just the faces: exploring the effects of facial features...

This article was downloaded by: [128.118.207.120] On: 08 December 2014, At: 12:19Publisher: Institute for Operations Research and the Management Sciences (INFORMS)INFORMS is located in Maryland, USA

Marketing Science

Publication details, including instructions for authors and subscription information:http://pubsonline.informs.org

Just the Faces: Exploring the Effects of Facial Features inPrint AdvertisingLi Xiao, Min Ding

To cite this article:Li Xiao, Min Ding (2014) Just the Faces: Exploring the Effects of Facial Features in Print Advertising. Marketing Science33(3):338-352. http://dx.doi.org/10.1287/mksc.2013.0837

Full terms and conditions of use: http://pubsonline.informs.org/page/terms-and-conditions

This article may be used only for the purposes of research, teaching, and/or private study. Commercial useor systematic downloading (by robots or other automatic processes) is prohibited without explicit Publisherapproval, unless otherwise noted. For more information, contact [email protected].

The Publisher does not warrant or guarantee the article’s accuracy, completeness, merchantability, fitnessfor a particular purpose, or non-infringement. Descriptions of, or references to, products or publications, orinclusion of an advertisement in this article, neither constitutes nor implies a guarantee, endorsement, orsupport of claims made of that product, publication, or service.

Copyright © 2014, INFORMS

Please scroll down for article—it is on subsequent pages

INFORMS is the largest professional society in the world for professionals in the fields of operations research, managementscience, and analytics.For more information on INFORMS, its publications, membership, or meetings visit http://www.informs.org

http://pubsonline.informs.org

http://dx.doi.org/10.1287/mksc.2013.0837

http://pubsonline.informs.org/page/terms-and-conditions

http://www.informs.org

Vol. 33, No. 3, May–June 2014, pp. 338–352ISSN 0732-2399 (print) � ISSN 1526-548X (online) http://dx.doi.org/10.1287/mksc.2013.0837

© 2014 INFORMS

Just the Faces: Exploring the Effects ofFacial Features in Print Advertising

Li XiaoSchool of Management, Fudan University, Shanghai, People’s Republic of China 200433, [email protected]

Min DingSmeal College of Business, Pennsylvania State University, University Park, Pennsylvania 16802; and

School of Management, Fudan University, Shanghai, People’s Republic of China 200433, [email protected]

Human faces are used extensively in print advertisements. In prior literature, researchers have studiedspokespersons in general, but few have studied faces explicitly. This paper aims to answer three ques-

tions that are important to both researchers and practitioners: (1) Do faces affect how a viewer reacts to anadvertisement on the metrics that advertisers care about? (2) If faces do have an effect, is it large enough towarrant careful selection of faces when constructing print advertisements? (3) If faces do have an effect andthe effect is large, what facial features elicit such differential reactions on these metrics, and are such reactionsdifferent across individuals and/or product categories? Relying on the eigenface method, a holistic approachwidely used in the computer science field for face recognition, we conducted an empirical study to answerthese three questions. The results show that different faces do have an effect on people’s attitude toward theadvertisement, attitude toward the brand, and purchase intention and that the effect is nontrivial. Multiple seg-ments were identified and substantial differences were found among people’s reactions to the faces in the adsacross those segments. We also found that the effect of faces interacts with product categories and is mediatedby various facial traits such as attractiveness, trustworthiness, and competence. Implications and directions forfuture research are discussed.

Keywords : face; facial features; advertising effectiveness; eigenfaceHistory : Received: July 9, 2012; accepted: November 12, 2013; Preyas Desai served as the editor-in-chief and

Michel Wedel served as associate editor for this article. Published online in Articles in AdvanceFebruary 10, 2014.

IntroductionPrint media plays an important role in advertis-ing practice. According to Nielsen’s (2013) report,global print advertising spending reached more than$150 billion in 2012, accounting for more than a quar-ter of total media advertising spending in the world.In print advertising, it is common practice to havea human endorser present the brand or product andadvocate its adoption. We collected more than 600print ads that appeared in the United States from2001 to 2012 from the AdForum database1 and foundthat over 50% of print ads feature faces. A recentGfK MRI (2011) report indicated that only a smallportion (less than 10%) of print ads use celebrityendorsers, a finding consistent with our search ofthe AdForum database. For many ads that use non-celebrity endorsers, consumers know nothing aboutthese endorsers other than what they can infer fromthe endorsers’ faces.

1 AdForum.com chooses what it deems the five best ads every weekfor inclusion in its database. Ads may be print, video, or online.

Although faces are widely used in print ads, theselection process in practice is ad hoc. We interviewedan advertising agency on how print ads are createdfor fashion brands. The normal procedures are as fol-lows: (a) the advertising agency receives a requestfrom a client company, (b) it sends a model request tomultiple modeling agencies, (c) the modeling agenciessend back pictures of recommended models (100+),(d) the advertising agency selects some models (about10 to 20, based the fashion brand and its request) andphotographs them wearing the same clothes, (e) theadvertising agency gives the pictures to the client,and (f) the client decides which models to use (oftenin conjunction with the advertising agency), at whichpoint copy testing may or may not be involved.

Although there is a large body of literature onthe effects of endorsers/spokespersons on advertisingeffectiveness from various aspects— for example, gen-der (e.g., Gilly 1988), ethnicity (e.g., Wheatley 1971),age (e.g., Freiden 1984), profession (e.g., Ohanian1991), fame (e.g., Heath et al. 1994), and overall attrac-tiveness (e.g., Bower and Landreth 2001)—few stud-ies deal directly with faces. Therefore, the effect of

338

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Xiao and Ding: Exploring the Effects of Facial Features in Print AdvertisingMarketing Science 33(3), pp. 338–352, © 2014 INFORMS 339

different faces on advertising effectiveness remainslargely unknown. In this paper, we focus on faces ofnoncelebrity endorsers about whom viewers know lit-tle but their faces.

Borrowing face recognition techniques from thefield of computer science, we empirically test how dif-ferent faces affect advertising effectiveness in the printadvertising context. Specifically, we aim to addressthree questions that are important to both researchersand practitioners:

1. Do faces affect how a viewer reacts to an adver-tisement (abbreviated as an “ad”) on the metrics thatadvertisers care about? While keeping certain aspectsof the spokesperson constant (gender, ethnicity, age,etc.), will different faces elicit different responses fromviewers? Here, we focus on three key metrics ofadvertising effectiveness that advertisers care mostabout, including attitude toward the ad (Aad), atti-tude toward the brand (Abrand), and purchase inten-tion (PI) (Miniard et al. 1990, Goldsmith et al. 2000).

2. If faces do have an effect, is it large enough towarrant a careful selection of faces when constructinga print advertisement? If the effect of face exists butis trivial, clearly advertisers should not spend muchtime or money on face selection. Only if the effectexists and is substantial should they be careful whenselecting the face to appear in an advertisement.

3. If faces do have an effect and the effect is large,what specific facial features elicit such differentialreactions on the metrics, and are such reactions dif-ferent across individuals and different product cate-gories? This question deals with the heterogeneity ofthe effect of faces on viewers’ responses on key admetrics. A critical step toward answering this ques-tion involves the decomposition of faces. As such, werelied on the eigenface method, which is a widelyused approach for face recognition in the computerscience field. The eigenface method is “the first reallysuccessful demonstration of machine recognition offaces” (Zhao et al. 2003, p. 412). Many face recognitionalgorithms have been derived from this approach. Theeigenface method will be explained in further detailin the next section.

The rest of the paper is organized as follows:We first review the literature on face research. We thendescribe the empirical study and discuss the dataanalysis and results. Finally, we present a discussionand conclusions.

Relevant LiteratureWe first provide a review of previous research on theeffects of faces and facial features. Then, we intro-duce the eigenface method and how it works toextract facial features from different faces. Finally, wereview literature indicating that substantial hetero-geneity exists in the way people make inferences fromfaces.

Existing Literature on Face ResearchThe effect of faces on people’s perceptions and behav-ior has been studied in multiple disciplines (seeZebrowitz 2006 for a review). In psychology, Todorovet al. (2005) showed that inferences of competencebased solely on facial appearance could predict theoutcomes of U.S. congressional elections to someextent. Dumas and Testé (2006) found that perceivedfacial maturity has an influence on juridic judgments.Collins and Zebrowitz (1995) found that appearanceis significantly related to the types of jobs peoplehold and partly determines job status. Pincott (2010)found that women use facial masculinity to inferthe attractiveness of males and form mating prefer-ences. In marketing, Gorn et al. (2008) found thatduring public relations crises, chief executive officerswith baby faces are perceived to be more trustworthyand creditable than those with mature faces. Solomonet al. (1992) showed that congruency between thefacial attractiveness of a spokesperson and productimage can elicit favorable communication responses.In computer science, Koda and Maes (1996) showedthat faces of virtual agents have an impact on human–computer interaction. And Brahnam (2002) focusedon techniques to customize virtual agents’ faces tomatch the personalities of users.

Despite the large body of literature, to our knowl-edge, the effect of face on the effectiveness of printads has not been empirically demonstrated. The mostrelevant literature in this domain is Bower andLandreth’s (2001) study on how the overall physi-cal attractiveness of models (but not specifically theirfaces) affects product evaluations. More important,the magnitude of a face effect, heterogeneity in reac-tions to a face, and the practical relevance to practi-tioners have never been addressed.

An important stream of face research is to iden-tify what specific facial features drive perceptionsor behavior. In psychology, researchers have used aphysiognomic approach to extract facial features fromdifferent faces. Physiognomic method represents aface with a set of facial distances. For example, Berryand McArthur (1985) decomposed a male face into 11facial distances and found that a face with bigger eyesand a wider chin is judged to be more baby-facedand thus more trustworthy. Cunningham et al. (1990)decomposed a male face into 26 facial distances andfound that a male face with bigger eyes and a longerchin is perceived as more attractive in a female’seyes. Other researchers focused on one feature, facialwidth-to-height ratio, and found a significant effect onperceived propensity for aggression (Carré et al. 2009)and cooperative behavior in a trust game (Stirrat andPerrett 2010).

As the above studies illustrate, the choice of phys-iognomic measurement is somewhat arbitrary and

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

Xiao and Ding: Exploring the Effects of Facial Features in Print Advertising340 Marketing Science 33(3), pp. 338–352, © 2014 INFORMS

only captures a subset of information from face. Forexample, Berry and McArthur (1985) chose 11 phys-iognomic features, and Cunningham (1986) chose 26,but it is unknown exactly how many features shouldbe studied. Furthermore, because of model identifica-tion issues and multicollinearity concerns, both stud-ies included some interaction terms in their models,such as eye area, eye roundness, and chin area, butthey did not include all possible interactions. Manyinteractions that might have an effect are missingfrom the models—e.g., the ratio of eye distance to facelength, the angle of the eyes to the nostril tip, etc. Thismethod especially is not useful if the faces studiedare similar on these typical physiognomic features,such as in the case of print advertisement where theactors used have similar measurements on most ofthese typical physiognomic features. People who donot fit the stereotypical look of a model are unlikely tobe selected. Thus very limited variations exist amongthese typical physiognomic features in this context.Based on these limitations, to decompose faces weused a holistic method, the eigenface method, whichcan capture all subtle differences across faces.

The Eigenface MethodEigenface is the gold standard method in computerscience for face recognition and was first proposed byTurk and Pentland (1991). It uses principal componentanalysis (PCA) to project training face images onto aset of dominant eigenvectors. Because each dominanteigenvector looks roughly like a face, these eigenvec-tors are also called eigenfaces. This method is easyto implement, is computationally efficient, achievesgood accuracy in face recognition tasks, and thus iswidely used in the computer science field (Zhao et al.2003). Its implementation is briefly described as fol-lows (for more details, see Turk and Pentland 1991):Step 1. Acquire a training set of face images. Ideally,

these images are all the same size, in a fixed position,and facing forward under constant lighting. The facesshow neutral emotion and do not wear accessories(e.g., glasses). Assume that there are M gray-scaledface images in the training set and the size of eachimage is H ×W pixels.Step 2. Stack each face image into a HW × 1 vector,

denoted as Xi. The training set of images is repre-sented by a HW ×M matrix, denoted as X.Step 3. Calculate the mean X̄ = 41/M5

∑

Xi and sub-tract the mean from X. The mean vector X̄ could beunstacked back to an H × W matrix and displayed,which looks roughly like a face with blurry contour.This is called the average face.

Step 4. Use PCA to calculate the eigenvectors andeigenvalues of the covariance matrix. At this step,we obtain M eigenvectors and the corresponding Meigenvalues.

Step 5. Keep the K most important/dominant eigen-vectors and project each image in the training set ontothese K eigenvectors, stacked together and denotedas EV4HW × K5. Thus, each image in the trainingset (Xi) can be approximated and represented by aunique loading vector, Li4K × 15. Each image can bereconstructed as Xi ≈ EV × Li + X̄. The choice of Kdepends on how many eigenfaces are needed to geta good reconstruction of training images (Turk andPentland 1991).Once the average face and K dominant eigenvectorsare stored, given a face image, either an image fromthe training set or a new image, it is easy to calculatethe loading vector of this face image. The dissimilar-ity between two face images can be well representedby the difference between two corresponding loadingvectors. An illustration of the eigenface decomposi-tion is shown in Figure 1.

Despite its wide use in face recognition, the eigen-face method has rarely been used to study the effect offacial features on viewers’ responses. Only Branham(2002) made the first attempt to apply the eigen-face method to her virtual agent study. However, thefaces used in her study were synthetic faces gener-ated by software, not real human faces. To date, nostudy has applied the eigenface method to real facesto study people’s perceptions or reactions to printadvertisements.

Unlike the physiognomic method, which requires apredefined set of facial distances, no prior informa-tion is needed for the eigenface method. The extractedeigenvectors contain the major differences amongtraining images; thus it is a holistic measurement.

Heterogeneity in Viewers’ Responses to FacesAll of the previous studies that link facial featuresto people’s reactions were conducted at the aggre-gate level, not the individual level, and attempted touse one model to explain how people make infer-ences from faces. People were treated as homoge-neous in these studies; individual differences wereignored. One underlying assumption is that peopleshare in common what facial features influence theirimpressions and how important these facial featuresare in forming their perceptions. For example, basedon Berry and McArthur’s (1985) findings, a trust-worthy face must have big eyes and a wide chin.The major theory underlying this assumption is thatthe perception of facial features has adaptive valueand that trait impressions are based on those facialqualities that demand the greatest attention for thesurvival of the species—namely, physical fitness, age,and emotional state (Zebrowitz 1998). Faces with fea-tures that are indicative of these qualities are believedto have pronounced overgeneralization effects. Thus,

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 1 Eigenface Decomposition

Average face

+

Eigenface 1

+

…

× loading on Eigenface 1: L1

Eigenface k

=

× loading on Eigenface k: Lk

Note. With the average face and k eigenfaces stored in the database, any face can be compressed and reconstructed by a k × 1 loading vector (L11 0 0 0 1 Lk )(Turk and Pentland 1991).

there is considerable consistency and universal con-sensus in the way people make inferences from faces(Cunningham et al. 1995).

However, this assumption has been greatly chal-lenged. Ample evidence shows that the heterogeneityamong people is not trivial and that it influences theirway of making inferences from faces. A comparisonbetween the results from the studies by Cunningham(1986) and Cunningham et al. (1990) reveals a gen-der difference in inferring attractiveness. Solomonet al. (1992) proposed six distinct types of attractivelooks. Pincott (2010) reported that women from areaswith good healthcare conditions (e.g., Sweden) pre-fer feminine-looking men, whereas women from areaswithout good healthcare systems (e.g., Russia) pre-fer masculine-looking men. Ekman and Friesen (2003)reported that culture and individual heterogeneityhas an effect on both the exhibited face side andthe recipient side. Based on these arguments, peopleshould be studied in separate segments defined byfactors such as gender, ethnicity, etc. Hence, multi-ple segments of people should be identified based onthe way they make inferences from faces. Within eachsegment, people share similar preferences towardfaces, whereas across segments, people show substan-tially different preferences toward faces.

Resolving this controversy for marketing practition-ers is important, because quite different marketingstrategies might need to be considered to optimizethe effects of advertisements on consumer percep-tions and behavior. If our results reveal that peo-ple are somehow similar and that the one best faceexists, then an advertising agency’s job is to finda spokesperson with a face that closely resemblesthe best face for the advertisement. However, if wefind that people are heterogeneous in their facialpreferences and multiple best faces exist with eachappealing to a different segment of consumers, whichmay or may not depend on product categories, thenan advertising agency would want to identify anddefine specific customer segments and create mul-tiple advertisements, each with a distinct face thatclosely matches the best face for that specific segmentof customers.

Experimental DesignStimulus AdvertisementsTo mimic advertising practice, we used real models’faces as stimuli in this study. As a result, whatevereffect we identify will not be artificial as a result ofthe poor choice of faces (e.g., using faces that will

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


never be selected by practitioners). We followed a rig-orous face selection process in this study. First, we col-lected thousands of faces of male models from fash-ion websites such as Macy’s, Brooks Brothers, andMyHabit.com; modeling agency websites such as Eliteand Ford; and other graphic websites such as GoogleImages, Flickr, and Getty Images. Then, we narrowedthe image pool down to a face database with 60distinctive faces that satisfied the following criteria:(a) Caucasian male, between 25 and 35 years old;(b) no accessories, no obvious moustache, moderatehairstyle; (c) mild smile, positioned right-side up, fac-ing forward; (d) not a celebrity (and no resemblanceto a celebrity); and (e) reasonable resolution. Finally,we randomly selected 12 faces from the face databaseand used them as stimulus faces.

The product categories were selected based onthe following criteria: (a) the category often featureshuman faces in print ads, (b) the general public hassome knowledge of the product category, and (c) it isappropriate to feature a man’s face in an advertise-ment in the category. Of the categories that fit thesecriteria, we selected 12 that also have substantial vari-ation in hedonism so that we could test whether peo-ple’s reaction to faces are related to the hedonismcharacteristic of a product category. Hedonism is avariable often used in the literature to represent differ-ent product categories (e.g., Raghunathan and Irwin2001, Carroll and Ahuvia 2006, Inman et al. 2009).It is indirectly indicated in the literature that hedo-nism might affect the relationship between facial fea-tures and people’s responses. For example, Todorovet al. (2005) found that competent faces are morelikely to be elected for political positions (less hedo-nic context); Englis et al. (1994) found that attrac-tive faces are frequently featured in ads for fashionproducts such as clothing and perfume (more hedoniccontext). However, to our knowledge, the interactionbetween hedonism and facial features has never beenstudied in the literature. The 12 categories selectedfor this study are beer, restaurant, job search agent,men’s cologne, coffee, computer, hotel, jeans, sportsutility vehicle (SUV), sports shoes, camcorder, and cardealer.

Based on the face selection results (12 stimulusfaces) and category selection results (12 categories),we created 12 × 12 = 144 stimulus ads. To create amore natural advertisement viewing experience forparticipants, the stimulus ads were created from realads where the original face was replaced with oneof the stimulus faces while holding all other ele-ments constant with the following exception: the orig-inal brand names were replaced with some genericnames and logos were removed. We hired profes-sional graphic designers to create the stimulus ads tomake sure that they looked reasonably natural at a

quick glance. At the end of the empirical study, weasked the participants whether they observed any-thing out of place or abnormal in the ads. Only a fewparticipants indicated that they sensed that some ofthe faces in the ads might have been photoshopped.

For four categories (restaurant, job search agent,computer, and sports shoes), we included the originalfaces in the stimulus ads as a baseline. The other eightoriginal faces were not used because they wore acces-sories such as glasses, had beards, or were celebrities.Altogether, we created 148 stimulus ads.

Extraction of Facial FeaturesUsing the eigenface method, we first extracted thefacial image from the original pictures and removedthe noise related to lighting and angle. Second, wenormalized each face and placed it into a 180 ×

180 pixel-sized plane. This allowed us to constructthe training set of 60 images. Third, we used thePCA method to obtain the eigenvectors and eigenval-ues. We retained 12 eigenvectors that explain almost75% of the variance among training images and canreconstruct the faces reasonably well. The root-mean-square pixel-by-pixel error in representing croppedversions of training face images, i.e., reconstructionerrors, was less than 10%. The average face and12 eigenfaces extracted from the 60 training imagesare shown in Figure 2.

Finally, the loading vector (12 × 1) of each stimulusface was calculated. Eigenface loadings ranged from−41130069 to 51467063 and thus were resized by divid-ing by 1,000. The resized eigenface loadings were thefacial features we extracted from the stimulus faces.These loadings were used as independent variables inour subsequent data analysis.

Experimental ProceduresA total of 989 participants participated in this study.Of those, 676 undergraduate students at a majorU.S. university completed the study in an on-campuscomputer laboratory for course credit. The remain-ing 313 participants were recruited from the AmazonMechanical Turk subject pool; each of them received$2 for their participation. In terms of gender, 472 weremale (48%) and 517 (52%) were female. In terms ofethnicity, 755 participants were Caucasian (77%), 25were African American (3%), 42 were Hispanic (4%),125 were Asian (12%), 22 were multiracial (2%), and20 were categorized as other (2%). Ages ranged from18 to 69 years old, with a mean of 24.36 and a stan-dard deviation of 10.27 years.

We applied a between-subject design. Each partici-pant was presented with 12 print advertisements, onefor each product category. Within each product cate-gory, one stimulus advertisement out of 12 (or 13 ifthe original face was used in the category) was ran-domly assigned to each participant. The sequence of

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 2 The Average Face and 12 Most Dominant Eigenfaces

Average face

12 most dominant eigenfaces

Notes. As shown above, the average face and the 12 most dominant eigenfaces were extracted from 60 training faces that were used in the present study,including the 12 stimulus faces used in the empirical study. The contribution of each eigenface to explain variances among 60 faces decreases from left toright and from the upper row to the lower row.

12 product categories was also randomized for eachparticipant. The average time spent on each advertise-ment evaluation was 11.4 seconds. After participantsviewed all 12 print advertisements, they began evalu-ating them one by one on attitude toward the ad, atti-tude toward the brand, and purchase intention. Afterevaluating all 12 ads, participants were asked to eval-uate the hedonism levels of the 12 product categories,followed by a survey about demographics, cognitivestyle, and value system. At the end of the empiricalstudy, they were asked to evaluate the faces they sawon six trait dimensions that have been identified in the

literature as facial characteristics that affect people’sresponses: baby-facedness (Berry and McArthur 1985,Dumas and Testé 2006, Gorn et al. 2008), masculin-ity (Pincott 2010), attractiveness (Solomon et al. 1992),trustworthiness (Berry and McArthur 1985, Gorn et al.2008, Stirrat and Perrett 2010), aggressiveness (Carréet al. 2009), and competence (Todorov et al. 2005).

We used the scale from Miniard et al. (1990) to mea-sure three key ad metrics. Three items (bad/good,uninteresting/interesting, dislike/like) were used tomeasure attitudes toward the advertisement ona seven-point scale. Attitudes toward the brand

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


were measured using a three-item seven-point scale(unfavorable/favorable, negative/positive, dislike/like). Purchase intention was measured using atwo-item seven-point scale (unlikely/likely, improba-ble/probable). Reliabilities for the three constructs are0.88, 0.89, and 0.91, respectively.

In our study, hedonism was measured using a four-item seven-point scale (unpleasant/pleasant, awful/nice, sad/happy, disagreeable/agreeable) (Crowleyet al. 1992). Reliability for hedonism is 0.89. A 36-itemseven-point scale was used to measure cognitive style(Hauser et al. 2009), and a 30-item nine-point scalewas used to measure value style (Schwartz 1992).Each of the six trait dimensions was measured ona seven-point bipolar scale (i.e., baby-faced/mature,feminine/masculine, unattractive/attractive, untrust-worthy/trustworthy, unaggressive/aggressive, andincompetent/competent).

Estimation and ResultsEffect of Faces on Key Ad Metrics at theAggregate LevelAs indicated in the experimental design section, theonly difference in the stimulus ads for each productcategory was the face appearing on the ad; therefore,any differences in key ad metrics can be attributed tothe face used in the ad. For each category, we calcu-lated the means for each face for each key ad metric(see Table 1).

The results show that faces do have an effect onadvertising effectiveness. For example, by comparingmeans, Face 6 in the beer ad significantly outper-formed Face 12 on all three key ad metrics. For eachcategory and each key ad metric, the face with highestmean significantly beats the face with lowest mean atthe p < 0005 level.

A comparison between the results for the 12 stim-ulus faces and the original faces justifies careful faceselection. For example, for sports shoes, by chang-ing from the original face to Face 7, the print adachieved about a 0.5-point increase on evaluations ofall three key ad metrics. As our face selection proce-dures show, these 12 stimulus faces were randomlyselected from among 60 male models’ faces. It is verypossible that by applying our method to a large sam-ple of faces, practitioners would achieve even greaterimprovement.

It is worth noting that the preferences for faceswere different across categories, which indicates astrong interaction between face effect and category.For example, for attitude toward ad, Face 11 achievedthe highest mean for the jeans ad but the lowest forthe car dealer ad.

To help answer the second question on whetherpractitioners should care about such a difference in

effect, we calculated the percentages of high-ratingparticipants (evaluation of key ad metric > 6) for eachcategory and each key ad metric. For example, inbusiness practice, practitioners care most about thecustomers who give high ratings to their productsor services (Urban and Hauser 1993). We report boththe percentage for each face for each ad (in Table 2)as well as absolute maximum improvement betweenthe best and worst face for each product category (inFigure 3).

The effect is quite striking. For example, Table 2shows that for the cologne ad, by merely changing theface from Face 2 to Face 1, advertisers could achievea more than 10% increase in attitude toward ad (from8% to 25.8%), attitude toward brand (from 5.3% to21.6%), and purchase intention (from 9.3% to 20.6%).In other words, such a change would double or eventriple the number of high rating participants. We plot-ted the maximum absolute percentage improvement(the percentage of high-rating participants for the bestface of high rating minus the percentage of high-rating participants for the worst face of high rating)in Figure 3.2 Figure 3 shows that by merely chang-ing from the worst face of high rating to the best faceof high rating, each ad could achieve an increase inthe number of high-rating people ranging from 4% to21% for each key ad metric. The interaction of faceeffect with product categories can also be seen in acomparison of high-rating people. For example, Fig-ure 3 shows that for the cologne ad, the face effecton all three key ad metrics is huge, whereas for thecomputer ad, the face effect is much smaller. For therestaurant ad, the face effect for Aad and Abrand issubstantial, but it is minor for PI.

These results indicate that different faces do havean effect on advertising effectiveness, and that theeffect is nontrivial. Hence, we conclude that adver-tisers should carefully select faces/spokespersons toappear in print advertisements.

Segmentation Analysis Using Facial FeaturesWe used 12 eigenface loadings and 12 interactionterms between eigenface loadings and hedonism asindependent variables and the three key ad metricsas dependent variables simultaneously in the segmen-tation analysis. To explore possible heterogeneity inhow people respond to facial features, we used thelatent class multivariate regression model proposedby Ramaswamy et al. (1993). Specifically,

Yi ∼Gi4Yi3w1Â1ã5=∑

k

wkf 4Yi �Xi1Âk1ãk51 (1)

2 We did not plot the maximum relative percentage improvement(the percentage of high-rating participants for the best face of highrating divided by the percentage of high-rating participants for theworst face of high rating) because, for several categories, the per-centage of high rating participants for worst face of high ratingwas zero.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 1 Mean of Key Ad Metrics by Product Category and Face

Category Face 1 Face 2 Face 3 Face 4 Face 5 Face 6 Face 7 Face 8 Face 9 Face 10 Face 11 Face 12 Original

BeerAad 3026 3056 2097 2081 3004 3066 3005 3003 3021 3012 2077 2 063 N/AAbrand 3037 3060 2099 3007 3021 3073 3035 3015 3043 3037 3028 2 090PI 3030 3065 2085 2086 2092 3044 3023 2092 3007 3012 2096 2 070

RestaurantAad 4075 4041 4042 4054 4063 4046 4041 4083 4 015 4039 4032 4059 4075Abrand 4088 4078 4069 4082 4090 4058 4078 5005 4 042 4069 4057 4095 5000PI 4081 4097 4077 4087 4096 4068 4080 5006 4057 4076 4 053 4089 4095

Job search agentAad 3025 3052 3022 3024 3025 3071 3006 3056 3 004 3035 3065 3021 3037Abrand 3064 3081 3048 3073 3056 3090 3029 3075 3 028 3064 3091 3062 3066PI 3024 3034 2083 3003 2095 3041 2 069 3015 2076 3003 3044 3009 3009

CologneAad 4093 3 091 4069 4036 4046 4065 4056 4014 4012 4065 4055 4002 N/AAbrand 4080 3 090 4057 4039 4038 4063 4034 4021 4010 4062 4049 3099PI 4021 3025 3095 3084 3081 4017 3077 3064 3045 4001 3099 3 023

CoffeeAad 3074 3010 3062 2 083 3030 3080 3070 3001 2093 3042 3012 2088 N/AAbrand 3081 3030 3071 3 007 3041 3084 3078 3031 3034 3045 3028 3029PI 3016 2085 3028 2 067 3019 3050 3039 2084 3015 3018 2099 2084

ComputerAad 3008 3034 3017 3004 2 054 3019 3039 2079 3001 3010 3032 3012 3030Abrand 3033 3054 3018 3036 2 082 3022 3053 3007 3009 3023 3041 3038 3020PI 2068 2088 2073 2073 2 019 2062 2085 2031 2039 2085 2088 2058 2044

HotelAad 3037 3051 3058 3018 3 002 3061 3097 3053 3008 3024 3042 3020 N/AAbrand 3052 3094 3089 3050 3051 3088 4023 3081 3 044 3075 3065 3065PI 3029 3072 3064 3019 3024 3045 4000 3048 3014 3042 3 013 3037

JeansAad 3070 3029 3066 3006 3 003 3081 3056 3059 3054 3018 3094 3030 N/AAbrand 3068 3019 3067 2 097 2099 3070 3051 3071 3039 3025 3080 3019PI 3026 2071 3019 2 061 2066 3042 3000 3013 2098 2076 3035 2083

SUVAad 4032 3092 4036 4019 3095 3069 4017 3093 3 063 4013 4026 3085 N/AAbrand 4020 3092 4017 4005 3081 3078 4007 3090 3 072 3099 4010 3095PI 3009 2081 2092 3023 2093 2083 3013 2091 2 073 2089 2099 3022

Sports shoesAad 4075 4 026 4059 4052 4069 4058 4099 4029 4042 4035 4058 4038 4046Abrand 4076 4038 4051 4057 4069 4053 4084 4030 4 028 4033 4057 4030 4037PI 4014 3098 3099 3092 4019 4013 4033 3078 3066 3 064 3092 3073 3084

CamcorderAad 3061 3039 3083 3 021 3045 3090 3048 3056 3046 3034 3052 3040 N/AAbrand 3075 3060 4005 3 050 3063 4007 3059 3078 3062 3052 3054 3056PI 2099 2090 3002 2080 3002 3007 2090 2089 2097 2096 2 068 2090

Car dealerAad 2049 2036 2025 2023 2004 2036 2008 2035 1089 2024 1073 2044 N/AAbrand 3000 2085 2067 2069 2052 2058 2053 2083 2040 2076 2 020 2076PI 2042 2042 2030 2018 2009 2020 2005 2044 2009 2048 1074 2051

Notes. For each category and each key ad metric, the face that achieves the highest mean is deemed the best face of mean and marked in bold, whereas theface with lowest mean is deemed the worst face of mean and marked in bold and italic. For each category and each key ad metric, evaluations from participantswho viewed the ad with best face of mean are significantly higher than evaluations from those who viewed the same ad with worst face of mean at the p < 0005level. In four categories, four original faces were also tested in the experiment and thus compared here. For each category, the original face refers to the facethat was featured in the real ad. So even though they are all called original face, they are unique for each category and different from each other.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 2 Percentage of High-Rating Participants (Key Ad Metric Evaluated Above 6) for Each Key Ad Metric by Product Category and Face

%

Category Face 1 Face 2 Face 3 Face 4 Face 5 Face 6 Face 7 Face 8 Face 9 Face 10 Face 11 Face 12 Original

BeerAad 1109 903 602 902 407 1207 603 2 05 803 708 404 603 N/AAbrand 803 400 2 01 507 407 805 705 308 803 708 202 205PI 1007 1407 4 01 1003 802 1207 1000 809 803 708 506 705

RestaurantAad 2808 1808 1600 2204 1704 2202 1804 3002 10 07 1905 1609 2507 1801Abrand 2808 2705 2103 3003 2107 2305 2108 3107 1900 2108 16 09 2209 3006PI 3500 3103 2903 3106 3303 3201 3202 3409 27 04 3100 2707 3000 3601

Job search agentAad 308 308 2 05 608 305 606 206 704 208 900 806 706 700Abrand 808 1103 205 905 701 309 206 704 104 705 1101 809 700PI 1000 603 205 905 407 902 0 00 404 208 405 704 706 700

CologneAad 2508 8 00 2809 2005 2005 1705 2205 1600 1209 2102 2707 905 N/AAbrand 2106 5 03 1701 1303 1100 1500 2101 1600 809 1605 2005 1109PI 2006 903 1907 1303 1203 1205 1207 1600 709 1808 1801 6 00

CoffeeAad 701 303 1208 308 501 1003 507 904 409 1104 803 101 N/AAbrand 507 404 407 103 501 1003 403 802 409 507 306 203PI 806 505 508 308 604 800 1104 701 601 1002 3 06 507

ComputerAad 306 605 504 2 04 304 805 503 402 208 308 405 501 705Abrand 204 309 504 407 101 208 309 402 104 205 405 308 405PI 306 103 608 305 101 0 00 206 0 00 104 205 300 205 405

HotelAad 701 909 607 305 409 804 803 409 504 2 06 507 403 N/AAbrand 701 909 506 2 04 409 804 803 1304 706 605 507 704PI 507 704 607 701 601 2 04 9 05 703 504 502 403 704

JeansAad 1100 1000 809 600 2 02 1200 708 900 308 506 907 803 N/AAbrand 808 1000 809 803 3 02 804 309 501 603 506 803 803PI 1100 1000 809 306 101 906 309 604 603 303 609 609

SUVAad 1704 1401 1908 1900 1504 1100 1608 1406 8 05 2207 1507 1203 N/AAbrand 1300 805 1005 803 606 1100 1109 3 04 700 1200 1102 909PI 403 104 305 600 303 307 509 202 402 607 202 102

Sports shoesAad 2205 1203 1804 1609 2801 2404 3008 9 06 1205 1500 2109 1403 2205Abrand 2308 1101 2101 1802 2500 1401 2404 804 6 03 1205 1708 1607 1907PI 1308 909 1302 1802 1702 1709 1902 8 04 904 1000 1203 1301 1609

CamcorderAad 304 2 07 901 603 407 1006 209 501 506 402 705 308 N/AAbrand 203 400 901 501 305 904 209 308 609 104 308 604PI 203 400 108 501 204 509 104 308 208 104 208 206

Car dealerAad 505 0 00 0 00 0 00 104 306 101 200 102 102 102 0 00 N/AAbrand 608 102 101 105 104 204 0 00 309 102 102 102 0 00PI 401 0 00 101 105 104 102 0 00 309 102 102 102 104

Notes. For each category and each key ad metric, the face that achieves the highest percentage of high-rating participants is deemed the best face of highrating and is marked in bold, whereas the face with the lowest percentage of high-rating people is deemed the worst face of high rating and is marked in boldand italic. For each category and each key ad metric, the percentage of high-rating participants for the best face of high rating is significantly higher than thepercentage for the worst face at the p < 0005 level. Note that for most categories and key ad metrics, the best and worst faces of high rating are the same withbest and worst faces in terms of mean (see Table 1) but are different for some cases.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 3 Maximum Absolute Percentage Improvement forHigh-Rating Participants

0.00

5.00

10.00

15.00

20.00

25.00B

eer

Res

taur

ant

Job

sear

ch a

gent

Col

ogne

Cof

fee

Com

pute

r

Hot

el

Jean

s

SUV

Spor

ts s

hoes

Cam

cord

er

Car

dea

ler

Max

imum

abs

olut

e im

prov

emen

t (%

)

Product categories

AadAbrandPI

Notes. For each category and each key ad metric, the maximum absolutepercentage improvement equals the percentage of high-rating participantsfor the best face of high rating minus the percentage of high-rating partici-pants for the worst face of high rating. For each category and each key admetric, the best face of high rating refers to the face that achieves the high-est percentage of high-rating participants, whereas the worst face of highrating refers to the face that achieves the lowest percentage of high-ratingparticipants.

where

f 4Yi �Xi1Âk1ãk5= 42�5−4T /25�ãk�

−41/25

exp{

−12 4Yi−XiÂk5

′ã−1k 4Yi−XiÂk5

}

(2)

and

�k ≥ 01K∑

k=1

�k = 10 (3)

Here, Yi = 6Yi11 0 0 0 1YiT 7′ is a T × 3 matrix with i =

11 0 0 0 1 I as an indicator of participant and T as thetotal number of replications, and 3 refers to the threekey ad metrics, i.e., attitude toward the ad, attitudetoward the brand, and purchase intention in our case;Xi is a T × J matrix of independent variables forparticipant i, where J refers to the number of inde-pendent variables; �k is the prior probability of seg-ment k; K is the predefined number of segments,an assumption to be imposed on the finite mixturemodel; Âk is the J ×1 parameter vector for segment k;and ã= 4ã11 0 0 0 1ãk5

′, where ãk is the T ×T variance–covariance matrix for segment k.

By introducing an indicator variable

zik =

{

1 iff participant i belongs to segment k10 otherwise1

the likelihood can be presented as

L=∏

i

∏

k

6wkf 4Yi �Xi1Âk1ãk57zik 0 (4)

We used the Latent GOLD Syntax Module forimplementation,3 which is mainly based on theexpectation-maximization algorithm for model esti-mation (Vermunt and Magidson 2005). Consideringthat we dealt with only 12 faces, although the modelincluding all 24 independent variables was identified,it was very unstable and the convergence was bad.Thus, we added a variable selection procedure to thesegmentation model. Given K, we used a stepwisegreedy search for variable selection (Bishop 1996).We used the Bayesian information criterion (BIC) forvariable selection (Andrews and Currim 2003). In thiscase, a variable will enter into the model if its coef-ficient is significant in at least one segment, and avariable will exit out of the model if its coefficient isinsignificant in all segments. The heuristic is brieflydescribed below.

Step 1. Select a K.Step 2. Run a K-segment finite mixture model with

empty set P0 = 8�9, i.e., only intercepts, no x’s.Step 3. Select the next best variable x+ = arg minxyPk

·

6BIC4Pk+x57 to enter into the K-segment finite mixturemodel.

Step 4. Update Pk+1 = Pk + x+; k = k+ 1.Step 5. Remove the worst variable x− = arg minx∈Pk

·

6BIC4Pk −x57 from the K-segment finite mixture modelif removing it will help decrease the BIC value of themodel.

Step 6. Update Pk+1 = Pk − x−; k = k+ 1.Step 7. Repeat Steps 5 and 6 until removing any

variable in the model will not decrease the BIC value.Step 8. Repeat Steps 3–7 until adding any variable

into the model or removing any variable from themodel will not help decrease the BIC value, or all24 variables are included in the model.

We tested for K = 112131415, and 6, and the fivesegment result is the best model based on the BIC cri-terion. The variables selected were hedonism×L1, hedo-nism×L4, hedonism×L5, hedonism×L6, hedonism×L7,hedonism × L10, hedonism × L11, and hedonism × L12.4

The five segments were composed of 40, 82, 386, 344,and 137 participants, respectively. The estimation forthe best model is shown in Table 3. Substantial het-erogeneity is shown across segments. For example, interms of attitude toward the ad, participants in Seg-ments 1 and 2 do not favor faces that load heavilyon Eigenface 1, especially in ads for hedonic prod-ucts; participants in Segments 4 and 5, by contrast, dofavor faces that load heavily on Eigenface 1, especiallyin ads for hedonic products; and participants in Seg-ment 3 do not care about face loadings on Eigenface 1

3 We thank Wayne DeSarbo for recommending the model andsoftware.4 To test the robustness of the result, we also ran the mixture modelwith the eight selected variables for K = 11 0 0 0 18. The model withK = 5 achieves the smallest BIC value as well.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Table 3 Preference of Facial Structure of Each Segment, by Different Key Ad Metrics

Segment 1 (40) Segment 2 (82) Segment 3 (386) Segment 4 (344) Segment 5 (137)

Variable Aad Abrand PI Aad Abrand PI Aad Abrand PI Aad Abrand PI Aad Abrand PI

Hedonism ∗ L1 −0026∗∗ −0001 −0011∗ −0021∗∗ −0018∗∗ −0018∗∗ −0001 0001 0005∗∗ 0019∗∗ 0019∗∗ 0024∗∗ 0014∗∗ 0012∗∗ −0008∗∗

Hedonism ∗ L4 0042∗∗ 0000 0041∗∗ 0031∗∗ 0028∗∗ 0021∗∗ 0000 −0003 −0009∗∗ −0020∗∗ −0016∗∗ −0023∗∗ −0012∗∗ −0008 0007Hedonism ∗ L5 0048∗∗ 0011 0052∗∗ 0031∗∗ 0032∗∗ 0019∗ −0008 −0008 −0012∗∗ −0041∗∗ −0034∗∗ −0044∗∗ −0029∗∗ −0014∗ 0031∗∗

Hedonism ∗ L6 1001∗∗ −0012 0070∗∗ 1046∗∗ 1031∗∗ 1042∗∗ −0003 −0011 −0024∗∗ −1041∗∗ −1027∗∗ −1061∗∗ −0063∗∗ −0056∗∗ 0068∗∗

Hedonism ∗ L7 −0040 −0006 −0045 −1000∗∗ −0084∗∗ −1005∗∗ 0014 0017∗ 0034∗∗ 1037∗∗ 1026∗∗ 1051∗∗ 0095∗∗ 0084∗∗ −0032∗∗

Hedonism ∗ L10 −0041∗∗ 0000 −0058∗∗ −0064∗∗ −0063∗∗ −0059∗∗ 0013∗∗ 0012∗∗ 0018∗∗ 0074∗∗ 0071∗∗ 0089∗∗ 0061∗∗ 0049∗∗ −0028∗∗

Hedonism ∗ L11 −0076∗∗ 0019 −0095∗∗ −0098∗∗ −0090∗∗ −0087 0011 0018∗∗ 0018∗∗ 0092∗∗ 0083∗∗ 1004∗∗ 0042∗∗ 0027∗∗ −0041∗∗

Hedonism ∗ L12 1007∗∗ −0012 0054∗∗ 1061∗∗ 1044∗∗ 1044∗∗ −0031∗∗ −0032∗∗ −0045∗∗ −1048∗∗ −1033∗∗ −1063∗∗ −0087∗∗ −0065∗∗ 0059∗∗

∗Significant at p < 0010; ∗∗significant at p < 0005.

at all, no matter how hedonic the featured productwould be. It is worth noting that all variables enter-ing into the best model are interaction terms, whichgives more evidence supporting the conjecture thatface effect interacts with product categories. Becausethe sizes of Segments 3 and 4 are substantially big-ger than those of Segments 1, 2, and 5, we focus ourfollowing discussion mainly on Segments 3 and 4.

To further explore the heterogeneity across seg-ments, we ran a classification and regression tree algo-rithm (a supervised learning method; see Bishop 2006for more details) of the segment labels on the demo-graphic information as well as the cognitive style andvalue system evaluation. Three demographic vari-

Table 4 Segment Characteristics

Trait Segment 1 Segment 2 Segment 3 Segment 4 Segment 5

Size 40 82 386 344 137Demographicsa More female More female Younger More male Older

Less Caucasian Mainly CaucasianCognitive stylea1b Prefer parts and details

More visually drivenLess analytical

Less visually driven More visually driven Prefer the wholeLess visually drivenMore analytical

Value systema1c Value authority less Value choosing own goalsmore

Value choosing own goalsmore

Value authority moreValue choosing own goals less

Value authority less

Baby-facedness Aad, PI Aad, AbrandMasculinity Aad, Abrand Aad, Abrand, PI Aad, Abrand Aad, Abrand, PIAttractiveness Aad Aad, Abrand, PI Aad, Abrand, PI Aad, Abrand, PI Aad, Abrand, PITrustworthiness Aad, PI Aad, Abrand, PI Aad, Abrand, PI Aad, Abrand, PI Aad, Abrand, PIAggressiveness PI Aad, Abrand, PI PICompetence Aad Aad, Abrand, PI Aad, Abrand, PI Aad, Abrand, PI Aad, Abrand, PI

aIf the feature is not included in the cell, it means that this feature for this segment is similar to the entire sample. For example, gender information isdescribed for Segments 1, 2, and 4, but not for Segments 3 and 5. This means the gender distribution in Segments 3 and 5 is similar to the entire sample.

bThe original scales for these three items are as follows: “Please indicate how much you agree or disagree with the following statements from 1 (stronglydisagree) to 7 (strongly agree).

• I’m usually more interested in the whole than in parts and details.• I will read an explanation of a graph/chart before I try to understand the graph/chart on my own.• I find that to adopt a careful, analytical approach to making decisions takes too long.”

cThe original scales for these two items are as follows: “Please rate the values below very carefully and then rate each value on an importance scale ‘AS AGUIDING PRINCIPLE IN MY LIFE,’ from 7 (of supreme importance) to 6 (very important), to 3 (important), to 0 (not important), to −1 (opposed to my values).

• Authority.• Choosing own goals.”

ables (age, gender, and ethnicity), three cognitiveitems (prefer the whole versus parts and details, visu-ally driven, and analytical), and two values (author-ity and choosing own goals) emerged as importantvariables to classify the participants into five identi-fied segments. A pairwise comparison on these eightvariables across five segments was run afterwards toexplain segment heterogeneity. We also conducted aset of mediation tests (Baron and Kenny 1986) withthe six trait variables we measured. Results are shownin Table 4.

It is shown that Segment 3 mainly comprisesyoung, visually driven people who value choosingtheir own goals. The effect of eigenface features on

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Figure 4 Best Face of Mean for Each Category and Segment

0

2

4

6

8

10

12

14B

eer

Res

taur

ant

Jobs

sea

rch

agen

t

Col

ogne

Cof

fee

Com

pute

r

Hot

el

Jean

s

SUV

Spor

ts s

hoes

Cam

cord

er

Car

dea

ler

Face

num

ber

Product categories

Best face of mean in terms of Aad

Aggregate

Segment 1

Segment 2

Segment 3

Segment 4

Segment 5

0

2

4

6

8

10

12

14

Bee

r

Res

taur

ant

Job

sear

ch a

gent

Col

ogne

Cof

fee

Com

pute

r

Hot

el

Jean

s

SUV

Spor

ts s

hoes

Cam

cord

er

Car

dea

ler

Face

num

ber

Product categories

Best face of mean in terms of Abrand

Bee

r

Res

taur

ant

Job

sear

ch a

gent

Col

ogne

Cof

fee

Com

pute

r

Hot

el

Jean

s

SUV

Spor

ts s

hoes

Cam

cord

er

Car

dea

ler

Product categories

0

2

4

6

8

10

12

14

Face

num

ber

Best face of mean in terms of PI

Note. The original face is denoted as Face 13 if applied.

all three key ad metrics appears to be mediatedby masculinity, attractiveness, trustworthiness, andcompetence; the effect of eigenface features on pur-chase intention is partially mediated by aggressive-

Figure 5 Maximum Absolute Percentage Improvement forHigh-Rating Participants, by Segment

0.0

10.0

20.0

30.0

40.0

50.0

60.0

70.0

Bee

r

Res

taur

ant

Job

sear

ch a

gent

Col

ogne

Cof

fee

Com

pute

r

Hot

el

Jean

s

SUV

Spor

ts s

hoes

Cam

cord

er

Car

dea

ler

Max

imum

abs

olut

e im

prov

emen

t (%

)

Product categories

Segment 4

0.0

5.0

10.0

15.0

20.0

25.0

30.0

35.0

40.0

45.0

50.0

Bee

r

Res

taur

ant

Job

sear

ch a

gent

Col

ogne

Cof

fee

Com

pute

r

Hot

el

Jean

s

SUV

Spor

ts s

hoes

Cam

cord

er

Car

dea

lerM

axim

um a

bsol

ute

impr

ovem

ent (

%)

Product categories

Segment 3

Aad

Abrand

PI

Note. Because Segments 1, 2, and 5 are of small size, here we only presentthe segment-level results for Segments 3 and 4.

ness, but aggressiveness does not play a role for atti-tude toward the ad or attitude toward the brand.Compared with Segment 3, Segment 4 is more maleand less white, and it comprises people with holis-tic views who are not visually driven, are analyt-ical, value authority, and do not care much aboutchoosing their own goals. In Segment 4, the effectof eigenface features on ad metrics is mediated byattractiveness, trustworthiness, aggressiveness, andcompetence. Baby-facedness plays a role in attitudetoward the ad and purchase intention, but not in atti-tude toward the brand; masculinity plays a role inattitude toward the ad and brand, but not in pur-chase intention. It is worth noting that even thoughattractiveness, trustworthiness, and competence areimportant mediating variables for both Segments 3and 4, the mediation effect is different across thetwo segments. For example, for Segment 3, com-petence partially mediates the effects of L6 and L7

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


on attitude toward the ad, whereas for Segment 4,competence partially mediates the effect of L1, L4,L5, L6, L7, L10, and L11 on attitude toward thead. Therefore, although competence is important forboth Segments 3 and 4, the faces that are regardedas competent are different for these two segments.The competence order of the 12 stimulus faces forSegment 3 is totally different from the order for Seg-ment 4. Face 9 is regarded as the most competent andFace 3 the least competent for Segment 3; Face 5 isregarded as the most competent and Face 1 the leastcompetent for Segment 4.

Based on these results, we conclude that peoplemake reasonably consistent inferences from facial fea-tures, and substantial heterogeneity exists in the waypeople make inferences.

Effect of Faces on Key Ad Metrics at theSegment LevelHow do the segmentation results help advertisersunderstand people’s heterogeneity in making infer-ences from facial features? To address this question,we compared the best faces of mean at the segmentlevel (see Figure 4).

We found substantial heterogeneity from segmentto segment. All three panels in Figure 4 show that dif-ferent faces are preferred for different segments anddiverge from the best face of mean at the aggregatelevel. For example, for the computer ad, participantsat the aggregate level favored Face 7 the most andFace 5 the least in terms of evaluation of attitudetoward the ad, but at the segment level, participantsin Segment 3 favored Face 1 the most and Face 4 theleast, whereas those in Segment 4 favored Face 2 themost and Face 1 the least.

Substantial heterogeneity can also be easily seen bycomparing high-rating people at the segment level,as shown by the maximum absolute improvement inpercentage of high-rating people at the segment level(see Figure 5). Different faces are preferred, and themagnitude of face effect also differs across segments.Interestingly, in contrast with the diverging trend forthe three key ad metrics at the aggregate level (seeFigure 3), the three key ad metrics move in a simi-lar direction at the segment level. This provides evi-dence that segmentation with three dependent vari-ables works.

Hence, ad agencies should not only be careful whenselecting faces for their ads, but they also should takeheterogeneity among target audiences into consider-ation. Ideally, multiple advertisements, each with adifferent face, should be created to appeal to differentsegments of the target audience.

Conclusions and DiscussionIn this paper, we empirically demonstrated the effectof different faces on advertising effectiveness for vari-ous product categories. To answer the three questionsthat were raised at the beginning of this paper, weconclude that (1) faces do affect how a viewer reactsto an ad on the metrics that advertisers care about,(2) the effect is substantial, and (3) people show rea-sonably consistent preferences toward faces, and sub-stantial heterogeneity among individuals and productcategories exists in how viewers react to advertise-ments. Moreover, eigenface features are practical forsegmenting people based on their preferences towarddifferent faces.

The present study contributes theoretically to theexisting literature in several ways. First, we focusedspecifically on faces in print advertisements andempirically demonstrated the effect of different faceson advertising effectiveness by using real ads and realfaces. Second, we introduced a quantitative method,the eigenface method, into marketing research, whichwill hopefully encourage future face studies in mar-keting. Third, we provided evidence for the interac-tion of face effects with product categories, and weempirically demonstrated that hedonism is a usefulcategory feature that differentiates product categoriesand interacts with face effects. Finally, we resolvedthe controversy over people’s heterogeneity in facepreferences while contributing to the face literature ingeneral.

In practice, this study has several implications foradvertisers and ad agencies. The substantial effect ofdifferent faces on advertising effectiveness indicatesthat ad agencies should be careful when selectingfaces to appear in print advertisements. Ad agen-cies should also pay attention to possible heterogene-ity in the preferences of the target audience and usedifferent faces to target different customer segments.In addition, the methods used in the present studyprovide a new approach for professionals interestedin conducting a quantitative study to assist in thescreening and selection of print media spokespersons.It is worth noting that for the categories that includedthe original faces, the original faces were very rarelyselected as the best face of either mean or high rat-ing. Thus, the ad hoc face selection process in practicedoes not generate the best result, showing the valueof our quantitative method.

We propose using the eigenface method for facescreening purposes. From an ad agency perspective,such face screening can be done with the follow-ing steps. First, ask modeling agencies to submit thefaces of the models they represent and pool all of thefaces into one database (containing perhaps 1,000 ormore faces of professional models). Second, decom-pose these faces into loadings on a set of eigenfaces.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


As a result, every face in the face database will be rep-resented by a vector of loadings on these eigenfaces.Third, measure the hedonism of a product categoryand identify the segment(s) that best match the seg-ments of the target customers. Fourth, identify thetop 5 or 10 faces from the face database using theresults from this paper. These top faces could be usedfor copy testing later. The ad agency need only dothe first and second steps once. The third step is spe-cific for each client and requires some very reasonableeffort. The fourth step is also specific for each clientbut can be done automatically. Note that an ad agencycan conduct a study similar to what we did in thispaper for a specific client and a specific product cate-gory using the client’s target segment as the sample.The results will be used for the third step above. Thismay yield a more precise choice of face to representthe client, and it obviously requires substantially moreeffort (but is still feasible).

The eigenface method is good for most practicesthat use faces extensively, such as print ads that fea-ture a real human face or a virtual agent that fea-tures a synthetic face. Unlike a real human face whoseeigenface loadings are fixed corresponding to a spe-cific face, a virtual agent’s face is synthetic so thatits eigenface loadings can be more easily changed soas to morph the face. For a virtual agent, the stepsdescribed above can be used not only for face screen-ing but also for face morphing to construct the opti-mal virtual agent face for a certain condition.

This paper is the first attempt to apply a quan-titative approach to studying faces in the market-ing field. We hope the introduction of this quantita-tive approach (specifically, the eigenface method) willencourage fruitful future research on faces. Here, welist several promising directions for future research.First, the eigenface method suffers from the limita-tion that it is hard to interpret and quite nonintu-itive. By looking at the eigenfaces, it is very hard totell exactly how each eigenface is different from theothers. Future research should focus on creating anintuitive way to interpret the eigenface method andits result. Second, this paper focuses on static faceswith constant expressions; future research could goone step further by studying the effect of differentfacial expressions. Third, in the current study, we con-trolled all elements except for the faces and productcategories. It might be interesting to study the inter-action of face effect with other elements (body shape,hairstyle, costume, gender, age, etc.) in the future.Fourth, we used one category feature, hedonism, torepresent product categories. Other category featuresmight also interact with face effect. In addition tocategory features, other variables in the ad context,such as brand personality (Solomon et al. 1992), might

interact with face effect as well. It might be worth-while exploring other variables for interactions withface effect. Fifth, we used greedy search for variableselection. Greedy search is a generic method that issometimes unable to generate the global optimalmodel. It might be worthwhile exploring other vari-able selection methods (e.g., the Metropolis-Hastingsmethod) in future research. Finally, this study focuseson print ads. Future researchers might consider study-ing the effects of faces in video ads, given the domi-nant role of TV advertising in total media ad spending(Nielsen 2013).

AcknowledgmentsThe authors thank Gary Lilien, Wayne DeSarbo, Jun Liu,Xinwei Deng, and participants from the 2011 MarketingScience Conference in Houston and the Annual Ph.D. Stu-dent Presentations at Pennsylvania State University for theirhelpful comments. This research is supported by a SmealSmall Research Grant and a National Natural Science Foun-dation of China fund [Grant 71232008].

ReferencesAndrews R, Currim I (2003) A comparison of segment reten-

tion criteria for finite mixture logit models. J. Marketing Res.40(2):235–243.

Baron R, Kenny D (1986) The moderator-mediator variable dis-tinction in social psychological research: Conceptual, strate-gic, and statistical considerations. J. Personality Soc. Psych.51(6):1173–1182.

Berry DS, McArthur LZ (1985) Some components and consequencesof a babyface. J. Personality Soc. Psych. 48(2):312–323.

Bishop CM (1996) Neural Networks for Pattern Recognition, 1st ed.(Oxford University Press, Oxford, UK).

Bishop CM (2006) Pattern Recognition and Machine Learning SpringerScience + Business Media, New York).

Bower A, Landreth S (2001) Is beauty best? Highly versus normallyattractive models in advertising. J. Advertising 30(1):1–12.

Brahnam S (2002) Modeling physical personalities for virtual agentsby modeling trait impressions of the face: A neural networkanalysis. Doctoral dissertation, Department of Computer Sci-ence, City University of New York, New York.

Carré JM, McCormick CM, Mondloch CJ (2009) Facial structure is areliable cue of aggressive behavior. Psych. Sci. 20(10):1194–1198.

Carroll B, Ahuvia A (2006) Some antecedents and outcomes ofbrand love. Marketing Lett. 17(2):79–89.

Collins M, Zebrowitz L (1995) The contributions of appearance tooccupational outcomes in civilian and military settings. J. Appl.Soc. Psych. 25(2):129–163.

Crowley A, Spangenberg E, Hughes K (1992) Measuring the hedo-nic and utilitarian dimensions of attitudes toward product cat-egories. Marketing Lett. 3(3):239–249.

Cunningham MR (1986) Measuring the physical in physical attrac-tiveness: Quasi-experiments on the sociobiology of femalefacial beauty. J. Personality Soc. Psych. 50(5):925–935.

Cunningham MR, Barbee AP, Pike CL (1990) What do womenwant? Facialmetric assessment of multiple motives in the per-ception of male facial physical attractiveness. J. Personality Soc.Psych. 59(1):61–72.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.


Cunningham MR, Roberts AR, Barbee AP, Druen PB, Wu CH(1995) Their ideas of beauty are, on the whole, the same asours: Consistency and variability in the cross-cultural percep-tion of female physical attractiveness. J. Personality Soc. Psych.68(2):261–279.

Dumas R, Testé B (2006) The influence of criminal facial stereotypeson juridic judgments. Swiss J. Psych. 65(4):237–244.

Englis BG, Solomon MR, Ashmore RD (1994) Beauty before theeyes of beholders: The cultural encoding of beauty typesin magazine advertising and music television. J. Advertising23(2):49–64.

Ekman P, Friesen WV (2003) Unmasking the Face: A Guide to Rec-ognizing Emotions from Facial Expressions (Malor Books, Cam-bridge, MA).

Freiden J (1984) Advertising spokesperson effects: An examinationof endorser type and gender on two audiences. J. AdvertisingRes. 24(5):33–41.

GfK MRI (2011) Despite risks, celebrity endorsers raiseprint advertising awareness. Press release (February 23).Retrieved January 2014, http://www.gfkmri.com/assets/PR/GfKMRI_022311PR_CelebrityEndorsers.htm.

Gilly M (1988) Sex roles in advertising: A comparison of televisionadvertisements in Australia, Mexico, and the United States. J.Marketing 52(2):75–85.

Goldsmith R, Lafferty B, Newell S (2000) The impact of corporatecredibility and celebrity credibility on consumer reaction toadvertisements and brands. J. Advertising 29(3):43–54.

Gorn G, Jiang Y, Johar G (2008) Babyfaces, trait inferences, andcompany evaluations in a public relations crisis. J. ConsumerRes. 35(1):36–49.

Hauser JR, Urban GL, Liberali G, Braun M (2009) Website morph-ing. Marketing Sci. 28(2):202–223.

Heath T, McCarthy M, Mothersbaugh D (1994) Spokesperson fameand vividness effects in the context of issue-relevant thinking:The moderating role of competitive setting. J. Consumer Res.20(4):520–534.

Inman JJ, Winer R, Ferraro RS (2009) The interplay among categorycharacteristics, customer characteristics, and customer activi-ties on in-store decision making. J. Marketing 73(5):19–29.

Koda T, Maes P (1996) Agents with faces: The effect of personal-ization. Proc. 5th IEEE Internat. Workshop Robot Human Comm.,189–194.

Miniard P, Bhatla S, Rose R (1990) On the formation and relation-ship of ad and brand attitudes: An experimental and causalanalysis. J. Marketing Res. 27(3):290–303.

Nielsen (2013) Global AdView Pulse Lite Quarter 4, 2012. Report(April 22). Retrieved January 2014, http://nielsen.com/us/en/reports/2013/global-adview-pulse-lite- - -q4-2012.html.

Ohanian R (1991) The impact of celebrity spokespersons’ perceivedimage on consumers’ intention to purchase. J. Advertising Res.31(1):46–54.

Pincott J (2010) Why women don’t want macho men. Wall StreetJournal (March 27). Retrieved January 2014, http://online.wsj.com/news/articles/SB10001424052748704100604575145810050665030.

Raghunathan R, Irwin J (2001) Walking the hedonic product tread-mill: Default contrast and mood-based assimilation in judg-ments of predicted happiness with a target product. J. Con-sumer Res. 28(3):355–368.

Ramaswamy V, DeSarbo WS, Reibstein DJ, Robinson WT (1993)An empirical pooling approach for estimating marketing mixelasticities with PIMS data. Marketing Sci. 12(1):103–124.

Schwartz SH (1992) Universals in the content and structure of val-ues: Theoretical advances and empirical tests in 20 countries.Adv. Experiment. Soc. Psych. 25:1–65.

Solomon M, Ashmore R, Longo L (1992) The beauty match-uphypothesis: Congruence between types of beauty and productimages in advertising, J. Advertising 21(4):23–34.

Stirrat M, Perrett D (2010) Valid facial cues to cooperationand trust: Male facial width and trustworthiness. Psych. Sci.21(3):349–354.

Todorov A, Mandisodza A, Goren A, Hall C (2005) Inferencesof competence from faces predict election outcomes. Science308(5728):1623–1626.

Turk M, Pentland A (1991) Face recognition using eigenface. J. Cog-nitive Neuroscience 3(1):71–86.

Urban G, Hauser J (1993) Design and Marketing of New Products,2nd ed. (Prentice Hall, Englewood Cliffs, NJ).

Vermunt J, Magidson J (2005) Technical Guide for Latent GOLD 4.0:Basic and Advanced (Statistical Innovations, Belmont, MA).

Wheatley J (1971) The use of black models in advertising. J. Mar-keting Res. 8(3):390–392.

Zebrowitz L (1998) Reading Faces: Window to the Soul? New Directionsin Social Psychology (Westview Press, Boulder, CO).

Zebrowitz L (2006) Finally, faces find favor. Soc. Cognition 24(5):657–701.

Zhao W, Chellappa R, Rosenfeld A, Phillips P (2003) Face recogni-tion: A literature survey. ACM Comput. Surveys 35(4):399–458.

Dow

nloa

ded

from

info

rms.

org

by [

128.

118.

207.

120]

on

08 D

ecem

ber

2014

, at 1

2:19

. Fo

r pe

rson

al u

se o

nly,

all

righ

ts r

eser

ved.

just the faces: exploring the effects of facial features...

Documents