well as category and household characteristics that drive such learning.We estimate brand-choicemodels with Bayesian learning on household purchases in over thirty CPG categories.This approach yields category- and household-specific estimates of the extent to which consumers update theirknowledge on the quality of specific brands, with new consumption-based information from these brands. Wethen link this degree of learning to the underlying household and category drivers. We find that learning ispresent and significant in almost all categories yet varies in strength across categories and households. Learningabout brand quality is negatively associated with variety seeking. Conversely, learning is stronger in categorieswhere consumers have higher monetary (expensive items) and especially non-monetary stakes (categorieswith higher performance risk and involvement). In line with the ‘enrichment’ hypothesis, familiarity with thecategory resulting from frequent category purchases increases the information extracted fromnew consumptionexperiences — but only up to a certain point. Interestingly, however, market mavens learn less, potentiallyreflecting their overconfidence. Whereas some households learn more than others across the board, categoryfactors are the strongest drivers of learning. Managerial implications are discussed.


1. Introduction

Traditional consumer behavior models (Pechmann & Ratneshwar,1992; Wright & Lynch, 1995) as well as studies in the field of informa-tion economics (Miller, 1984;Nelson, 1970) posit that consumer choicesover time are driven by uncertainty and learning. Consumption-basedlearning about brands is a commonly acknowledged phenomenon insettings with ‘complex decision making’ (Narayanan, Chintagunta, &Miravete, 2007) and in high-involvement or strongly evolving catego-ries such as insurance, pharmaceuticals, and telephone services(Akçura, Gönül, & Petrova, 2004; Narayanan & Manchanda, 2009;Narayanan et al., 2007). However, there is growing evidence that evenin mature consumer packaged goods (CPG) categories such as ketchup,liquid laundry detergent, toothpaste and toothbrushes, consumers maybe uncertain about the product attribute levels that underlie the qualityof a brand (Erdem, 1998; Erdem, Keane, & Sun, 2008; Mehta, Rajiv, &Srinivasan, 2004). This phenomenon is especially true for experienceattributes, such as the teeth-whitening ability of a toothpaste brand orthe cleansing power of a detergent, which may only reveal themselvesover time or vary across usage occasions (Ching, Erdem, & Keane,

2012; Erdem, 1998). In the face of such uncertainty, consumers basetheir choices on beliefs about the brands' quality2 and rely on informa-tion obtained through usage to update these beliefs over time.

Not only is such consumption-based quality learning interesting ac-ademically, it is important from a managerial perspective. Every year,CPG manufacturers and retailers invest billions of dollars in enhancingthe quality of their products and ensuring their consistency.3 Whetherthis is money well spent depends on the extent to which consumersupdate their brand beliefs in response to quality improvements (Mitra& Golder, 2006) or breakdowns (van Heerde, Helsen, & Dekimpe,2007). Moreover, experience-based brand-quality learning influencesthe impact of price promotions (Akçura et al., 2004; Erdem et al.,2008), free samples (Erdem, 1998), product innovation (Narayanan &Manchanda, 2009), and umbrella branding (Erdem, 1998; Erdem &Sun, 2002) and thus should play a role in the planning of these activities.Nonetheless, the available evidence on such learning in mature pack-aged goods is limited to only a few product categories, and little isknown empirically about category or household characteristics thatdrive it.


⁎ Corresponding author. Tel.: +31 104082935; fax: +31 104089011.E-mail addresses: [email protected] (M. Szymanowski), [email protected]

(E. Gijsbrechts).1 Tel.: +31 13 466 8224; fax: +31 13 466 8354.

2 Following extant literature on consumer brand choice under uncertainty, we inter-pret (perceived) quality as “a summary statistic that captures any intangible and tan-gible attributes of a product that may be imperfectly observable by consumers”(Erdem et al., 2004, p. 87).

3 For instance, in 2009 alone, Unilever invested 891 million Euros in R&Dworldwide.





Our study aims to reduce this knowledge gap. In particular, wefocus on the following questions. First, to what extent do consumerslearn about the quality of specific brands from consumption experi-ences with those brands, in different CPG categories? Second, whatare the drivers of such consumption-based learning? Specifically,which household and/or category characteristics determine the extentof brand-quality learning, and what is the direction and magnitude ofthis effect? The answers to these questions are important for bothmar-keting scholars and practitioners. Froman academic perspective, identi-fying categories where consumers learn the most about brand qualityfrom consumption and assessing the category- or household-relatedantecedents enhances our understanding of this learning phenomenon.For managers, these insights may prove helpful in designing marketingprograms. For instance,managersmay focus learning-based tactics suchas free samples for consumers on categories where consumption-basedlearning is stronger.

To address our research questions, we present a framework of thecategory- and household-factors that drive learning. This frameworkis empirically tested on data from a stable panel of 584 households,covering brand purchases in 33 product categories for a period of3.5 years. We combine these purchase data with information from twosurveys, documenting the panel member and category characteristics.We estimate choice models with Bayesian learning (see, e.g., Erdem &Keane, 1996; Narayanan & Manchanda, 2009) to obtain category- andhousehold-specific learning parameters, which are then linked to theirunderlying factors. The results shed light on the prevalence and magni-tude of learning for packaged goods, the variation in learning acrosscategories and across households, and the underlying factors.

The paper is organized as follows. In the next section,we present ourconceptual framework, building on the available literature on consumerdecision making and learning. Section 3 describes the data andoperationalization of the household and category drivers. The method-ology is outlined in Section 4, followed by the results in Section 5.Section 6 summarizes thefindings, discusses implications, and indicateslimitations and suggestions for future research.

2. Conceptual framework

2.1. Background

Experience-based learning has long been a topic of interest in theconsumer behavior and marketing modeling literature. Most of thisliterature has studied the early stages of learning in new or turbulentenvironments4 (Hutchinson & Eisenstein, 2008). In such settings,learning primarily involves the development of ‘category knowledge’:consumers have yet to uncover the relevant product attributes andtheir preferences regarding these attributes (Alba & Hutchinson,1987). In this paper, we focus on mature consumer packaged goods(CPG) categories, for which such category knowledge is typically welldeveloped. Instead, consumers primarily build ‘brand knowledge’;that is, they learn about the quality levels of specific brands (Kardes &Kalyanaram, 1992; Zhou, 2002). As in previous studies (e.g., Erdemet al., 2008; Narayanan et al., 2007), we assess such learning as thedegree to which consumers update their beliefs about the quality ofa brand based on consumption experiences with that brand, where‘quality’ is some overall assessment of the experience attributes under-lying the brand's ‘performance’ or ‘appeal’ (Erdem, 1998; Erdem &Keane, 1996).

Even in mature markets, consumers may not be perfectly knowl-edgeable about the quality of brands because the quality effects maytake time or multiple consumptions to materialize (Erdem, 1998) or

because it may be difficult to isolate the quality of the brand fromother confounding factors (Hoch & Deighton, 1989). Furthermore,consumers generally buy and consume products sequentially ratherthan simultaneously, which hampers effective brand comparisons andlearning (Hoch & Ha, 1986; Warlop, Ratneshwar, & van Osselaer,2005) and causes the memory of brand quality to deteriorate betweenconsumption and purchase occasions (Mehta et al., 2004; Warlopet al., 2005). Finally, consumers may need to update or revisit theirknowledge as their consumption patterns evolve (Du & Kamakura,2006).

Although these phenomena explain why consumers may graduallylearn about brand quality from their experiences in CPG markets, theyalso suggest that the degree of learningmaywell vary across consumersand categories. Extant studies in consumer behavior and cognitivepsychology conclude that, indeed, brand learning is both person- andproduct-dependent (see, e.g., Hoch & Deighton, 1989; Hoch & Ha,1986; Hutchinson & Eisenstein, 2008; Warlop et al., 2005 for an over-view). However, because their main focus is on uncovering (specificaspects of) the learning process as such, these papers are mostlyconceptual in nature and/or use highly controlled laboratory settingsto make their point. Conversely, analysis of scanner panel data hasshown that consumption-based learning about brand quality is anempirically important phenomenon that may substantially affectbrand choice over time (see, e.g., Ackerberg, 2003; Erdem & Keane,1996; Mehta et al., 2004; Shin, Misra, & Horsky, 2012). Nonetheless,this evidence is confined to a limited number of categories and doesnot pursue household differences in learning.5 We aim to bridgethese two literature streams by (i) empirically assessing consumers'experience-based brand-quality learning in a wide range of CPGs and(ii) testing a number of propositions on what drives the degree oflearning.

2.2. Drivers of consumption-based learning

Buildingon the available literature on consumer decisionmaking andlearning from experience (e.g., Hoch & Deighton, 1989; Hutchinson &Eisenstein, 2008; Kardes, 1994), we postulate that the strength of learn-ing about the quality of specific brands in a category depends on threedimensions: the consumers' domain familiarity, their incentives tolearn, and their variety seeking. Below, we conceptualize how andwhy these dimensions trigger learning.

2.2.1. Domain familiarityThedegree towhich consumers learn fromnewconsumption signals

of specific brands depends on their ‘domain familiarity’ — that is, theiraccumulated experiences with the category (Alba & Hutchinson, 1987;Hoch & Deighton, 1989; Hutchinson & Eisenstein, 2008). The generalpremise – also known as the ‘enrichment hypothesis’ (Johnson &Russo, 1984) – is that category familiarity creates ‘expertise,’ which inturn allows consumers to better comprehend new consumption signalsand “retain more information with lower levels of error” (Hutchinson &Eisenstein, 2008, p. 103). For one, consumers who are more familiarwith the product category have acquired more ‘cognitive structure’and thus are better able to organize information from new experiences(Hutchinson & Eisenstein, 2008). Moreover, greater knowledge ofrelevant product attributes and better developed preferences for theseattributes greatly reduce their cognitive load and the (mental) cost ofprocessing new signals (Broniarzcyk, 2008). Experts have also beenfound to engage inmore ‘analytic processing’, i.e., to bemore influencedby brand-specific attributes rather than by salient but irrelevant overallimpressions (Hutchinson & Alba, 1991; Hutchinson & Eisenstein, 2008).

4 Examples of such studies are Carpenter and Nakamoto (1989, 1996), who studyhow experiences with pioneer brands shape consumers' preferences in a new productclass, or Heilman, Bowman, and Wright (2000), who analyze how category experiencedrives choice dynamics for consumers that are new to a market.

5 Moreover, the results of these studies may be difficult to compare because of (i)differences in the definition of consumption units and categories, (ii) variations incountry/type of brand analyzed, and (iii) methodological differences (i.e., inclusion offorgetting and consumer heterogeneity).

220 M. Szymanowski, E. Gijsbrechts / Intern. J. of Research in Marketing 30 (2013) 219–235

This finding suggests that consumers who are more familiar with theproduct category extract more information from a given consumptionsignal. However, there are also counter-indications. Experienced buyersmay be less likely to encode new signals that make earlier informationobsolete (Wood & Lynch, 2002), and their ’overconfidence’ mightslow the dissemination of new information (Hoch & Deighton, 1989;Hutchinson & Eisenstein, 2008). Such a finding would imply that con-sumers highly familiar with the category learn less about brand qualityfrom consumption than moderately familiar consumers (Johnson &Russo, 1984). Given these countervailing forces, we postulate thatproduct category familiarity affects the informativeness of consumptionsignals but leave the direction of the effect an empirical issue.

2.2.2. Incentives to learnWe expect consumer learning to be positively related to the antic-

ipated gains from learning. This claim stems from the notion in infor-mation economics that learning comes at a cost, which is justified ifconsumers expect to obtain sufficient gains (Stigler, 1961). Greatergains from learning may incite consumers to pay more attention tothe actual performance of a brand (i.e., intentional learning increasesthe likelihood of analytic information processing, Hutchinson & Alba,1991) and thus learn more (Hawkins & Hoch, 1992). Incentives tolearn may stem from monetary as well as non-monetary gains. First,we expect consumers to have higher learning stakes in ‘expensive’categories that require a high discretionary sum spent per purchaseoccasion (Narasimhan, Neslin, & Sen, 1996). Second, incentives tolearn may be related to the performance risk or the expected utilityloss in case of the wrong brand choice in the category (Beatty & Smith,1987). More performance risk in a category will lead stronger house-holds to bemore motivated to learn from experience to avoid mistakes.Finally, consumers will be more keen to monitor brand quality experi-ences and more likely to use accessible content in categories in whichthey are more involved (Celsi & Olson, 1988; Schwarz, 2004).

2.2.3. Brand loyalty/variety seekingThe degree of brand loyalty or, conversely, variety seeking is the third

dimension we associate with the magnitude of brand-quality learning.Whereas consumption-based learning is driven by the desire to maxi-mize the expected quality of brands consumed, variety seeking is associ-ated with the desire to mitigate boredom and satiation (Gijsbrechts,Campo, & Nisol, 2000; McAlister & Pessemier, 1982). Previous researchhas documented that consumers with an intrinsic desire for stimulationmay be willing to give up quality in return for variety (Kahn & Raju,1991; Trivedi & Morgan, 2003). Therefore, we expect such variety-seeking consumers to take less interest in learning about the quality ofspecific brands.

3. Data and variable operationalizations

3.1. Data

3.1.1. Scanner panel dataOur main data source consists of GfK scanner panel data covering

households' supermarket purchases, over 3.5 years, across awide rangeof product categories. Only households that are active in the panelthroughout the entire period are retained. We use the first year ofdata as an initialization period, maintaining the remaining 2.5 yearsfor model estimation. Given our interest in purchase dynamics, house-holds are only included in a category if they have at least 6 purchasesin that category in the estimation sample. The final estimation samplethus covers 584 different households engaging in 154,077 purchasesacross 33 categories, as defined by GfK.

Our data include brand-level information on the quantities boughtand the corresponding prices and promotions. In each category, weconsider the top brands covering at least 50% of total category sales(with a minimum of four brands) for inclusion in the brand choice

model. Table 1 provides descriptions for the brands in our sample cat-egories. This table indicates that categories vary considerably in termsof concentration (the share of sampled purchases of the top 2 brandsranging from 41% in the powdered laundry detergent category to 85%for sugar) and price spread (the highest-to-lowest price ratio, rangingfrom 1.06 for dairy products to 3.86 for flavor additives). In addition,the data show considerable variation in households' brand choicesover time — a desirable feature for the assessment of learning.6

3.1.2. Survey dataIn addition to the scanner panel data, we have access to information

from two surveys, which we use to operationalize household andcategory drivers. The first survey, administered to the panel membersby GfK, records households' time invariant scores on a number of attitu-dinal and shopping-related items. The second survey, administered7 toDutch consumers who are not members of the panel, yields summarymeasures on a number of category characteristics that do not varyacross households or over time. Below, we indicate how these dataare used to operationalize the drivers of learning.

3.2. Driver operationalizations

Table 2 provides an overview of the learning drivers, togetherwith the underlying dimensions and the hypothesized effects. Foreach driver, the table also provides measurement details and indicatesthe corresponding data source. We briefly discuss the learning driversbelow.

3.2.1. Domain familiarityWe measure a household's ‘domain familiarity’ or accumulated

category knowledge (CatKnowch) by the average number of category

purchases per month in the year preceding our estimation sample.As indicated above, this driver's impact on learning may be positive(e.g., better comprehension of new brand-specific signals) or negative(e.g., less encoding of such signals), and the direction of the effect maydepend on the knowledge level already attained. To flexibly capture itsimpact on learning, we include both the variable and its square in themodel.

3.2.2. Incentives to learnCategory expensiveness (Expensivenessch) is measured as the aver-

age amount spent by the household per purchase occasion in the cat-egory (Narasimhan et al., 1996). Performance risk in a given category(Riskc) and category involvement (Involvementc) are capturedthrough category-specific item scores from the second survey (seeTable 2 for details). Based on our conceptualization, we expect allthree variables to be positively associated with learning.

3.2.3. Variety seekingBuilding on the available literature, we adopt three indicators of

brand loyalty/variety seeking. The first, Brand_Loyaltyh, is a householdtrait obtained from the panel-member survey data (see Table 2 fordetails). This indicator reflects the household's commitment to his/herfavorite brand in different categories (see, e.g., Baumgartner &Steenkamp, 1996) — with an expected positive impact on learning.The second indicator, Need_Varietyc, measures the need for variety as acategory characteristic (Givon, 1984; Pessemier & Handelsman, 1984).We expect higher values for this variable to entail less learning. Third,we include the number of different brands purchased by a householdin a given category (Number_Brandsch), which we obtain for the yearprior to our estimation sample and mean-center against the categoryaverage. This measure is behavioral in nature (van Trijp & Steenkamp,

6 Please see Appendix B for details on model identification.7 These data were collected in the context of a study by Steenkamp et al. (2004). We

thank the authors for making part of the information available to us for this research.

221M. Szymanowski, E. Gijsbrechts / Intern. J. of Research in Marketing 30 (2013) 219–235

Table 1Descriptive statistics for brands in sample categories.

Category Brand Choice share Temporal variation inhouseholds' choice sharesa

Number of purchases Average price(in Eurocent)

Brand type

Flavor additives 1 0.08 0.06 194 0.41/g PL2 0.20 0.12 483 1.43/g NB3 0.15 0.10 350 0.46/g NB4 0.10 0.07 249 0.37/g PL5 0.15 0.10 365 0.79/g NB6 0.31 0.19 733 0.38/g NB

Periodic care\diapers 1 0.13 0.05 490 21.96/piece PL2 0.13 0.10 506 20.15/piece NB3 0.31 0.16 1188 20.05/piece NB4 0.42 0.15 1594 21.03/piece NB

Dairy products 1 0.16 0.04 1521 0.43/ml PL2 0.33 0.10 3247 0.44/ml NB3 0.24 0.08 2333 0.46/ml NB4 0.27 0.09 2619 0.45/ml NB

Air refreshers 1 0.29 0.16 733 2.08/g NB2 0.50 0.19 1244 2.85/g NB3 0.13 0.12 314 2.84/g NB4 0.08 0.04 194 1.42/g PL

Dairy drinks 1 0.18 0.05 1711 0.11/ml PL2 0.32 0.09 3087 0.12/ml NB3 0.15 0.07 1479 0.13/ml NB4 0.19 0.04 1800 0.11/ml PL5 0.16 0.04 1538 0.1/ml PL

Liquid laundry detergent 1 0.39 0.18 1132 0.71/ml NB2 0.16 0.13 447 0.69/ml NB3 0.15 0.05 433 0.44/ml NB4 0.14 0.13 394 0.85/ml NB5 0.17 0.12 477 0.69/ml NB

Salads 1 0.19 0.10 1146 0.82/g PL2 0.21 0.13 1267 0.86/g NB3 0.12 0.11 720 0.63/g NB4 0.48 0.20 2903 0.81/g NB

Meal decorators 1 0.33 0.14 2714 2.08/g NB2 0.39 0.13 3272 1.75/g NB3 0.14 0.08 1124 2.09/g NB4 0.14 0.10 1190 1.94/g NB

Breakfast cereals 1 0.26 0.12 1120 0.83/g NB2 0.20 0.11 880 0.75/g NB3 0.12 0.05 526 0.58/g PL4 0.31 0.08 1334 0.64/g NB5 0.11 0.05 487 0.62/g NB

Mouth hygiene 1 0.50 0.14 1374 2.28/ml NB2 0.14 0.08 390 4.55/ml NB3 0.22 0.10 611 2.58/ml NB4 0.14 0.08 373 3.3/ml NB

Pet food 1 0.09 0.04 690 0.44/g PL2 0.08 0.05 612 0.49/g NB3 0.23 0.09 1744 0.47/g NB4 0.14 0.05 1079 0.69/g NB5 0.28 0.11 2060 0.52/g NB6 0.17 0.09 1274 0.38/g NB

Fabric softeners 1 0.09 0.06 311 0.31/ml PL2 0.44 0.20 1447 0.36/ml NB3 0.19 0.07 630 0.2/ml PL4 0.28 0.17 907 0.31/ml NB

Spices and herbs 1 0.14 0.08 610 6.06/g PL2 0.63 0.18 2714 8.16/g NB3 0.06 0.07 267 3.89/g NB4 0.17 0.11 712 6.02/g NB

Alcoholic drinks 1 0.54 0.15 3413 0.61/ml NB2 0.16 0.09 1004 0.23/ml NB3 0.12 0.08 751 0.24/ml NB4 0.18 0.09 1098 0.24/ml NB

Dish detergent (tablets) 1 0.19 0.04 259 1.45/g NB2 0.09 0.04 132 0.71/g PL3 0.48 0.18 670 1.14/g NB4 0.24 0.15 330 1.07/g NB

Powdered laundry detergent 1 0.11 0.07 297 0.45/g NB2 0.19 0.15 497 0.43/g NB3 0.21 0.13 557 0.43/g NB4 0.17 0.05 432 0.26/g PL5 0.12 0.08 313 0.59/g NB6 0.20 0.16 514 0.38/g NB

Margarine\oil 1 0.09 0.05 811 0.51/g PL


Table 1 (continued)



Brand type

2 0.12 0.03 1107 0.28/g NB3 0.13 0.05 1276 0.36/g NB4 0.26 0.06 2494 0.52/g NB5 0.40 0.09 3798 0.31/g NB

Liquid dish detergent 1 0.12 0.06 273 0.28/ml PL2 0.52 0.12 1229 0.34/ml NB3 0.19 0.08 454 0.26/ml NB4 0.16 0.07 385 0.22/ml PL

Milk substitutes 1 0.12 0.06 780 0.52/ml PL2 0.44 0.11 2881 0.61/ml NB3 0.15 0.05 999 0.48/ml PL4 0.16 0.04 1080 0.58/ml NB5 0.13 0.05 881 0.45/ml PL

Sweet-and-sour canned products 1 0.23 0.10 731 0.24/g PL2 0.12 0.06 367 0.22/g PL3 0.36 0.16 1109 0.19/g NB4 0.15 0.10 478 0.19/g PL5 0.14 0.12 429 0.23/g NB

Ready-to-eat meals 1 0.21 0.11 1123 0.86/g PL2 0.35 0.17 1843 0.92/g NB3 0.19 0.14 1021 0.87/g NB4 0.24 0.13 1291 0.79/g NB

Toilet and kitchen tissues 1 0.13 0.07 826 23.83/piece NB2 0.29 0.15 1816 41.96/piece NB3 0.21 0.17 1291 43.59/piece NB4 0.22 0.07 1398 29.91/piece PL5 0.15 0.07 919 33.59/piece PL

Pastries 1 0.26 0.15 550 0.91/g PL2 0.11 0.08 233 0.78/g PL3 0.08 0.07 163 0.41/g NB4 0.09 0.08 185 0.78/g PL5 0.46 0.24 954 0.87/g NB

Crackers 1 0.21 0.16 700 0.92/g NB2 0.20 0.14 657 0.76/g NB3 0.09 0.05 308 0.77/g NB4 0.27 0.15 872 0.84/g NB5 0.11 0.06 352 0.62/g PL6 0.12 0.11 383 1.01/g NB

Canned soup 1 0.19 0.11 703 0.39/ml PL2 0.05 0.04 177 0.31/ml PL3 0.63 0.17 2343 0.37/ml NB4 0.14 0.10 504 0.62/ml NB

Salty snacks 1 0.10 0.05 892 1.19/g PL2 0.14 0.07 1277 1.02/g NB3 0.54 0.12 5091 0.91/g NB4 0.11 0.06 1016 0.97/g NB5 0.05 0.03 444 0.87/g PL6 0.07 0.05 622 0.83/g NB

Eggs 1 0.29 0.08 1668 23.34/piece PL2 0.12 0.06 690 22.9/piece PL3 0.27 0.09 1545 21.35/piece PL4 0.22 0.11 1286 20.35/piece NB5 0.10 0.06 601 20.91/piece NB

Hair care 1 0.36 0.18 1091 1.64/ml NB2 0.16 0.11 473 1.31/ml NB3 0.27 0.18 816 1.66/ml NB4 0.21 0.11 615 1.34/ml NB

Baking products 1 0.30 0.19 1392 1.28/g NB2 0.37 0.22 1683 1.38/g NB3 0.08 0.08 367 2.12/g NB4 0.08 0.09 347 2.14/g NB5 0.06 0.05 272 1.39/g NB6 0.11 0.07 504 0.81/g NB

Cleaning materials 1 0.12 0.10 302 0.11/g NB2 0.20 0.08 519 0.07/g NB3 0.51 0.15 1289 0.19/g NB4 0.17 0.09 440 0.07/g NB

Sugar 1 0.06 0.04 313 0.23/g PL2 0.09 0.05 434 0.21/g NB3 0.54 0.14 2705 0.24/g NB4 0.31 0.12 1559 0.24/g NB

Cleaners 1 0.12 0.13 324 0.39/ml NB2 0.10 0.10 282 0.38/ml NB3 0.38 0.22 1057 0.42/ml NB4 0.40 0.23 1090 0.41/ml NB

(continued on next page)


1990) and reflects the household's purchase variety in the categoryrelative to other households. We expect higher levels of this variableto be associated with lower levels of learning.

3.2.4. Control variablesApart from these hypothesized drivers, we include several control

variables. First, we distinguish between food and non-food categories(Foodc equals 1 for food categories and zero otherwise). Next, weinclude two household characteristics. Mavenismh reflects the extentto which the consumer acts as an opinion leader and shares his/herproduct information and experience with other consumers (Feick &Price, 1987) and is obtained as a summated scale from the panelmembers' self-reported items in the first survey (see Table 2). Becausewe expect shopping mavens to be more involved grocery shoppers(who are more eager to learn) but also more knowledgeable shoppers

(who have little left to learn), we offer no directional hypotheses ontheir degree of consumption-based learning. Planningh captures thehousehold's degree of purchase planning through the use of a mentalor written shopping list (Inman, Winer, & Ferraro, 2009) and is alsoobtained from the panel-member survey (see Table 2). Buyers whoplan their specific purchases in advance are likely to be more certainabout their choices and, as indicated by Urbany (1986) and Urbany,Dickson, and Wilkie (1989), tend to be less open to new informationand less responsive to changes in utility. This characteristic wouldimply a negative relationship between planning and subsequent,experience-based learning. However, such a link only holds to theextent that consumers plan the specific brand within the category.Because ourmeasure does not contain this level of detail (i.e., it pertainsto the consumers' degree of planning across product categories andshopping trips), it enters as a control only.

Table 1 (continued)



Brand type

Deodorant 1 0.18 0.07 232 95.39/piece PL2 0.23 0.12 293 90.28/piece PL3 0.21 0.07 268 83.49/piece PL4 0.17 0.08 218 99.18/piece NB5 0.21 0.14 276 176.59/piece NB

a Standard deviation of the household's brand share, calculated over four 32-week subperiods, and then averaged over households visiting the chain.

Table 2Overview of learning drivers.

Underlying dimension Variable: Notation Expectedeffect onlearning

Data source Source ofvariation

Operationalization

Domain familiarity Category purchasefrequency: CatKnowc

h+/! Scanner data Households

and categoriesHousehold h's average number of purchases/ month in category c, ina one-year initialization period.

Incentives to learn Category expensiveness:Expensivenessch

+ Scanner data Householdsand categories

Average amount spent by household h per purchase occasion incategory c, in a one-year initialization period

Incentives to learn Performance risk: Riskc + Non-panel-membersurvey

Categories Score average of the following items for category ca:- There is much to lose if you make the wrong choice in category c- It matters a lot when you make the wrong choice in category c- In category c, there are big differences in quality between thevarious brands(Scale Average obtained from Steenkamp, Geyskens, Gielens, & Koll,2004)

Incentives to learn Category involvement:Involvementc

+ Non-panel-membersurvey

Categories Score average of the following items for category ca:-The category c is very important to me- The category c interests me a lot(Scale Average obtained from Steenkamp et al., 2004)

Variety seeking Number of brandsconsumed:Number_Brandsch

! Scanner data Householdsand categories

Number of different brands bought by household h in category c, in aone-year initialization period (mean-centered relative to the categorymean across households)

Variety seeking Brand loyalty:Brand_Loyaltyh

+ Panel-membersurvey

Households Score average for household h on the following itemsb:- I prefer to buy the brand that I usually buy- I consider myself a brand-loyal consumer- I really feel connected with brands that I buy(Cronbach's Alpha = .953)

Variety seeking Need for variety:Need_Varietyc

! Non-panel-membersurvey

Categories Score of the following item for category ca:In category c, I want a good variety of product variants to choose from

Control variable Dummy for foodcategories: Foodc

+/! Scanner data Categories Dummy equal to 1 if c is a food category and to 0 if c is a non-foodcategory

Control variable Mavenism: Mavenismh +/! Panel-membersurvey

Households Score average for household h on the following itemsb:- I talk with friends about products I buy- Friends and neighbors often ask me for advice- People often ask me what I think about new products(Cronbach's Alpha = .969)

Control variable Purchase planning:Planningh

+/! Panel-membersurvey

Households Score average for household h on the following itemsb:- When I go shopping I know exactly what I want to buy- I plan my purchases carefully so that supermarket visits will costless time- I make a list before I go shopping and stick to it(Cronbach's Alpha = .953)

a Based on a five-point scale (1 = strongly disagree to 5 = strongly agree).b Based on a ten-point scale (1 = strongly disagree to 10 = strongly agree).


4. Methodology

4.1. Stage 1: Degree of consumption-based quality learning

In the first stage, we estimate the extent to which households learnabout brand quality from consumption signals in different categories.

4.1.1. Basic model specificationOur approach to quantifying the amount of consumption-based

quality learning closely follows the available literature and is basedon structural dynamic brand choice models with Bayesian learning(see, e.g., Crawford & Shum, 2005; Erdem & Keane, 1996; Mehta et al.,2004; Narayanan & Manchanda, 2009). In these models, consumersselect the brand from a category that maximizes their current utility,the utility for household h of brand j in category c on purchase occasiont being given by the following:

Uhjct ! f Qh

jct

! "" Xjctβ

hc " εhjct ; #1$

where Qjcth indicates household h's beliefs on purchase occasion t

regarding brand j's ‘quality’, i.e., a composite of the brand's experienceattributes aboutwhich the household is uncertain.8 Xjct, in turn, is a vec-tor of utility determinants other than perceived quality that are directlyobservable to both the researcher and the consumer. Similar to previousbrand choice models, these determinants include price and promotionsupport (feature and display).9 βc

h are parameters, and εjcth are i.i.d.extreme-value utility components unobserved by the researcher butobserved by the consumer (for a complete overview of the notation,see Table 3). Like previous authors (e.g., Crawford & Shum, 2005;Narayanan & Manchanda, 2009), we allow consumers to be risk aversewith respect to their uncertainty about the true quality of brands andaccommodate this by bringing quality beliefs into the utility expressionthrough a negative exponential function:

f Qhjct

! "! ! exp !rhcQ

hjct

! "; #2$

where rch is a non-negative risk aversion coefficient.Household h's quality belief about brand j in category c on purchase

occasion t, Qjcth , is normally distributed with mean μjcth and variance (or

uncertainty) σjcth2. Given the distribution of Qjct

h , and using Eq. (1), theexpected utility for household h of brand j in category c on purchaseoccasion t may be written as (Narayanan & Manchanda, 2009):

E Uhjct

###Ihth i

! ! exp !rhc μhjct!rhc

σh2jct

2

! !" Xjctβ

hc " εhjct : #3$

where Iht is the information available to the household at time t.In line with extant learning models, we assume that consumers do

not know brands' ‘true quality’, and we model their learning fromconsumption. Similar to previous authors (see, e.g., Crawford & Shum,2005; Narayanan & Manchanda, 2009; Shin et al., 2012; Szymanowski& Gijsbrechts, 2012), we allow this ‘true’ brand quality, qjch , to beconsumer-specific — accommodating the fact that the attributes of a

given brand may be better tuned to the needs of some consumersthan others.10

Information about these attributes is accumulated through actualbrand experience. Specifically, consumption by household h of eachunit of brand j in category c in period t ! 1 provides a new qualitysignal gjcth , which is assumed to be i.i.d. normally distributed with amean equal to the true brand quality qjch and variance λc

h2, or:gjcth ~ N(qhjc, λc

h2). We denote a series of Mcth consumption units for

household h of brand j in category c consumed at time t ! 1 as

Ghjct !

XMhct

m!1

ghjctm

Mhct e N qhjc;

λhc2

Mhct

$ %. Consumers' uncertainty is reduced grad-

ually as they learn, but it may also increase again due to forgetting(Mehta et al., 2004).11 In the absence of consumption at t ! 1, we

expect σjcth2 to increase exponentially: σh2

jct ! σh2jct!1 % e

bhc whct!wh

ct!1# $,where bch is an estimated decay parameter for household h and cate-gory c and wct

h ! wct ! 1h refers to the time elapsed between purchase

occasions t and t ! 1.12 As in previous studies, we assume that oneach occasion t in category c, the consumer adopts only one brand,

such that "jdhjct ! 1, where dcjth = 1 if brand j in category c was

bought by h in t, and 0 otherwise. Similar to those studies, we alsoassume that consumption of the brand bought at t ! 1 occurs imme-diately before the purchase in t, such that at the time of the update int, the consumer has not forgotten the Mct

h consumption signals gcjtmh .Consumption-based learning is thenmodeled using the conventional

expression (e.g., de Groot, 1970). Specifically, the consumers' meanquality belief of brand j in category c on purchase occasion t (μjcth ) andthe variance of this quality belief (σjct

h 2) are given by:

μhjct !

μhjct!1

σh 2jct!1 % eb

hc % wh

ct!whct!1# $ "

dhjct!1Ghjct

λhc2

Mhct

0

B@

1

CA

% 1

σh 2jct!1 % eb

hc % wh

ct!whct!1# $ "

dhjct!1λhc2

Mhct

0

B@

1

CA

!1

#4$

and

σhjct

2 ! 1

σhjct!1

2 % ebhc % wh

ct!whct!1# $ "

Mhct % d

hjct!1

λhc2

0

@

1

A!1

: #5$

The updated mean quality belief in Eq. (4) is a weighted sum of theprior quality belief (μjct ! 1

h ) and the average consumption signal (Gjcth ),

with weights inversely proportional to the prior quality uncertainty(σjct ! 1

h 2 ) and the consumption signal variance λch2, respectively.13 Weaccommodate unobserved household heterogeneity by using a random

8 As also indicated by Erdem (1998), quality is a multi-dimensional construct, andconsumers may have different levels of uncertainty regarding the different quality di-mensions. However, given that experience attributes (such as a toothpaste's ability towhiten teeth and fight cavities) are typically not directly observable, extant (scanner-based) papers on learning collapse them into one ‘overall quality’ attribute that con-sumers are uncertain and learn about (Erdem, 1998, Erdem & Keane, 1996).

9 Like previous authors of papers on learning, we assume that consumers are fullyinformed about these marketing mix activities. Also, following previous papers onbrand-choice learning, the focus of this paper is on the choice of brands, not SKUs. IfSKUs were the unit of analysis, other ‘directly observable’ search attributes of the prod-uct (on which there is no uncertainty, such as, e.g., color or size) would also fall underthese ‘X’ variables.

10 The true qualities qjch (and ensuing mean brand quality beliefs μjcth ) thus combinethe ‘objective’ level of the brand's experience attributes and how this level matchesthe consumer's preferences. However, to the extent that consumers' preferences inmature CPG categories are typically well developed (see, e.g., Kardes & Kalyanaram,1992), the uncertainty is most likely in the ‘objective attributes’ part, and it is this partthat consumers primarily learn about.11 Note that in Mehta et al. (2004), forgetting not only impacts the variance of thequality beliefs (a feature that we also incorporate) but also affects the prior mean ina random-walk fashion over time (an aspect not taken up in our model).12 To avoid numerical problems, we restrict the uncertainty σjct

h2 to always remain be-low the initial uncertainty.13 With Mct

h , individual consumption signals in Gcth , the variance for Gct

h becomes λch/Mcth2.

Note that, as in previous studies, we assume that the consumption signal variance, λch2, is

pooled across brands. This finding does not imply, however, that the magnitude of learn-ing at a given point in time is the same for all brands. For instance, brands that are moreestablished will experience weaker learning, to the extent that consumers' uncertaintyabout their quality, σjct ! 1

h 2 , has become lower than for new, less established brands —

thereby placing more weight on the first term in Eq. (4).


effects specification. The parameters qjch and βch are normally distributed,

whereas λch2, bch, and rch have log-normal distributions (to ensure posi-tive values). The initial quality belief is assumed to be normally distrib-uted, with mean μjc0h and variance σjc0

2 . We assume that at time t = 0,consumers are uncertain about their own ‘true quality’ but know thebrand's mean true quality across the population (vqjc) and rationallyform their initial mean beliefs to be the same as this population-level

mean, such that μhjc0 ! vqjc (see Crawford & Shum, 2005; Narayanan &

Manchanda, 2009 for a similar approach). The initial belief uncertaintyis estimated as a separate parameterσc0

2 ,which, for reasons of parsimony,we set to be homogeneous across brands (see also Erdem&Keane, 1996;Mehta et al., 2004). We estimate the parameters using simulated likeli-hood. Appendix A provides details pertaining to the log likelihood func-tion and the estimation procedure.

4.1.2. Degree of learningAs shown in Eq. (4), when the variance of the consumption signal

λch2 is smaller, the impact of a new consumption signal gjcth on theconsumer's quality belief μjcth increases, and vice versa. Therefore, asmaller (larger) consumption signal variance implies stronger (weaker)learning from consumption. However, Eq. (4) also reveals that thedegree of updating depends on the prior brand quality uncertainty,which amounts to σc0

2 at the beginning of the observation period.Because this initial variance also differs across categories,we ‘normalize’the consumption signal variance against this initial uncertainty. There-fore, the estimated value of the consumption signal variance relativeto the initial variance (λch2/σc0

2 ) captures the magnitude of learningand thus constitutes the focal parameter in our study. For convenience,we refer to this ratio as the ‘learning parameter’, with higher levels ofthis parameter implying less learning.

The learning parameters thus reflect the informativeness of oneconsumption signal,14 where each consumption signal in a categoryrepresents the experience contained in one consumption unit in thatcategory. We normalize the measurement units such that a category's‘consumption unit’ corresponds to the average purchase volume, acrossconsumers, per purchase occasion in that category. More specifically,we specify the number of normalized units adopted by the householdin category c at time t as Mh

ct ! Mh;%ct =M

%c , where Mct

h,⁎ is the quantityadopted by the household at that time and M

%c is the average cate-

gory purchase quantity across households and purchase occasions(which largely coincides with a ‘standard’ or most commonly adoptedpackage size). For instance, if household h bought 500 g of coffee attime t (Mct

h,⁎ = 500 g) and the average purchase quantity equals 250 g(M

%c = 250 g), then the normalized number of units for the household

at that time becomes 2 (Mcth = 2). This normalizationwill enablemean-

ingful cross-category comparisons.

4.2. Stage 2: Linking consumption-based learning to household andcategory drivers

In stage 2,we examinehowhousehold and category drivers influencethe degree of learning. The dependent variable in this stage is theopposite of the logarithm of the learning parameter, obtained fromthe first-stage equation. Specifically, for each household and category,we calculate the posterior estimate of the (logarithm of the) learningparameter as aweighted average of draws15 from the parametermixingdistribution in that category, using the household-specific likelihoodsas weights (see Train, 2003, p. 266). To facilitate interpretation ofthe coefficients, we use the negative of the log-learning parameter,! log(λch2/σc0

2 ), for our dependent variable, such that positive coeffi-cients associatedwith a driver point to stronger learning. As explanatoryvariables, we use the household characteristics (Brand_Loyaltyh,Mavenismh, Planningh), category characteristics (Riskc, Need_Varietyc,Foodc, Involvementc), and household- as well as category-specific

14 Mehta, Rajiv, and Srinivasan (2003) refer to it as the inverse of the ‘signal-to-noiseratio’, which measures the extent to which the initial uncertainty of a consumer differsfrom the “inherent product variability.”15 Specifically, for each category and household, we consider multiple draws from themixing distribution of the consumption signal variance. Each of these draws is then di-vided by the (homogeneously) estimated initial belief variance and log-transformed toobtain the log-learning parameter. We then ‘weigh’ this log-learning parameter withthe household's likelihood function evaluated at the corresponding consumption sig-nal variance draw and initial variance estimate before aggregating across draws.

Table 3Notation overview for the first stage model.

Symbol Description

Indices and (vectors of) indicatorsc Index for product categoryC Number of sample categoriesj, k Indices for brandt, h Indices for purchase occasion and householddjcth Indicator equal to 1 if brand j in category cwas consumed by household h

at purchase occasion t and 0 otherwiseDcth Vector of purchase indicators dhjct up to purchase occasion t

Ehct Vector of consumption signals Ghjct up to purchase occasion t

Jc, Nc Number of sample brands, and number of households, respectively, incategory c

m Index for unit of a productMct

h Number of product units purchased by household h in category c on thepurchase occasion t

Tch Number of purchases in category c made by household h during thesample period

wcth Week index for purchase t of household h in category c

Zjcth Indicator equal to 1 if brand j in category c was available to household hat purchase occasion t and 0 otherwise.

Utility and its determinantsrch Risk aversion parameter of household h in category cUjcth Utility of household h for brand j in category c at purchase occasion t

Xjct Utility determinants for brand j in category c at purchase occasion t otherthan quality belief

βch Household's h sensitivity parameters to utility determinants in Xjt

εjcth i.i.d. extreme value component of utility observed by the household h,unobserved by the researcher

Quality and quality beliefqjch True quality of brand j in category c for household hQjcth Quality belief of household h for brand j in category c on purchase

occasion tμjcth Mean quality belief of household h for brand j in category c on purchase

occasion tσjcth2 Uncertainty or variance in quality belief of household h for brand j in

category c at purchase occasion t(Initial belief variance: σ c0

2)

Learning and forgettingbch Parameter capturing the rate of forgetting for household h in category cItch Household h's information set in category c at purchase occasion tgjctmh Consumption signal of one unit of brand j in category c on purchased by

household h on occasion tGjcth Average over a series of Mct

h consumption signals gjcth for brand j incategory c on purchase occasion t

λch Standard deviation of the consumption signal in category c for household

h

Household heterogeneityvbc ; ςbc Mean and variance across households of the parameter capturing the

extent of forgetting, i.e., bhce log!N vbc ; ςbc

& '

νqjc ; ςqjc Mean and variance across households of the true quality parameter, i.e.,

qhjce log!N vqjc ; ςqjc

! "

vrc ; ςrc Mean and variance across households of the risk aversion parameter, i.e.,rch ~ log ! N(vrc,ςrc)

vβc; ςβc

Vectors of mean and variance across households of the parameters βch,

i.e., βhc e N vβc

; ςβc

& '

vλc ; ςλc Mean and variance across households of the standard error of theconsumption signal, i.e., λh

c e log!N vλc ; ςλc# $

Other notationαc Set of choice model parameters to be estimated (including the means and

variances of the mixing distributions):

αc# νqjc ; ςqjc ;νβc; ςβc

;νλc ; ςλc ;νbc ; ςbc ;νrc ; ςr c ;σ2c0

n o


variables (CatKnowch, Expensivenessch,Number_Brandsch) fromour concep-

tual framework. This approach leads to the following equation:

!log λhc2=σ2

c0

! "! θ0 " θ1 % BrandXLoyalty

h " θ2 %Mavenismh " θ3 % Planningh

" θ4 % Riskc " θ5 % NeedXVarietyc " θ6 % Foodc " θ7% Involvementc " θ8 % CatKnow

hc " θ9 % Expensiveness

hc

" θ10 % NumberXBrandshc "whc ;

#6$

where log(λch2/σc02 ) is the posterior log-learningparameter for household

h and category c (obtained from stage 1), θ are parameters capturing itssensitivity to the household-, category-, or household- and category-specific drivers, andw is a household- and category-specific error.

To ensure the efficiency of the estimates, we must consider thatour dependent variable is a posterior of an estimated parameter,which is inherently uncertain. To assess the sampling error, we computefor each category and household a set of posterior values of the learningparameter using draws frommodel estimates' sampling distributions.16

We then compute, for each category and household, the standard devi-ation of the resulting posteriors of the learning parameter and use theirinverse as weights in the second-stage regression (see Steenkamp, Nijs,Hanssens, & Dekimpe, 2005 for a similar approach).

5. Results

5.1. Stage 1: Degree of consumption-based quality learning

5.1.1. Model fitIn each category, we compare the fit of the proposed Bayesian

learning model to a benchmark model with no purchase dynamics(Appendix B reports the fit statistics). Except in one category (mealdecorators), the model with Bayesian quality updating results in animprovement in the AIC statistic, indicating that consumption-basedlearning about brand quality does occur.

5.1.2. Learning parameter estimatesTable 4 lists our sample categories, together with their estimated

learning parameters, sorted from lowest to highest (the complete setof stage-1 estimation results, for all categories, is given in Appendix C).As the parameter follows an asymmetric (log-normal) distributionacross households, we focus on the median rather than the mean as asummary statistic. For each category, we also report the lower andupper quartile of the learning parameter to provide a feel for the differ-ences between households. Themean value of the learning parameters,pooled across categories and households, is .62. Given that this figurereflects the variance in the consumption signals relative to the initialquality uncertainty, its low value implies that new information obtainedthrough consumption is highly influential. Our observed rate of learningappears within the ‘ballpark’ range from earlier studies (see Erdem,1998; Erdem & Sun, 2002; Shin et al., 2012) and exerts a strong influ-ence on the updated beliefs. Specifically, the consumer's first consump-tion experience determines 62% of his updated quality belief (the priorbelief determining the remaining 38%) — be it that this figure levels offwith subsequent brand purchases. This finding suggests that for theaverage of household and category combinations, learning is substantial.

At the same time, Table 4 points to considerable variation bothacross categories and across households within categories. Categorieswith strong learning (low learning parameter) appear to be mostly

hedonic or personal items (e.g., flavor additives, dairy products anddiapers), whereas categories with slow learning include basic care(e.g., cleaners, deodorants) and secondary food items (e.g., bakingproducts, sugar). However, the pattern is not really clear-cut: forsome of these categories, the variation across households is substantial(e.g., toilet and kitchen paper, or deodorants, where the inter-quartilerange moves from rather strong learning, .26, to virtually no learning,18.05). The question remains what drives this pattern of effects — anissue addressed by the second stage of our analysis.

5.2. Stage 2: Linking consumption-based learning to household andcategory drivers

5.2.1. Regression coefficientsTable 5 reports the results of the second-stage regression. With an

adjusted R2 of 32.3%, the model fit is very good, and, taken together,the drivers explain a significant portion of the learning parametervariation (p b .001). The variance inflation factors all remain below 6,indicating that we do not have collinearity problems. For interpretationof the results, it is important to recall that our dependent variable is thenegative of the posterior log-learning parameter such that variablesassociated with positive coefficients imply more learning. To facilitatethe comparison of effects across variables, we focus on the standardizedcoefficients.

Zooming in on the drivers underlying incentives to learnfirst, wefindthat learning is significantly higher in categories with high performancerisk (! = .094, p b .01) and high involvement (! = .477, p b .01). Thisfinding matches expectations: consumers closely monitor their

16 We thus use two ‘layers’ of draws: one layer to assess the ‘mean’ posterior learningparameter for each household and category (as described in footnote 12), and another(higher) layer to assess the uncertainty of this posterior estimate. This higher layer in-volves draws of σc0

2 as well as draws for the means and standard deviations of themixing distribution of λc

h2, from their sampling distribution.

Table 4Learning magnitude per category, across households.a

Category Learning parameter: distribution withincategory across households

Median Lower quartile Upper quartile

Flavor additives .000 .000 .000Periodic care\diapers .001 .000 .002Dairy products .001 .001 .002Air refreshers .002 .002 .003Dairy drinks .008 .004 .013Liquid laundry detergent .019 .013 .030Salads .019 .001 .457Meal decorators .020 .009 .048Breakfast cereals .022 .016 .033Mouth hygiene .025 .019 .033Pet food .026 .022 .029Fabric softeners .044 .023 .081Spices and herbs .049 .029 .077Alcoholic drinks .054 .048 .060Dish detergent (tablets) .066 .030 .241Powdered laundry detergent .096 .074 .146Margarine\oil .097 .050 .208Liquid dish detergent .136 .136 .137Milk substitutes .426 .349 .527Sweet-and-sour canned products .541 .499 .582Ready-to-eat meals .682 .461 1.076Toilet and kitchen tissues .761 .079 3.691Pastries .779 .736 .849Crackers .785 .304 3.484Canned soup .877 .697 1.092Salty snacks .945 .441 1.943Eggs 1.256 1.102 1.447Hair care 1.519 1.367 1.771Baking products 1.648 .888 2.630Cleaning materials 1.792 .297 10.172Sugar 1.938 1.458 2.701Cleaners 2.259 2.258 2.259Deodorant 3.546 .263 18.054a The table reports the (posterior) learning parameter, i.e. the ratio of consumption

signal variance to initial variance. Higher levels of the learning parameter imply lesslearning.


consumption experiences and update their quality beliefs in categorieswhere the risk of making a wrong choice is higher and/or has personalrelevance to them. We also find consumption-based learning to bestronger in more expensive categories (! = .082, p b .01). Althoughthis effect was anticipated, it is surprisingly weak (partial eta-squaredof .009, compared to .143 for involvement17), suggesting that consumers'motivation to learn stems from non-monetary rather than monetarystakes.

As expected, variety seeking reduces the degree of consumption-based quality updating. Learning is significantly lower in categorieswith a highneed for variety (! = ! .396, p b .01) and among consumerswho buy many different brands within a given category (! = ! .079,p b .01). In a similar vein, consumers' general tendency to be brandloyal exerts a positive impact on learning (! = .067, p b .01)— complet-ing the consistent pattern of results for this dimension.

With regard to domain familiarity, we obtain a significant positivemain effect for category purchase frequency (! = .391, p b .01). Atthe same time, the coefficient of the squared term is also significantwith the opposite sign (! = ! .158, p b .01). This finding suggeststhat households with more extensive category expertise (are) better(able to) keep track of the performance of specific brands based onnew incoming consumption signals. However, this effect occurs onlyup to a certain point, after which it levels off, perhaps because thereis little remaining uncertainty to be resolved.

Turning to the control variables, we find significantly more learningin food categories (! = .265, p b .01). Consumers who generally plantheir purchases in advance are no different from others in theirconsumption-based learning (! = .003, p = .794). Interestingly,households that act as shopping mavens (based on the self-reportsurvey data) appear to learn slightly less (! = ! .024, p = .024) — apossible sign of overconfidence on their part (Hutchinson & Eisenstein,2008).

Taking the findings together, we find that all three dimensionssignificantly contribute to the degree of learning. Whereas most of theexplained variation comes from consumers' incentives to learn (48%),variety seeking (19%) and domain familiarity (14%) also exert anon-negligible influence.18 Zooming in on the role of category- versus

household-specific factors, however, it appears that the former play afar more important role, which we turn to next.

5.2.2. Do some households learn more than others?A somewhat surprising outcome of the second-stage regression is

that household characteristics that are independent of the categoryexplain only a small portion of the variation in learning. This findingraises the following question: are there households that learn more(less) than others across categories, and can we profile them? Toexplore this issue, we proceed as follows.19 For each category, we firstdetermine the median level of the learning parameter across house-holds. Next, for each household, we assess the fraction of categorieswhere it is present, in which it exhibits above-median learning(i.e., has a learning parameter below the category median). Householdswith a fraction of zero would be ‘weak learners’; that is, they learn lessthan others in all the categories that they buy. Conversely, householdswith a fraction of one would learn more strongly than others for allpurchased items.

Fig. 1 displays the histogram of this metric. A first observation isthat many households are situated in the middle, indicating thatalthough they are relatively strong learners in half of the categories,they are revealed to be weak learners in the other half. This findingsuggests that consumption-based quality learning is not simply ahousehold trait; it fits in with the notion that consumers (i) often‘trade down’ (go for cheap alternatives) in some categories to be ableto ‘trade up’ (focus on quality) in others (Costa, Dalens, Jacobsen, &Ramsö, 2006; Nielsen, 2008; Silverstein & Butman, 2006; Steenkamp,van Heerde, & Geyskens, 2010) and (ii) update their brand qualitybeliefs predominantly in the latter categories, where quality is of focalimportance.20

Meanwhile, Fig. 1 does point to substantial variation across house-holds, with a non-negligible subgroup being located at the extremesof the scale. To better understand the profile of these households, wesplit the sample into a lower-quartile segment (fraction below .4: theweak learners), a higher-quartile segment (fraction above .6: the stronglearners), and a mid-segment. We then link segment membership tothe households' socio-demographic characteristics through a multino-mial logit model (see Table 6 for the estimation results). We find that

17 We note that involvement and expensiveness are not correlated in our data (cor-relation coefficient: ! .003, p > .10), indicating that they capture separate constructs.Similarly, the correlation between performance risk and expensiveness, though signif-icant, remains very low (correlation coefficient .05; p b .05).18 Figures based on the ANOVA results associated with our second-stage regression.Control factors account for the remaining portion of explained variation (15%).

19 Because most households purchase only a subset of categories, cluster analysis (i.e.,grouping households based on the category-learning parameters) is not appropriatehere, as it would suffer from a major missing values problem.20 See also Erdem, Mayhew, and Sun (2001), who find use-experience-sensitive con-sumers to be less price-sensitive.

Table 5Second stage results: impact of learning drivers.

Variable Variable notation Unstandardized coefficients Standardized coefficients p-Value

Constant 3.895 0.00Purchase frequency CatKnowc

h 0.175 .391 0.00Purchase frequency_squared ! .003 ! .158 0.00Category expensiveness Expensivenessch 0.001 0.082 0.00Performance risk Riskc 0.922 0.094 0.00Involvement Involvementc 3.490 .477Number of brands consumed Number_Brandsch !0.058 !0.079 0.00Brand loyalty Brand_Loyaltyh 0.219 0.067 0.00Need for variety Need_Varietyc !2.773 !0.396 0.00Dummy for food categories Foodc 1.315 .265 0.00Mavenism Mavenismh !0.099 !0.024 0.02Purchase planning Planningh !0.006 !0.003 0.79N = 6668R2 = .570, R2

adj = .323F test: p b .001

Note: Dependent variable: ! log of learning parameter (=!log (λch2/σc0

2 )), weighted regression.


low-incomehouseholds tend to be stronger learners: they have a signif-icantly lower probability of being in the mid- or low-learning segment(p b .07 and p b .04, resp.). In contrast, the low-learning segment com-prises more two-person households (rather than singles or householdswith children) — be it that this effect is only marginally significant(p b .10). Interestingly, the age of the household head has aninverted-U effect (p b .05), suggesting that the propensity to be aweak learner is lower for both young and elderly shoppers. What cate-gories do these households learn differently about? Comparing thecategory-specific learning parameters, we find that weak learners failto update their quality beliefs especially in pastries, deodorant, mouthhygiene and breakfast cereals — categories where strong learnerslearn significantly more. Items such as margarine, salads, and dairydrinks, where learning is strong to begin with, trigger comparablelearning among the different household groups.

6. Discussion, limitations and future research

Emerging empirical evidence has already suggested that consumption-based brand-quality learning not only occurs in ‘consequential’ or evolving

categories but also prevails for CPG products. However, the studies todate are restricted to only a few categories – which begs the question ofgeneralizability– anddonot examinewhich consumers aremore inclinedto learn. In this study, we explored patterns in the development ofconsumption-based brand-quality knowledge, across 584 householdsand 33 packaged goods categories, and investigated household and cate-gory characteristics that drive learning.

6.1. Findings

Our main findings are as follows. First, we obtain evidence oflearning for almost all products: except in one category (where learn-ing does not improve fit), consumers are found to update their beliefsabout brand quality based on their consumption experiences. Thisfinding underscores the importance of such learning even in mature,packaged goods settings and justifies the use of marketing tacticsdesigned to benefit from consumption-based quality assessmentsover time.

Second, the degree of learning varies substantially across categoriesand across households within categories. In some cases, learning isminimal. This is, for instance, the case for cleaners, where even forstrong-learning consumers (upper quartile), only 30% of the brand'squality belief is shaped by the initial consumption experience. Inother instances, households' quality beliefs are predominantly shapedby recent consumption signals, as is the case for, e.g., diapers. In addi-tion, cross-household differences are significant and substantial – incategories where learning is strong or weak overall (e.g., for toilettissue, the inter-quartile range of the posterior learning parameterincreases from a low of .79 to a high of 3.691; for deodorants: from.263 to 18.05).

Third, our results uncover category-related characteristics andhousehold-category features that drive the degree of consumption-based learning. As expected, the need for variation and the tendencyto buy several brands detract from the consumption-based updating ofspecific brand quality beliefs. Conversely, consumers attach far moreweight to experienced consumption quality for higher-involvementCPG items and in categories with high performance risk. Higher ‘stakes’in these categories imply that households have extra incentives to keeptrack of the products' quality. Our results thus empirically corroboratethat, even in CPG categories, consumer motivation is a key determinantof experience-based learning (Hoch & Deighton, 1989). Although wealso observe more quality-belief updating for products where a typicalpurchase is expensive, this effect is much less pronounced. Therefore,non-monetary rather than monetary stakes play a pivotal role in con-sumers' motivation to learn. We also find that learning is stronger incategories for which the consumer has a history of frequent purchasesand thus is more familiar overall. This finding is in line with the ‘enrich-ment’ hypothesis, that prior knowledge reduces the cost of processingnew information and enhances the ability to learn (Hoch & Deighton,1989; Urbany et al., 1989). Following this hypothesis, frequent categorybuyers would be better able to interpret consumption signals andextract (brand) quality information from these signals. Although wesee evidence of such an effect, we also find that it levels off for veryhigh levels of purchase frequency — in line with the premise thatencoding deteriorates at very high levels of familiarity (Johnson &Russo, 1984).

Fourth, we find some differences in consumers' overall tendency tolearn, with consumption-based updating of brand-quality beliefs beingless prevalent amonghigh-income,middle-aged, and two-personhouse-holds. The overarching finding, however, is that ‘stable’ (category-independent) household characteristics explain only a small portionof the degree of learning. The vastmajority of households exhibit stronglearning in some categories but weak learning in others. This findingsuggests not only that consumption-based quality learning is notsimply a household trait but also that households tend to update theirquality knowledge more or less strongly depending on the product.

Fig. 1. Households' degree of learning across categories.

Table 6Link between segment membership and socio-demographics.

Variable Segment 1(weaklearners)

Segment 2(mediumlearners)

Segment 3(stronglearners)a

Intercept 8.81⁎⁎ 2.54Low income ! .663⁎⁎ ! .659⁎⁎ –

Two-person household .314⁎ .229 –

Age of head-of-householdb !2.317⁎⁎ ! .542 –

Age of head-of-household_squared .149⁎⁎ .029 –

Model fit: !2Loglikelihood 133.0, Chi-Square statistic 16.4 (p b .05)a Reference segment.b Based on 15 subsequent age categories.

⁎⁎ p b .05.⁎ p b .10.


Such an implication appears in line with the premise that consumershave limited attention and cognitive abilities and closely keep track ofbrand quality only in product categories that have personal relevancefor them (Celsi & Olson, 1988). It also appears consistent with thenotion that many consumers ‘trade up and down’ (Costa et al., 2006;Nielsen, 2008; Silverstein & Butman, 2006), exhibiting a strong interestin quality (learning) in some categories while focusingmore on price inothers.

6.2. Implications

6.2.1. Consumer welfareThe commonly accepted view is that learning enhances consumer

welfare. As indicated by Hutchinson and Eisenstein (2008), learningfrom experience enhances consumer expertise (they become betterand more efficient in their role as consumers). Moreover, experience-based learning reduces consumers' perception bias for specific brands,i.e., brings their quality perceptions of the brands closer to their truequality (Erdem, 1998). Therefore, learning allows consumers to selectproducts that better match their preferences at lower cost. In addition,there exists an indirect effect through ‘supply-side’ factors. Categoriesin which consumers learn leave less room for ‘cheating’ and undue(misleading) quality signaling. These categories also entail strongerincentives for manufacturers to continuously monitor the quality(consistency) of their products and provide free samples, allowingconsumers to assess the brand's quality ‘firsthand’ rather than havingto rely on ad messages. Therefore, learning leads to higher consumerwelfare. From this perspective, it is particularly encouraging thatconsumers learn more in categories where they are more vulnerableor have higher stakes, that is, expensive categories, frequently boughtcategories, and categories featuring high performance risk. Similarly,the finding that more ‘vulnerable’ segments (low-income, large-familyor single, young and elderly households) exhibit stronger learning ispositive from a consumer welfare perspective.

6.2.2. Managerial implicationsOur results reveal that even for CPG products, consumers continue

to learn from experience with (and keep on forgetting about) incum-bent brands. This finding has important implications for the managersof these brands. First, it provides a rationale for continued investmentsin quality, even for established brands. Although such investments maynot yield an immediate boost in brand share, their effect is boundto materialize after several periods as consumers gradually update

their quality perceptions across subsequent consumption occasions.A second implication is that marketing actions that stimulate trialand consumption – such as free samples –may remain powerful instru-ments for incumbent brands in ‘mature’ CPG categories. By highlightingcategories where learning is stronger, our results guide multiproductmanufacturers in the allocation of free sampling promotions: samplingappears most promising in frequently purchased, expensive, high-involvement categories with high performance risk but may pay offless in categories where variety seeking prevails. Of course, somecaution is warranted here, as the learning parameter – which capturesthe amount of information that consumers extract from an initialconsumption experience with the brand21 – is not the only driver of abrand's free trial effectiveness. First, though learning reduces qualityuncertainty, it also affects consumers' mean quality perception of thebrand, and this adjustment (and the ensuing benefits from learning)is obviouslymore positive for brandswhose true quality is high. Second,even for a given learning parameter, quality updating diminishes asconsumers familiarize themselves with the brand. It follows thatwithinthe boundaries of a given category, the effectiveness of free samplingwill be higher for brands that ‘deliver on their promise’ (generatetruly satisfactory quality experiences) and lower for brands that alreadyoccupy a dominant market position — especially among their regularbuyers.

To gain an understanding of these effects, we run a small simulation,in which we consider a manufacturer that operates in multiple catego-ries, and focus on the different brands\categories in our analysisproduced by that manufacturer.22 For each of these brands, we useour estimated model to dynamically simulate the effect of a free trialpromotion, using our actual data set as a backdrop. The promotion‘runs’ for one week per year, during which each consumer engaging ina category purchase receives a small version (one tenth of the regularpackage size) of the focal brand for free. Similarly to Erdem (1998),we assume that consumers who receive the sample actually use it. Wethen compare the brand shares predicted by the model with and with-out the sample promotion. Because these share(s) (differences) dependon the estimated parameters, which have some inherent uncertainty,

21 That is, for a given level of prior brand uncertainty.22 For confidentiality reasons, the data provider does not allow us to reveal the man-ufacturer and brand names.

Table 7Free trial effects: illustration.

Product category Brand Free trial effectb

(brand share increase)Category learning parameter:Median [IQR]c

Brandbase share

True brandqualitya

Free trial effectcompared tocompany average(t-statistic)Mean SD

Canned soup Brand 3 0.002⁎⁎ 0.000 0.877 [.697–1.092] 0.640 + !18.29Hair care Brand 1 0.004⁎⁎ 0.001 1.519 [1.367–1.771] 0.337 + !8.50Powdered laundry detergent Brand 1 0.006⁎⁎ 0.002 0.096 [.074–.146] 0.130 ! !2.23Flavor additives Brand 2 0.007⁎⁎ 0.002 .000 [.000–.000] 0.095 -! !1.65Fabric softeners Brand 2 0.008⁎⁎ 0.003 0.044 [.023–.081] 0.390 ! !0.98Dish detergent (tablets) Brand 3 0.010⁎⁎ 0.002 0.066 [.030–.241] 0.333 ! !0.35Cleaning materials Brand 3 0.013⁎⁎ 0.004 1.792 [.297–10.172] 0.486 0 0.78Meal decorators Brand 1 0.016⁎⁎ 0.001 0.020 [.009–.048] 0.361 ! 8.72Liquid laundry detergent Brand 1 0.027⁎⁎ 0.007 0.019 [.013–.030] 0.368 + 2.54Average across company products .01

⁎⁎ Effect significantly different from zero (p b 0.05).a Based on parameter estimates of the brand's true quality relative to the leading competitors in the category (see Appendix C): + = higher than average, ! = lower than

average, !! = far lower than average, 0 = average.b Average increase in brand share, across all (promotion and non-promotion) weeks in the data set, resulting from the free sample promotion. Increase expressed as a fraction

(e.g. .002 refers to .2% points). In the free sample promotion week (once every 52 weeks), category buyers receive a small (10% of regular pack size) sample of the promoted brand.c IQR = inter-quartile range (see also Table 4).


we repeat the simulations for 300 parameter draws from their samplingdistributions and use these draws to derive the standard errors of thefree sample effects.

Table 7 displays the average increase in share for the focal brandsresulting from the promotion, listed from low to high. First, the freetrial generates a significant positive effect for each of the brands,suggesting that such sampling indeed may play out well. Given thatthe promotion is ‘implemented’ only once every 52 weeks and thatits impact is calculated across all (promotion and non-promotion)weeks, the effects are quite sizable. On average, the sample leads toa one percentage point higher brand share, the effects being muchhigher in some cases. For instance, for the company's liquid laundrydetergent brand, a share increase of 2.7 percentage points is noted —approximately 7% of the brand's original share. These benefits willhave to be weighed against the costs of the sample promotion. Second,the pattern of effects reflects the estimated category differences indegree of learning. As shown by the last column of Table 7, categoriesfor which the share increase is significantly below the company (onepercentage point) average are those with a consistently high learningparameter (low learning) (canned soup, hair care, powdered laundrydetergent), whereas above-average gains are obtained for productswhere the learning parameter remains low (meal decorators, liquidlaundry detergent). Categories where sampling does not significantlyover- or underperform are those with ‘medium’ learning (fabric soft-eners, dish detergent tablets) or strongheterogeneity across households(cleaning materials). Therefore, our observed category differences inlearning offer guidance to manufacturers in determining which prod-ucts to offer free samples of. Third, notable deviations from the patternare observed. Even though it belongs to the strongest-learning category(flavor additives), the free sample effect for the company brand remainsrelatively small (.7% points)—which may be ascribed to the fact that itstrue quality is far below that of leading competitors (see the true qualityestimates in Appendix C). In addition, interesting reversals occur:though quality updating is stronger (i.e., the category learning parame-ter is lower) for canned soup than hair care, the sampling effect forthe company's brand is only half, possibly as a result of its alreadydominantmarket position (64%market share, indicating thatmany con-sumers are already familiar with the brand). Therefore, as mentionedearlier, the effect of sampling hinges on the brand's true quality anduntapped market potential — something for brand managers to reckonwith.

Targeted samplingmay be the ideal approach in this case. Specifically,in categories with a high learning parameter, brands that already enjoy alarge customer base will especially benefit from ‘competitive’ targetedpromotions, i.e., offering free samples to previous non-buyers (ratherthan buyers) of the brand (Zhang & Wedel, 2009), who still stronglylearn from the free trial experience. In a similar vein, in categorieswhere the learning parameter itself is heterogeneous (in the illustrationabove, cleaning materials), the sampling effect may be boosted bytargeting the stronger learners (i.e., frequent category buyers or house-holds that buy few different brands).

6.3. Limitations and future research

Our study has limitations that open interesting opportunities forfuture research. First, we focused on the (four to six) top brands ineach category. To the extent that leading brands are better known toconsumers than less popular brands, they may be subject to weakerlearning. As such, our estimate of the degree of learning is likely to bea conservative one, something to be verified in future studies. Second,our set of household and category drivers (which we needed to linkwith the panel households' estimated learning parameters) was limitedto the items available from the surveys. Future studies could extend thislist with other perceptual and attitudinal variables. One such variablecould be the ‘ambiguity’ of experience information in the product cate-gory (Hoch & Deighton, 1989), i.e., the degree to which consumers find

the brand's quality difficult to pin down because, for instance, it is easilyconfoundedwith contextual factors. Incorporating these variables couldfurther our understanding of when and why households learn fromconsumption. Third, as in previous choice models with Bayesianupdating, our measure of learning was an implicit one, as brand qualitybeliefs are unobserved. To ensure identification, we thus had to imposerestrictions (and set the initial belief uncertainties and consumption sig-nal variances to be equal across brands within a category). As indicatedby Shin et al. (2012), combining observed brand choices with statedmeasures of brand familiarity and liking may remove the need forsuch restrictions and allow for a more refined assessment of learningby brand or brand-type (national brand versus private label). Finally,given our broad range of categories, we could not estimate the brandchoice models simultaneously across categories, which may havedampened the cross-category effects. In addition, we focused on brandlearning within categories. Analyzing how consumers update theirbeliefs about umbrella brands across categories is a challenging topicfor future study.

Acknowledgments

The authors gratefully acknowledge AiMark for providing the data.They also thank Marnik Dekimpe, Kusum Ailawadi, Puneet Manchandaand Ralf van der Lans for their excellent suggestions. Finally, theythank the IJRM guest-editor, Hubert Gatignon, and the (anonymous)members of the review team, for their very constructive and helpfulcomments on earlier versions of the paper.

Appendix A. Estimation issues in the first stage of the analysis

A.1. Identification

Assuming that at time t = 0, consumers know the brands' meantrue quality across the population and begin with rational expectations,we let μh

jc0 ! vqjc , i.e., set the initial mean quality beliefs for a brandequal to its true quality population mean (see Crawford & Shum,2005; Narayanan &Manchanda, 2009 for a similar approach). To obtainidentification, we fix in each of the category-specific models the popu-lation mean of one of the true quality parameters to zero (i.e., vqjc ! 0for j = 1). The remaining (mean) true brand qualities are identifiedrelative to the reference brand based on the choice probability of con-sumers who are well experienced with a given brand and thus havelow quality uncertainty for that brand. Variations within the samebrand across these experienced consumers allow us to assess the stan-dard error of the true qualities' mixing distribution. We acknowledgethat this assessment of true brand qualities rests on two assumptions.First, the consumers' rate of forgetting should be substantially lowerthan their rate of learning, such that repeated consumption indeedlets their quality beliefs converge to their true qualities. Second, theobserved consumption histories should be sufficiently long for experi-enced consumers' beliefs to converge to the true qualities. Previousstudies suggest that it typically takes consumers several periods toforget what they learned from consumption (see, e.g., Mehta et al.,2004). Moreover, our estimation period covers 2.5 years, duringwhich consumers may accumulate experience. Therefore, it is reason-able to expect that, at least for a subset of households, we observeconvergence to the true brand qualities relative to the referencebrand.

The response parameters for the price and promotion variables(i.e., mean and standard error of themixing distributions) are identifiedbased on the fact that these variables change over time both across andwithin brands, with each consumer facing several changes in thesevariables.

The initial brand quality uncertainty, degree of learning and rate offorgetting are assessed based on consumers' purchase patterns overtime for the different brands. Like previous authors (see e.g., Mehta


et al., 2004; Erdem & Keane, 1996), we restrict the initial variance tobe equal across households and brands and set the consumption sig-nal variance to be the same for all brands.23 The learning rate, i.e., theratio of the consumption signal variance to the initial belief uncertainty,is identified based on the degree to which the (systematic) utility (andensuing choice probability) of a brand changes following brand con-sumption. The initial uncertainty (variance of the prior brand beliefs)may be separately identified from the evolution of the stochastic com-ponent in the utilities following brand consumption. As shown by themean updating equation (Eq. (4)), this random component stemsfrom the learning errors in the consumption signals, which weighmore heavily as the initial variance increases.

Conversely, the forgetting parameter is identified based on thedegree to which a brand's choice probability changes following aperiod in which the brand was not consumed. Learning may be sep-arated from forgetting by comparing the change in choice probabilityafter consumption for occasions where the previous brand purchaseoccurred a long time ago or was very recent. If a consumer stronglylearns but also strongly forgets, the ratio of (i) the brand's purchaseprobability after a long interpurchase time, relative to (ii) the brand'spurchase probability after a short interpurchase time, will be muchlower than if he hardly learns but also hardly forgets. The dataallow us to separate out these two patterns based on the variationin interpurchase times for a given brand within consumers' purchasehistories.

The risk aversion parameter and the brands' true qualities, both ofwhich are allowed to vary across households, also must be separatedout. The risk aversion parameter is not simply multiplied with thebrand's mean quality belief (which depends on the true brand quality);it also enters the utility in a separate term involving the uncertainty inthis belief, which is independent of the true quality but decreases (orincreases) depending on brand purchases (forgetting). Because con-sumers are observed to switch brands (e.g., as a result of promotions),and because purchases deterministically reduce the brand's beliefuncertainty but do not systematically increase or decrease its meanquality belief,24 wemay separate true quality effects from risk aversion.As indicated by Crawford and Shum (2005) and Chintagunta, Jiang, andJin (2009), the separate identification of the risk aversion parameterand the consumption signal variance is more subtle. High risk aversiongoes along with a low overall degree of switching, i.e., a high overallpersistence of brand choices. The consumption signal variance, inturn, is identified by the degree to which switching changes over timeas a function of the number of previous brand consumptions. Thispattern of changes over time is restricted by the Bayesian variance-updating expression (Eq. (5)), which allows the separation of the riskaversion and learning parameters.

A.2. Dealing with left truncation

Wedo not observe consumers from the beginning of their consump-tion history and thus face the problem of left truncation. Following pre-vious authors (e.g., Mehta et al., 2004), we include an initializationperiod. Specifically, we use the first four purchases of a household tobuild μjcth and σjct

h 2 and use these levels to initialize the estimation ofthe remaining data. We set the initial quality beliefs at the beginningof the initialization period equal to the mean of the true quality acrossthe population, such that μh

jc0 ! vqjc .

A.3. Likelihood function

For any household h, the likelihood of the entire purchase history

in a categoryDhTh , conditional on Ωh; ΕhTh

n o; where Ωh summarizes the

set of (household-specific) parameters related to prior beliefuncertainty, true brand qualities, sensitivity to utility determinants(i.e., price, feature/display), risk aversion, forgetting, and consumption-based learning (i.e., consumption signal variance), and where ΕhTh isthe vector of consumption signals Gh

jt received by household h up topurchase occasion t, is as follows:

Lh DhTh

###Ωh; ΕhTh

! "! $

Th

t!1$J

j!1Pr dhjt ! 1

###Ωh; Εht

! "dhjt:

We denote the p.d.f. of parameters in Ωh as uΩ(Ωh) and the p.d.f. ofΕhTh as uE(ΕhTh). The unconditional likelihood of household h's purchasehistory Dh

Th is as follows:

Lh DhTh

###αc

! "! %

Ωh %ΕhtLh Dh

Th Ωh; ΕhTh

###"uE ΕhTh

! "dΕhTh

! "uΩ Ωh

! "dΩh

:!

Because this likelihood involves multidimensional integrals,numerical computation is prohibitively expensive. Therefore, in linewith previous studies, we resort to simulated likelihood. Using Fsets of scrambled Halton draws (Train, 2003) for the coefficientsin Ωh and the consumption signals in ΕhTh , we obtain an estimateof Lh:

L̂h DhTh

###αc

! "!XF

f

Lh DhTh Ω f

; Ε f###

"! ":

!

We set F = 100 because larger values are not feasible given thecomputational demands of the model. As a robustness check, werun the model with different sets of draws, and the results do notchange.

The log-likelihood for N households is as follows:

LL DhTh

n oN

h!1

###αc

! "!XN

h!1

lnL̂h DhTh

###αc

! ":

We estimate the parameters in αc by maximizing this likelihood:

αc;MLE ! arg maxα

LL DhTh

n oN

h!1

###αc

! ":

23 From the updating expressions, it is clear that the initial perception variance andthe quality signal variance are not separately identified for each brand and consumer(only their ratio is identifiable; see Shin et al., 2012). Therefore, we must impose re-strictions. One possibility is to set one element of the ratio (either the consumption sig-nal variance, or the initial variance of the quality belief) equal to one and estimate theother. Another option is to impose homogeneity restrictions on each component, e.g.,make the initial variance the same across consumers, and the consumption signal var-iance homogeneous across brands, an approach followed by Narayanan andManchanda (2009). We use the latter approach here. Moreover, similar to Mehta etal. (2004) and Erdem and Keane (1996), we restrict the initial variance to be homoge-neous across brands to further facilitate identification.24 Because the consumption signals are random draws (from a distribution with amean equal to the true brand quality), the value of them may lift or reduce the previ-ous mean quality belief.


Appendix B. Fit of first stage model, compared with a static model

Appendix C. Stage 1 estimation results for the different categories

Category Number of brands Static model Bayesian learning model

Number of parameters Loglikelihood AIC Number of parameters Loglikelihood AIC

Flavor additives 6 17 1577.06 3188.12 22 1508.86 3061.73Periodic care\diapers 4 13 2246.74 4519.49 18 1620.99 3277.99Dairy products 4 13 7475.44 14,976.89 18 7426.81 14,889.61Air refreshers 4 13 874.60 1775.19 18 744.23 1524.46Dairy drinks 5 15 2914.15 5858.29 20 2860.84 5761.69Liquid laundry detergent 5 15 1218.76 2467.53 20 1087.58 2215.16Salads 4 13 3326.95 6679.89 18 3252.14 6540.29Meal decorators 4 13 7323.15 14,672.29 18 7342.91 14,721.82Breakfast cereals 5 15 1638.43 3306.86 20 1506.56 3053.13Mouth hygiene 4 13 1060.31 2146.62 18 914.76 1865.52Pet food 6 17 5224.73 10,483.47 22 4816.30 9676.60Fabric softeners 4 13 1055.36 2136.71 18 988.36 2012.71Spices and herbs 4 13 2162.85 4351.70 18 2082.12 4200.24Alcoholic drinks 4 13 3477.71 6981.43 18 3367.44 6770.88Dish detergent (tablets) 4 13 309.17 644.34 18 218.43 472.86Powdered laundry detergent 6 17 2041.71 4117.43 22 1199.77 2443.54Margarine\oil 5 15 4828.34 9686.68 20 4605.69 9251.37Liquid dish detergent 4 13 518.96 1063.91 18 395.04 826.07Milk substitutes 5 15 1267.60 2565.20 20 1124.12 2288.24Sweet-and-sour canned products 5 15 1010.01 2050.03 20 959.40 1958.80Ready-to-eat meals 4 13 3236.55 6499.10 18 3130.49 6296.99Toilet and kitchen tissues 5 15 5016.56 10,063.12 20 4643.18 9326.37Pastries 5 15 620.49 1270.98 20 556.56 1153.11Crackers 6 17 1958.34 3950.68 22 1871.84 3787.68Canned soup 4 13 1057.03 2140.06 18 1029.64 2095.28Salty snacks 6 17 7670.29 15,374.58 22 7575.76 15,195.52Eggs 5 15 855.84 1741.69 20 621.10 1282.21Hair care 4 13 1776.59 3579.18 18 1628.40 3292.80Baking products 6 17 4538.97 9111.94 22 4445.11 8934.23Cleaning materials 4 13 808.63 1643.27 18 716.94 1469.89Sugar 4 13 1303.40 2632.80 18 1207.41 2450.83Cleaners 4 13 1623.11 3272.22 18 1543.05 3122.11Deodorant 5 15 403.39 836.79 20 191.58 423.15

Category Flavor additives Periodic care\diapers

Dairy products Air refreshers Dairy drinks Liquid laundrydetergent

Salads

Mean Brand1 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 FixedBrand2 0.06 2.25 ⁎⁎⁎ !3.76 ⁎⁎⁎ 1.32 ⁎⁎⁎ !4.83 ⁎⁎⁎ 0.74 ⁎⁎ !2.02 ⁎⁎⁎

Brand3 0.67 ⁎⁎⁎ !0.71 ⁎⁎⁎ !4.28 ⁎⁎⁎ 0.34 !5.23 ⁎⁎⁎ 0.23 !2.14 ⁎⁎⁎

Brand4 15.68 ⁎⁎⁎ 0.07 !3.98 ⁎⁎⁎ 0.91 ⁎⁎ !2.83 ⁎⁎⁎ 0.36 !1.06 ⁎⁎⁎

Brand5 0.47 ⁎ !3.72 ⁎⁎⁎ 0.81 ⁎⁎⁎

Brand6 2.61 ⁎⁎⁎

Belief variance at t = 1 1.05 ⁎⁎⁎ 1 ⁎⁎⁎ 2.18 ⁎⁎⁎ 1.89 ⁎⁎⁎ 2.32 ⁎⁎⁎ 1.98 ⁎⁎⁎ 0.96 ⁎⁎⁎

Signal st. error !16.29 ⁎⁎ !4.35 ⁎⁎⁎ !2.52 ⁎⁎⁎ !3.10 ⁎⁎⁎ !1.65 ⁎⁎⁎ !1.82 ⁎⁎⁎ !7.13 ⁎⁎⁎

Forgetting !6.90 ⁎⁎ !1.15 ⁎ !0.14 † !0.02 !2.97 ⁎⁎⁎ !2.72 ⁎⁎⁎ !0.65 ⁎⁎⁎

FD !14.41 ⁎⁎ 5.12 ⁎⁎⁎ 0.52 ⁎⁎⁎ 6.28 ⁎⁎⁎ 0.11 6.20 ⁎⁎⁎ 3.02 †

Price !0.05 !0.05 0.40 ⁎⁎⁎ 0.11 4.33 ⁎⁎⁎ !0.39 ⁎ !0.10Risk aversion 0.04 !0.03 !1.48 ⁎⁎⁎ !0.19 !1.22 ⁎⁎⁎ !0.35 ⁎ !0.61 ⁎⁎⁎

Variance (log) Brand1 !4.15 0.31 2.18 ⁎⁎⁎ !0.98 2.58 ⁎⁎⁎ !17.32 0.83 ⁎⁎⁎

Brand2 !4.36 † 3.59 ⁎⁎⁎ 1.37 ⁎⁎⁎ !1.18 ⁎⁎⁎ !0.97 ⁎⁎⁎ !0.74 ⁎ !0.92 †

Brand3 !16.29 0.82 ⁎⁎⁎ 0.89 ⁎⁎⁎ !0.80 ⁎ 1.12 ⁎⁎⁎ !6.43 !2.43 ⁎⁎

Brand4 4.88 ⁎⁎⁎ !1.97 ⁎⁎⁎ 0.71 † !0.13 2.08 ⁎⁎⁎ !0.02 0.16Brand5 !2.07 ⁎⁎⁎ 0.85 ⁎⁎⁎ !0.04Brand6 1.66 ⁎⁎⁎

Signal st. error !0.74 1.09 ⁎⁎⁎ !1.03 !0.05 0.11 0.71 ⁎⁎ 1.02Forgetting 3.81 ⁎⁎⁎ 1.59 ⁎⁎⁎ !1.84 ⁎⁎⁎ !0.07 1.08 ⁎⁎⁎ 2.77 ⁎⁎⁎ 1.05 ⁎⁎⁎

FD 0.68 1.93 ⁎⁎ 3.38 ⁎⁎⁎ 4.15 ⁎⁎⁎ 4.51 ⁎⁎⁎ !0.33 !0.97Price !1.94 ⁎⁎⁎ !4.24 ⁎⁎⁎ 1.30 ⁎⁎⁎ !2.50 ⁎⁎⁎ 5.96 ⁎⁎⁎ !0.20 !1.33 ⁎⁎⁎

Risk aversion !2.25 ⁎⁎⁎ 2.17 ⁎⁎⁎ !5.07 ⁎⁎⁎ !3.07 ⁎⁎⁎ !3.26 ⁎⁎⁎ !2.50 ⁎⁎⁎ !4.89 ⁎⁎⁎

Category Meal decorators Breakfastcereals

Mouth hygiene Pet food Fabric softeners Spices and herbs Alcoholicdrinks

Mean Brand1 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 FixedBrand2 0.73 ⁎⁎⁎ 0.26 0.58 ⁎ !0.70 ⁎⁎⁎ !0.51 ⁎⁎ 3.55 ⁎⁎⁎ !8.46 ⁎⁎⁎

Brand3 !1.03 ⁎⁎⁎ 1.24 ⁎⁎⁎ 0.17 0.06 !0.17 !1.82 ⁎⁎⁎ !9.21 ⁎⁎⁎

(continued on next page)


(continued)

Category Meal decorators Breakfastcereals

Mouth hygiene Pet food Fabric softeners Spices and herbs Alcoholicdrinks

Brand4 !0.83 ⁎⁎⁎ 0.82 ⁎⁎⁎ !0.13 !0.53 ⁎⁎⁎ !0.48 ⁎ !0.40 ⁎ !8.18 ⁎⁎⁎

Brand5 !0.17 0.15Brand6 0.03Belief variance at t = 1 1.06 ⁎⁎⁎ 2.61 ⁎⁎⁎ 2.60 ⁎⁎⁎ 2.70 ⁎⁎⁎ 1.00 ⁎⁎⁎ 2.45 ⁎⁎⁎ 4.54 ⁎⁎⁎

Signal st. error !1.87 ⁎⁎⁎ !1.09 ⁎⁎⁎ !1.08 ⁎⁎⁎ !0.51 ⁎⁎⁎ !4.00 ⁎ !0.84 ⁎⁎⁎ 0.75 ⁎⁎⁎

Forgetting !1.87 ⁎⁎⁎ !2.56 ⁎⁎⁎ !3.12 ⁎⁎⁎ !1.88 ⁎⁎⁎ 2.12 ⁎ !2.90 ⁎⁎⁎ !3.28 ⁎⁎⁎

FD 2.43 ⁎⁎⁎ 5.25 ⁎⁎⁎ 4.58 ⁎⁎⁎ 3.18 ⁎⁎⁎ 5.09 ⁎⁎⁎ !1.47 ** 0.57 ⁎⁎⁎

Price !0.04 0.03 !0.02 !0.27 ⁎⁎⁎ !0.16 !0.02 !0.88 ⁎⁎⁎

Risk aversion !0.62 ⁎⁎⁎ !0.53 ⁎⁎⁎ !0.57 ⁎⁎ !0.80 ⁎⁎⁎ !0.29 !0.87 ⁎⁎⁎ !2.13 ⁎⁎⁎

Variance (log) Brand1 1.59 ⁎⁎⁎ !0.03 0.38 !2.41 ⁎ !4.79 3.63 ⁎⁎⁎ 3.48 ⁎⁎⁎

Brand2 1.17 ⁎⁎⁎ !0.30 0.27 !41.42 !2.29 ⁎ 1.76 ⁎⁎⁎ !0.70 ⁎⁎

Brand3 0.10 !0.37 !3.64 ⁎ !0.09 !0.39 1.34 ⁎⁎⁎ 1.61 ⁎⁎⁎

Brand4 !0.23 0.00 0.10 !0.97 ⁎ !1.71 ⁎⁎⁎ 0.04 !9.52 ⁎

Brand5 !0.09 0.03Brand6 0.03Signal st. error 0.10 0.13 0.13 !2.43 ⁎⁎⁎ 0.18 0.14 !2.27 ⁎⁎⁎

Forgetting !0.23 0.92 ⁎⁎⁎ 1.04 ⁎⁎⁎ 0.76 ⁎⁎⁎ !5.95 0.94 ⁎⁎⁎ 0.94 ⁎⁎⁎

FD 1.19 ⁎⁎ !0.01 0.86 !0.11 !7.70 5.29 ⁎⁎⁎ 2.18 ⁎⁎⁎

Price !3.09 ⁎⁎⁎ 0.47 !4.21 ⁎⁎⁎ !0.95 ⁎⁎⁎ 1.42 ⁎⁎ !7.22 ⁎⁎⁎ !0.66 ⁎⁎⁎

Risk aversion !3.05 ⁎⁎⁎ !2.49 ⁎⁎⁎ !2.60 ⁎⁎⁎ !4.72 ⁎⁎⁎ !1.13 ⁎ !2.05 ⁎⁎⁎ !5.79 ⁎⁎⁎

Category Dish detergent(tablets)

Powderedlaundrydetergent

Margarine\oil Liquid dishdetergent

Milksubstitutes

Sweet-and-sourcannedproducts

Ready-to-eatmeals

Mean Brand1 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 FixedBrand2 1.41 ⁎ 0.86 ⁎⁎⁎ !1.25 ⁎⁎⁎ 0.27 0.39 ⁎⁎ 1.38 ⁎⁎ !0.98 ⁎⁎⁎

Brand3 0.45 0.90 ⁎⁎⁎ !0.73 ⁎⁎⁎ !0.12 0.03 !0.85 ⁎⁎⁎ !1.19 ⁎⁎⁎

Brand4 1.25 ⁎⁎ 1.64 ⁎⁎⁎ !0.21 ⁎ 1.62 ⁎⁎⁎ !0.42 ⁎ !0.53 ⁎ !1.09 ⁎⁎⁎

Brand5 1.02 ⁎⁎⁎ 0.44 ⁎⁎⁎ 0.49 ⁎⁎ !1.27 ⁎⁎⁎

Brand6 1.09 ⁎⁎⁎

Belief variance at t = 1 2.71 ⁎⁎⁎ 1.42 ⁎⁎⁎ 1.60 ⁎⁎⁎ 1 ⁎⁎⁎ 1 ⁎⁎⁎ 1.07 ⁎⁎⁎ 2.14 ⁎⁎⁎

Signal st. error !1.26 ⁎⁎ !0.82 ⁎⁎⁎ !0.57 ⁎⁎⁎ !0.55 ⁎ !0.13 0.13 0.72 ⁎⁎⁎

Forgetting !2.72 ⁎⁎⁎ !3.02 ⁎⁎⁎ !3.88 ⁎⁎⁎ !1.89 ⁎⁎⁎ !3.89 ⁎⁎⁎ !55.87 !16.90FD 10.22 ⁎⁎⁎ 5.65 ⁎⁎⁎ 0.37 ⁎⁎ 5.10 ⁎⁎⁎ 3.78 ⁎⁎⁎ 4.48 ⁎⁎⁎ 4.31 ⁎⁎⁎

Price 0.07 0.00 !0.75 ⁎⁎⁎ !0.40 !0.83 ⁎⁎⁎ 0.78 ⁎⁎⁎ !0.06Risk aversion !0.40 0.08 !0.31 ⁎⁎⁎ 0.65 ⁎⁎ 0.22 !0.24 !0.77 ⁎⁎⁎

Variance (log) Brand1 !3.42 !5.28 † 0.93 ⁎⁎⁎ !3.83 1.04 ⁎⁎⁎ 0.74 ⁎ !1.29 ⁎⁎⁎

Brand2 !0.27 !14.33 † !0.22 !4.87 ⁎⁎ !1.46 ⁎⁎⁎ !0.98 !1.21 ⁎

Brand3 !1.70 ⁎⁎ !13.85 1.03 ⁎⁎⁎ !4.97 ⁎⁎⁎ !0.39 !1.94 ⁎⁎⁎ !28.75 †

Brand4 !1.06 !3.47 !0.46 ⁎⁎⁎ !0.21 !1.88 ⁎⁎⁎ !1.85 ⁎⁎⁎ !2.43 ⁎⁎

Brand5 !2.66 ⁎⁎ 0.31 ⁎⁎ 0.61 ⁎ !2.08 *Brand6 !9.20 †

Signal st. error 0.13 !0.02 !0.15 !7.48 ⁎⁎ !0.61 ⁎⁎⁎ !3.33 ⁎ !0.47 ⁎⁎

Forgetting 0.72 1.60 ⁎⁎ 0.81 ⁎⁎⁎ 0.63 0.55 !0.52 !0.25FD !1.72 !0.96 † 2.42 ⁎⁎⁎ !8.82 1.89 † 1.28 !3.76Price !1.08 ⁎⁎ !0.66 ⁎⁎ 1.61 ⁎⁎⁎ !1.80 ⁎ 2.49 ⁎⁎⁎ 3.15 ⁎⁎⁎ !0.90 ⁎⁎⁎

Risk aversion !2.20 ⁎⁎⁎ !3.93 ⁎⁎⁎ !3.41 ⁎⁎⁎ !0.71 !3.34 ⁎⁎⁎ !2.85 ⁎⁎⁎ !4.89 ⁎⁎⁎

Category Baking products Cleaningmaterials

Sugar Cleaners Deodorant

Mean Brand1 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 Fixed 1.00 FixedBrand2 1.59 ⁎⁎⁎ 3.37 ⁎⁎⁎ !0.14 0.80 ⁎⁎⁎ 9.92 †

Brand3 !0.83 ⁎⁎⁎ 1.78 ⁎⁎⁎ 0.43 ⁎ 1.31 ⁎⁎⁎ 7.13Brand4 !0.66 ⁎⁎⁎ 0.38 !0.11 1.43 ⁎⁎⁎ 7.45Brand5 !1.38 ⁎⁎⁎ !0.56Brand6 !1.40 ⁎⁎⁎

Belief variance at t = 1 3.41 ⁎⁎⁎ 2.54 ⁎⁎⁎ 2.35 ⁎⁎⁎ 1.20 ⁎⁎⁎ 1.05Signal st. error 1.60 ⁎⁎⁎ !1.11 ⁎⁎⁎ 1.26 ⁎⁎⁎ 0.97 ⁎⁎⁎ !9.97Forgetting !33.77 !2.78 ⁎⁎⁎ !14.94 ⁎ !17.04 !1.50FD 2.43 ⁎⁎⁎ 3.29 ⁎⁎⁎ !0.19 5.58 ⁎⁎⁎ !0.50Price !0.11 !0.97 ⁎⁎⁎ !0.59 ⁎⁎⁎ !0.05 0.81 ⁎⁎⁎

Risk aversion !1.34 ⁎⁎⁎ !0.52 ⁎ !0.53 ⁎⁎⁎ 0.17 1.54 ⁎⁎⁎

Variance (log) Brand1 !6.43 ⁎ !0.22 !0.29 !580.69 0.03Brand2 !0.79 0.37 !2.60 ⁎⁎⁎ !3.20 ⁎⁎ 3.60Brand3 !54.10 0.45 !5.18 ⁎ !1.98 † !1.61Brand4 !38.16 !1.26 0.32 !5.07 ⁎⁎ !2.96Brand5 0.89 ⁎⁎⁎ !0.32Brand6 !0.55 ⁎

Signal st. error !0.41 0.13 !1.05 !18.58 !0.58Forgetting 2.71 1.03 2.73 !41.84 0.66FD !66.78 !0.10 !1.16 !217.38 1.40Price !3.78 ⁎⁎⁎ 1.94 ⁎⁎⁎ 2.80 ⁎⁎⁎ 0.45 ⁎ 0.71 ⁎⁎

Risk aversion !3.89 ⁎⁎⁎ !2.60 ⁎⁎⁎ !3.46 ⁎⁎⁎ !4.63 ⁎⁎⁎ !0.42 †

†p b .1., ⁎p b .05., ⁎⁎p b .01., ⁎⁎⁎p b .001. FD = feature\display.

Appendix C. (continued)


References

Ackerberg, D. A. (2003). Advertising, learning, and consumer choice in experience goodmarkets: An empirical examination. International Economic Review, 44(3),1007–1040.

Akçura, M. T., Gönül, F. F., & Petrova, E. (2004). Consumer learning and brand valuation:An application on over-the-counter drugs. Marketing Science, 23(1), 156–169.

Alba, J. W., & Hutchinson, J. W. (1987). Dimensions of consumer expertise. Journal ofConsumer Research, 13, 411–454.

Baumgartner, H., & Steenkamp, J. B. (1996). Exploratory consumer buying behavior:Conceptualization and measurement. International Journal of Research in Marketing,13, 121–137.

Beatty, S., & Smith, S. (1987). External search effort: An investigation across severalcategories. Journal of Consumer Research, 14, 83–95.

Broniarzcyk, S. (2008). Product assortment. In C. P. Haugtvedt, P. M. Herr, & F. R. Kardes(Eds.), Handbook of consumer psychology (pp. 755–780). New York: PsychologyPress.

Carpenter, G. S., & Nakamoto, K. (1989). Consumer preference formation andpioneering advantage. Journal of Marketing Research, 26, 285–298.

Carpenter, G. S., & Nakamoto, K. (1996). Impact of consumer preference formation onmarketing objectives and competitive second mover strategies. Journal of ConsumerPsychology, 5(4), 325–358.

Celsi, R., & Olson, J. (1988). The role of involvement in attention and comprehensionprocesses. Journal of Consumer Research, 15, 210–224.

Ching, A. T., Erdem, T., & Keane,M. P. (2012). Learningmodels: An assessment of progress,challenges and new developments.Working Paper.

Chintagunta, P., Jiang, R., & Jin, G. (2009). Information, learning, and drug diffusion: Thecase of Cox-2 inhibitors. Quantitative Marketing and Economics, 7, 399–443.

Costa, C., Dalens, F., Jacobsen, R., & Ramsö, J. (2006). The battle for Europe's grocery shoppers.Boston consulting group focus report.

Crawford, G. S., & Shum, M. (2005). Uncertainty and learning in pharmaceuticaldemand. Econometrica, 73(4), 1137–1173.

Du, R. Y., & Kamakura, W. A. (2006). Household life cycles and lifestyles in the UnitedStates. Journal of Marketing Research, 43(1), 121–132.

Erdem, T. (1998). An empirical analysis of umbrella branding. Journal of MarketingResearch, 35(3), 339–351.

Erdem, T., & Keane, M. P. (1996). Decision-making under uncertainty: Capturingdynamic brand choice processes in turbulent consumer goods markets. MarketingScience, 15(1), 1–20.

Erdem, T., Zhao, Y., & Valenzuela, A. (2004). Performance of store brands: across-country analysis of consumer store-brand preferences, perceptions andrisk. Journal of Marketing Research, XLI, 86–100 (February).

Erdem, T., Keane, M. P., & Sun, B. H. (2008). A dynamic model of brand choice whenprice and advertising signal product quality. Marketing Science, 27(6), 1111–1125.

Erdem, T., Mayhew, G., & Sun, B. H. (2001). Understanding reference-price shoppers: Awithin- and cross-category analysis. Journal of Marketing Research, 38, 445–457.

Erdem, T., & Sun, B. H. (2002). An empirical investigation of the spillover effects ofadvertising and sales promotions in umbrella branding. Journal of Marketing Research,39(4), 408–420.

Feick, L., & Price, L. (1987). The market maven: A diffuser of marketplace information.Journal of Marketing, 51, 83–97.

Gijsbrechts, E., Campo, K., & Nisol, P. (2000). Towards a theory-based measure ofpurchase variation. Academy of Marketing Science Review, 2000(6), 1–29.

Givon, M. (1984). Variety seeking through brand switching. Marketing Science, 3(11),1–22.

Groot, de M. H. (1970). Optimal statistical decisions. New York: McGraw-Hill.Hawkins, S. A., & Hoch, S. J. (1992). Low-involvement learning — Memory without

evaluation. Journal of Consumer Research, 19(2), 212–225.Heilman, C. M., Bowman, D., & Wright, G. P. (2000). The evolution of brand preferences

and choice behaviors of consumers new to a market. Journal of Marketing Research,37(2), 139–155.

Hoch, S. J., & Deighton, J. (1989). Managing what consumers learn from experience.Journal of Marketing, 53(2), 1–20.

Hoch, S. J., & Ha, Y. W. (1986). Consumer learning from advertising and the ambiguityof processing experience. Journal of Consumer Research, 13(2), 221–233.

Hutchinson, J. W., & Alba, J. W. (1991). Ignoring irrelevant information: Situationaldeterminants of consumer learning. Journal of Consumer Research, 18, 325–346.

Hutchinson, J. W., & Eisenstein, E. M. (2008). Consumer learning and expertise. In C. P.Haugtvedt, P. M. Herr, & F. R. Kardes (Eds.), Handbook of consumer psychology(pp. 103–132). New York: Psychology Press.

Inman, J., Winer, R., & Ferraro, R. (2009). The interplay among category characteristics,customer characteristics, and customer activities on in-store decisionmaking. Journalof Marketing, 73, 19–29.

Johnson, E. J., & Russo, E. (1984). Product familiarity and learning new information.Journal of Consumer Research, 11, 542–550.

Kahn, B. E., & Raju, J. S. (1991). Effects of price promotions on variety-seeking andreinforcement behavior. Marketing Science, 10(4), 316–337.

Kardes, F. (1994). Consumer judgement and decision processes. Handbook of socialcognition. In R. Wyer, & T. Srull (Eds.), Applications, Volume 2, . Hillsdale, New Jersey:Lawrence Erlbaum Ass. Inc.

Kardes, F., & Kalyanaram, G. (1992). Order-of-entry effects on consumer memory andjudgment: An information integration perspective. Journal of Marketing Research,29(3), 343–357.

McAlister, L., & Pessemier, E. A. (1982). Variety seeking behavior: An interdisciplinaryreview. Journal of Consumer Research, 9, 311–322.

Mehta, N., Rajiv, S., & Srinivasan, K. (2003). Price uncertainty and consumer search: Astructural model of consideration set formation. Marketing Science, 22(1), 58–84.

Mehta, N., Rajiv, S., & Srinivasan, K. (2004). Role of forgetting in memory-based choicedecisions: A structural model. Quantitative Marketing and Economics, 2(2), 107.

Miller, R. A. (1984). Job matching and occupational choice. Journal of Political Economy,92(6), 1086–1120.

Mitra, D., & Golder, P. N. (2006). How does objective quality affect perceived quality?Short-term effects, long-term effects, and asymmetries. Marketing Science, 25(3),230–247.

Narasimhan, C., Neslin, S., & Sen, S. (1996). Promotional elasticities and category char-acteristics. Journal of Marketing, 60(2), 17–30.

Narayanan, S., Chintagunta, P. K., & Miravete, E. J. (2007). The role of self selection, usageuncertainty and learning in the demand for local telephone service. QuantitativeMarketing and Economics, 5(1), 1–34.

Narayanan, S., & Manchanda, P. (2009). Heterogeneous learning and the targeting ofmarketing communication for new products. Marketing Science, 28(3), 424–441.

Nelson, P. (1970). Information and consumer behavior. Journal of Political Economy, 78,311–329.

Nielsen (2008). Grocery store choice and value for money. A Global Nielsen Consumerreport.

Pechmann, C., & Ratneshwar, S. (1992). Consumer covariation judgments — theory ordata driven. Journal of Consumer Research, 19(3), 373–386.

Pessemier, E. A., & Handelsman, M. (1984). Temporal variety in consumer behavior.Journal of Marketing Research, 21, 435–444.

Schwarz, N. (2004). Metacognitive experiences in consumer judgment and decisionmaking. Journal of Consumer Psychology, 14, 332–348.

Shin, S., Misra, S., & Horsky, D. (2012). Disentangling preferences and learning in brandchoice models. Marketing Science, 31(1), 115–137.

Silverstein, M., & Butman, J. (2006). Treasure hunt: Inside the mind of the new consumer.Portfolio Hardcover.

Steenkamp, J. B., Geyskens, I., Gielens, K., & Koll, O. (2004). A global study into drivers ofprivate label success. Report. Brussels: AIM.

Steenkamp, J. B., Nijs, V., Hanssens, D. M., & Dekimpe, M. G. (2005). Competitive reactionsto advertising and promotion attacks.Marketing Science, 24(1), 35–54.

Steenkamp, J. B., van Heerde, H., & Geyskens, I. (2010). What makes consumers willing topay a price premium for national brands over private labels? Journal of MarketingResearch, 47(6), 1011–1024.

Stigler, G. J. (1961). The economics of information. Journal of Political Economy, 69, 213.Szymanowski, M., & Gijsbrechts, E. (2012). Conditional cross-brand learning: when are

private labels really private? Journal of Marketing Research, 49(2), 231–247.Train, K. (2003). Discrete choice methods with simulation. Cambridge: Cambridge

University Press.Trivedi, M., & Morgan, M. (2003). Promotional evaluation and response among variety

seeking segments. The Journal of Product and Brand Management, 12(6), 408–423.Urbany, J. E. (1986). An experimental examination of the economics of information.

Journal of Consumer Research, 13, 257–272.Urbany, J., Dickson, P., & Wilkie, W. (1989). Buyer uncertainty and information search.

Journal of Consumer Research, 16, 208–215.van Heerde, H., Helsen, K., & Dekimpe, M. G. (2007). The impact of a product-harm

crisis on marketing effectiveness. Marketing Science, 26(2), 230–245.van Trijp, H., & Steenkamp, J. B. (1990). An investigation into the validity of measures

for variation in consumption used in economics and marketing. European Review ofAgricultural Economics, 17(1), 19–41.

Warlop, L., Ratneshwar, S., & van Osselaer, S. (2005). Distinctive brand cues and memoryfor product consumption experiences. International Journal of Research in Marketing,22(1), 27–44.

Wood, S. L., & Lynch, J. G., Jr. (2002). Prior knowledge and complacency in new productlearning. Journal of Consumer Research, 29, 416–426.

Wright, A. A., & Lynch, J. G. (1995). Communication effects of advertising versus directexperience when both search and experience attributes are present. Journal ofConsumer Research, 21(4), 708–718.

Zhang, J., & Wedel, M. (2009). The effectiveness of customized promotions in onlineand offline stores. Journal of Marketing Research, 46(2), 190–206.

Zhou, Z. (2002). Achieving late-mover advantage: The effects of enhancing and distinctivestrategies. Doctoral Dissertation. Virginia Polytechnic Institute and State University.


Does private-label production by national-brand manufacturers creatediscounter goodwill?

Anne ter Braak a,⁎, Barbara Deleersnyder b,1, Inge Geyskens b,2, Marnik G. Dekimpe b,a,3

a University of Leuven, Naamsestraat 69, 3000 Leuven, Belgiumb Tilburg University, Warandelaan 2, 5000 LE Tilburg, the Netherlands


Article history:First received in 21, June 2012 and was underreview for 5 monthsAvailable online 7 July 2013

Area Editor: Peter J. Danaher

Keywords:Private labelShelf allocationSourcing decisionsDiscountersChannel relationships

Discount stores have a private-label dominated assortment where national brands have only limited shelfaccess. These limited spots are in high demand by national-brand manufacturers. We examine whetherprivate-label production by leading national-brand manufacturers for two important discounters (one hardand one soft) creates discounter goodwill. We estimate a selection model that is based on a sample of 450manufacturer-category combinations from two leading discounters (Aldi in Germany and Mercadona inSpain), and we show that private-label production is indeed rewarded: national-brand manufacturers thatare involved in such practices have a higher likelihood of procuring shelf presence for their brands. Moreover,while powerful manufacturers are intrinsically more likely to obtain shelf presence with soft discounters,manufacturers with less power can compensate for this by producing private labels. No such dependenceon power exists for hard discounters.However, not all national-brand manufacturers are equally likely to produce private labels for discounters.We find that national-brand manufacturers are less likely to do so when: (a) they experience more salesgrowth, (b) it is more difficult to produce high-quality products in a specific category, (c) they invest moreadvertising support into their brands, and (d) they introduce more innovations. Moreover, a higher price dif-ferential relative to the discounter's private labels makes national-brand manufacturers less likely to engagein private-label production for hard discounters.


1. Introduction

Private labels are becoming increasingly important in grocery retail-ing. They already account for 43% of the total consumer packaged goods(CPG) consumption in the U.K., 32% in Germany, 31% in Spain, and 17%in the U.S. (ACNielsen, 2011). Absolute sales numbers are equallyimpressive. Wal-Mart's U.S. private-label sales, for example, wereexpected to generate approximately $90 billion in 2012 (PlanetRetail,2012a). In 2011, the value of the European private-label food marketreached €436 billion (PlanetRetail, 2011).

Private-label production opportunities have arisen in response tothese large volumes. A substantial part of this production is accountedfor by dedicated private-label manufacturers. However, severalnational-brand manufacturers – such as Alcoa (which owns Reynolds

Wrap aluminum foil), Parmalat, and H.J. Heinz – have adopted an “ifyou can't beat them, join them” attitude, as they now engage inprivate-label production (Dunne & Narasimhan, 1999; Quelch &Harding, 1996). In the U.S., it has been estimated that over half ofthe national-brand manufacturers pursue a dual strategy, engagingin private-label production in addition to their national-brand activi-ties (Kumar & Steenkamp, 2007).

A major motivation for national-brand manufacturers to engagein private-label production is, as the former CEO of Ontario Foods tes-tified, “to cultivate a better relation with retailers” (Littman, 1992,p. 2). According to Dunne and Narasimhan (1999), private-label pro-duction represents a neglected opportunity for manufacturers thatseek closer ties with their retailers, offering a chance to create retailergoodwill. IGD's European Private Label Survey revealed that 47% of allof the surveyed suppliers believe that strengthening their relation-ship with the retailer could be a major advantage of supplying privatelabels (IGD, 2006). However, there is substantial variability in thewillingness of brand manufacturers to produce private labels. Whilesome manufacturers, such as Dole and Kraft, are eager to do so,other manufacturers, such as Coca-Cola and Heineken, have explicitlystated that they will never engage in private-label production. Thisvariation in willingness may be driven by differences in the manufac-turers' sales growth (Kumar & Steenkamp, 2007), in their ability to


⁎ Corresponding author. Tel.: +32 16 326 903; fax: +32 16 326 732.E-mail addresses: [email protected] (A. ter Braak),

[email protected] (B. Deleersnyder), [email protected] (I. Geyskens),[email protected] (M.G. Dekimpe).

1 Tel.: +31 13 466 8252.2 Tel.: +31 13 466 8083.3 Tel.: +31 13 466 3423.


Contents lists available at ScienceDirect



influence the produced private-label quality (e.g., by loweringprivate-label quality or by producing a private label that imitates acompetitive national brand to preserve their own sales) (Dunne,1999), or in the level of brand image that is at stake for them (deJong, 2007). Moreover, a retailer may be more inclined to collaboratewith some national-brand manufacturers to produce its private labelsthan with others, as some may provide more quality assurance(Sethuraman, 2009) or have higher innovative capacities (Kumar &Steenkamp, 2007) than others. In other instances, a retailer mayprefer to assign its private-label production to a dedicated producer(that does not have any national brands of its own), rather than to adual brander, as the former are known to be more cost-focused.Whether a given national-brand manufacturer produces private la-bels for a given retailer is therefore driven by manufacturer as wellas retailer considerations. This issue will be reflected in our subse-quent theorizing.

In this study, we look at the antecedents and consequences ofprivate-label production by national-brand manufacturers for re-tailers. This topic has recently been recognized as being of great man-agerial interest but seriously lacking in empirical research (Sayman &Raju, 2007, p. 147; Sethuraman, 2009, p. 771; Sethuraman & Raju,2012, p. 331). The lack of empirical insights can be attributed to thesecrecy surrounding the question of which manufacturer producesretailers' private labels, which makes it notoriously difficult to obtainthe required data (Sethuraman & Raju, 2012). We undertook a mas-sive data collection effort, covering detailed information on over400 manufacturer-category combinations and two retailers, to empir-ically test whether private-label production by national-brand manu-facturers indeed leads to retailer goodwill, while also considering thepotential drivers of private-label production.

We test our hypotheses in a discount setting, where we focus onthe prototypical hard discounter, Aldi, and on Europe's largest softdiscounter, Mercadona. Discounters are characterized by a limited as-sortment that is dominated by private labels, relatively small shop-ping areas, and very competitive prices.4 To offer lower prices, theyuse a simplified ‘no-frills’ store format. Hard discounters typicallyoffer fewer than 1400 SKUs in stores of approximately 1000 squaremeters. Soft discounters, in contrast, have a more extended range ofbetween 1400 and 7000 SKUs, which are sold in stores of approxi-mately 1500 square meters. In addition, the store environment issomewhat more attractive, and the proportion of national brands inthe assortment is larger compared to that of hard discounters (IGD,2002).

We chose the discount setting for two reasons. First, it has clearcontemporary value. Discounters are the fastest growing retail for-mat, with worldwide revenues that have been forecasted to increaseby approximately 60% over the next five years (PlanetRetail, 2010).Discounters de-emphasize national-brand offerings in their assort-ment, and they focus predominantly on private-label products. Dis-counters have only recently begun to allow a limited number ofnational brands into their assortment (Deleersnyder, Dekimpe,Steenkamp, & Koll, 2007).5 Given their impressive market sharegrowth, brand manufacturers have come to realize that they cannotafford the luxury not to conduct business with discounters (ThomsonReuters, 2009). This has resulted in a very competitive market setting,with numerous candidate brand manufacturers competing for a lim-ited number of spots (BusinessWeek, 2005). One way in which man-ufacturers hope to gain a discounter's goodwill, and thereby improvethe odds of obtaining shelf presence for their national brands, is byproducing private labels for the discounter.

Second, the discount setting allows us to test our ideas in a morecontrolled, natural experiment-like setting. While many alternativemanifestations of retailer goodwill (in addition to shelf presence)should be taken into account at traditional retailers such as reducedslotting allowances, lower promotional fees, a higher extent ofpass-through, and/or more promotional feature and display support(most of which also constitute data that are difficult to obtain), thisis not the case for discounters. Discounters typically do not chargeslotting allowances or promotional fees, and their format is char-acterized by very limited promotional activity (PlanetRetail, 2010).Because one form of retailer goodwill (e.g., shelf presence) is espe-cially plausible for brand manufacturers that work with discounters,the discount setting offers a cleaner and more controlled environmentfor studying the effects of private-label production by national-brandmanufacturers on retailer goodwill.

We develop a model that allows us to test whether brand manu-facturers that are involved in private-label production for a discount-er have a higher likelihood of procuring shelf presence for theirnational brands. The paper is organized as follows. First, we reviewthe limited literature on private-label production by brand manufac-turers. Then, we present our conceptual framework and hypotheses.Next, we discuss the research method, data, and empirical findings.The final section discusses the implications for researchers and man-agers and provides suggestions for further research.

2. Literature review

Despite “the richness of the phenomenon and the high level of man-agerial interest,” surprisingly little research has studied private-label(PL) production by national-brand (NB) manufacturers (Sethuraman,2009, p. 771). An initial set of papers has developed game-theorymodels to study the issue. Kumar, Radhakrishnan, and Rao (2010), forexample, consider the retailer's decision toworkwith either a dedicatedPL supplier or a dual brander. They focus on a two-level supply chain in-cluding a retailer, a NB manufacturer, and a dedicated (independent)manufacturer. Depending on whether the dedicated supplier or theNB manufacturer produces the retailer's PL products, a two-vendoror one-vendor regime emerges. Based on the size of the retailer'squality- versus price-sensitive customer segment, the retailer isshown to be better off with either the NB manufacturer or a dedicatedPL manufacturer. Retailers with a large price-sensitive customer seg-ment would not prefer a NB manufacturer to supply their PL products:NB manufacturers would lower PL quality (to preserve their ownsales), and this is only profitable for the retailer if the quality-sensitive(NB-prone) customer segment is large enough. Interestingly, Kumaret al. (2010, p. 156) note that their analysis does not apply to discountstores, as these are very focused on the low-end, highly price-sensitive,customer segment.

Other studies have approached the PL-production decision fromthe point of view of a manufacturer that has to decide whether to en-gage in PL production for a given retailer. Using a model of verticaldifferentiation, Gomez-Arias and Bello-Acebron (2008) derive that,depending on the PL's quality positioning, a NB manufacturer mightbe more or less willing to supply the PL product. Gomez-Arias andBello-Acebron distinguish between high-quality and low-quality NBmanufacturers. The pressure to produce PLs is highest when thePL enters the market at a quality level that is similar to themanufacturer's NB quality. A high-quality NB manufacturer thenwants to produce a high-quality PL to pre-empt competition and tobe able to position the PL such that it competes with other NBs andnot its own. Wu and Wang (2005), in turn, show analytically howPL production by one NB manufacturer can be used to mitigate pro-motional competition with other NB manufacturers. However, thisstudy may be less applicable in a typical EDLP-based discount settingin which promotions are largely absent to start with.

4 As such, they are distinct from large-scale, every-day-low-price (EDLP) retailerssuch as Wal-Mart in the U.S. and from large supermarkets as Carrefour and Tesco inEurope (Cleeren et al., 2010).

5 Aldi relied exclusively on its private labels for many years (Brandes, 2005).

344 A. ter Braak et al. / Intern. J. of Research in Marketing 30 (2013) 343–357

Moreover, none of the aforementioned studies provides empiricalsupport for their various contentions. This is not surprising, given thatmost manufacturers keep their involvement in PL production confi-dential (de Jong, 2007) out of fear for the impact that such knowledgemay have on their main brands (Gomez-Arias & Bello-Acebron, 2008)and on their other retail relationships. This situation makes empiricalresearch on the topic difficult, and it has prompted Sethuraman(2009, pp. 771–773) to call for more empirical research on both theantecedents and the consequences of dual branding (i.e., the practiceof NB manufacturers to also engage in PL production). More recently,Sethuraman and Raju (2012, p. 351) again concluded that “we need abetter understanding of why a manufacturer would supply private la-bels and why a retailer would accept the same.” In line with thesecalls, Chen, Narasimhan, John, and Dhar (2010) have estimated astructural model to derive the profit implications for various PL sup-ply arrangements, and they have shown through a number of policysimulations, in the context of a single market (fluid milk), how en-gagement in PL production may be beneficial to NB manufacturers.

Our analysis differs from the above-referenced studies in a num-ber of ways. First, while previous studies (e.g., Chen et al., 2010;Gomez-Arias & Bello-Acebron, 2008; Kumar et al., 2010) consideredthe (expected) profit consequences from various PL-supply schemes,we focus on the potential goodwill creation of a PL-production deci-sion, which has been recognized as a key strategic considerationin both the academic literature (Dunne & Narasimhan, 1999) and inthe business press (IGD, 2006). Second, we empirically examinewhy certain NB manufacturers are more likely to be involved in theproduction of a discounter's PLs than their competitors, and wepursue this question across a broad set of grocery categories andmanufacturers. We study this problem at two leading European dis-counters. To the best of our knowledge, no such large-scale empiricalresearch is available on the determinants of a manufacturer's engage-ment in PL production. Finally, several of the aforementioned analyticalmodels (e.g., Kumar et al., 2010; Wu & Wang, 2005) are less suited foruse in a (EDLP-based) discount context. We focus on discounters, andwe therefore contribute to the (thus far) limited empirical literature(Cleeren, Verboven, Dekimpe, & Gielens, 2010; Deleersnyder et al.,2007) on this fast-growing retail format. As such, our study also differssubstantially from ter Braak, Dekimpe, and Geyskens (2013), who studythe implications of dual branding on retail profit margins in the context

of a conventional supermarket, which follows a very different businessmodel than discount stores.

3. Conceptual framework and hypotheses

Fig. 1 summarizes our theorizing, both on the relationship be-tween PL production and shelf presence, and on the determinants ofa NB manufacturer's participation in such production.

3.1. The effect of PL production on shelf presence

PL production by a NB manufacturer may be seen by the discount-er as a pledge by the NB manufacturer. “Pledges” are actions that areundertaken by channel members that demonstrate good faith andthat bind channel members to the relationship (Anderson & Weitz,1992, p. 20). The offering of pledges functions as a signal of goodwilland invites reciprocal actions. By making the pledge of PL production,a NBmanufacturer effectively signals its cooperative behavior (Dunne& Narasimhan, 1999). This may lead to retailer goodwill, which canoriginate during the negotiation process and/or can become morepronounced as the relationship develops. This is consistent with thereciprocity literature (see, e.g., Uhl-Bien & Maslyn, 2003, p. 514),which has argued that the time span of reciprocation can rangefrom high immediacy (i.e., instantaneous) to low immediacy. Follow-ing this logic, PL production may represent a good opportunity for NBmanufacturers to enhance the likelihood of obtaining access to thediscounter's most valued, yet scarce, resource, i.e., shelf presence.Hence:

H1. PL production by a NB manufacturer increases the likelihood ofshelf presence.

3.2. The effect of NB manufacturer market power on shelf presence

When testing H1, we control for the intrinsic market power of theNB manufacturer. Following Shervani, Frazier, and Challagalla (2007),we definemanufacturer market power as amanufacturer's ability to in-fluence the actions of others in a product market. The industrial-organization literature has argued that a firm's market power is basedon itsmarket position, as reflected by itsmarket share. If a firm operates

Fig. 1. Conceptual framework.

345A. ter Braak et al. / Intern. J. of Research in Marketing 30 (2013) 343–357

in multiple product markets, its market power can vary considerablyacross them.6 Empirical research has shown that powerful firms areable to secure relatively high levels of influence on the behavior of relat-ed channel members (e.g., Anderson, Lodish, & Weitz, 1987; Shervaniet al., 2007; Sudhir & Rao, 2006). In a similar vein, we argue that morepowerful manufacturers – with a larger volume share in the categorythat is obtained frommore and stronger brands – are generally in a bet-ter position to secure shelf space. The addition of their NB is expected toimprove a discounter's perceived assortment quality and variety, asthese brands will stand out more against an otherwise PL-dominatedassortment (Deleersnyder et al., 2007). Because of their expected con-tribution to the discounter'sfinancial goals, discounterswill bemore re-ceptive to brands that are offered by powerful NB manufacturers. Wetherefore postulate:

H2. NB manufacturer market power increases the likelihood of shelfpresence.

Moreover, this variable may interact with PL production. The pos-itive effect of PL production on the likelihood of shelf presence isexpected to be more pronounced for less powerful NB manufacturers.Relationship quality has been shown to have a larger impact onnew-product acceptance when products are moderately attractivethan when they are very attractive (as when they are offered by themost successful manufacturer) (Kaufman, Jayachandran, & Rose,2006). In a similar vein, retailers have been found to prefer the lead-ing NB manufacturer as category captain unless smaller players can(or are willing to) offer a special service (Subramanian, Raju, Dhar,& Wang, 2010). As such, PL production can be used by smaller NBmanufacturers as a tool to overcome an inherent market-powerdisadvantage. In other words, when a NB manufacturer's marketpower is lower, the likelihood that a discounter will grant shelf pres-ence to that NB manufacturer increases considerably when the man-ufacturer establishes a strong relationship with the discounter (i.e., itproduces PL products for the discounter). When its market power isvery high, a NB manufacturer is already likely to be granted shelfpresence, regardless of PL production. Therefore, the latter has less in-cremental impact. Following this line of reasoning:

H3. NB manufacturer market power diminishes the effect of PLproduction in increasing the likelihood of shelf presence for the NBmanufacturer.

3.3. Drivers of PL production

We include three major drivers of PL production by a NB manufac-turer for a discounter, reflecting three categories of motives: the extentof sales growth (economic motive), the relative ease of producinghigh-quality products in the category (quality/positioning motive),and the marketing tools that are used to support the manufacturer'sbrands (brand-equity motive).

3.3.1. Manufacturer sales growthPL production often starts on an opportunistic basis when a man-

ufacturer has idle capacity (Gomez-Arias & Bello-Acebron, 2008;Kumar & Steenkamp, 2007), which may, for example, be caused byincreased competition from other manufacturers or by reduced de-mand during tough economic times (Sethuraman & Raju, 2012). APL order is then used to complement the lower NB sales. PL produc-tion can increase or maintain NB manufacturers' total sales volumesand thereby safeguard profits (Quelch & Harding, 1996; Soberman &Parker, 2006).

For a discounter, we could argue that (all else being equal) it mayalso prefer to work with a manufacturer with lower NB sales growth.Indeed, such manufacturers could be perceived as easier targetsthrough which to achieve good supply conditions, thereby increasingthe discounter's profitability.

We therefore hypothesize that higher sales growth decreases amanu-facturer's likelihood of engaging in PL production. Negative NB salesgrowth, in contrast, may lead to costly excess capacity for amanufacturerand hence increase a manufacturer's likelihood of producing PLs.

H4. Manufacturer sales growth decreases the likelihood of observingPL production by the manufacturer for the discounter.

3.3.2. Ease of producing high-quality productsA second key consideration for NB manufacturers is the possibility

of managing and influencing PL quality (Dunne, 1999; Kumar &Steenkamp, 2007).When it is difficult to produce high-quality productsin a category, discounterswill have a harder timematching the intrinsicquality of theNBs,which gives theNBs a competitive advantage. To pre-serve that advantage, NBmanufacturers will be reluctant to enter into aPL-production relationshipwith a discounter. By contrast,when it is rel-atively easy for discounters to match NB quality, NB manufacturers willbe more willing to engage in PL production, as there is no sustainablecompetitive advantage at stake. Gomez-Arias and Bello-Acebron(2008) showed that a NB manufacturer's incentive to produce PLs isparticularly high if the PL matches a manufacturer's NB quality, whichis more likely when the ease of producing high-quality products ishigh. NB manufacturers then have the ability to influence the position-ing of PLs (Gomez-Arias & Bello-Acebron, 2008; Kumar et al., 2010),for example by producing a PL that imitates a competitive NB to pre-serve their own sales (Sayman, Hoch, & Raju, 2002). Moreover, if theydo not produce a discounter's PL, a competitor is likely to do so, whichgives NB manufacturers an incentive to pre-empt the competition(Steenkamp & Dekimpe, 1997).

However, to come to actual PL production for a given discounter,both the manufacturer and the discounter must agree. Indeed, thelatter must decide whether to work with a dedicated PL manufacturer(which does not sell any NBs) or to select a dual brander. Dual brandersare said to have more innovative capacity (Kumar & Steenkamp, 2007)and to offermore quality assurance (Sethuraman, 2009) than dedicatedmanufacturers. Clearly, the second issue is especially relevant when it isdifficult to produce high-quality products. Hence, when it is difficult toproduce high-quality products, a discounter will more likely opt for aleading NB manufacturer for its PL production. In contrast, when pro-duction is relatively easy, a discounter may well opt for a dedicatedmanufacturer that is more price-focused (Kumar & Steenkamp, 2007).

Depending on the relative strength of both arguments, the ob-served choice will be driven more by manufacturer (positive associa-tion between ease of producing and PL production) or discounter(negative association between ease of producing and PL production)considerations. If the effect is positive, the manufacturer considerationsdominate; if it is negative, the discounter rationale prevails. In sum:

H5a/b. The higher the ease of producing high-quality products in thecategory, the higher/lower the likelihood of observing PL productionby a NB manufacturer for the discounter.

3.3.3. Manufacturer marketing toolsNumerous studies discuss how NB manufacturers can successfully

compete against PLs. According to Kumar and Steenkamp (2007),there are three distinct tools that are used by NB manufacturers tocreate winning value propositions for their NBs. First, NB manufac-turers may try to differentiate themselves from cheaper PL imitationsby conveying unique emotional benefits and signaling future demandthrough advertising (Desai, 2000; Steenkamp & Dekimpe, 1997).However, this carefully built-up image may be damaged if their

6 Power has been examined at different levels in the strategy literature. In contrastto a firm's overall power position within the manufacturer-discounter interfirm rela-tionship, our focus is on a firm's market power in a product market or product category(see Shervani et al., 2007 for a similar approach).


customer base discovers that they also produce PLs (Gomez & Benito,2008; Hoch, 1996), making NB manufacturers that rely heavily on ad-vertising, reluctant to engage in PL production. Second, charging ahefty price premium in the market over PLs is another way to positionthe NBs away from PLs (Ailawadi, Lehmann, & Neslin, 2003). Manu-facturers, again, may not be willing to risk this competitive advantageby producing a cheaper variant for the discounter. Finally, they canintroduce innovations and thereby offer products that are distinctfrom any existing PL. Maintaining this quality edge curbs PL growth,as it puts retailers in the position of imitating yesterday's favorites(Lamey, Deleersnyder, Steenkamp, & Dekimpe, 2012). NB manufac-turers that are heavily involved in developing new products will beless inclined to produce PLs, as it may put them in a position wherethe discounter can exert pressure to share the latest technologies(Dunne & Narasimhan, 1999), which would undermine their compet-itive advantage. Each of the marketing tools is therefore predicted todecrease the likelihood of PL production by the NB manufacturer forthe discounter. As indicated above, some manufacturers motivatetheir willingness to participate in the PL production process throughan “if you can't beat them, join them” reasoning (Sethuraman &Raju, 2012). The three above-stated arguments reflect this idea inthat they describe the conditions under which NB manufacturersare more/less able to withstand the impact of PL growth on theirown performance.

Given the secrecy surrounding the identity of the actual PL pro-ducers, discounters should not be concerned about the amount of ad-vertising support that is given by the NB manufacturer to its ownbrands nor about the price differential between the manufacturer'sNBs and the discounter's PLs when deciding on their PL sourcing.However, given that dual branders are thought to provide more inno-vative capacity when producing PLs (Kumar & Steenkamp, 2007), theinnovativeness of a NB manufacturer may play an important part inthe discounter's preference (i) for a dedicated supplier versus a dualbrander and (ii) if the latter option is selected, its preference forsome NB manufacturers over others. The more innovative a manufac-turer is, the higher a discounter's interest in that manufacturer as thepreferred option to produce its PLs. In sum:

H6. The more a NB manufacturer advertises, the lower the likelihoodof PL production by the manufacturer for the discounter.

H7. The higher the price premium a NB manufacturer charges over adiscounter's PLs, the lower the likelihood of PL production by themanufacturer for the discounter.

H8a/b. The more innovative a NB manufacturer is, the lower/higherthe likelihood of observing PL production by the manufacturer forthe discounter.

For hypothesis H8a/b, depending on the relative strength of botharguments, the observed choice will be driven more by manufacturer(negative association between a manufacturer's innovativeness and PLproduction) or discounter (positive association between amanufacturer'sinnovativeness and PL production) considerations.

3.3.4. Manufacturer market powerFinally, we also consider the effect of manufacturer market power

on the PL-production decision. We propose that manufacturer marketpower may have an impact on a NB manufacturer's likelihood of pro-ducing PLs. According to Dunne and Narasimhan (1999), producingPLs is especially interesting for small manufacturers of non-leadingNBs that seek to increase their sales volume. Alternatively, onecould argue that PLs have become so successful that NB manufac-turers that want to produce PLs for discounters need sufficient capac-ity (size) to handle the large volumes (IGD, 2005). Because goodarguments are available to support an increased likelihood of PL

production for lower/higher manufacturer market power, we offercompeting hypotheses:

H9a/b. NB manufacturer market power decreases/increases thelikelihood of observing PL production by the manufacturer for thediscounter.

4. Research setting and measures

4.1. Setting

Consumer packaged goods (CPG) companies regard Germany andSpain as two key European markets with respect to discounters. Theyare both among the largest consumer markets in Western Europe. Notonly does the discount format originate fromGermany, its currentmar-ket share in Germany already exceeds 40% (PlanetRetail, 2012b). Fur-ther, discounter share is rapidly increasing in Spain, where it sums upto close to 30% (PlanetRetail, 2012c).

In Germany, we study Aldi, the “mother of all discounters.”Aldi's market share in Germany was approximately 15% in 2012(PlanetRetail, 2012d). Aldi operates a total of over 8000 stores acrossEurope, in 16 countries. With current global sales of $80 billion, Aldiis ranked among the top ten grocery retailers in the world. AlthoughAldi's PL assortment differs across countries, it follows a centralized,national approach to its assortment management, with the sameoffering present in each outlet within a country (ACNielsen, 2007).7

Traditionally, Aldi did not carry any NBs in its assortment. However,it recently started accepting NBs on a limited basis. Although morethan 90% of Aldi's range still consists of PLs, recent NB additions inGermany include Snickers candy bars (Masterfoods), Del Montecanned fruit (Del Monte), and Quality Street sweets (Nestlé).

In Spain, we study the leading discounter Mercadona. Mercadonahas been the most successful store in the country in recent years. It isthe largest grocery retailer in Spain, operating approximately 1400outlets, and it was ranked the ninth most reputable company in theworld in 2009 by Forbes Magazine. The firm's strategy is built aroundconsistently offering value for money rather than short-term pricepromotions. Its sales in Spain have almost doubled over the last fiveyears, amounting to $24 billion (PlanetRetail, 2012c). Currently,Mercadona sells approximately 2000 PL products. Shelf space forNBs is limited, as Mercadona tries to achieve a 50–50% mix betweenPLs and popular brands (Bain & Company, 2008), with typically justone or two NBs facing each PL (Fernandez Nogales & Gomez Suarez,2005). As a result, 30% fewer product varieties tend to be availableat Mercadona than can be found in traditional supermarkets (Bain &Company, 2008). While Mercadona can be classified as a soft dis-counter, Aldi is a prototypical hard discounter.

Clearly, NB shelf space is a scarce resource for both discounters,which is sought after by many NB manufacturers. Interestingly, dis-counters typically do not charge slotting allowances or promotionalfees (PlanetRetail, 2010). Hence, these issues do not enter channel(goodwill) negotiations. By considering both Europe's leading hard(Aldi) and soft (Mercadona) discounter, we can infer to what extentour findings differ (or are common) across both format types.

4.2. Sample and measures

Our unit of analysis is the NB manufacturer in a certain category atthe discounter. We only consider categories that contained at leastone NB and thus, where the discounter made the deliberate decisionto extend its assortment with a NB offering. For Aldi, we collecteddata on 37 grocery categories. For each category, we determined thetop-five NB manufacturers in terms of their 2005 category purchasesin Germany, based on panel data from GfK. For Mercadona, we

7 This feature is important when collecting data through store visits (cf. infra).


obtained data on 53 categories in which at least one NB was availablefrom Kantar Worldpanel. We retained the top-five NB manufacturersin terms of their 2009 category sales volume in Spain.8 To illustratethe range of products that are available in our dataset, we groupedthe categories into broader product groups. Table 1 shows thesegroups, along with some illustrative examples of the included catego-ries. For each of the resulting 450 manufacturer-category combina-tions (the top-five NB manufacturers for 90 categories, with 37*5observations for Aldi and 53*5 observations for Mercadona), we com-bined a wide variety of data sources to operationalize our variables.

4.2.1. Shelf presenceWe measure shelf presence on the basis of GfK/Kantar consumer

panel data for Germany and Spain, which cover all of the grocery pur-chases that were made by a representative national sample of20,000+ German and 8000+ Spanish households. Shelf presencecaptures whether a NB that was owned by one of the top-five NBmanufacturers in a category was available in that category at agiven discounter (1 = yes, 0 = no). For Aldi, this information wasavailable between January of 2002 and June of 2008. For the 185(37*5) manufacturer-category combinations, we checked whether theNBmanufacturers ownaNB forwhich a purchase record in the categoryat Aldi could be found during the time frame of our data. We observedshelf presence for 46 (out of 185)manufacturer-category combinations.For Mercadona, shelf presence was measured in the year 2009. In thiscase, shelf presence was observed for 139 (out of 265) manufacturer-category combinations. Proportionally, more instances of shelf presencefor the top-five NB manufacturers are seen with Mercadona, which isconsistent with its positioning as a soft discounter. Table 2 providesmore details on the distribution of the number of top-five NB manufac-turers with shelf presence at the discounter across categories.

4.2.2. PL productionWe measure “PL production” as a binary variable that is coded 1 if

the NB manufacturer engaged in PL production for the discounter inthe category and 0 otherwise. In spite of the secrecy surroundingAldi (Brandes, 2005) and the reluctance to acknowledge PL produc-tion by NB manufacturers (de Jong, 2007), we were able to obtain in-formation on PL production for Aldi through extensive field research.Between January of 2002 and June of 2008, approximately 650 PLswere sold in the 37 categories that were under investigation at Aldi.For each of those 650 PLs, we determined the producer. Becausethese data were not readily available through conventional channels,various sources were consulted. As a starting point, we obtained four

books with information on the manufacturers of 200 popular PLsthat are sold at Aldi (Bertram, 2006; Schaab & Eschenbek, 2008;Schneider, 2005, 2006). For the remaining 400+ products, we repli-cated the procedure described in these books to uncover the manu-facturers of all of the PLs in the 37 categories that were studied. Tothat extent, we exploited the fact that Aldi is one of the rare retailersthat prints the address of the manufacturer on its packages (de Jong,2007). To extend and/or validate this information, we also consultedvarious websites that are devoted to consumer product reviews, com-pany profiles, and/or discounter products (including www.ciao.de,www.yopi.de, www.discounter-archiv.de, and www.wer-zu-wem.de). Specifically, for each PL, we recorded (through numerous storevisits and online searches) the manufacturer's address. We thenchecked whether the manufacturer's address matched the addressof one of the top-five NB manufacturers in the category. In case ofno address match, we looked up the name of the manufacturer thatwas located at the stated address and assessed whether it was a subsid-iary of one of the top-five NBmanufacturers. In this way, we account forsituations inwhich NBmanufacturers may produce PLs for a discounterat a different address or location and/or at a subsidiary with a differentfirm name. For 68 (out of 185) manufacturer-category combinations,we found evidence of PL production for Aldi (see Table 3, panel A). For63 (93%) of those manufacturer-category combinations, more thanone data source was found confirming PL production activity. In 32(or 86%) of the 37 categories that were examined at Aldi, at least oneof the top-five NB manufacturers was involved in the production ofAldi's PLs.9 Examples for Aldi include Campina (yogurt), Dr. Oetker(dessert), and Bonduelle (canned vegetables).

For Mercadona, we were fortunate to obtain, with the help ofKantar Worldpanel, internal category-level PL producer informationfor the year 2009. In this case, we again implemented an extensiveonline search to detect any less obvious dependence between thelisted producer and the 53*5 NB manufacturers in our sample. For24 (out of 265) manufacturer-category combinations, we found evi-dence of PL production for Mercadona (see Table 3, panel B). In 23(or 43%) of the 53 categories that were examined at Mercadona, atleast one of the top-five NB manufacturers was involved in the pro-duction of Mercadona's PLs.10 For reasons of confidentiality, noMercadona examples can be revealed. A simple bivariate !2 test pro-vides initial evidence of a relationship between PL production andshelf presence (!2(1) = 21.33, p b .01 for Aldi and !2(1) = 16.27,

8 For Aldi, PL production data covered the 2002 to 2008 period (cf. infra) — we de-termined the top-five NB manufacturers for the midpoint of this time frame (2005).Because PL production data for Mercadona pertained to 2009, we determined thetop-five NB manufacturers for 2009. No PL production data on preceding years wereavailable for Mercadona.

Table 1Category coverage.

Product group Examples

Assorted foods Noodles, jam, riceBeverages Beer, mineral water, juiceCandy Candy bars, chocolate, bonbonsCanned/bottled foods Fish, beans, tomatoCare products Shampoo, diapers, toilet tissueCleaning products Bleach, detergent, fabric conditionerCooking fats Butter, olive oil, margarineDairy products Milk, yogurt, ice creamHousehold supplies Basket bin bags, toilet tablets, cellulosesInstant meals Ready desserts, ready meals, saladPastry Cakes, sweet biscuitsPet products Wet dog food, dry dog foodTaste enhancers Ketchup, salad dressing, mayonnaise

Table 2Distribution of number of top-5 NB manufacturers with shelf presence at thediscounter.

Number of top-5 NBmanufacturers withshelf presence inthe category

% ofcategoriesat Aldi

Representativeexamples

% of categoriesat Mercadona

Representativeexamples

0 38% Jam, ketchup 2% Honey1 27% Yogurt, tinned

soup21% Instant coffee,

basket bin bags2 14% Cream, canned

vegetables30% Bleach, tea

3 16% Butter, cannedfish

19% Canned tuna,pasta

4 5% Sweet biscuits,chocolate

17% Deodorant,fruit juice

5 0% n.a. 11% Biscuits, cereals

Note: n.a. = not applicable.

9 In the other five categories, where none of the top-five players produced any of thediscounter's PLs, we found evidence of PL production by a non-top-five NBmanufacturer.10 In 12 out of the 30 remaining categories, Mercadona chose a non-top-five NB man-ufacturer to produce its PLs.


p b .01 for Mercadona), a finding that is validated in our subsequentmultivariate analyses.

4.3. Predictors

We measure the covariates for Aldi in the year preceding shelfpresence. When no brand of the NB manufacturer was listed at Aldi,we use 2005 as the base year for the covariates. For Mercadona, wemeasure the covariates in the year 2008 (i.e., the year before we ob-serve shelf presence). For a complete overview of the timing, source,and operationalization of our variables, we refer to Table 4.

4.3.1. Manufacturer sales growthManufacturer sales growth is measured as themaximum growth in a

manufacturer'sNB volume sales across three consecutive years in the cat-egory and obtained from the panel data. Whereas positive sales growthsignals sufficient NB demand and perhaps the need for investment in ad-ditional production capacity, negative sales growth signals idle produc-tion capacity (Lieberman, 1987). We also specify sales growth as theaverage growth in a manufacturer's NB volume sales in the categoryacross the three consecutive years. Our results remain the same.

4.3.2. Ease of producing high-quality productsWe measure the perceived ease of producing high-quality prod-

ucts through a 5-point reverse-scored survey item: “In the categoryXXX, making good quality products is difficult.” We combined thiscountry-specific information (i.e., German consumers for the Aldi ob-servations and Spanish consumers for the Mercadona observations)from regular users of the category (each consumer could provide in-formation on at most four categories),11 with information from fourindustry experts (who had to rate all of the categories).12 Both groupshad similar assessments, as evidenced in a highly significant correla-tion (r = .41, p b .01) between their respective averages.

4.3.3. Manufacturer marketing toolsWe measure the extent of “advertising” by means of the annual

national advertising expenditures (in €) in the category by the NB

manufacturer.13 The German advertising data were obtained fromThomson Media Control, a provider of market data on advertisingspending across all media in Germany. The Spanish advertising datawere obtained from InfoAdex, a company that measures advertisinginvestments across various media in Spain. Second, a manufacturer's“price premium over PLs”was obtained from the panel data and mea-sured as the market-share-weighted average national unit price of allof a manufacturer's NBs in the category compared to the market-share-weighted average unit price of the discounter's PLs in the cate-gory (Ailawadi et al., 2003). Finally, to measure a NB manufacturer'sinnovativeness, we counted the number of “innovations” that werelaunched in the category by the manufacturer in the German market(for the Aldi observations) and the Spanish market (for the observa-tions from Mercadona) using data from Product Launch Analytics.Product Launch Analytics, formerly known as Productscan, is asubscription-based database that tracks CPG introductions (see, e.g.,Sorescu & Spanjol, 2008, for an in-depth discussion on this datasource). The vast majority of the NB manufacturer-category combina-tions in our sample are characterized by either no innovation (62% ofthe sample) or one innovation (23%) in the relevant category peryear. Only 15% of the sample had more than one innovation peryear in a given category.

4.3.4. Manufacturer market powerWe measure manufacturer market power as the national “catego-

ry share” of all of a NB manufacturer's brands in terms of volume,using the panel data (Gielens & Steenkamp, 2007). Given that we in-teract this variable with PL production, we grand-mean-center it tofacilitate interpretation (Irwin & McClelland, 2001).14

4.3.5. Control variablesApart from the above-stated variables, we also control for the differ-

ences between the two discounters by including a discounter dummy.Moreover, we control for PL production by the NB manufacturer forthe discounter beyond the focal category (1 = yes, 0 = no PL produc-tion is observed in another category in our sample at that discounter).We expect a NB manufacturer to be more likely to produce PLs in thefocal category if it also does so in one of the other categories. Moreover,we allow for PL production by the NB manufacturer in one category toalso affect NB shelf presence in another category. On the one hand,one could expect that producing a PL in one category would also facili-tate shelf presence in another category. On the other hand, as retailersembrace category management practices (e.g., Basuroy, Mantrala, &Walters, 2001), the pledge of PL production may not lead to goodwillin another category (for which a different category manager is respon-sible) or only to a lesser extent. Perhaps, if PL production in a category isalready rewarded in that same category, it could even lead to a lowerlikelihood of shelf presence in another category.

In addition, we also control for the relative strength of the specificdiscounter in a given category. This factor may affect a discounter'swillingness to accept stronger, or to prefer weaker, NB opponents inits assortment. Moreover, it may affect a manufacturer's willingnessto produce PLs for the discounter in that category. To control forthe relative strength of the discounter, we include the discounter's(volume) share in the total (national) sales in the category. Giventhat we want to control for the discounter's relative strength in agiven category (compared to the other categories), we center this co-variate within the discounter.

11 Consumer perceptions on the difficulty of producing high-quality products in thecategory were also used in Aaker and Keller (1990) and Steenkamp et al. (2010),among others.12 Specifically, participants include the chairman of IPLC (a consultancy firm that spe-cializes in international PL production issues), the director Strategic Initiatives fromGfK Benelux, the director of the EFMI Business School (a knowledge center for the foodsector), and a knowledgeable practitioner with substantial experience working on thePL supply side.

13 The German advertising data were only available from 2005 onwards, so for someof the Aldi observations, we could not use the year before the NB listing for this covar-iate. Instead, we used the year 2005 for all of the observations. However, in the years2005 to 2007, the average correlation between the NB manufacturers' advertisingspending in two consecutive years was N .95.14 We also used mean-centering within a discounter, and all of the findings remainedperfectly stable.

Table 3Relation between PL production and NB shelf presence.

Panel A: ALDI

NB shelf presence

PL production Yes No TotalYes 30 38 68No 16 101 117Total 46 139 185

Note: Each cell presents the number of manufacturer-category combinations;!2(1) =21.33 (p b .01).

Panel B: MERCADONA

NB shelf presence

PL production Yes No TotalYes 22 2 24No 117 124 241Total 139 126 265

Note: Each cell presents the number of manufacturer-category combinations; !2(1) =16.27 (p b .01).


Finally, we include two product-class dummies (one to indicatehousehold care and personal care, and one to indicate beverages, withfood products as the base-level category) to control for unobserved cat-egory effects (see, e.g., Lamey et al., 2012 or Steenkamp, van Heerde, &Geyskens, 2010 for a similar practice). The covariates are not the focusof our study, but controlling for them provides a stronger test of our hy-potheses (Greene, 2000).

We conducted formal tests to assess the skewness of all of the con-tinuous covariates in our model. Following Tabachnick and Fidell(1996), we determined whether the skewness statistic exceededtwice the standard error of the statistic. Based on these tests, all ofthe continuous variables were found to be skewed and were log-transformed.15 None of the VIF statistics exceeded 3, suggesting thatmulticollinearity is not an issue (Cohen, Cohen, West, & Aiken, 2003).

5. Model

PL production is not a purely exogenous variable, but rather theresult of strategic considerations from both the NB manufacturer'sand the discounter's side. A failure to statistically correct for thisendogeneity may not only lead to biased estimates but also to faultyconclusions about our key propositions (Hamilton & Nickerson,2003). If manufacturers that produce PLs self-select themselves and/or are selected by the discounter on the basis of unobservable charac-teristics that affect both shelf presence and PL production, a problemof selection bias may arise. Selection bias can be corrected throughthe traditional two-step estimation technique that was proposed byHeckman (1979) or, alternatively, by using a Maximum Likelihood(ML) estimation on a system of equations. Because ML estimates aremore efficient and have smaller standard errors than the two-stageestimates (Breen, 1996), we opt for that procedure.

To allow for the potential inter-correlation among observations(within a discounter) from multiple manufacturers within the same

category and among observations from the samemanufacturer acrossmultiple categories, we use a robust two-way clustered-error termestimation on the system of equations (Cameron, Gelbach, & Miller,2011).

Preliminary pooling tests failed to reject the assumption of homoge-neity across both discounters (p N .05) for all but four effects: theimpact of (i) manufacturer market power and (ii) its interaction effectwith PL production on shelf presence, and the impact of (iii) pricepremium and (iv) PL production in a non-focal category on thePL-production decision. Hence, we pool the data across the two dis-counters, but we allow for discounter-specific effects in these four in-stances (along with a discounter-specific fixed-effects correction).

We estimate the following system of equations, where the firstand the second equation are referred to as the outcome equationand the strategy-selection equation, respectively:

PRESENCEijd ! β0 " β00 D MERCd " β1 PLPRODijd " β2 M POWERijd # D ALDId

! "

"β20 M POWERijd # D MERCd

! "

"β3 PLPRODijd #M POWERijd # D ALDId! "

"β30 PLPRODijd #M POWERijd # D MERCd

! "" ΓZijd " εijd

$1%

PLPRODijd ! δ0 " δ00D MERCd " δ1 GROWTHijd " δ2 EASEjd " δ3 ADVijd

"δ4 PPREMijd # D ALDId! "

" δ40 PPREMijd # D MERCd

! "

"δ5 INNOVijd " δ6 M POWERijd "ΩZijd " εijd0

$2%

with D_MERCd as a discounter-specific fixed effect and Zijd as a vectorthat contains all of the other control variables (PL production in anon-focal category, discounter category share, and two product-classdummies), and Γ and Ω the corresponding vector of coefficients in theoutcome and strategy-selection equation, respectively.

In the outcome equation, PRESENCEijd captures whether NB man-ufacturer i obtained shelf presence in category j with discounter d.PLPRODijd reflects whether NBmanufacturer i produces PLs in category

15 Because innovations could have a value of zero, the transformation ln (innova-tions + 1) was used. A similar transformation was used for advertising expenditures.

Table 4Overview variables.

Variable Operationalisation Sources Timing

Aldi Mercadona Aldi Mercadona

Shelf presence Is a NB of the NB manufacturer present on thediscounter shelves in the category? yes/no

GfK panel Kantar Worldpanel 2002–2008 2009

PL production Does the NB manufacturer engage in PL production forthe discounter in the focal category? yes/no

Various (see Section 4.2.2) Kantar Worldpanel 2002–2008 2009

Manufacturer sales growth Maximum [volume sales growth of the NBmanufacturer in the category in year t/year t ! 1,volume sales growth of the NB manufacturer in thecategory in year t ! 1/year t ! 2]

GfK panel Kantar Worldpanel 2003–2004 and2004–2005 or3 years beforeNB shelf presence

2007–2008 and2008–2009

Ease of producinghigh-quality products

5-point reverse-scored survey item: “In the categoryXXX, making good quality products is difficult”

- Consumers: Steenkamp et al. (2010)- Experts: primary data

– –

Advertising Total national advertising expenditures of the NBmanufacturer in the category (in €)

Thomson Media Control InfoAdex 2005 2008

Price premium Market-share-weighted average national unit price ofall NBs of the NB manufacturer in the category/market-share-weighted average PL unit price of thediscounter in the category

GfK panel Kantar Worldpanel 2005 or the yearbefore NB shelfpresence

2008

Innovations Number of innovations of the NB manufacturer in thecategory

Product Launch Analytics 2005 or the yearbefore NB shelfpresence

2008

Manufacturer market power Volume sales of all NBs of the NB manufacturer in thecategory/national category volume sales


2008

Discounter category share Volume sales of the discounter in the category/nationalcategory volume sales


2008

PL production in anon-focal category

Does the NB manufacturer engage in PL production forthe discounter in a different category in our samplethan the focal category? yes/no

Various (see Section 4.2.2) Kantar Worldpanel 2002–2008 2009


j for discounter d. Moreover, this equation includes discounter-specificparameters for both the main effect (β2 and β2′) of a manufacturer'smarket power in the country of discounter d (M_POWERijd) and for itsinteraction (β3 and β3′) with PL production.

The strategy-selection equation includes the observable drivers of PLproduction by a NB manufacturer, namely a manufacturer's categorysales growth (GROWTHijd) in the country of discounter d, the perceivedease of producing high-quality products in the category (EASEjd) in thatcountry, and the marketing tools that were used by a NB manufacturerto position its NBs in each country: a manufacturer's advertising(ADVijd), price premium (PPREMijd), and innovations (INNOVijd) inthe category. Based on the preliminary pooling tests, this equation hasdiscounter-specific parameters for a manufacturer's price premium(δ4 and δ4′) and its PL production in a non-focal category. We further in-clude a manufacturer's market power (M_POWERijd).

The strategy-selection equation takes the form of a probit model.To reflect the dichotomous nature of shelf presence, the outcomeequation is also estimated through a probit specification, so that thesystem of equations becomes a bivariate probit model. This specifica-tion is often referred to as a recursive bivariate probit model with thetreatment variable (in our case, the PL production) as an endogenousregressor in the outcome equation (see, e.g., Guo & Fraser, 2010;Jones, 2007). To allow for the possibility that unobserved characteris-tics drive both shelf presence and PL production, no restrictions areimposed on the correlation (ρ) between εijd and εijd′ . Therefore, ρcan take on any value between !1 and 1. A significant ρ underscoresthe importance of explicitly considering selection. A significant posi-tive (negative) ρ implies that the PL-production effect that is estimat-ed without the correction would be biased upward (downward)(Briggs, 2004). When ρ is not significant, there is no selection prob-lem (Breen, 1996).

6. Estimation results

The significance of the error correlation ρ indicates that un-observable variables exist that affect the dependent variable inboth of the equations and thus that a selection correction is needed(ρ = ! .87, p b .01). The negative sign of ρ indicates that thePL-production coefficient without the correction would be biaseddownward and that unobserved factors that make selection more(less) likely, tend to have an opposite effect in the outcome equation.The negative sign therefore shows that, on average, unobserved factorsthat make a NB manufacturer intrinsically less likely to produce PLs in-crease the likelihood that the discounter will reciprocate by grantingshelf presence if this production nevertheless takes place. Under suchcircumstances, PL production is seen as a larger pledge from the manu-facturer, whichmakes itmore likely to be rewarded by the discounter inreturn. In the context of our study, an unobserved factor that mayhave such opposite effects is brand quality: manufacturers that ownhigh-quality brands may be less interested in producing PLs, but thesebrands are intrinsically more valuable to a discounter to have in its as-sortment. However, we can only speculate on the potential causes of ρas these factors are by nature unmeasured.16 Table 5 reports the param-eter estimates for the outcome and strategy-selection equation. In bothinstances, a satisfactory hit rate is obtained, as the model is able to cor-rectly identify PL production in 85% and shelf presence in 73% of all ofthe manufacturer-category combinations. In case our hypotheses aredirectional, we report one-sided significance levels.

6.1. Outcome equation estimates

Turning first to the outcome equation, we find that PL productionby a NB manufacturer for a discounter leads, at the average level ofmanufacturer market power, to a significantly higher likelihood ofobtaining shelf presence in the category at that discounter (β1 =2.26, p b .01). Hence, H1 is supported. Also when a manufacturer al-ready produces PLs in another category, its acceptance probability inthe focal category remains increased (β1 + γ1 = 1.53, p b .01), albeitto a lesser extent. Hence, while production in the focal category stillcreates additional goodwill, the incremental effect if the manufactur-er already did so in another category is smaller (as evidenced by thenegative γ1), as there is less signaling value to this additional pledge.Similarly, when the manufacturer already produces in another cate-gory, a refusal to do so in the focal category sends quite a negative sig-nal to the discounter, and it jeopardizes the probability of gainingshelf access from that discounter.

As for H2 and H3, no significant effect of manufacturer marketpower is found for Aldi, neither in the produce condition (β2 + β3 =.04, p N .10) nor in the no-produce condition (β2 = ! .07, p N .10).For Mercadona, in contrast, we do find a significant effect of manufac-turer power in the no-produce condition (β2′= .51, p b .01). In thecase of PL production, however, NB manufacturer power no longer hasa significant effect on the likelihood of obtaining shelf presence withMercadona (β2′ + β3′ = ! .10, p N .10). Hence, manufacturer marketpower diminishes the effect of PL production on the likelihood of NBshelf presence, supporting H3. As for the control variables, we alsofind a significant positive discounter-specific fixed effect forMercadona(β0′ = 1.21, p b .01). This result is in linewith the fact thatMercadona isa soft discounter, and thus it is more inclined to accept NBs than Aldi,which is a hard discounter. Finally, we find no effect of discounter cate-gory share or of the product-class dummies on shelf presence.

6.2. Strategy-selection equation estimates

We find evidence that a NB manufacturer's sales growth negativelyaffects the PL production decision (δ1 = ! .37, p b .05). Hence, we cansupport H4. Ease of producing high-quality products in a given categoryalso has an impact. The easier it is to produce high-quality products, thehigher the likelihood of observing PL production by a NB manufacturerfor the discounter (δ2 = 2.69, p b .10). As such, H5a is supported. Thisfinding provides support for the manufacturer's rationale, as opposedto the discounter's rationale.

Concerning the marketing tools that are used by the NBmanufactur-er, we find that advertising expenditures have the expected negative ef-fect. The likelihood that a NB manufacturer engages in PL productiondecreases as it spends more on advertising (δ3 = ! .02, p b .10). Thus,H6 is supported. As for H7, a higher price premium over hard discounterAldi's PLs reduces a manufacturer's probability of producing PLs(δ4 = ! .80, p b .01). For soft discounterMercadona, the price premiumdoes not have an effect. Concerning H8, we find that the innovativenessof a NB manufacturer on the likelihood of PL production has a negativeeffect (δ5 = ! .18, p b .10). As such, we find that the manufacturer's ra-tionale for PL production prevails over the discounter's rationale.

To obtain insights into the relative importance of the three mo-tives of PL production by a NB manufacturer (economic, quality/posi-tioning, and brand-equity), we conduct a “what-if” analysis (seeInman, Winer, & Ferraro, 2009 for a similar practice) that examinesthe impact on a NB manufacturer's PL-production propensity whenincreasing each of the variables by one standard deviation above itsmean,17 while the other variables are held fixed at their baselinelevels (i.e., the mean for continuous variables and zero for dummyvariables). Among the marketing-mix variables (the brand-equity16 While there is a general agreement on the importance of controlling for sample se-

lection issues, some authors (see e.g., Lemke & Reed, 2001; Guo & Fraser, 2010) havewarned against over-interpreting the actual size of the correction parameter (in oursetting, the correlation ρ).

17 Because of the skewness of the data, we use the mean and the standard deviationof the log-transformed variables.


motive), NB manufacturers dealing with Aldi are primarily concernedwhen a price premium is at stake (a one standard deviation increasein price premium reduces their propensity to produce with 11.8%, i.e.,from 24.2% to 12.4%). Advertising and innovation-related concernshave a roughly equivalent, yet smaller, effect (advertising:!4.4%; in-novation: !3.3%). Marketing mix concerns are much less of an issuewhen dealing with soft discounter Mercadona: the price-premiumargument becomes insignificant, while the impact of advertising(!1.7%) and innovation (!1.3%) becomes substantially smaller.Sales growth (the economic motive) and ease of production (thequality/positioning motive) have an impact of similar magnitude,yet it is again more pronounced for Aldi than it is for Mercadona(Aldi: !7.0% and +5.4%; Mercadona: !2.6% and +2.4%).

The effect of manufacturer market power provides some addition-al insights: the likelihood of observing that a NB manufacturer pro-duces PLs increases when the NB manufacturer's category shareincreases (δ6 = .21, p b .10). Hence, H9b is supported.18 This findingprovides support for the discounter's rationale (see Table 6 for a sum-mary of the arguments and the corresponding empirical evidence).

For the control variables, we observe that PL production byleading NB manufacturers is more likely for Aldi than for Mercadonaδ00 ! !1:16;p b :01# $

. Second, we find that PL production in one

category increases the likelihood of producing in another. The effect issignificantly positive for both Aldi (ω1 = .83, p b .05) and Mercadonaω1

0 ! 2:69; p b :01# $

, although it differs inmagnitude. Finally, the covar-iate that is related to discounter category share shows that the likeli-hood that a NB manufacturer will engage in PL production in a certaincategory increases when the discounter has a higher share of the mar-ket in that category (ω2 = .99, p b .01), as there is more at stake forthe manufacturer in its efforts to obtain shelf presence.19

6.3. Robustness checks

Several tests were performed to substantiate the robustness of ourfindings.

6.3.1. Brand versus manufacturer powerConsistent with our conceptual framework, we measured a firm's

market power by its total market share in a given category. However,a manufacturer's high market power can be derived through multiplemoderately strong brands or through one very popular main brand.Retailers may be especially interested in listing the latter. We there-fore re-estimated our model using brand rather than manufacturerpower, with brand power defined as the volume share of themanufacturer's largest brand.20 All of the results remain substantivelythe same.

6.3.2. Absolute versus relative measuresIn line with Ailawadi and Harlam (2004) and Slotegraaf and

Pauwels (2008), we used absolute measures to operationalize the ex-tent of a manufacturer's advertising and innovation activity. Howev-er, some authors (see, e.g., Lamey et al., 2012; Reibstein & Wittink,2005) have advocated the use of relative measures, given the impor-tance of knowing how a manufacturer performs relative to competi-tors that operate under the same economic conditions. Therefore,we re-estimated our model, measuring the extent of a manufacturer'sadvertising and innovations relative to the total level of all five lead-ing NB manufacturers in that category. Our findings remain stable.The only exception is that the effect of the extent of advertising on

18 We also examined whether the non-directional effects of ease of producing high-quality products and innovations are dependent on the level of manufacturer marketpower. Specifically, we added the interaction terms to Equation 2 (again taking into ac-count whether these interactions could be pooled based on preliminary pooling tests).The interactions were not significant, neither individually nor jointly (p N .10).

19 We also tested for an interaction effect of PL production and discounter categoryshare in the outcome equation. However, this effect was not found to be significant(p = .33). Moreover, all of the other results remained stable. For ease of interpretationand because we had no theoretical foundation for the interaction with discounter cat-egory share, we report the model with just the interaction of PL production and man-ufacturer market power.20 Even though the correlation between brand and manufacturer power equals .89,we find some noticeable differences in power for some NB manufacturers. For exam-ple, while Campina's largest brand has a market share of 3.5% in the German yogurtcategory, its total share in the category amounts to 7.3%. Its competitor, Ehrmann, incontrast, has a similar market share (7.9%), but with only one brand.

Table 5Parameter estimates for the shelf-presence and PL-production equations.

Variable Coefficient (t-value)

Outcome equation (shelf presence)Intercept (β0) !1.28 ††† (!6.63)PL production (β1) 2.26 *** (11.25)Manufacturer market power Aldi (β2) ! .07 (! .46)Manufacturer market power Mercadona (β′2) .51 *** (5.86)PL production * Manufacturer market power Aldi (β3) .11 (.56)PL production * Manufacturer market powerMercadona (β′3)

! .61 *** (!4.97)

Control variablesDiscounter dummy (0 = Aldi, 1 = Mercadona) (β′0) 1.21 ††† (5.93)PL production in a non-focal category (γ1) ! .73 †† (!2.48)Discounter category share (γ2) .05 (.23)Beverages dummy (food = baseline) (γ3) ! .02 (! .13)Non-food dummy (food = baseline) (γ4) .06 (.25)

Strategy-selection equation (PL production)Intercept (δ0) !3.35 † (!1.74)Manufacturer sales growth (δ1) ! .37 ** (!1.91)Ease of producing high-quality products (δ2) 2.69 † (1.70)Manufacturer marketing tools

Advertising (δ3) ! .02 * (!1.43)Price premium over PLs Aldi (δ4) ! .80 *** (!3.45)Price premium over PLs Mercadona (δ′4) .19 (1.64)Innovations (δ5) ! .18 † (!1.67)

Manufacturer market power (δ6) .21 † (1.82)Control variables

Discounter dummy (0 = Aldi, 1 = Mercadona) (δ′0) !1.16 ††† (!5.07)PL production in a non-focal category Aldi (ω1) .83 †† (2.22)PL production in a non-focal categoryMercadona (ω′

1) 2.69 ††† (6.30)Discounter category share (ω2) .99 ††† (2.78)Beverages dummy (food = baseline) (ω3) ! .37 (!1.32)Non-food dummy (food = baseline) (ω4) ! .27 (!1.09)

Selection parameter (") ! .87 ††† (!2.62)Log-likelihood !377.63

*** p b .01, ** p b .05, * p b .10 (one-sided). ††† p b .01, †† p b .05, † p b .10 (two-sided).Significant effects are indicated in bold.

Table 6Manufacturer versus discounter rationale for PL production.

Manufacturer'srationale

Discounter'srationale

Empiricalfindings

Aldi Mercadona

Manufacturer sales growth (H4) ! ! ! !Ease of producing high-qualityproducts (H5)

+ ! + +

Manufacturer marketing toolsAdvertising (H6) ! ! !Price premium over PLs (H7) ! ! nsInnovations (H8) ! + ! !

Manufacturermarket power (H9) ! + + +

Note: For the manufacturer, a + (!) signifies the manufacturer is more (less) willing toproduce PLs for the discounter. For the discounter, a + (!) signifies the discounter ismore (less) interested to work with a particular leading NB manufacturer to produce itsPLs. ‘ns’ means that no significant effect was observed.


PL production, while similar in sign and magnitude, fails to reach sig-nificance (p = .19).

6.3.3. Timing of PL production and shelf presenceIn our theorizing and empirical analysis, we implicitly assumed

that PL production influences shelf presence (rather than the otherway around). Hence, PL production should precede (or coincidewith) NB shelf presence. Given the aforementioned secrecy surround-ing PL production, it was impossible to identify the exact starting dateof that activity. However, the temporal ordering of both events is notan issue for 398 (out of 450) manufacturer-category combinations(i.e., observations), as these NB manufacturers did not engage in PLproduction and/or did not obtain shelf presence in the category. Ofthe remaining 52 (12%) observations, where both PL productionand shelf presence occurred, 30 (22) observations involved Aldi(Mercadona). For the 30 Aldi observations, we adopted an extensiveprocedure to eliminate any ambiguity surrounding the temporalordering between both events as much as possible. For the 22Mercadona observations, however, given prior agreements with ourdata provider that we would not contact the manufacturers, thiswas not possible. By comparing the date of the first PL-production doc-umentation for Aldi (based on the above-listed sources, among whichare Bertram, 2006; Schaab & Eschenbek, 2008; and Schneider, 2005,2006) with the date that the NB manufacturers managed to obtainnew shelf presence for at least one of their brands in the category(obtained from the GfK data), we observed that for ten of the observa-tions, PL production unambiguously preceded the listing of a brandwithin the category. For the remaining 20 observations, we contactedthe manufacturers. For seven observations, we could deduce from themanufacturers' information that PL production definitely precededshelf presence within the category, while in two instances, the reverseorder was confirmed. For the remaining 11 observations, the manufac-turers refused to reveal the required information.

To test the robustness of our insights to this timing issue, wedropped the two observations for which we obtained the confirmationof a reverse order. This action did not affect any of our substantive in-sights, with the only exception being that the effect of innovations be-came less significant (p-value of .17 rather than .09). In addition, toaccount for the fact thatwe could not unambiguously resolve the timingissue for 11 Aldi observations, we also implemented a more conserva-tive test. Assuming that the same proportion (2/19) would apply tothose 11 observations, we dropped an additional observation.21 Wedid this 11 times, each time dropping a different observation (on topof the two that were consistently excluded). The results are once morevery robust. In every instance, the parameter for the PL-productiondummy remains highly significant (p b .01) in the outcome equation,while the other findings also remain very stable. Looking at the medianp-value across the 11 estimations, the same substantive conclusions asbefore are obtained, except for the effect of innovations, which was sig-nificant at a p-value of .09 in the main estimation but now obtains ap-value of .19. Across all of the 26 parameters that are listed inTable 5, the correlation between the original p-values and the medianp-value that was obtained in this exercise was a high .98. Hence, eventhough the aforementioned secrecy surrounding PL production pre-cluded us from obtaining the exact timing, we feel confident that thislimitation did not substantively affect our results.

6.3.4. Shelf presence versus shelf breadthWe also explored whether our results remain stable when looking

at the number of NBs with which a NB manufacturer obtains shelfpresence, i.e., the extent of its presence, instead of only looking atwhether a NB manufacturer was admitted. Manufacturers that offerbetter services to retailers may not only be rewarded with shelf

presence per se (i.e., a 0/1 decision) but also with a more extensivepresence (Mangold & Faulds, 1993). To that extent, we considered“shelf breadth,” which captures the number of NBs that are listedfrom a given manufacturer in a given category as an alternative de-pendent variable in the outcome equation. Again, using ML estima-tion, we estimated a recursive model with the treatment variable(PL production) as an endogenous (binary) regressor in the outcomeequation, but the dependent variable in this outcome equation nowtakes the form of a continuous performance metric (see, e.g., Guo &Fraser, 2010; Jones, 2007). The major advantage of shelf breadthover shelf presence is that no discretization (which may involve aninformation loss) is needed. However, shelf presence has the attrac-tive feature of being more stable over time, a non-negligible advan-tage given the cross-sectional (snapshot) nature of our study. In 47out of the 185 manufacturer-category combinations, a NB manufac-turer obtains shelf presence for more than one NB in a category,with a maximum of six brands. The results for the shelf breadth equa-tion largely corroborate the shelf presence findings, with the excep-tion of the effects of sales growth, ease of producing high-qualityproducts, and manufacturer market power: while they are of thesame sign, these variables now fail to reach statistical significance(p = .19, p = .30, and p = .36, respectively). Most importantly,however, we find that PL production, which is evaluated at the aver-age level of manufacturer market power, positively affects shelfbreadth in the category at the discounter (β1 = 1.53, p b .01).Hence, although it is not likely that multiple NBs from the same man-ufacturer will be accepted, the odds are better for a PL producingmanufacturer than for a manufacturer that does not do so.

6.3.5. Including direct pull effectsWe also tested for a direct ‘pull’ effect of the three marketing tools

on shelf presence. No such evidence is found for a NB manufacturer'sprice premium (p = .99) and innovations (p = .27). Only for adver-tising, a direct effect on the likelihood to obtain NB shelf presence isobtained (β = .02, p b .10). When allowing for potential pull effects,all of the other results are again very stable. The one change is that theeffect of innovations in the production equation, while having thesame sign as before, is no longer significant (p = .39).

6.3.6. Correlated errors across countriesWe allowed for dependencies between observations using

two-way clustering (i.e., for manufacturers and for categories withina country) in our main model. It is also possible that certain NB man-ufacturers are active in both countries. Ten such international NBmanufacturers were found in our dataset (including Unilever andNestlé). To allow for a potential correlation among observationsfrom the same manufacturer across the two countries, we adoptedan alternative two-way clustering approach (i.e., for categories withina country as before, but for manufacturers across both countries). Ourresults remain perfectly stable.

To summarize, across the various robustness checks, our resultsare found to be very stable, as detailed in Table 7. Foremost, thefocal effect that PL production leads to a higher likelihood ofobtaining shelf presence at discounters is stable across all of the ro-bustness tests. Additionally, all of the other parameters in the out-come (shelf presence) equation consistently provide the samesubstantive insights. Not only the main effect of manufacturer marketpower, but also its interaction effect with PL production, and all of thecontrol variables have similar effects across all of the checks. For thestrategy-selection (PL production) equation, we achieve very robustresults (with only one exception) for the effect of a manufacturer'ssales growth, the ease of producing high-quality products in a catego-ry, and a manufacturer's market power. For the manufacturer market-ing tools, we obtain equally robust results for a manufacturer's pricepremium and for advertising. For the number of innovations, somevariation in the significance of the effect is found, even though the21 For 19 (=10 + 7 + 2) observations, the ambiguity could be resolved.


sign of the parameter remains the same across all of the validationchecks. Finally, the control variables also show very consistent results.Across the 168 parameters that were estimated in the various robust-ness checks, over 95% result in the same substantive insight as in ourfocal model.22

7. Discussion

Faced with soaring PL shares and volumes, PL production has be-come increasingly attractive to CPG manufacturers, particularly inlight of the current economic downturn. The decision of whether a NBmanufacturer should engage in PL production is, however, not an easyone. Hence, it is not surprising that most NB manufacturers are stillstruggling with the issue (de Jong, 2007). In this article, we empiricallyinvestigated the determinants, as well as the effect on discounter good-will, of NB manufacturers' engagement in PL production. We tested ourhypotheses on a sample of 450 NB manufacturer-category combina-tions for two leading discounters in the German and Spanish market,using a uniquely assembled dataset that combines extensive field re-search with various secondary data sources.

Our analyses offer important insights to both NB manufacturersand discounters. First, while many NB manufacturers now regard thediscount sector as “an essential channel they proactively approach”(IGD, 2005, p. 34), they should realize that gaining access to thatchannel is far from straightforward or automatic. Although we fo-cused on the leading manufacturers in the German and Spanish mar-ket for close to 60% of all of the manufacturer-category combinations,no shelf presence was obtained for any of their brands. Given that

managers of convenience goods strive to distribute their brands as in-tensively as possible (Coughlan, Anderson, Stern, & El-Ansary, 2001,p. 288), missing out on some of the fastest-growing retailers repre-sents a substantial set-back. To make matters worse, discounters(especially hard discounters) tend to be influenced less by the manu-facturers' power status than conventional supermarkets; additionally,the tools that manufacturers are most familiar with to ensure shelf ac-cess, such as the payment of slotting allowances (Sudhir & Rao, 2006)and/or increased trade promotions (Narasimhan, 2009), are rarelyused by discounters (PlanetRetail, 2010).

However, we find strong support for the (thus-far untested) notionthat PL production for a discounter in a given category substantially in-creases the likelihood of obtaining shelf presence in that category.While quite some NB manufacturers are still reluctant to produce PLs,they should be aware that a refusal to engage in PL production for adiscounter may seriously jeopardize their chances of gaining access tothe limited spots on that discounter's shelves. Moreover, unlike theaforementioned tools (such as slotting allowances and promotionalsupport), which entail a direct monetary transfer from a manufacturerto a retailer (Scott Morton & Zettelmeyer, 2000), PL production, wheninstrumental in securing NB shelf access, contributes positively to amanufacturer's bottom line in various ways: (i) revenues from theNB(s) that are sold through the discount channel, (ii) revenues fromthe PLs (which could potentially take place through otherwise unused,and therefore costly, over-capacity), and (iii) positive margin implica-tions thatmay result from themore intense relationshipwith the retail-er (ter Braak et al., 2013).

In spite of these potential benefits, many managers remainconcerned with the potential downsides to such a dual-branding(PL production along with NB development) strategy. Foremost,they fear the reaction of their customer base if it were to become

Table 7Robustness checks.

Focal modelresults

Alternativeoperationalization

Alternativesample

Shelfbreadth

Direct 'pull'effects

Alternativetwo-wayclustering

Brandpower

Relativemeasures

N = 448 N = 447a

Outcome equation PL production ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Manufacturer market power AldiManufacturer market power Mercadona ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

PL production * Manufacturer market power AldiPL production * Manufacturer market power Mercadona ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Control variablesDiscounter dummy (0 = Aldi, 1 = Mercadona) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

PL production in a non-focal category ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Discounter category shareBeverages dummy (food = baseline)Non-food dummy (food = baseline) ✓

Advertising n.a. n.a. n.a. n.a. n.a. n.a. ✓ n.a.Price premium over PLs n.a. n.a. n.a. n.a. n.a. n.a. n.a.Innovations n.a. n.a. n.a. n.a. n.a. n.a. n.a.

Strategy-selectionequation

Manufacturer sales growth ✓ ✓ ✓ ✓ ✓ ✓ ✓

Ease of producing high-quality products ✓ (✓) (✓) (✓) (✓) ✓ ✓

Manufacturer marketing toolsAdvertising ✓ (✓) ✓ ✓ ✓ ✓ ✓

Price premium over PLs Aldi ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Price premium over PLs MercadonaInnovations ✓ (✓) (✓) ✓ ✓

Manufacturer market power ✓ ✓ ✓ ✓ ✓ ✓ ✓

Control variablesDiscounter dummy (0 = Aldi, 1 = Mercadona) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

PL production in a non-focal category Aldi ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

PL production in a non-focal category Mercadona ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Discounter category share ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Beverages dummy (food = baseline)Non-food dummy (food = baseline)

Selection parameter (ρ) ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

A ‘✓’ indicates that a significant effect is found. ‘n.a.’ means that the variable is not applicable.A ‘(✓)’ indicates that even though the parameter for this effect is robust in both sign and magnitude, its p-value is b .15.a The p-values reported are the median p-values across 11 estimations (i.e. when randomly deleting one additional observation for which the timing issue could not be resolved).

22 Detailed results are available from the first author upon request.


known that they also produce (cheaper) PLs. This fear is reflected notonly in the lower PL-production likelihood that we obtained for man-ufacturers that currently charge a higher premium for their brands(especially when dealing with very price-oriented hard discounters),and/or that invest heavily in advertising in an effort to differentiatetheir brands, but also in the efforts that those that do produce taketo conceal this practice. Many companies go even one step further,and separate both activities not only commercially, but also inthe location of their physical production facilities. Firms as ReckittBenckiser (a leading manufacturer of household products) andMcCain (manufacturer of, among other categories, frozen potatoproducts), for example, have completely separated all of their PL ac-tivities from those for their own brands, to the extent of creating sep-arate companies (de Jong, 2007). Not only does this restrain theflexibility on how to utilize the total production capacity (Bergès &Bouamra-Mechemache, 2011), manufacturers should also be carefulthat this strategy does not hurt potential relational (goodwill-related)benefits.

Apart from these customer-equity concerns, we find that manufac-turers are also afraid that they will become pressured to share their lat-est technologies and innovations with the retailer/discounter, whichbecomes even more likely when only a few players can assure high-quality production. In spite of this concern, PL production can be an at-tractive opportunity for innovation-oriented manufacturers as well.Apart from the aforementioned financial benefits, they can use suchproduction as “a chance to try out product ideas at much lower costs”(Dunne & Narasimhan, 1999, p. 42; ter Braak et al., 2013). Indeed,launching a new product is risky and costly. When manufacturers arenot sure if a new product will work (which is the case for many newproducts), they can limit their risk through an upfront involvement ofa discounter (while also bringing shelf presence for their NBs into thenegotiations) and subsequently supply a product that primarily substi-tutes for their competitors' NBs (Dunne, 1999).

Finally, many manufacturers operate in multiple categories andmay be more/less reluctant to engage in PL production in some cate-gories than in others, if only because their advertising intensity differsacross categories or because they have more intellectual property toprotect in some categories than in others. Such firms should beaware of the fact that the impact of PL production in one category(which they themselves may perceive as an initial pledge towards abetter relationship with the discounter) may actually amplify thenegative implications of refusing to do so in other categories.

Our analyses also offer some key insights for discounters. Given theirhistorically almost exclusive focus on PLs, they may have very limitedknowledge on (or even interest in) NB manufacturers. Deleersnyderet al. (2007) have discussed that it may pay off for discounters to mon-itor the NB scene more closely to identify which brands to add to theirassortment to maximally increase their overall category performance.However, apart from these more demand-driven considerations, itmay also be useful to identify which NB manufacturers will be mosteager to obtain shelf space for their NBs, and/or least reluctant to engagein PL production, as this will reflect on their bargaining position whennegotiating supply terms. Declining sales for the other party's NBs, forexample, may signal over-capacity, and a higher willingness to notonly produce PLs, but also to do so at lower wholesale costs. Similarly,when looking to add more premium PLs to the assortment (which in-volves vertical and horizontal differentiation from existing alternatives;Kumar & Steenkamp, 2007), theymay primarilywant to approachman-ufacturers that are known for their innovative capabilities. To overcomethese manufacturers' initial reluctance to do so, offering shelf presencefor their NBs may be an appealing option.

7.1. Limitations and further research

In line with Kaufman et al. (2006) and Ailawadi, Pauwels, andSteenkamp (2008), we considered two different retailers to test our

hypotheses. It was encouraging to see that most of our findings gen-eralized across both discounters, even though they have a differentpositioning (hard vs. soft discounter), and they come from two differ-ent countries (Germany vs. Spain).23 Moreover, the observed differ-ences (i.e., the role of manufacturer market power in the outcomeequation and the role of price premium in the selection equation)had considerable face validity in light of the discounters' differentpositioning. Still, it would be interesting to investigate to whatextent our findings can be generalized to (i) other product categories,(ii) other countries, and (iii) other retail formats. Additionally, the re-cent economic crisis may contribute both to further growth (andhence, a stronger negotiation position) for the discounters, while NBmanufacturers may become more inclined to engage in PL production.

Because discounters operate with a limited assortment, shelfspace that is given to NBs is presumably the most important dimen-sion of retailer goodwill. For traditional retailers, shelf space mightnot be such a good measure of goodwill. To them, slotting allowancesprovided by NB manufacturers play a key role in a retailer's decisionto carry a NB (Sudhir & Rao, 2006), while they are rarely used in a dis-counter setting (PlanetRetail, 2010). Future research could investi-gate to what extent our findings are generalizable to traditionalretail formats on other dimensions of retailer goodwill, such as the ex-tent of pass-through of trade promotions (Ailawadi & Harlam, 2009),the size of the slotting allowances (Sudhir & Rao, 2006), a NBmanufacturer's influence on a retailer's promotional calendar (Dunne&Narasimhan, 1999), and/or the likelihood of being appointed “catego-ry captain” (Subramanian et al., 2010).

Similar to previous studies (see, e.g., Nijs, Dekimpe, Steenkamp, &Hanssens, 2001; Pauwels, Srinivasan, & Franses, 2007), we focused onthe top players in each category to keep the data-collection effortmanageable. Due to our focus on the top-five NB manufacturers inthe category, the more limited variation in the manufacturer market-power variable may have contributed to the non-significant effect ofthat variable for Aldi. Future research could also investigate non-top-five manufacturers. Our data are further limited in that it offersonly a historical snapshot. While we took great care to ensure that PLproduction precedes shelf presence, we did not have precise informa-tion on the starting dates of both events. A longitudinal dimensionwould undoubtedly offer additional insights.

Furthermore, future research could investigate potential cross-country effects. For example, a top NB manufacturer could supply PLsto a discounter in a country in which the NBmanufacturer is not active.Such PL production would create goodwill with the focal discounter,while the negative implications (reputational and competitive dis-advantages) for the NB manufacturer would be limited. In addition tothese cross-country effects, future research could also investigatecross-retailer effects of NB manufacturers' PL production decisions.IGD (2005) reported that a concern that wasmentioned by PL suppliersfor the discount channel was how this engagement would impact theirrelationwith other retailers. Although the goodwill of the discounter forwhich the NB manufacturer starts to produce may increase, this couldcome at the expense of deteriorated relations with other retailers. Sim-ilarly, should PL production be restricted to a single discounter? Suchexclusivity could create even more goodwill with the focal discounter,but the negative implications with other retailers (discounters) maythen become more pronounced. Finally, while we provided initial evi-dence that the implications of PL production may extend beyond thefocal category (through the γ1 parameter), it may be interesting toquantify the net implications across all of the categories in which amanufacturer is active. In doing so, it would pay off to also considerthe moderating influence of a manufacturer's power across all of thecategories.

23 In addition, the economic situation may have been different, given the later year ofthe Spanish data collection. We thank an anonymous reviewer for this observation.


We have only begun to scratch the surface of research possibilitiesin an area that warrants more attention. We hope that this article willprovide a stimulus for more research on PL production.

Acknowledgments

We are grateful for the constructive comments that we receivedfrom the anonymous IJRM reviewers. We would also like to thankKaren Gedenk, Els Gijsbrechts, Koen Pauwels, Raj Sethuraman, andJan-Benedict Steenkamp for their insightful comments on an earlierdraft. We further acknowledge AiMark for providing access to theGfK consumer panel data and thank Kantar Worldpanel for their helpin obtaining the private label supplier information forMercadona. Final-ly, we are indebted to the Netherlands Organization for Scientific Re-search (NWO) for its financial assistance.

References

Aaker, D. A., & Keller, K. L. (1990). Consumer evaluations of brand extensions. Journal ofMarketing, 54(1), 27–41.

ACNielsen (2007). The hard discounter report — an overview of Aldi and Lidl inEurope. http://dk.nielsen.com/reports/ConsInsightsEuropeJune07_media.pdf (RetrievedJanuary 10, 2013).

ACNielsen (2011). The rise of the value conscious shopper. http://hk.nielsen.com/documents/PrivateLabelGlobalReport.pdf (Retrieved January 10, 2013).

Ailawadi, K. L., & Harlam, B. A. (2004). An empirical analysis of the determinants of re-tail margins: The role of store-brand share. Journal of Marketing, 68(1), 147–165.

Ailawadi, K. L., & Harlam, B. A. (2009). Retailer promotion pass-through: A measure, itsmagnitude, and its determinants. Marketing Science, 28(4), 782–791.

Ailawadi, K. L., Lehmann, D. R., & Neslin, S. A. (2003). Revenue premium as an outcomemeasure of brand equity. Journal of Marketing, 67(4), 1–17.

Ailawadi, K. L., Pauwels, K., & Steenkamp, J. B. E. M. (2008). Private-label use and storeloyalty. Journal of Marketing, 72(6), 19–30.

Anderson, E., Lodish, L. M., & Weitz, B. A. (1987). Resource allocation behavior in con-ventional channels. Journal of Marketing Research, 24(1), 85–97.

Anderson, E., & Weitz, B. A. (1992). The use of pledges to build and sustain commit-ment in distribution channels. Journal of Marketing Research, 29(1), 18–34.

Bain, Company (2008). Finding growth in Europe's shifting grocery landscape. http://www.bain.com/bainweb/PDFs/cms/Public/European_grocer_BB.pdf (Retrieved January 10,2013).

Basuroy, S., Mantrala, M. K., & Walters, R. G. (2001). The impact of category managementon retailer prices and performance: Theory and evidence. Journal of Marketing, 65(4),16–32.

Bergès, F., & Bouamra-Mechemache, Z. (2011). Is producing a private label counterpro-ductive for a branded manufacturer? European Review of Agricultural Economics,39(2), 213–239.

Bertram, H. J. (2006). Der Discounter Marken-guide: Die bekannten Marken hinterno-name Artikeln. Berlin: Ullstein Taschenbuch.

Brandes, D. (2005). Bare essentials — The Aldi way to retail success. Frankfurt: CampusVerlag.

Breen, R. (1996). Regressionmodels— Censored, sample selected, or truncated data. London:Sage Publications.

Briggs, D. C. (2004). Causal inference and the Heckman model. Journal of Educationaland Behavioral Statistics, 29(4), 397–420.

BusinessWeek (2005). The big brands go begging in Europe. http://www.businessweek.com/stories/2005-03-20/the-big-brands-go-begging-in-europe (Retrieved January10, 2013).

Cameron, A. C., Gelbach, J. B., & Miller, D. L. (2011). Robust inference with multiwayclustering. Journal of Business & Economic Statistics, 29(2), 238–249.

Chen, J. X., Narasimhan, O., John, G., & Dhar, T. (2010). An empirical investigation of pri-vate label supply by national label producers. Marketing Science, 29(4), 738–755.

Cleeren, K., Verboven, F., Dekimpe, M. G., & Gielens, K. (2010). Intra- and inter-formatcompetition among discounters and supermarkets. Marketing Science, 29(3),456–474.

Cohen, J., Cohen, P., West, S. G., & Aiken, L. S. (2003). Appliedmultiple regression/correlationanalysis for the behavioral sciences. Mahwah, NJ: Lawrence Erlbaum Associates.

Coughlan, A. T., Anderson, E., Stern, L. W., & El-Ansary, A. I. (2001). Marketing channels.Upper Saddle River, NJ: Prentice Hall.

de Jong, K. A. M. (2007). Private labels in Europe — Trends and challenges for retailers andmanufacturers. Vught: International Private Label Consult.

Deleersnyder, B., Dekimpe, M. G., Steenkamp, J. B. E. M., & Koll, O. (2007). Win-winstrategies at discount stores. Journal of Retailing and Consumer Services, 14(5),309–318.

Desai, P. S. (2000). Multiple messages to retain retailers: Signaling new product de-mand. Marketing Science, 19(4), 381–389.

Dunne, D. (1999). Should grocery manufacturers supply private labels? Working Paper.Joseph L. Rotman School of Management, University of Toronto.

Dunne, D., & Narasimhan, C. (1999). The new appeal of private labels. Harvard BusinessReview, 77(3), 41–52.

Fernandez Nogales, A., & Gomez Suarez, M. (2005). Shelf space management of privatelabels: A case study in Spanish retailing. Journal of Retailing and Consumer Services,12(3), 205–216.

Gielens, K., & Steenkamp, J. B. E. M. (2007). Drivers of consumer acceptance of newpackaged goods: An investigation across products and countries. International Journalof Research in Marketing, 24(2), 97–111.

Gomez, M., & Benito, N. R. (2008). Manufacturer's characteristics that determine thechoice of producing store brands. European Journal of Marketing, 42(1/2), 154–177.

Gomez-Arias, J. T., & Bello-Acebron, L. (2008). Why do leading brand manufacturerssupply private labels? The Journal of Business and Industrial Marketing, 23(4),273–278.

Greene, W. (2000). Econometric analysis. Upper Saddle River, NJ: Prentice Hall.Guo, S., & Fraser, M. W. (2010). Propensity score analysis: Statistical methods and appli-

cation. London: Sage Publications.Hamilton, B. H., & Nickerson, J. A. (2003). Correcting for endogeneity in strategic man-

agement research. Strategic Organization, 1(1), 51–78.Heckman, J. J. (1979). Sample selection bias as a specification error. Econometrica,

47(1), 153–161.Hoch, S. J. (1996). How should national brands think about private labels? Sloan

Management Review, 37(2), 89–102.IGD (2002). European discount retailing. Future focus. Watford: IGD.IGD (2005). Winning in a discounter's world. Watford: IGD.IGD (2006). European private label growth: Strategic responses for suppliers and retailers.

Watford: IGD.Inman, J. J., Winer, R. S., & Ferraro, R. (2009). The interplay among category characteristics,

customer characteristics, and customer activities on in-store decision making. Journalof Marketing, 73(5), 19–29.

Irwin, J. R., & McClelland, G. H. (2001). Misleading heuristics and moderated multipleregression models. Journal of Marketing Research, 38(1), 100–109.

Jones, A. M. (2007). Applied econometrics for health economists: A practical guide.Abingdon: Radcliffe Publishing.

Kaufman, P., Jayachandran, S., & Rose, R. L. (2006). The role of relational embeddedness inretail buyers' selection of new products. Journal of Marketing Research, 43(4), 580–587.

Kumar, N., Radhakrishnan, S., & Rao, R. C. (2010). Private label vendor selection in asupply chain: Quality and clientele effects. Journal of Retailing, 86(2), 148–158.

Kumar, N., & Steenkamp, J. B. E. M. (2007). Private label strategy. Cambridge, MA: HarvardBusiness School Press.

Lamey, L., Deleersnyder, B., Steenkamp, J. B. E. M., & Dekimpe, M. G. (2012). The effectof business cycle fluctuations on private-label share: What has marketing conductgot to do with it? Journal of Marketing, 76(1), 1–19.

Lemke, D., & Reed, W. (2001). War and rivalry among great powers. American Journal ofPolitical Science, 45(2), 457–469.

Lieberman, M. B. (1987). Excess capacity as a barrier to entry: An empirical appraisal.The Journal of Industrial Economics, 35(4), 607–627.

Littman, M. (1992). House brands hit home. http://www.highbeam.com/doc/1G1-12471761.html (Retrieved January 10, 2013).

Mangold, W. G., & Faulds, D. J. (1993). Service quality in a retail channel system. Journalof Services Marketing, 7(4), 4–10.

Narasimhan, C. (2009). Trade promotions. In V. R. Rao (Ed.), Handbook of pricing re-search in marketing (pp. 283–318). Edward Elgar Publishing.

Nijs, V. R., Dekimpe, M. G., Steenkamp, J. B. E. M., & Hanssens, D. M. (2001). Thecategory-demand effects of price promotions. Marketing Science, 20(1), 1–22.

Pauwels, K., Srinivasan, S., & Franses, P. H. (2007). When do price thresholds matter inretail categories? Marketing Science, 26(1), 83–100.

PlanetRetail (2010). Discounters: Rapid growth or myth. London: Planet Retail.PlanetRetail (2011). Country report Western Europe. London: Planet Retail.PlanetRetail (2012a). Retailer profile Walmart. London: Planet Retail.PlanetRetail (2012b). Germany market profile. London: Planet Retail.PlanetRetail (2012c). Mercadona in Spain. London: Planet Retail.PlanetRetail (2012d). Aldi in Germany. London: Planet Retail.Quelch, J. A., & Harding, D. (1996). Brands versus private labels: Fighting to win. Harvard

Business Review, 74(1), 99–110.Reibstein, D., & Wittink, D. (2005). Competitive responsiveness. Marketing Science,

24(1), 8–11.Reuters, Thomson (2009). Food groups face up to life with deep discounters. http://www.

reuters.com/article/2009/03/18/us-food-summit-discounters-idUSTRE52H7L620090318(Retrieved January 10, 2013).

Sayman, S., Hoch, S. J., & Raju, J. S. (2002). Positioning of store brands. Marketing Science,21(4), 378–397.

Sayman, S., & Raju, J. S. (2007). Store brands — From back to the future. In N. K.Malhotra (Ed.), Review of Marketing Research (pp. 132–151). Emerald GroupPublishing.

Schaab, S., & Eschenbek, W. (2008). Aldi – welche Marke steckt dahinter? 100 neueTop-Artikel und Ihre prominenten Hersteller. Munchen: Sudwest Verlag.

Schneider, M. (2005). Welche Marke steckt dahinter? No-Name-Produkte und Ihrenamhaften Hersteller. Munchen: Sudwest Verlag.

Schneider, M. (2006). Aldi – welche Marke steckt dahinter? 100 Aldi-Top-Artikel und Ihreprominenten Hersteller. Munchen: Sudwest Verlag.

Scott Morton, F., & Zettelmeyer, F. (2000). The strategic positioning of store brandsin retailer-manufacturing bargaining, Yale School of Management. Working PaperSeries H, #10.

Sethuraman, R. (2009). Assessing the external validity of analytical results from nationalbrand and store brand competition models. Marketing Science, 28(4), 759–781.

Sethuraman, R., & Raju, J. S. (2012). Private label strategies — Myths and realities. In V.Shankar, & G. S. Carpenter (Eds.), Handbook of marketing strategy (pp. 318–335).Edward Elgar Publishing.


Shervani, T. A., Frazier, G., & Challagalla, G. (2007). The moderating influence of firmmarket power on the transaction cost economics model: An empirical test in a for-ward channel integration context. Strategic Management Journal, 28(6), 635–652.

Slotegraaf, R. J., & Pauwels, K. (2008). The impact of brand equity and innovation on thelong-term effectiveness of promotions. Journal of Marketing Research, 45(3), 293–306.

Soberman, D. A., & Parker, P. M. (2006). The economics of quality-equivalent storebrands. International Journal of Research in Marketing, 23(2), 125–139.

Sorescu, A. B., & Spanjol, J. (2008). Innovation's effect on firm value and risk: Insightsfrom consumer packaged goods. Journal of Marketing, 72(2), 114–132.

Steenkamp, J. B. E. M., & Dekimpe, M. G. (1997). The increasing power of store brands:Building loyalty and market share. Long Range Planning, 30(6), 917–930.

Steenkamp, J. B. E. M., van Heerde, H. J., & Geyskens, I. (2010). What makes consumerswilling to pay a price premium for national brands over private labels? Journal ofMarketing Research, 47(6), 1011–1024.

Subramanian, U., Raju, J. S., Dhar, S. K., & Wang, Y. (2010). Competitive consequences ofusing a category captain. Management Science, 56(10), 1739–1765.

Sudhir, K., & Rao, V. R. (2006). Do slotting allowances enhance efficiency or hindercompetition? Journal of Marketing Research, 43(2), 137–155.

Tabachnick, B. G., & Fidell, L. S. (1996). Using multivariate statistics. New York, NY:Harper Collins.

ter Braak, A., Dekimpe, M. G., & Geyskens, I. (2013). Retailer private label margins:The role of supplier and quality-tier differentiation. Journal of Marketing, 77(4),86–103.

Uhl-Bien, M., & Maslyn, J. M. (2003). Reciprocity in manager–subordinate relation-ships: Components, configurations, and outcomes. Journal of Management, 29(4),511–532.

Wu, C. -C., & Wang, C. -J. (2005). A positive theory of private-label: A strategic role of pri-vate label in a duopoly national-brand market. Marketing Letters, 16(2), 143–161.


The influence of ad-evoked feelings on brand evaluations:Empirical generalizations from consumer responses to morethan 1000 TV commercials

Michel Tuan Pham a,⁎, Maggie Geuens b,c, Patrick De Pelsmacker d,b

a Columbia University, Graduate School of Business, 3022 Broadway, New York, NY 10027, United Statesb Ghent University, Tweekerkenstraat 2, 9000 Ghent, Belgiumc Vlerick Business School, Reep 1, 9000 Ghent, Belgiumd University of Antwerp, Prinsstraat 13, 2000 Antwerp, Belgium


Article history:First received in 25, January 2012and was under review for 5 monthsAvailable online 20 June 2013

Guest Editor: Marnik G. DekimpeArea Editor: Hans Baumgartner

Keywords:AdvertisingEmotionsFeelingsInvolvementMotivationEmpirical generalizations

It has been observed that ad-evoked feelings exert a positive influence on brand attitudes. To investigate theempirical generalizability of this phenomenon, we analyzed the responses of 1576 consumers to 1070 TVcommercials frommore than 150 different product categories. The findings suggest five empirical generaliza-tions. First, ad-evoked feelings indeed have a substantial impact on brand evaluations, even under conditionsthat better approximate real marketplace settings than past studies did. Second, these effects are both directand indirect, with the indirect effects largely linked to changes in attitude toward the ad. Third, these effectsdo not depend on the level of involvement associated with the product category. However, fourth, the effectsare more pronounced for hedonic products than utilitarian products. Finally, these effects do not depend onwhether the products are durables, nondurables, or services, or whether the products are search goods or ex-perience goods.


1. Revisiting the effects of ad-evoked feelings onbrand evaluations

Twenty-five years ago, an influential series of controlled lab stud-ies seemed to indicate that feelings evoked by advertisements have apositive influence on consumers' brand attitudes (Aaker, Stayman, &Hagerty, 1986; Batra & Ray, 1986; Edell & Burke, 1987; Holbrook &Batra, 1987). Everything else being equal, participants in these stud-ies typically reported more favorable brand attitudes after viewingads that elicited pleasant feelings than after viewing ads that elicitedless pleasant feelings. These early findings have been replicated in alarge number of studies with both television commercials and printadvertisements (e.g., Baumgartner, Sujan, & Padgett, 1997; Burke &Edell, 1989; Derbaix, 1995; MacInnis & Park, 1991; Miniard, Bhatla,Lord, Dickson, & Unnava, 1991; Morris, Woo, Geason, & Kim, 2002).In a meta-analysis, Brown, Homer, and Inman (1998) found that theaverage correlation between ad-evoked feelings and brand evalua-tions was around r = .35 to .40, an effect that typically would be con-sidered “medium to large” (Cohen, 1988).

The finding that brand attitudes can be substantially influenced bythe mere emotional pleasantness of brand advertisements has obvi-ous managerial implications: It would suggest that everything elsebeing equal, marketers should generally try to improve the emotionalpleasantness of their advertising. However, before this finding and itsmarketing implications can be accepted at face value and acted uponby managers, it is important to assess its empirical generalizabilitywithin the real world of brand advertising. To this end, two sets ofissues need to be addressed.

First, because previous studies have focused primarily on thetheoretical explanation of the phenomenon and therefore priori-tized concerns for internal and construct validity, these studieshave tended to compromise the external validity of their method-ologies. For example, an important limitation of previous studiesis that they were usually conducted among college students asopposed to more broadly representative samples of consumers(e.g., Machleit & Wilson, 1988; Madden, Allen, & Twible, 1988;Moore & Hutchinson, 1983; Stayman & Aaker, 1988). However, ithas been observed that the correlation between ad-evoked feelingsand brand attitudes tends to be stronger among students thanamong nonstudents (Brown & Stayman, 1992).

In addition, most of the classic studies involved rather limited poolsof advertisements. For example, Edell and Burke (1987) examined ten


⁎ Corresponding author. Tel.: +1 212 854 3472; fax: +1 212 316 9214.E-mail addresses: [email protected] (M.T. Pham), [email protected]

(M. Geuens), [email protected] (P. De Pelsmacker).


Contents lists available at ScienceDirect



and six commercials across studies. The limited pool of advertisementsused in prior studies raises issues of potential selection bias. In par-ticular, if the ad stimuli were specifically selected because of theiremotional richness, prior studies may overstate the influence ofad-evoked feelings on brand evaluations in the real world. This po-tential selection bias is compounded by the fact that a number ofprevious studies relied on fictitious or unfamiliar ads and brands(e.g., MacKenzie & Lutz, 1989; Miniard, Bhatla, & Rose, 1990; Olney,Holbrook, & Batra, 1991; Park & Young, 1986), which tends to furtheramplify the observed influence of ad-evoked feelings on brand eval-uations (e.g., Brown et al., 1998; Derbaix, 1995; see also Bakamitsos,2006; Ottati & Isbell, 1996).

Finally,most of the previous studies (e.g., Batra & Ray, 1986; Burke &Edell, 1989; Edell & Burke, 1987; Stayman&Aaker, 1998) solicitedmea-sures of ad-evoked feelings and brand evaluations from the samerespondents. Such repeated measurements raise significant issues ofshared method variance, again likely to exaggerate the true effects ofad-evoked feelings on brand evaluations (see Feldman & Lynch,1988). In summary, given the questionable external validity of mostprevious studies of the phenomenon, it is difficult to gauge the truemagnitude of the effects of ad-evoked feelings on brand evaluations inthe real world, with a distinct possibility that these effects may havebeen overstated.

A second set of issues pertains to the boundary conditions of the phe-nomenon. In particular, an important question for marketing profes-sionals is whether these effects are generally true across productcategories or are instead limited to certain product categories. For exam-ple, does an emotional ad have similar effects on attitudes toward a brandof automobiles as opposed to a brand of financial services? Surprisinglyenough, this type of question has yet to receive an adequate empirical an-swer. This is because the relatively small, convenience samples of adver-tisements—and hence brands and product categories—used in previousstudies preclude a rigorous analysis of the generalizability of the phenom-enon across product categories. As a result, there is an important gap be-tweenwhatmarketing professionals need to know about the scope of thephenomenon and what previous research is able to substantiate.

The purpose of our research is to provide managerially relevantempirical generalizations about the effects of ad-evoked feelingson brand evaluations by (a) addressing key external-validity short-comings of previous studies, and (b) assessing the generalizabilityof the phenomenon across product categories. Our investigationdeparts from previous studies in four major respects. First, insteadof soliciting responses from college students, we collected re-sponses from 1576 consumers who were broadly representativeof the Belgian population. Second, instead of using a small, conve-nience sample of advertisements, we used a total of 1070 TV com-mercials for existing brands, constituting a near census of allcommercials shown on Dutch-speaking Belgian television over athree-year period. Third, we used a design that does not requirethe measurement of ad-evoked feelings and brand evaluationsfrom the same respondents. Finally, we explicitly examined poten-tial product-category-level moderators of the effects of ad-evokedfeelings on brand evaluations.

Our results show that evenwhenmajor issues of external validityare addressed, ad-evoked feelings do have substantial positive ef-fects on brand evaluations, with an effect size that is roughly compa-rable to that uncovered in previous studies. These effects hold evenafter controlling for attitude toward the ad (Aad) and cognitive be-liefs, suggesting that ad-evoked feelings have direct effects onbrand evaluations. The effects appear to be stronger for productstypically associated with hedonic motives than for products typical-ly associated with utilitarian motives. However, we found little evi-dence that these effects depend on whether the product is typicallyassociated with high versus low consumer involvement, whetherthe product is a durable, a nondurable, or a service, and whetherthe product is a search or experience good.

2. Conceptual background

2.1. Major theoretical explanations

Fourmajor theoretical explanations can be advanced for the effectsof ad-evoked feelings on brand evaluations. The first is that these ef-fects are mediated by changes in consumers' attitude toward the ad.Ads that evoke more pleasant feelings are typically preferred to adsthat evoke less pleasant feelings. Favorable attitudes toward the ad(Aad) translate into favorable attitudes toward the brand (Ab)through a carryover process known as “affect transfer” (MacKenzie,Lutz, & Belch, 1986). Consistent with this explanation, some studieshave found that the effects of ad-evoked feelings on brand evaluationsare fully mediated by their effects on Aad (e.g., Batra & Ray, 1986;Holbrook & Batra, 1987; MacInnis & Park, 1991). Other studies, however,have found only partial mediation (e.g., Burke & Edell, 1989; Stayman &Aaker, 1988), suggesting that additional processes may be at work.

The second explanation posits that these effects are mediated bydifferences in the beliefs and thoughts that consumers have aboutthe brand. Compared to ads that elicit less pleasant feelings, adsthat elicit more pleasant feelings may trigger more positive beliefsand thoughts about the brand, which, when integrated into summa-ry evaluations, would result in more favorable brand attitudes(e.g., Burke & Edell, 1989; Fishbein & Middlestadt, 1995). This sec-ond explanation is consistent with research suggesting that immedi-ate feelings toward a target tend to bias subsequent thoughts aboutthe target in the direction of these feelings (e.g., Batra & Stayman,1990; Pham, Cohen, Pracejus, & Hughes, 2001; Yeung & Wyer,2004), and with research showing that affective states tend to acti-vate affect-consistent materials in memory (Bower, 1981; Isen,Shalker, Clark, & Karp, 1978). In line with this explanation, somestudies have found that the effects of ad-evoked feelings on brandevaluations are largely mediated by changes in brand beliefs andthoughts (e.g., Burke & Edell, 1989; Cho & Stout, 1993; Edell &Burke, 1987), although other studies indicate that the effects ofad-evoked feelings remain substantial after controlling for brand be-liefs (Morris et al., 2002).

The third explanation is that the effects are due to a more automaticprocess of evaluative conditioning (De Houwer, Thomas, & Baeyens,2001). Specifically, themere pairing of a brandwith the feelings evokedby an ad may result in the valence of these feelings being associativelyincorporated into the brand evaluations (Gorn, 1982; Jones, Olson, &Fazio, 2010). This explanationwould be consistentwith studies indicat-ing direct effects of ad-evoked feelings on brand evaluations after con-trolling for Aad and brand beliefs (e.g., Burke & Edell, 1989; Cho &Stout, 1993; Homer & Yoon, 1992).

A final explanation is a more inferential process of “affect-as-information” (Schwarz & Clore, 1983, 2007). Given that peopleoftenmake judgments by inspecting their momentary feelings and ask-ing themselves, “How do I feel about it?” (see Pham, 2004, and Schwarz& Clore, 2007, for reviews), it is possible that consumers interpret theirmomentary feelings experienced during ad exposure as being indica-tive of howmuch they like or dislike the advertised brand. This processis distinct from affect transfer and evaluative conditioning in that thehow-do-I-feel-about-it (HDIF) process is inferential, whereas affecttransfer and evaluative conditioning are purely associative. The HDIFprocess is also different from a belief- and thought-priming process inthat feelings enter judgments directly, rather than indirectly through achange in beliefs and thoughts.

In summary, ad-evoked feelings can influence brand evaluationsthrough at least four different processes. Given that these processesare probably not mutually exclusive and are instead likely to operatejointly, one would predict that ad-evoked feelings are likely to havesubstantial positive effects on brand evaluations in the real world, evenwhen major threats to external validity are addressed. Moreover,these effects are likely to be both direct and indirect.

384 M.T. Pham et al. / Intern. J. of Research in Marketing 30 (2013) 383–394

2.2. Potential product-category-level moderators

Given that the purpose of our research is to identify empiricalgeneralizations that are actionable from a managerial standpoint,we focus our analysis of boundary conditions on potential modera-tors of the phenomenon that are at the product-category level ratherthan at the individual-consumer level. This is because marketingprofessionals typically do not know with precision the psychologicalstates of individual consumers, but do know the general characteris-tics of the product category being advertised. In our study we inves-tigate four potential category-level moderators of the effects ofad-evoked feelings on brand evaluations: (a) whether the productcategory is typically associated with low or high levels of consumerinvolvement; (b) whether the product category is typically associat-ed with hedonic versus utilitarian consumption motives; (c) wheth-er the advertised product is a durable, a nondurable, or a service; and(d) whether the product is a “search good,”whose quality can be de-termined by examination prior to purchase (e.g., curtains, creditcards), or an “experience good,”whose quality can only be determinedthrough actual consumption (e.g., diet programs, pre-prepared meals)(Nelson, 1970).

With respect to a potential moderating role of involvement typicallyassociated with the product category, some studies suggest that the ef-fects of ad-evoked feelings on brand evaluationsmay be stronger underconditions of lower consumer involvement than under conditions ofhigher consumer involvement (Batra & Stephens, 1994; Brown &Stayman, 1992; MacInnis & Park, 1991; Madden et al., 1988). Ifad-evoked feelings operate as “peripheral cues,” this finding would beconsistent with a popular prediction in the persuasion literature thatlow involvement tends to increase the influence of peripheral or heuris-tic cues on attitudes (Eagly & Chaiken, 1993; Petty & Cacioppo, 1986).However, other considerations would suggest that when assessed atthe product-category level (as done in popular planning frameworks)and examined under conditions of greater external validity, in-volvement may not moderate the effects of ad-evoked feelings onbrand evaluations. First, the studies that found increased effectsof ad-evoked feelings under low involvement have typically usedstrong instruction-based manipulations of consumer involvement(e.g., explicit instructions to evaluate the advertised brand vs. in-structions to simply watch the ad). It is not clear that “natural” var-iations in involvement across product categories would be strongenough to produce the type of moderation effects observed inthese studies. Second, while certain product categories (e.g., cars)are on average typically more involving than others (e.g., soaps),individual consumers may still vary considerably in their level ofinvolvement with a given product category (e.g., cooking oil for acasual cook versus a professional chef) (Bloch & Richins, 1983).The considerable heterogeneity across consumers in their level ofinvolvement with a given product category would tend to attenu-ate differences in typical involvement across product categories.Finally, some research suggests that feelings may influence judg-ments under conditions of both low involvement and high involve-ment, albeit through different processes. Whereas under lowinvolvement, ad-evoked feelings may influence brand evaluationsthrough heuristic mechanisms such as affect transfer, evaluativeconditioning, or the HDIF heuristic, under high involvement,ad-evoked feelings may influence brand evaluations by shapingthe beliefs and thoughts that consumers have about the brand(Albarracín & Wyer, 2001; Batra & Stayman, 1990; Forgas, 1995;Petty, Schumann, Richman, & Strathman, 1993). In other words, in-volvement may affect the process by which ad-evoked feelings in-fluence brand evaluations rather than the extent to which theyinfluence brand evaluations. In summary, it is not clear whether inreal-world settings the effects of ad-evoked feelings on brand evalua-tions would depend on the level of involvement typically associatedwith the product category.

With respect to the type of consumption motive, hedonic versusutilitarian, typically associated with the product category, several con-siderations lead us to predict that it will moderate the effects ofad-evoked feelings on brand evaluations. Specifically, the effects ofad-evoked feelings on brand evaluations are likely to be more pronouncedfor products that are typically associated with hedonic motives than forproducts that are typically associated with utilitarian motives. Althoughthis prediction makes intuitive sense and was the rationale for the dis-tinction made in the FCB grid between “think” and “feel” products(Vaughn, 1980, 1986), it has received scant empirical testing in the ac-ademic literature. Nevertheless, indirect evidence consistent with thisproposition can be found in the literature on incidental mood effectson consumer judgments, which consistently shows that consumersare more influenced by their mood states when they have hedonic mo-tives than when they have utilitarian motives (e.g., Chang & Pham,2013; Pham, 1998; Pham, Meyvis, & Zhou, 2001). This finding is alsoobserved when comparing the evaluation of hedonic products to theevaluation of utilitarian products (Adaval, 2001; White & McFarland,2009; Yeung & Wyer, 2004). Therefore, unlike with involvement, weexpect to observe this greater influence of feelings among consumerswith hedonic as opposed to utilitarian motives extended to settingswhere these motives are defined at the category level rather than atthe individual consumer level. This is because we expect greater homo-geneity across consumers in type of motives that they associate withdifferent product categories than in level of involvement with thesesame product categories.

Two additional product-category characteristics are examined here aspotential moderators of the effects of ad-evoked feelings on brand evalu-ations: whether the product is a durable, a nondurable (e.g., FMCG), or aservice; and whether the product is a search good or an experiencegood. While the literature does not suggest strong a priori predic-tions about these two product characteristics as moderators of theeffects of ad-evoked feelings, they are worth studying on an explor-atory basis because they have proven to be important moderators ofthe impact of advertising in general (Vakratsas & Ambler, 1999; seealso Hanssens, 2009). Indeed, it has been found that the impact of ad-vertising is greater for experience goods than for search goods, andthat this impact may be 50% higher for durable goods than fornondurable goods (see Vakratsas in Hanssens, 2009). It has alsobeen found that marketers tend to use different types of advertisingfor search goods than for experience goods (Nelson, 1970, 1974). It istherefore conceivable that the effects of ad-evoked feelings on brandevaluations might be different depending on these two product-category characteristics.

3. Empirical study

Our study examines (a) how emotional feelings evoked by a largepool of TV commercials influence the brand evaluations of a representa-tive set of adult consumers, and (b) how this influence is moderated byfour different product-category-level characteristics. Three importantaspects of our investigation should be noted. First, our investigationseeks to better approximatemarketplace settings than previous studiesgenerally did. Second, our investigation focuses on the ads themselves,rather than the consumers, as themain units of analysis (as in Holbrook& Batra, 1987). This is because advertisers have greater control over thecontents of their ads than over consumers' responses to these ads. Third,our investigation focuses on product-category-level moderators of thephenomenon, rather than on consumer-level moderators within a cat-egory. This is becausemarketers typically know the general characteris-tics of the product category but not the specific states of individualconsumers. Thus, our research questions are considered at the followinglevel: “Does the emotional pleasantness of ad X influence attitudes to-ward brand Y, given that Y belongs to product category Z?,” which isthe way brand managers and advertising executives would contem-plate such questions.

385M.T. Pham et al. / Intern. J. of Research in Marketing 30 (2013) 383–394

3.1. Method

As in Holbrook and Batra (1987), our study relied on anaggregate-covariation design in which the units of analysis were dif-ferent TV commercials to which the consumers were exposed. Alarge sample of consumers watched a large number of TV commer-cials, then reported their attitude toward each advertised brand(Ab) and toward each ad (Aad). The emotions evoked by each com-mercial were coded by an independent set of judges, and the majorcharacteristics of each product category (involvement, motive, dura-bility, search/experience) were coded by another set of judges. Thevarious responses were averaged across respondents and judges,and the covariation across these responses was modeled across com-mercials. In addition to being more relevant from a managerialstandpoint, this design significantly reduces problems of sharedmethod variance that arise when all the measures are collectedfrom the same respondent (Holbrook & Batra, 1987; see also Pham,Cohen, et al., 2001).

3.1.1. Advertising stimuliThe data were collected in two waves, with the help of a market re-

search firm. For the first wave, we secured a census of all the differentbrand commercials that aired on a major Dutch-speaking Belgian TVchannel over a two-year period. For the second wave, conducted threeyears later, we secured another census of all brand commercials thataired on a different Dutch-speaking Belgian TV channel over aone-year period. From each census, we excluded commercials directedat children (because respondents were adult consumers), as well asshort promotional messages of less than 30 seconds Ads for “umbrella”brands (e.g., P&G) that were not linked to a specific product categorywere also excluded. The final stimulus set consisted of a total pool of1070 commercials (407 for wave 1 and 663 for wave 2), featuring 318different brands (national and international) across 153 different prod-uct categories (e.g., beer, credit cards, diapers, coffee, laundry deter-gents, cars, computers, etc.). This pool of ads is broadly representativeof the full spectrum of brand commercials directed at adult Belgian con-sumers. The dataset's summary characteristics are presented in Table 1.

3.1.2. RespondentsRespondents were 1576 Dutch-speaking Belgian consumers who

were recruited via TV ads on the same channels and received about!25 for their participation. The recruiting ads generated more than3000 initial responses for each wave. Of the initial respondents,1000 were selected in each wave and invited to participate in thestudy to form a broadly representative sample of adult Belgian con-sumers. Of those invited, 854 consumers eventually participated inwave 1 and evaluated the first pool of 407 commercials, and 839 con-sumers participated in wave 2 and evaluated the second pool of 663commercials. However, due to a variety of factors explained below,

only 722 of the 839 wave-2 respondents were retained for the analy-ses. The demographics of the 1576 eventual consumer respondentsare summarized in Table 1.

3.1.3. Procedure and consumer-response measuresThe procedure and measures were essentially parallel across the

two waves, except for slight differences. In both waves groups of 20 to30 respondents were invited at regular intervals to a research facility.Each group was shown a subset of the stimulus commercials, one com-mercial at a time, and asked to rate their responses after viewing eachcommercial. Two sequences of each subset of ads were used across ses-sions.Wave-1 respondents were shown about 20 commercials on aver-age (M = 20.14, SD = 5.64), whereas wave-2 respondents wereshown about 50 commercials on average, with the number of commer-cials varying considerably across sessions (M = 48.73; SD = 23.33).(The substantially greater number of commercials shown to wave-2 re-spondents, which was beyond our control, resulted in lower data qual-ity for the second wave, which necessitated some data purification, asexplained below.)

As detailed in Table 2, the consumer respondents completedthree main measures for each ad: (a) their attitude toward the ad(Aad), (b) their cognitive assessment of the ad (CogAss) to controlfor cognitive-belief effects of the ad, and (c) their attitude towardthe brand (Ab) as the main dependent measure in this study. Eachad was rated by an average of 43 consumers. These Aad, cognitive as-sessments (CogAss), and Ab ratings were averaged across respon-dents to form 1070 aggregate ad-level observations. To verify thatthe responses were sufficiently homogeneous across respondents,we computed !-coefficients of inter-respondent agreement foreach individual item (Holbrook & Batra, 1987; Pham, Cohen, et al.,2001). As shown in Table 2, the inter-respondent agreement coeffi-cients were high for all items in each of the two waves, thereby jus-tifying an aggregation of the Ab, Aad, and CogAss responses acrossrespondents.

3.1.4. Ad emotional content and creativityTwo independent groups of judges (12 judges for wave 1 and 24

judges for wave 2)—graduate students in marketing who were blindto the study's hypotheses—rated the emotional content and creativityof each ad (see Table 3 for details). There were six judges per ad. Eachwave-1 judge coded half of the wave-1 ads, whereas each wave-2judge coded one quarter of the wave-2 ads. The ad sequence was rotat-ed across judges. After viewing each ad, the judges first rated the extentto which the ad made them feel various emotions (e.g., excited, senti-mental) on a series of 1 (not at all) to 7 (very much) scales, with theorder of the items rotated across judges. Given that the vast majorityof TV commercials in Belgium are positively valenced, we measuredonly positive emotions (i.e., warmth-type, excitement-type, andhappiness/cheerfulness-type feelings, cf. Burke & Edell, 1989; Edell &Burke, 1987; Holbrook & Batra, 1987). To control for the possibilitythat ratings of emotional responses to the ads may reflect some otheraspects of the ads, such as their originality or creativity, the judgeswere also asked to rate the ads in terms of creativity.

Again, there was substantial inter-judge agreement in terms ofhow the different judges rated both the emotional content and thecreativity of the ads, justifying their averaging across judges foreach ad. Although the emotional-content items were expected tocapture different types of positive emotions (warmth, excitement,and happiness), a factor analysis of the judges' average responsesto the individual emotional items suggested a single dominantpositive-emotion factor, accounting for 71.1% of the variance acrossitems in wave 1 and for 77% of the variance across items in wave 2.Consequently, we computed a single-factor score of positive emotionfor each ad, which served as themain independent variable at the ag-gregate level. With respect to the ad creativity items, the ratings

Table 1Dataset characteristics.

Wave 1 Wave 2 Total

# ads 407 663 1070# products 93 122 153# brands 197 312 318# respondents 854 722 1576Gender % men 46.7 44.7 45.8

% women 53.3 55.3 54.2Age % 15–24 46.9 23.5 36.3

% 25–34 20.0 22.2 21.0% 35–44 21.6 34.7 27.5% 45–55 11.5 19.6 15.2

Education % vocational school degree 40.3 23.7 31.2% high school degree 37.8 55.2 47.3% technical or community college degree 17.5 12.9 15.4% university degree 4.4 8.1 6.2


Table 2Consumer respondent measures.

Wave 1 Wave 2

Inter-respondentagreement (range)

Inter-itemconsistency

Inter-respondentagreement (range)


Aad .91 .94“I like this ad”(Wave1: 1 = not at all; 10 = very much; Wave2: !3 = totally disagree, +3 = totally agree)

.80–.96 .77–.98

“The ad is well made”(Wave1: 1 = totally disagree; 7 = totally agree; Wave2: !3 = totally disagree, +3 = totally agree)

.81–.96 .76–.94

“My general evaluation of the ad is…”

(Wave1: 1 = very negative; 7 = very positive; Wave2: !3 = very negative, +3 = very positive).81–.97 .77–.93

Cognitive assessment .81“The ad gives useful information”(Wave1: 1 = totally disagree; 7 = totally agree; Wave2: !3 = totally disagree, +3 = totally agree)

.81–.94 .79–.93

“The ad is believable”(Wave1: 1 = totally disagree; 7 = totally agree; Wave 2: item not measured)

.77–.94

Ab .74 .85“My evaluation of the brand is…”

(Wave 1: 1 = very negative; 7 = very positive; Wave 2: !3 = very negative, +3 = very positive).68–.94 .68–.93

“If I need the product, I would buy this brand” a

(Wave 1: 1 = certainly not; 4 = certainly; Wave 2: !3 = totally disagree, +3 = totally agree).77–.95 .79–.96

a Although this item can be seen as a measure of behavioral intention, we treat it as a measure of brand attitude because of its high correlation with the other Ab item and thenotion that attitudes also have a behavioral (conative) component.

Table 3Ad-level and product category-level codings.

Wave 1 Wave 2

Interjudgereliabilitiesa


Mean (SD) Interjudgereliabilitiesa


Mean (SD)

Ad emotional content .97 2.40 (.81) .96 2.84 (.95)Warmthb

“Sentimental” .76–.78 .69–.77“Emotional” .75–.78 .67–.75“Moves me” .78–.81 .69–.75“Warm-hearted” .76–.76“Touches me” .73–.75

Excitementb

“Energetic” .61–.79 .69–.79“Excited” .72–.77 .68–.78“Enthusiastic” .65–.74 .68–.79“Upbeat” .69–.75“Stimulated” .66–.77

Happiness/cheerfulnessb

“Cheerful” .74–.83 .70–.79“Joyful” .72–.82 .71–.81“In a good mood” .72–.84 .71–.81“Happy” .75–.82“Delighted” .74–.84

Ad creativityb .99 2.90 (1.87) .99 3.54 (1.29)“Unique” .80–.83 .75–.87“Original” .77–.86 .82–.85“Creative” .81–.82 .78–.86“Novel” .79–.84“Unlike other ads” .83–.84

Product involvement .93 3.85 (.98) .94 3.76 (1.04)“[x] means a lot/nothing to consumers” .72 .82“[x] is of great/little concern to consumers” .68 .80“Choosing [x] is/is not an important decision” .80 .90“There is substantial/no risk involved with [x]” .74 .83“It is/is not a big deal if consumers make a mistake when buying [x]” .78 .89

Hedonic vs. utilitarian product .96 4.01 (1.42) .95 3.99 (1.42)“[x] is more a luxury than a necessity/more a necessity than a luxury” .81 .93“The benefits are primarily hedonic/functional” .87 .92“With [x] sensations and sensory stimulations play an important/minor role” .81 .86“The motivation for using [x] is mainly emotional/rational” .87 .93“This is a product/service consumers use for pleasure or to impress others/to address oravoid problems”

.86 .92

a Entries represent the range of interjudge reliabilities across the different sets of judges for the measures of Ad emotional content and Ad creativity. For Product involvement andHedonic vs. utilitarian product only one set of judges was used.

b In wave 2, only three of the five items used in wave 1 were used to assess Ad emotional content and Ad creativity.


across judges were also internally consistent across items and weretherefore averaged to form a single score of creativity for each ad.

3.1.5. Product-category-level involvement and motivation typeTo test potential product-category-level moderators of the effects of

ad-evoked emotional feelings, two other sets of six judges (who werealso blind to the hypotheses) coded the 153 product categories repre-sented in the ads. The first set of judges coded the 93 product categoriesfeatured in the wave-1 ads, whereas the second set of judges coded anadditional 60 categories featured in wave-2 ads that were not includedin the wave-1 ads. Unlike the other judges, these judges did not watchthe ads. Instead, they were simply given the names of the product/service categories (e.g., “cell phones,” “employment agency,” “pizza,”“paper towels”) and asked to rate each category on five semantic differ-ential items assessing product involvement (based on Laurent &Kapferer, 1985; Zaichkowsky, 1985) and five semantic differentialitems assessing hedonic versus utilitarian motives (adapted from Dhar& Wertenbroch, 2000; Voss, Spangenberg, & Grohmann, 2003) (seeTable 3). The order of both the product category names and the codingitems was rotated across judges. Inter-judge agreement on each ofthese items was again high, justifying an aggregation across judges. Inboth wave 1 and wave 2 a factor analysis of the 10 averaged items re-vealed two separate factors: the five hedonic-utilitarian items loadedon the first factor (wave 1: 45.7% and wave 2: 51.86% of the variance),and the five involvement items loaded on the second factor (wave 1:38.8% and wave 2: 31.57% of the variance). The factor scores served asindependent variables at the aggregate level.

3.1.6. Durable–nondurable–service and search–experience productcategorization

Finally, another set of four judges was asked to categorize each ofthe 153 product categories into (1) “a durable product, that is, a tan-gible product that lasts some time, such as refrigerators,” (2) “anondurable product, that is, a tangible product that usually is fullyconsumed in one or a few uses, such as beer or soap,” or (3) “a ser-vice, that is, an intangible, inseparable, variable, and short-term pro-vided product such as a haircut or a phone subscription” (inter-judgeagreement = 97.4%; disagreements resolved by majority rule). Thesame four judges additionally rated the extent to which each of the153 product categories was a search good or an experience goodusing two 5-point semantic differential scales inspired by Nelson(1970, 1974): (1) “This is a product for which it is easy for the con-sumer to evaluate the quality before the purchase by inspection(e.g., clothing, cookware, luggage, furniture, etc.)” versus “This is aproduct for which consumers (who have not tried the brand before)can only evaluate the quality after the purchase by consuming and

experiencing the product (e.g., beer, food, movies, tennis rackets)”;and (2) “The quality of this product or service is pretty obviouseven without trying it” versus “The quality of this product or servicecan only be determined by trying it and experiencing it.” The order ofboth the product category names and the coding items was rotatedacross judges. Inter-judge agreement was ! = .78 for both items,again justifying an aggregation.

3.1.7. Data purification of wave 2Although the overall pattern of data was generally consistent across

data-collection waves, preliminary analyses indicated a substantiallygreater level of noise in the second wave. We attribute this greaternoise to the fact that respondents in the second wave had to watchand evaluate on average more than twice as many commercials as re-spondents in the first wave (for example, many wave-2 respondentsevaluated between 75 and 101 commercials). Because different respon-dents saw a different number of ads and because the order in which theads were seen by individual respondents was not available to us, it wasnot possible to restrict our analyses to the first 20 or so advertisementsthat eachwave-2 respondent evaluated. Therefore, in order tomake thewave-2 datamore comparable in terms of qualitywith thewave-1 data,the former were purified as follows. First, we eliminated any responsesto a given ad by a given respondent that indicated a clear lack of carethrough either two or more missing values across items or identical re-sponses across all items (e.g., 5, 5, 5, 5…). This resulted in the elimina-tion of 8325 out of 41,320 total observations (20.2%). Next, weestimated whether respondents were diligent in their responses orresponded quasi-randomly by fitting, for each respondent, a regressionmodel in which the respondent's attitude toward a given brand was tobe explained by the respondent's attitude toward the ad (Aad) and therespondent's brand familiarity. Any respondent whose regression R2

was less than .10 was considered to be a quasi-random responder andthus dropped from the data. This resulted in the further elimination of2794 observations (6.8%) from 77 respondents. Finally, we restrictedthe data to those respondents who showed a strong degree of internalconsistency of .80 in Aad responses across advertisements. This resultedin the further elimination of 1522 observations (3.7%). The purifiedwave-2 data set thus consists of 28,679 observations (69.4%) from 722respondents evaluating a total of 663 commercials.1

1 If all wave-2 respondents are retained in the analyses, the results are directionallythe same as reported in Tables 6 and 7, but are statistically weaker. The main differencein the results is that the emotion ! hedonic/utilitarian motive interaction does notreach significance in Models 7 and 8 using hierarchical linear model regression, al-though it does reach significance in OLS regression.

Table 4Correlations (wave 1, n = 407).

Ab Aad Ad emotionalcontenta

Cognitiveassessment

Ad creativity Productinvolvement

Hedonic vs.utilitarianproductb

Durable or service (0)vs non-durable (1)

(Non-)Durable(0) vs service (1)

Search vs.experienceproductc

Ab 1 .607⁎⁎⁎ .453⁎⁎⁎ .590⁎⁎⁎ .222⁎⁎⁎ ! .352⁎⁎⁎ .177⁎⁎⁎ .266⁎⁎⁎ ! .107⁎ .422⁎⁎⁎

Aad 1 .676⁎⁎⁎ .692⁎⁎⁎ .609⁎⁎⁎ ! .133⁎⁎ .274⁎⁎⁎ .010 .011 .172⁎⁎⁎

Ad emotional contenta 1 .358⁎⁎⁎ .650⁎⁎⁎ ! .058 .334⁎⁎⁎ .050 ! .039 .243⁎⁎⁎

Cognitive assessment 1 .206⁎⁎⁎ ! .033 .023 ! .084 .173⁎⁎ .110⁎

Ad creativity 1 ! .001 .126⁎ ! .148⁎⁎ .077 ! .033Product involvement 1 .000 ! .617⁎⁎⁎ .345⁎⁎⁎ ! .430⁎⁎⁎

Hedonic vs. utilitarianb 1 .242⁎⁎⁎ ! .270⁎⁎⁎ .467⁎⁎

Durable/service (0) vs.non-durable (1)

1 ! .795⁎⁎ .573⁎⁎⁎

Non-durable/durable(0) vs. Service (1)

1 ! .378⁎⁎⁎

Search vs. experiencec 1

***p b .001; **p b .01; *p b .05 (two-tailed significance tests).a A higher score indicates more positive emotions.b A higher score indicates more hedonic motives.c A higher score indicates relatively more experience goods.


3.2. Results

Tables 4 and 5 report descriptive statistics for all the variables,along with the simple correlations among them. Because a significantproportion of the brands had more than one commercial and becausedifferent brands would often share the same product category (e.g.,Nissan andMercedes), the data were analyzed in a series of multilevelregression models in which emotional responses and their potentialmediators and moderators were treated as fixed effects, and brandand product-category effects were modeled as random interceptswith the brand effects nested in product category. These three-levelmodels, with ads nested in brands and brands nested in product cat-egories, adjust for any data dependencies that may exist across ads forthe same brands and across brands within the same product category

(simpler OLS regression models produced largely similar results). Allpredictors in the models were mean-centered, with the effects ofad-evoked feelings, involvement, and hedonic versus utilitarian mo-tives being additionally standardized as factor loadings.

3.2.1. Effects of ad-evoked feelingsTable 6 summarizes the results of four models testing the basic ef-

fects of ad-evoked feelings on brand evaluations and the potential me-diators of these effects. The results of Model 1, the “basic-feeling-effectmodel,” show that even under conditions of greater external validity,ad-evoked feelings indeed have a substantial influence on brand evalu-ations (! = .486, t = 13.69, p b .001). Interestingly, the simple corre-lation between ad-evoked feelings and brand evaluations was r =.331, which is roughly of the same magnitude as what had been

Table 5Correlations (wave 2, n = 663).

Ab Aad Ad emotionalcontenta

Cognitiveassessment

Ad creativity Productinvolvement

Hedonic vsutilitarianproductb

Durable or service (0)vs non-durable (1)

(Non-)Durable(0) vs service (1)

Search vs.experienceproductc

Ab 1 .501⁎⁎⁎ .256⁎⁎⁎ .316⁎⁎⁎ .116⁎⁎ ! .351⁎⁎⁎ ! .005 .262⁎⁎⁎ ! .227⁎⁎⁎ .184⁎⁎⁎

Aad 1 .496⁎⁎⁎ .130⁎⁎ .441⁎⁎⁎ ! .015 .090⁎ ! .032 ! .036 ! .120⁎⁎

Ad emotional contenta 1 ! .162⁎⁎⁎ .584⁎⁎⁎ ! .056 .120⁎⁎ ! .091⁎ .058 ! .111⁎⁎

Cognitive assessment 1 ! .324⁎⁎ ! .052 ! .344⁎⁎⁎ .155⁎⁎⁎ ! .077⁎ .129⁎⁎

Ad creativity 1 ! .086⁎ .011 ! .333⁎⁎⁎ ! .245⁎⁎⁎ ! .312⁎⁎⁎

Product involvement 1 .000 ! .574⁎⁎⁎ .297⁎⁎⁎ ! .387⁎⁎⁎

Hedonic vs. utilitarianb 1 .229⁎⁎⁎ ! .261⁎⁎⁎ .241⁎⁎⁎

Durable/service (0) vs.non-durable (1)

1 ! .712⁎⁎⁎ .697⁎⁎⁎

Non-durable/durable(0) vs. service (1)

1 ! .347⁎⁎⁎

Search vs. experiencec 1

***p b .001; **p b .01; *p b .05 (two-tailed significance tests).a A higher score indicates more positive emotions.b A higher score indicates more hedonic motives.c A higher score indicates relatively more experience goods.

Table 6Standardized regression coefficients and z-values (for random effects) and t-values (for fixed effects) (n = 1070 ads).

Model 1 Model 2 Model 3 Model 4

Basic-feeling-effect model Aad-mediation model Cognitive-assessment-and-Aad mediation model Ad-creativity-confound model

Model fitAIC 3073.9 2781.8 2712.3 2705.2BIC 3083.0 2790.9 2721.4 2714.3Chi2 437.19⁎⁎⁎ 493.58⁎⁎⁎ 459.63⁎⁎⁎ 446.33⁎⁎⁎

Random effectsProduct .750⁎⁎⁎ .523⁎⁎⁎ .442⁎⁎⁎ .433⁎⁎⁎

(5.83) (5.90) (5.73) (5.65)Brand .557⁎⁎⁎ .486⁎⁎⁎ .466⁎⁎⁎ .463⁎⁎⁎

(8.88) (10.23) (10.53) (10.56)Fixed effects

Emotional contenta .486⁎⁎⁎ .108⁎⁎ .162⁎⁎⁎ .199⁎⁎⁎

(13.69) (2.81) (4.31) (4.83)Waveb .030 .047 .038 ! .019

(.72) (1.27) (1.05) (! .50)Emotional contenta ! Waveb .279⁎⁎⁎ .080⁎ .113⁎⁎ .180⁎⁎⁎

(8.16) (2.11) (3.04) (4.46)Aad .692⁎⁎⁎ .445⁎⁎⁎ .510⁎⁎⁎

(18.07) (9.46) (10.03)Aad ! Waveb .216⁎⁎⁎ .022 .107⁎

(5.71) (.50) (2.14)Cognitive assessment .599⁎⁎⁎ .552⁎⁎⁎

(8.83) (7.90)Cognitive assessment ! Waveb .235⁎⁎⁎ .150⁎

(3.71) (2.28)Ad creativity ! .081⁎⁎

(!2.32)Ad creativity ! Waveb ! .139⁎⁎⁎

(!4.15)

Note. ***p b .001; **p b .01; *p b .05 (two-tailed significance tests).a A higher score indicates more positive emotional feelings.b Wave 1 = 1; wave 2 = !1.


Table 7Standardized regression coefficients and z-values (for random effects) and t-values (for fixed effects) (n = 1070 ads).

Model 5 Model 6 Model 7 Model 8 Model 9 Model 10

Involvement-moderationmodel

Hedonic/utilitarian-moderationmodel

Involvement-and-hedonic/utilitarian-moderation model

Product-durability-moderationmodel

Search/experience-moderationmodel

Full model

Model fitAIC 2630.6 2682.1 2637.2 2646.7 2665.7 2634.1BIC 2682.1 2733.7 2713.0 2710.1 2717.2 2745.8Chi2 326.63⁎⁎⁎ 432.70⁎⁎⁎ 310.09⁎⁎⁎ 339.89⁎⁎⁎ 344.37⁎⁎⁎ 290.87⁎⁎⁎

Random effectsProduct .253⁎⁎⁎ .425⁎⁎⁎ .244⁎⁎⁎ .315⁎⁎⁎ .327⁎⁎⁎ .237⁎⁎⁎

(4.93) (5.63) (4.80) (5.04) (4.99) (4.76)Brand .452⁎⁎⁎ .458⁎⁎⁎ .451⁎⁎⁎ .457⁎⁎⁎ .462 .454

(10.66) (10.60) (10.68) (10.63) (10.59) (10.77)Fixed effects

Emotional contenta .218⁎⁎⁎ .198⁎⁎⁎ .217⁎⁎⁎ .098 .174⁎⁎⁎ ! .010(5.14) (4.78) (5.09) (1.08) (4.22) (!0.06)

Aad .479⁎⁎⁎ .522⁎⁎⁎ .472⁎⁎⁎ .484⁎⁎⁎ .490⁎⁎⁎ .468⁎⁎⁎

(9.57) (10.18) (9.27) (9.58) (9.67) (9.18)Cognitive assessment .579⁎⁎⁎ .538⁎⁎⁎ .586⁎⁎⁎ .591⁎⁎⁎ .579⁎⁎⁎ .601⁎⁎⁎

(8.47) (7.62) (8.41) (8.50) (8.34) (8.55)Ad creativity ! .081⁎ ! .085⁎ ! .083⁎⁎ ! .057 ! .059 ! .074⁎

(!2.38) (!2.43) (!2.37) (!1.65) (!1.68) (!2.13)Product involvement ! .408⁎⁎⁎ ! .405⁎⁎⁎ ! .350⁎⁎⁎

(!7.69) (!7.33) (!4.58)Emotional contenta ! Involvement ! .021 ! .041 ! .043

(! .82) (!1.49) (.165)Hedonic vs. utilitarianb ! .063 .019 ! .015

(! .94) (.33) (! .24)Emotional contenta ! Hedonic/utilitarianb .054(⁎) .078⁎⁎ .060+

(1.75) (2.49) (1.72)Emotional ! Involvement ! Hedonic/utilitarianb ! .038 ! .049

(!1.20) (!1.51)Durable/service (0) vs. nondurable (1) .601** ! .054

(3.15) (! .23)Durable/non-durable (0) vs. service (1) ! .032 ! .275

(! .15) (!1.32)Emotional contenta ! Dur/serv vs. nondur .111 .091

(1.21) (.78)Emotional contenta ! (Non)Dur vs. service .025 .051

(.25) (.47)Search/experience productc .250⁎⁎⁎ .015

(3.40) (.23)Emotional contenta ! Search/experiencec .047 .039

(1.30) (.86)Waved ! .0162 ! .036 ! .017 ! .219* ! .010 ! .213

(! .38) (! .90) (! .39) (!2.28) (! .27) (!1.03)

Notes. ***p b .001; **p b .01; *p b .05; +p b .10 (two-tailed significance tests).a A higher score indicates more positive feelings.b A higher score indicates more hedonic motives.c A higher score indicates relatively more experience goods.d Wave 1 = 1; wave 2 = !1. Although not included in the table, all interactions with wave were also modeled.

390M.T.Pham

etal./Intern.J.ofResearch

inMarketing

30(2013)

383–394

observed in a meta-analysis of previous studies (Brown et al., 1998).Therefore, it appears that the basic effects are of genuinely substantialsize and are not driven by methodological artifacts such as the use ofstudent respondents, a selection bias in the ads used, and shared meth-od variance due to the repeated measurement of respondents. Howev-er, a significant interaction with wave (! = .279, t = 8.16, p b .001)indicated that the effects of ad-evoked feelings were substantiallystronger in wave 1 than they were in wave 2 (see also Tables 4 and5). The difference between the twowaves could be due to a genuine dif-ference in the strength of the effects depending on the pool of advertise-ments studied or tomethodological differences between the twowaves(e.g., in wave 2 respondents evaluated amuch larger number of ads andad-evoked feelings were assessed with fewer items). (As reported inTable 6, several other interactionswithwavewere uncovered inmodels2–4. These interactions are not discussed here because of their lowertheoretical and substantive importance.)

The results of Model 2, the “Aad-mediation model,” show that re-spondents' attitudes toward the ads (Aad) are strong predictors oftheir attitudes toward the brands (Ab) (! = .692, t = 18.07,p b .001) (cf. McKenzie et al., 1986; Mitchell & Olson, 1981). The re-sults additionally show that the effect of ad-evoked feelings is sub-stantially reduced when respondents' attitudes toward the ads arecontrolled for (! = .486 ! ! = .108), though the effect remainssignificant (t = 2.81, p = .005). Thus, according to Alwin andHauser's (1975) simple formula, as much as 78% of the effects ofad-evoked feelings on brand evaluations ([.486 ! .108] / .486 =0.78) may be mediated by changes in Aad. This result is consistentwith the notion that effects of ad-evoked feelings on brand evalua-tions are largely mediated by their effects on ad attitudes, as sug-gested by various authors (e.g., Batra & Ray, 1986; Holbrook &Batra, 1987; MacInnis & Park, 1991), but the mediation is not com-plete, as suggested by other authors (e.g., Burke & Edell, 1989;Stayman & Aaker, 1988).

The results of Model 3, the “cognitive-assessment and Aad-mediationmodel,” show that while respondents' cognitive assessments of the adsinfluenced their brand evaluations (! = .599, t = 8.83, p b .001), theseassessments did not attenuate the effects of ad-evoked feelings onbrand evaluations (! = .162, t = 4.31, p b .001). Although these cogni-tive assessments of the ads are not actual measures of brand beliefs, thisfinding seems somewhat inconsistent with a pure belief-based explana-tion of the phenomenon.

Finally, the results of Model 4, the “ad-creativity-confoundmodel,” show that the direct effects of ad-evoked feelings on brandevaluations remain (! = .199, t = 4.83, p b .001) even after con-trolling for differences in ad creativity (! = ! .081, t = !2.32,p = .02). This finding suggests that the observed effects of ad-evoked feelings are not confounded by executional elements of theads such as their creativity.

Overall, the results of these initial analyses support two empiricalgeneralizations. First, even under conditions of greater external validity,ad-evoked feelings exert a substantial positive influence on brand evalu-ations (EG1). Second, ad-evoked feelings have both direct and indirecteffects on brand evaluations, with the indirect effects being strongerand largely linked to changes in Aad (EG2).

3.2.2. Product-category moderators of the effects of ad-evoked feelingsTable 7 summarizes the results of five models testing the poten-

tial product-category-level moderators of the effects of ad-evokedfeelings on brand evaluations. The results of Model 5, the“involvement-moderation model,” indicate that while category-levelinvolvement has a “main” effect on brand evaluations—brand evalua-tions were less favorable for high-involvement products than forlow-involvement products (! = ! .408, t = !7.69, p b .001)—it didnot moderate the effects of ad-evoked feelings on brand evaluations(! = ! .021, t = !0.82, p = .414).

The results of Model 6, the “hedonic/utilitarian-moderation model,”indicate that the effects of ad-evoked feelings on brand evaluationswere marginally stronger for product categories typically associatedwith hedonic motives than for product categories typically associatedwith utilitarian motives (! = .054, t = 1.75, p = .081). This findingwas not qualified by an interaction with wave (! = .033, t = 1.09,p = .275).

Because popular planning models such as the FCB grid and theRossiter–Percy grid (Rossiter, Percy, & Donovan, 1991) generally con-ceptualize the advertising effects of product involvement and productmotive (hedonic vs. utilitarian) in a two-dimensional space, Model 7,the “involvement- and hedonic/utilitarian-moderation model,” testsjointly for the moderating effects of category-level involvement andtype of motive. The results again indicate that product-category-levelinvolvement does not moderate the effects of ad-evoked feelings onbrand evaluations (! = ! .041, t = !1.49, p = .137). However, thetype of motive does moderate the effects of ad-evoked feelings onbrand evaluations,with the effects being stronger for product categoriestypically associated with hedonic motives than for product categoriestypically associated with utilitarian motives (! = .078, t = 2.49, p =.013) (see Fig. 1). Spotlight analyses (Aiken & West, 1996) showthat the effect (slope) of ad-evoked feelings was significant for therelatively more hedonic product categories (slope = .350, SE =.053, t = 6.57, p b .001), but only marginally significant for the rela-tively more utilitarian product categories (slope = .098, SE = .056,t = 1.74, p = .082). Again, the effects were not qualified by an inter-action with wave (! = .049, t = 1.59, p = .112). (There was nothree-way interaction among feelings, involvement, and type of mo-tive [! = ! .038, t = !1.10, p = .231].)

Overall, the results of Models 5 through 7 support two additionalempirical generalizations. First, the effects of ad-evoked feelings on

-0.4

-0.3

-0.2

-0.1

0

0.1

0.2

0.3

0.4

0.5

-1 SD +1 SD

Att

itude

tow

ard

the

bran

d

Positivity of Emotional Content

"-1SD" Relatively MoreUtilitarian Product Categories

"+1 SD" Relatively MoreHedonic Product Categories

Fig. 1. Emotional content ! Type of motive interaction on brand attitude.


brand evaluations do not appear to depend on the level of involvementtypically associated with the product category (EG3). Second, the effectsof ad-evoked feelings on brand evaluations are more pronounced forproducts typically associated with hedonic motives than for productstypically associated with utilitarian motives (EG4).

The results of Model 8, the “product-durability-moderation model,”indicate that while brand evaluations were more favorable fornondurable products than for durable products (! = .601, t = 3.15,p = .002), the effects of ad-evoked feelings on brand evaluations didnot depend on whether the product was a durable, a nondurable, or aservice (! = .111, t = 1.21, p = .226; ! = .025, t = 0.25, p = .805).Similarly, the results of Model 9, “the search/experience-moderationmodel,” indicate that while brand evaluations were more favorable forexperience products than for search products (! = .250, t = 3.40,p = .001), the effects of ad-evoked feelings on brand evaluations didnot depend onwhether the product was a search good or an experiencegood (! = .047, t = 1.30, p = .195). Thus, even though the two gener-al product characteristics have been found to be important moderatorsof the overall effectiveness of advertising (Vakratsas inHanssens, 2009),product durability and the search-versus-experience nature of the good donot appear to moderate the effects of ad-evoked feelings on brand evalua-tions (EG5).2

The results of the final model, Model 10 (the “full model”), supportthe inferences suggested by the more restricted models. Category-level involvement did not moderate the effects of ad-evoked feelingson brand evaluations (! = ! .043, t = !1.39, p = .165), but theeffects of ad-evoked feelings were marginally stronger for hedonicproducts than for utilitarian products (! = .060, t = 1.72, p = .085).(The three-way interaction among ad-evoked feelings, involvement,and motive was not significant: p = .135.) The interactions betweenad-evoked feelings and the product-type dummies were not significant(t's b 1), suggesting that the effects of ad-evoked feelings did not de-pend onwhether the product was a durable, a nondurable, or a service.Finally, there was no interaction between ad-evoked feelings andwhether the product was a search good or an experience good (t b 1).

4. General discussion

Considering the practical significance of the effects of ad-evokedfeelings on brand evaluations, it is somewhat surprising that the empir-ical generalizability of this phenomenon—in terms of both external va-lidity and boundary conditions across product categories—had yet to besystematically investigated. Our study addresses this void by analyzingconsumer responses to a total of 1070 brand TV commercials frommorethan150 different product categories. Unlike previous studies that ofteninvolved student respondents, limited samples of often fictitious ads,and repeated measurement of respondents, our study (a) examinedthe evaluation responses of a large and broadly representative sampleof actual consumers, (b) was based on a virtual census of all ads (forreal brands) shown by two national TV channels during a three-year

period, and (c) used a design that reduces issues of shared methodvariance.

The results show that even under conditions that are closer to mar-ketplace settings than most previous academic studies, ad-evoked feel-ings indeed have a substantial impact on brand evaluations (EG1). Theeffect size was in fact quite comparable to that found in a meta-analysis of these earlier studies (Brown et al., 1998). The results addi-tionally show that the effects of ad-evoked feelings on brand evalua-tions are both direct and indirect, with the indirect effects beingsubstantially stronger and largely linked to a change in Aad (EG2).The finding that ad-evoked feelings have both direct and indirect effectshelps reconcile previously conflicting results that documented eitherone or the other.

Important additional results pertain to the category-level moderatorsof the phenomenon. First, we found little evidence of amoderating role ofproduct-category-level involvement (EG3). Given the very large numberof observations in our study and the recorded reliability of the measures,we believe that this lack of moderating effect of category-level involve-ment is more than an artifact of poor measurement or low statisticalpower. Rather, it is a substantive and generalizable result. We suspectthat natural variations in consumer involvement as a function of productcategory are not very strong in the real world due to substantial inter-consumer heterogeneity in involvementwithin a product category. In ad-dition, it is possible that feelings can influence judgments under condi-tions of both low involvement and high involvement, albeit throughdifferent mechanisms.

Second, while the level of involvement with the product categorydoes not appear to significantly moderate the effects of ad-inducedfeelings on brand evaluations, the type of motive typically associatedwith the category does. The effects of ad-induced feelings on brandevaluations appear to be significantly more pronounced when theproduct category is more hedonic than when it is more utilitarian(EG4). This is consistent with research in the affect-as-informationliterature indicating that consumers are more likely to rely on theirmomentary feelings in judgments and decisions when they haveexperiential motives than when they have instrumental motives(Pham, 1998; see also Adaval, 2001; Yeung & Wyer, 2004).

Although our study design does not allow for strong process infer-ences, this finding has potential theoretical implications for researchon both the basic phenomenon and the affect-as-information frame-work. With respect to the former, our findings suggest that the robusteffects of ad-evoked feelings on brand evaluations may not be drivensolely by affect transfer, evaluative conditioning, and differences inbrand beliefs and thoughts, as previously suggested. The phenome-non may also be driven by HDIF inferences during ad exposure,whereby consumers may interpret their feeling responses to the adas indicative of how much they like or dislike the brand. With respectto the affect-as-information literature, it is important to note that inour studies consumers who were not explicitly asked to assess theirfeelings toward the various brands nevertheless appeared to incorpo-rate ad-evoked feelings selectively as a function of the type of motivetypically associated with the advertised product category. This find-ing may suggest that the selective reliance on feelings as a functionof their relevance for the judgment at hand is a very spontaneousand flexible process that is much more flexible than conceptualizedin early affect-as-information research (Schwarz & Clore, 1983).

Finally, while product durability and the search-versus-experiencenature of the good have been found to be important moderators ofthe overall impact of advertising, we found no evidence that thesetwo general product characteristics play any role in moderating theeffects of ad-evoked feelings on brand evaluations (EG5).

One obvious limitation of the study is that because the units ofanalysis were at the aggregate ad level, rather than at the individualrespondent level, the results do not allow strong theoretical infer-ences about the actual psychological processes at work. In addition,it would have been useful to control for prior brand attitude, which

2 Following a reviewer's suggestion, we also examined whether the effects of ad-evoked feelings depended on the familiarity of the brand. One would generally predicta greater influence of feelings for unknown or unfamiliar brands (e.g., Miniard et al.,1990; Park & Young, 1986). However, it has also been observed that affective ads areparticularly effective for well-known, mature brands (MacInnis, Rao, & Weiss, 2002),leaving the possibility of a curvilinear effect of brand familiarity. Because brand famil-iarity was not assessed in wave 1, we restricted our analyses to the second wave of da-ta. We tested an extension of Model 1 that additionally included a (centered) term forbrand familiarity, the square of that term, plus the interaction between familiarity andad-evoked feelings and the interaction between familiarity-square and ad-evoked feel-ings. The results show that brand familiarity had a positive effect on brand evaluations(! = 0.414, t = 9.75, p b .001). However, neither the familiarity ! feelings interac-tion (! = 0.056, t = 1.52, p = .13) nor the familiarity-square ! feelings interaction(! = 0.025, t = 1.07, p = .28) emerged as significant. These null effects may be dueto a lack of statistical power, or to a genuine possibility that the effects of ad-evokedfeelings are equally pronounced for familiar and less familiar brands, which would ex-plain the mixed results observed in previous studies.


we were unable to do with these data. Moreover, given evidence ofqualitative difference among distinct emotions (e.g., Batra & Ray,1986; Raghunathan, Pham, & Corfman, 2006), it would also havebeen interesting to distinguish among subtypes of ad-evoked feelings,which was not possible here because the subtypes of feelings that weassessed were too correlated.

Our findings have obvious managerial implications. Advertisers dobenefit substantially from advertisements that elicit pleasant emo-tional feelings, not only in terms of greater ad liking, but more impor-tantly in terms of more favorable brand attitudes. The total effects ofad-evoked feelings on brand evaluations are substantial. In addition,they appear to be rather general across product categories. Theyapply equally to (a) low- and high-involvement products, (b) durableproducts, nondurable products, and services, and (c) search and ex-perience products. However, the effects are more likely to occur inproduct categories typically associated with hedonic motives than inproduct categories typically associated with utilitarian motives.While one could think that the latter proposition should be obviousto advertisers, there is evidence that it is not. For example, it hasbeen found that although food products are typically associatedwith hedonic and experiential motives in consumers' minds, foodadvertisers still tend to rely on informational appeals rather thanmore emotional ones (Dube et al., 1996). Hence the importance ofrevisiting what is assumed to be known, even after 25 years.

References

Aaker, D. A., Stayman, D. M., & Hagerty, M. R. (1986). Warmth in advertising: Mea-surement, impact and sequence effects. Journal of Consumer Research, 12(4),365–381.

Adaval, R. (2001). Sometimes it just feels right: The differential weighting ofaffect-consistent and affect-inconsistent product information. Journal of ConsumerResearch, 28(1), 1–17.

Aiken, L. S., & West, S. G. (1996). Multiple regression: Testing and interpreting interac-tions. Newbury Park, CA: Sage Publications.

Albarracín, D., & Wyer, R. S. (2001). Elaborative and non-elaborative processing of abehavior-related communication. Personality and Social Psychology Bulletin, 27(6),691–705.

Alwin, D. F., & Hauser, R. M. (1975). Decomposition of effects in path analysis. AmericanSociological Review, 40(1), 37–47.

Bakamitsos, G. A. (2006). A cue alone or a probe to think? The dual role of affect inproduct evaluations. Journal of Consumer Research, 33(3), 403–412.

Batra, R., & Ray, M. L. (1986). Affective responses mediating acceptance of advertising.Journal of Consumer Research, 13(2), 234–249.

Batra, R., & Stayman, D. M. (1990). The role of mood in advertising effectiveness. Jour-nal of Consumer Research, 17(2), 203–214.

Batra, R., & Stephens, D. (1994). Attitudinal effects of ad-evoked moods and emotions —The moderating role of motivation. Psychology and Marketing, 11(3), 199–215.

Baumgartner, H., Sujan, M., & Padgett, D. (1997). Patterns of affective reactions to ad-vertisements: The integration of moment-to-moment responses into overall judg-ments. Journal of Marketing Research, 34(2), 219–232.

Bloch, P. H., & Richins, M. L. (1983). A theoretical model for the study of product impor-tance perceptions. Journal of Marketing, 47(3), 69–81.

Bower, G. H. (1981). Mood and memory. American Psychologist, 36(2), 129–148.Brown, S. P., Homer, P. M., & Inman, J. J. (1998). A meta-analysis of relationships be-

tween ad-evoked feelings and advertising responses. Journal of Marketing Research,35(1), 114–126.

Brown, S. P., & Stayman, D. M. (1992). Antecedents and consequences of attitudetoward the ad — A meta analysis. Journal of Consumer Research, 19(1), 34–51.

Burke, M. C., & Edell, J. A. (1989). The impact of feelings on ad-based affect and cogni-tion. Journal of Marketing Research, 26(1), 69–83.

Chang, H.H., & Pham,M. T. (2013). Affect as a decision-making systemof the present. Jour-nal of Consumer Research, 40(1), 42–63.

Cho, H. G., & Stout, P. A. (1993). An extended perspective on the role of emotion inadvertising processing. Advances in Consumer Research, 20, 692–697.

Cohen, J. (1988). Statistical power analysis for the behavioral science. Hillsdale, NJ:Lawrence Erlbaum Associates, Inc.

De Houwer, J., Thomas, S., & Baeyens, F. (2001). Associative learning of likes anddislikes: A review of 25 years of research on human evaluative conditioning. Psy-chological Bulletin, 127(6), 853–869.

Derbaix, C. M. (1995). The impact of affective reactions on attitudes toward the adver-tisement and the brand. A step toward ecological validity. Journal of MarketingResearch, 32(4), 470–479.

Dhar, R., & Wertenbroch, K. (2000). Consumer choice between hedonic and utilitariangoods. Journal of Marketing Research, 37(1), 60–71.

Dube, L., Chattopadhyay, A., & Latarte, A. (1996). Should advertising appeals match thebasis of consumers' attitudes? Journal of Advertising Research, 36(6), 82–89.

Eagly, A. H., & Chaiken, S. (1993). The psychology of attitudes. Orlando, FL: HarcourtBrace Jovanovich College Publishers.

Edell, J. A., & Burke, M. C. (1987). The power of feelings in understanding advertisingeffects. Journal of Consumer Research, 14(3), 421–433.

Feldman, J. M., & Lynch, J. G. (1988). Self-generated validity and other effects of mea-surement on belief, attitude, intention, and behavior. Journal of Applied Psychology,73(3), 421–435.

Fishbein, M., & Middlestadt, S. (1995). Noncognitive effects on attitude formation andchange: Fact or artifact? Journal of Consumer Psychology, 4(2), 181–202.

Forgas, J. P. (1995). Mood and judgment: The affect infusion model (AIM). PsychologicalBulletin, 117(1), 39–66.

Gorn, G. J. (1982). The effects of music in advertising on choice behavior: A classicalconditioning approach. Journal of Marketing, 46(1), 94–101.

Hanssens, D. M. (2009). Empirical generalizations about marketing impact. Cambridge,MA: Marketing Science Institute, Relevant Knowledge Series.

Holbrook, M. B., & Batra, R. (1987). Assessing the role of emotions as mediators of con-sumer responses to advertising. Journal of Consumer Research, 14(3), 404–420.

Homer, P. M., & Yoon, S. G. (1992). Message framing and the interrelationships amongad-based feelings, affect, and cognition. Journal of Advertising, 21(1), 19–33.

Isen, A. M., Shalker, T. E., Clark, M., & Karp, L. (1978). Affect, accessibility of material inmemory, and behavior — Cognitive loop. Journal of Personality and Social Psycholo-gy, 36(1), 1–12.

Jones, C. R., Olson, M. A., & Fazio, R. H. (2010). Evaluative conditioning: The “how”question. In J. M. Olson, & M. P. Zanna (Eds.), Advances in experimental social psy-chology, 43. (pp. 205–255)San Diego: Academic Press.

Laurent, G., & Kapferer, J. -N. (1985). Measuring consumer involvement profiles. Jour-nal of Marketing Research, 22(1), 41–53.

Machleit, K. A., & Wilson, R. D. (1988). Emotional feelings and attitude toward the ad-vertisement — The roles of brand familiarity and repetition. Journal of Advertising,17(3), 27–35.

MacInnis, D. J., & Park, C. W. (1991). The differential role of characteristics of music onhigh- and low-involvement consumers' processing of ads. Journal of Consumer Re-search, 18(2), 161–173.

MacInnis, D. J., Rao, A. G., & Weiss, A. M. (2002). Assessing when increased mediaweight of real-world advertisements helps sales. Journal of Marketing Research,39(4), 391–407.

MacKenzie, S. B., & Lutz, R. J. (1989). An empirical examination of the structural ante-cedents of attitude toward an ad in an advertising pretesting context. Journal ofMarketing, 53(2), 48–65.

MacKenzie, S. B., Lutz, R. J., & Belch, G. E. (1986). The role of attitude toward the ad as amediator of advertising effectiveness: A test of competing explanations. Journal ofMarketing Research, 23(2), 130–143.

Madden, T. J., Allen, C. T., & Twible, J. L. (1988). Attitude toward the ad: An assessmentof diverse measurement indexes under different processing sets. Journal of Market-ing Research, 25(3), 242–252.

Miniard, P. W., Bhatla, S., Lord, K. R., Dickson, P. R., & Unnava, H. R. (1991).Picture-based persuasion processes and the moderating role of involvement. Jour-nal of Consumer Research, 18(1), 92–107.

Miniard, P. W., Bhatla, S., & Rose, R. L. (1990). On the formation and relationship of adand brand attitudes — An experimental and causal analysis. Journal of MarketingResearch, 27(3), 290–303.

Mitchell, A. A., & Olson, J. C. (1981). Are product attribute beliefs the only mediator ofadvertising effects on brand attitude? Journal of Marketing Research, 18(3),318–332.

Moore, D. L., & Hutchinson, J. W. (1983). The effects of ad affect on advertising effec-tiveness. Advances in Consumer Research, 10, 526–531.

Morris, J. D., Woo, C. M., Geason, J. A., & Kim, J. (2002). The power of affect: Predictingintention. Journal of Advertising Research, 42(3), 7–17.

Nelson, P. (1970). Information and consumer behavior. Journal of Political Economy,78(2), 311–329.

Nelson, P. (1974). Advertising as information. Journal of Political Economy, 82(4),729–754.

Olney, T. J., Holbrook, M. B., & Batra, R. (1991). Consumer responses to advertising —The effects of ad content, emotions, and attitude toward the ad on viewing time.Journal of Consumer Research, 17(4), 440–453.

Ottati, V. C., & Isbell, L. M. (1996). Effects of mood during exposure to target informa-tion on subsequently reported judgments: An on-line model of misattributionand correction. Journal of Personality and Social Psychology, 71(1), 39–53.

Park, C. W., & Young, S. M. (1986). Consumer response to television commercials— Theimpact of involvement and background music on brand attitude formation. Journalof Marketing Research, 23(1), 11–24.

Petty, R. E., & Cacioppo, J. T. (1986). The elaboration likelihood model of persuasion. In L.Berkowitz (Ed.), Advances in experimental social psychology, 19. (pp. 123–205)Orlando,FL: Academic Press.

Petty, R. E., Schumann, D. W., Richman, S. A., & Strathman, A. J. (1993). Positive moodand persuasion: Different roles for affect under high- and low-elaboration condi-tions. Journal of Personality and Social Psychology, 64(1), 5–20.

Pham, M. T. (1998). Representativeness, relevance, and the use of feelings in decisionmaking. Journal of Consumer Research, 25(2), 144–159.

Pham, M. T. (2004). The logic of feeling. Journal of Consumer Psychology, 14(4),360–369.

Pham, M. T., Cohen, J. B., Pracejus, J. W., & Hughes, G. D. (2001). Affect monitoring and theprimacy of feelings in judgment. Journal of Consumer Research, 28(2), 167–188.

Pham, M. T., Meyvis, T., & Zhou, R. R. (2001). Beyond the obvious: Chronic vividness ofimagery and the use of information in decision making. Organizational Behaviorand Human Decision Processes, 84(2), 226–253.


Raghunathan, R., Pham, M. T., & Corfman, K. P. (2006). Informational properties of anxietyand sadness, and displaced coping. Journal of Consumer Research, 32(4), 596–602.

Rossiter, J. R., Percy, L., & Donovan, R. J. (1991). A better advertising planning grid. Jour-nal of Advertising Research, 31(5), 11–21.

Schwarz, N., & Clore, G. L. (1983). Mood, misattribution, and judgments of well-being—Informative and directive functions of affective states. Journal of Personality andSocial Psychology, 45(3), 513–523.

Schwarz, N., & Clore, G. L. (2007). Feelings and phenomenal experiences. In E. T.Higgins, & A. W. Kruglanski (Eds.), Social psychology: Handbook of basic principles(pp. 385–407) (2nd ed.). New York: Guilford.

Stayman, D. M., & Aaker, D. A. (1988). Are all the effects of ad-induced feelings medi-ated by Aad? Journal of Consumer Research, 15(3), 368–373.

Vakratsas, D., & Ambler, T. (1999). How advertising works: What do we really know?Journal of Marketing, 63(1), 26–43.

Vaughn, R. (1980). How advertising works: A planning model. Journal of AdvertisingResearch, 20(5), 27–33.

Vaughn, R. (1986). How advertising works: A planning model revisited. Journal ofAdvertising Research, 26(1), 57–66.

Voss, K. E., Spangenberg, E. R., & Grohmann, B. (2003). Measuring the hedonic and util-itarian dimensions of consumer attitude. Journal of Marketing Research, 40(3),310–320.

White, K., & McFarland, C. (2009). When are moods most likely to influence consumers'product preferences? The role of mood focus and perceived relevance of moods.Journal of Consumer Psychology, 19(3), 526–536.

Yeung, C. W. M., & Wyer, R. S. (2004). Affect, appraisal, and consumer judgment. Jour-nal of Consumer Research, 31(2), 412–424.

Zaichkowsky, J. L. (1985). Measuring the involvement construct. Journal of ConsumerResearch, 12(3), 341–352.


Intern. J. of Research in Marketingportal.idc.ac.il/en/main/research/ijrm/documents/pdf of...

Documents

Transcript of Intern. J. of Research in Marketingportal.idc.ac.il/en/main/research/ijrm/documents/pdf of...