Personality judgments from natural and composite facial images: more evidence for a “kernel of truth” in social perception

8/3/2019 Personality judgments from natural and composite facial images: more evidence for a kernel of truth in social perception.

1/35

PENTON-VOAK ETAL.PERSONALITY JUDGMENTS FROM FACIALIMAGES

PERSONALITY JUDGMENTS FROM NATURAL

AND COMPOSITE FACIAL IMAGES: MORE

EVIDENCE FOR A KERNEL OF TRUTH

IN SOCIAL PERCEPTION

Ian S. PentonVoakUniversity of Bristol

Nicholas PoundBrunel University

Anthony C. LittleUniversity of Liverpool

David I. PerrettUniversity of St Andrews

In addition to signaling identity, sex, age, and emotional state, people fre-quently use facial characteristics as a basis for personality attributions. Typi-

cally, there is a high degree of consensus in the attributions made to faces.Nevertheless, the extent to which such judgments are veridical is unclear andsomewhat controversial. We have examined the relationship between selfre-port and perceived personality using both faces of individuals and computergraphic composites. Photographs were takenof 146menand 148womenwhoeachalsocompleteda selfreport personality questionnaire fromwhich scoreson thebigfive personality dimensionswere derived. In study 1, we identified arelationship between selfreported extraversion and perceived extraversion inindividual faces. For male faces alone, we also found some accuracy in theperception of emotional stability andopenness toexperience. In study 2, com-posite faces were made from individuals selfreportinghigh andlowscoresoneach of the five dimensions. These composites were rated for personality andattractiveness by independent raters. Discriminant analyses indicated that,controlling for attractiveness, independent ratings on congruent personality

dimensions were bestable todiscriminatebetween composite faces generated

607

Social Cognition, Vol. 24, No. 5, 2006, pp. 607-640

Correspondence to Ian PentonVoak, Email: [email protected]


2/35

from individuals high or low on the selfreport dimensions of agreeableness,extraversion, and, for male faces only, emotional stability.

It would be easy to start this article with a quote from a novel or playrelating an individuals facial features to hisorherpersonality attrib-utes or character, since the belief that faces portray character is ubiq-uitous. This persistence of lay theories of physiognomy is hard tocredit, given the longstanding assumption that such attributionsare considered to be erroneous by orthodox science. Nonetheless,in 1974 Liggetreported that 90%of undergraduate students believedthat the face is a valid guide to character. By2000, this figurewas still

around75%(Hassin & Trope,2000),andthe importance of this beliefcan hardly be overstated. A recent article demonstrated that infer-ences of competence made from politicians faces predictedU.S. con-gressional election outcomes at levels far above chance (the mostcompetent looking candidatewonin nearly70%of the2004 senateraces; Todorov, Mandisodza, Goren, & Hall, 2005). Within psychol-ogy, it is generally thought that the trait inferences made from faces,although consistent, have very low validity. Here, however, we

briefly review the growing literature suggesting that somewhat ac-curate perception of personality is possible from facial characteris-tics alone. In the two studies reported here, we assess the accuracy ofresponses made to individual faces and to computer composite

faces.There is a considerable literatureonthenatureof theattributions wemake when encounteringother people.Someofthis workhasinvesti-gated the inferences that are made to target faces associated with cer-tain behaviors,typicallybypairing face stimuli with brief descriptionsof behavior (Todorov & Uleman, 2002, 2004). These studies suggestthat these spontaneous trait inferences become an integral part of therepresentation of a person, rather than mere associations caused bythecooccurrence of faces andbehaviors, assuggestedbyearlier work(e.g.,Skowronski, Carlston, Mae, & Crawford,1998).Given thatsocialperception functions to guide our own actions and predict theactionsof others, incorporating trait attributions into representations of indi-

vidual conspecifics seems judicious.Nevertheless, the willingness of people to make personality attri-butions in the absence of behavioral cues (i.e., from facial informa-tion alone) is perhaps somewhat surprising. We seem to make suchinferences about personality traits spontaneously and without re-

608 PENTON-VOAK ET AL.


3/35

flection or deliberation (Hassin & Trope, 2000; Todorov et al., 2005).In short, many aspects of person perception could be considered to

be intuitive in a two-system framework (Kahneman, 2003).Socialcognitive approaches to person perception often stress

economy in information processing. Perceptual cues mayact as trig-gers for categorization of individuals, which may in turn lead to theactivation of stereotypes (Cloutier, Mason,& Macrae, 2005). In manycases, the cues to category membership (sex or age, for example) arethemselves accurate, even if the stereotypes thenactivated arenot. Itis intriguing to consider the possibility that there may be perceptualcues that accurately inform judgments of personality. Should suchcues exist, a number of questions emerge in person perception, be-

cause trait attribution ceases to be a generally inaccurate cognitiveshortcut and becomes instead a potentially adaptive perceptualability.

Physiognomy has a bad reputation in psychology, and this islargely well deserved: most studies in the area have been reso-lutely unscientific, leading physiognomists to be dismissed as char-latans (or worse). Studies carried out in the first half of the 20thcentury demonstrated that measurements of individual facial fea-tures seldom, if ever, correlate with psychological characteristics(e.g., Cleeton & Knight, 1924). Since the 1940s, although work in thisarea has been sporadic, it has generally yielded similar negative re-sults (e.g., Alley, 1988, for review). More recent work, however, hassuggestedthat theremayin fact besome observable relationships be-tween physical appearance and personality that could have formedthe basis for the beliefs, if not the methods, of the physiognomists.One difference between more recent and earlier approaches to judg-mental accuracy is that later research concentrates less on fea-ture-based relationships (e.g., do long noses accurately indicate anyaspect of personality?) and more on configural properties of faces(Hassin & Trope, 2000). Configural approaches are more consistentwith current theories of face processing in generalcognitive repre-sentations of faces seem to be more holistic than representations ofother objects such as houses (Tanaka & Farah, 1993). Here the terms

configural and holistic should be taken as referring to the diver-sity of possiblecues to personality, includingtherelative positionsofthe different parts of the face. A configural or holistic approach doesnot ignore individual facial features; rather, it does not specify inadvance exactly what constitutes a feature. Previous featurebased

PERSONALITY JUDGMENTS FROM FACIAL IMAGES 609


4/35

approaches have considered particular structures as features (e.g.,the nose or the lips). However, a feature can be more broadly consid-ered as anything that is visual and potentially diagnostic. For exam-ple, the spaces between and arrangement of structures moretraditionally considered as features (e.g., the space between eyesand eyebrows) may be a cue to some aspects of perceivedpersonality.

Ratherthancorrelating physical measurements withselfreportedpersonality traits, more recent studies concentrate on the relation-ship between strangers impressions of a person, configural proper-ties (suchasthemultiple facialcues that contribute toa babyfacedappearance), and the personality that that person selfreports

(Zebrowitz, 1998).

STUDIES AT ZERO ACQUAINTANCE

Studies using zero acquaintance paradigms (in which participantsrate the personality of strangers) have found a surprising degree ofcorrelation between selfratings and stranger ratings on personalitydimensions, often using five-factor models of personality (e.g., Nor-man, 1963). Many trait theorists agree that five trait dimensions (de-rived from factor analyses of ratings of trait adjective pairs) seem torepresent a reasonable compromise that is at least a partially accu-

rate measure of peoples personalities (e.g., Barrett & Pietromonaco,1997; McCrae & Costa, 1987; Watson, 1989).Of particular relevance to thework presented in this article, is pre-

vious research employing a fivefactor personality model that hasdemonstrated a surprising correlation between stranger ratings andselfratings of an individuals personality. Zero acquaintance stud-ies show that people are able to generate consensual impressions ofothers personalities. Not only do different judges generate similarresponsesto a given target individual, butalso these impressions aremore accurate than would be expected by chance (i.e., these impres-sions agree with the targets selfratings). Whilethese studies dem-onstrate that trait attribution on zero acquaintance can be

surprisingly accurate, few have studied specifically the role of theface alone in these judgments.In Passini and Normans 1966 experiment, small groups of under-

graduates were put together fora 15-minute periodandasked,with-out verbal interaction, to rate each other using scales corresponding



5/35

to the big five personality factors. They found correlations greaterthan 0.3between self andothers ratings forExtraversion, Consci-entiousness, and Culture/Openness to experience factors. Corre-lations for Agreeableness (r = 0.15) and Emotionalstability/Neuroticism (r = 0.02) failed to reach significance (Passini& Norman, 1966). Clearly, however, the 15 minutes that the groupspent together provided opportunities for personality judgments to

be influenced by nonfacial characteristics, such as clothing, posture,and other nonverbal behavior (eye contact, smiling etc.).

Albright, Kenny, and Malloy (1988) replicated Passini and Nor-mans findings forthe traits sociable (extraversion) andresponsi-

ble (conscientiousness), and found that the physical attractiveness

of subjects correlated positively with ratings of the traits sociable,talkative, and goodnatured, suggesting a halo effect. In addi-tion, there wasa strongpositive association between raters opinionsof targets neatness of dress and ratings of conscientiousness, indi-cating that nonfacial cues could be employed to make judgments(Albright et al., 1988). Kenny, Horner, Kashy, and Chu (1992) repli-cated Albrightet al.s findings usingvideotapes of individuals ratherthan thesmall-group paradigm, indicating that consensusat zero ac-quaintance is notdependenton interactionsbetween theindividualsrating each other.

Passini and Normans findings were also replicated by Watson(1989), who used a small-group experiment, but allowed subjects tosay one sentence introducing himself or herself to the rest of thegroup. For individual judges, there was significant selfpeer agree-ment for ratings of Extraversion and Conscientiousness, but not forother personality dimensions. By averaging ratings across all judgesthat had rated a target individual, Watson also determined that in-creasing the number of judges increased the size of the selfpeer cor-relations for Extraversion (r = .43), Agreeableness (r = .31),Conscientiousness (r = .28), and Culture (r = 0.20, ns) but not Emo-tional Stability ratings (r = 0.04).

Individual differences have been found in both how accurate peo-plearewhen judging othersandinaccuracy when being judgedone-

self. Ambady, Hallahan, and Rosenthal (1995) found thatindividuals who rate themselves highly on extraversion and agree-ableness aremore accurately perceivedbyothers. They also reportedthat women aremore accurate judgesofstrangers personalities thanare men, consistent with other literature on nonverbal behavior.



6/35

Although individual differences in zero acquaintance judgmentsdo exist, Albright et al. (1997) demonstrate the crosscultural natureof consensus of personality attributions. Using a small-group para-digm with Chinese students in Beijing, they reported a pattern of re-sults similar to those found in Western studies. In a further study,American participants judged personality traits of Chinese studentsfrom photographs, and vice versa. They found that Chinese andAmerican participants showed much crosscultural agreement intheir judgments of personality at zero acquaintance (Albright et al.,1997).

Borkenau and Liebler (1992) attempted to identify the sources ofvalidity for zero-acquaintance judgments using Brunswiks (1956)

lens model, which attempts to define which objective cues (such as asmile or thin lips) are employed (correctly or incorrectly) to perceivepersonality. Judgeswere presentedwith film with sound,silent film,a still photo,or just theaudiotapeof the target reading. Thestill-pho-tograph condition (most relevant to the studies reported in this arti-cle) shows that even with extremely limited information and notransient facial expression, there were significant correlations be-tween judges and selfratings of Extraversion (r = 0.33) and Consci-entiousness (r = 0.32) while ratings of Agreeableness showed anonsignificant but positive relationship (r = 0.19). The application ofthe lens model was, however, somewhat unsuccessfulmeasuringdiscrete characteristics seems a less productive research strategy

than concentrating on configural cues in this area (Zebrowitz &Collins, 1997).

Other studies essentially replicate the findings reported here (seeKenny, Albright, Malloy, & Kashy, 1994, for a review) suggestingthat, from limited information (even still photographs), judges tendto agree on the personality of other individuals. Furthermore, be-cause these ratings appear to correlate reliably with selfratedextraversion and conscientiousness in all studies, and selfratedagreeableness and opennessintellect in a subset of the literature,they are to some extent valid.

ECOLOGICAL PERSPECTIVES ON ACCURATEPERSONALITY PERCEPTION

Studies have also found some evidence foraccuracyin socialpercep-tion using models of personality other than the big five." Typically,



7/35

these studies are based within a Gibsonian/ecological theoreticalframework(see Zebrowitz & Collins, 1997, forreview). Forexample,Berry and Brownlow (1989) found that ratings of male facialbabyishness were positively correlated with the face ownersselfreported approachability and warmth, but negatively related toselfreported aggression. For female faces, babyishness ratingswere associated with low selfreported levels of physical power andassertiveness. Bond, Berry,andOmar (1994)demonstrated that indi-viduals whose faces are rated as appearing less honest are morelikely to volunteer for experiments that involve them in deceivingothers than peoplewho are judged to look more honest. Selfreportsof personality attributes are subject to social desirability influences,

so it is interesting to note that facialappearance is related to behavioras well as self-reported traits.

Fewstudies cangenuinelyclaimto bebased onphysiognomiccuesalone since in many experiments thin slices of behavior are pre-sented, providing a multitude of nonfacial cues. Even in studies ofstatic facial photographs, nonfacial cues such as clothing and hair-style are often visible. It is not difficult to see how a neat,wellgroomed appearance could accurately signal conscientious-ness, for example, while fashionable hairstyles may be associatedwith extraversion. Thecurrent studyaims to overcome some ofthesepotential problems by employing computer graphics techniques toeliminate nonface-based cues to personality.

USING COMPUTER GRAPHICS TO STUDYPERSONALITY ATTRIBUTION

Computer graphic composites (averages) can be used to extract thedefining characteristics of a group while losing the characteristicsthat make each face look individual (Rowland & Perrett, 1995). Forexample, the average of several young males looks both young andmale, whereas the average of older females looks older and fe-maleyet neither image bears great resemblance to any of theconstituent faces.

Average faces are constructed by delineating feature points (e.g.,

the outline of the eyes, the mouth) on each individual face in thegroup. From the delineation data of all the faces in the set, the aver-age position ofeachfeature point onthe facecan be calculated. Com-puter graphics techniques allow the rendering of average color and



8/35

texture information onto the average face shape, thus generatingphotorealistic composites. These techniques are related in concep-tion, if not in execution, to the photographic composite techniquesused by Galton toward the end of the 19th century (Galton, 1878).

Within an apparentlyhomogeneous population, subgroupscanbeaveraged to generate visibly different composites. Such techniqueshave been used in studies that test biologically based theories of fa-cial attractiveness (e.g., PentonVoak et al., 2001; Perrett, May, &Yoshikawa,1994). Although the differences between subgroup com-posites aresubtle, they reliably generate differingsocialperceptions.This key property of composites suggests that subgroups of individ-uals with particular selfreported personality traits (e.g., extraverts)

canbe averagedtogether togeneratenewfacesthat embody that per-sonality trait. The facial characteristics shared by individuals whopossess a particular trait should remain in the composite, leading toattributions of that trait (e.g., extraversion) to the composite face. Inthis article, we attempt to generate composite faces that embody thecharacteristics of people who self-report being at the extremes ofeach of the dimensions in the fivefactor model of personality (i.e.,extravert and introvert composites; neurotic and stable composites;agreeable and disagreeable composites; conscientious and notconscientious composites and open and closed to experiencecomposites).

The major advantage of using composites generated using this

computer graphics technique is that individuating or idiosyncraticcharacteristics not linked to attributions of the personality trait inquestion should be averaged out and not be appreciable in the finalimage. Since the technique is empirical, it does not presuppose anysingle feature or configuration underlying attributions, but it doesassume that information related to personality will be selectivelymaintained by averaging. As such, the technique can potentially lo-cate either individual features that are characteristic of personalitytraits or multiple features that support configurational/holistic

judgments of a trait. Another benefit is that characteristics that arepotentially linked to personality attribution but are temporary (i.e.,hairstyle; facial jewelry; blemishes, not all of which are removablewith careful image editing) will not be present in the final compos-ites. As such, computer graphics techniques are a novel and poten-tially informative way to assess the accuracy of personality



9/35

judgmentsbased on facialcharacteristics. In study1, we replicatetheaccuracy found in earlier studies of personality judgment to individ-ual faces, using a large sample of nearly 300 faces. In study 2, we cre-ated compositefacesfromindividuals who self-reported themselvesto be at the extremes of each of thebig five dimensions, and assessedwhether raters could accurately we assess the personality of theindividuals who made up the composites.

STUDY 1: ASSESSING ACCURACY OFRESPONSES TO INDIVIDUAL FACES

METHODS

Volunteers (146 male and148female) were photographed(inneutralpose) andeach completed a 40 item selfreport personality question-naire (Botwin, Buss, & Shackelford, 1997). All participants were un-dergraduate students at the University of Stirling aged between 18and 22 years old. A confirmatory factor analysis (independently foreach sex, varimax rotation, constrained to five factors) was used togenerate scores on the big five personality dimensions. For femaleparticipants, the five factors explained 46.3% of the variance in re-sponses; agreeableness (8.8%), conscientiousness (9.7%),extraversion (10%), emotional stability (11.4%) and openness (6.4%).For male participants, the five factors accounted for 39.9% of vari-

ance in the questionnaire responses: agreeableness (7.3%), conscien-tiousness (8.4%), extraversion (7.5%), emotional stability (9.6%) andopenness (7.1%)

Photographs of the individuals were then rated on each of theper-sonality dimensions (rating scale = 17 where 7 = highest) by groupsof 10 raters. The faces were presented sequentially on a computermonitor in random order, and each rater group rated all the individ-uals of onesexononeof thefive personality dimensions. Raterswerealso undergraduate students at the University of Stirling (63 femaleand 37 male; mean age 20.8 years) and were instructed to press thespace bar (rather than submit a rating) if they happened to be famil-iar with any of the presented faces. Reliability of these personality

ratings was good(Cronbachs>

0.70 foralldimensions exceptmaleconscientiousness [ = 0.63]; Table 1).



10/35

RESULTS

For males, there were significant positive correlations between per-sonality ratings based on photographs and selfreport scores forextraversion, emotional stability,and openness to experience. For fe-males, there was a significant positive correlation between ratingsand selfreport scores for extraversion (Tables 2a & 2b). In addition,for both male and female faces, there were also many significant cor-relations between ratings on the five personality dimensions (Table3). The correlations are significantly positive for socially desirabletraits (high emotional stability, high extraversion, agreeableness,contentiousness, and openness to experience).

Since scores on some rated personality dimensions were not just

correlated with scores on the congruent selfreport dimensions,stepwise linear regressions were carried out to determine whichselfreport personality dimensions were the best predictors of rat-ings byobserverson each of thebigfivepersonalitydimensions. Sep-arate analyses were carried out for male and female faces, and foreach of the big five rated personality dimensions. The results of thestepwise linear regressions indicated that for males selfreportscores on congruent dimensions were the best predictors of ratingsof extraversion, emotional stability, and openness. For females,selfreported extraversion was the best predictor of extraversionratings (Table 4).

DISCUSSION

The most robust evidence for accuracy in social perception comesfrom selfstranger agreement in ratings of extraversion across sev-


TABLE 1. Reliability (Cronbachs ) of ratings of 148 female faces and 146 male facesby groups of 10 raters

Female Male

Agreeableness 0.705 Agreeableness 0.747

Conscientiousness 0.775 Conscientiousness 0.625

Extraversion 0.888 Extraversion 0.736

Emotional Stability 0.782 Emotional Stability 0.753

Openness 0.736 Openness 0.735


11/35

eral studies and diverse personality instruments (see Zebrowitz &Collins, 1997, for a tabulated review of this literature). Our resultsreplicate theseearlier findings,withsignificant, positivecorrelations

between otherrated and selfreported extraversion of both maleand female targets. Stepwise regression indicated that selfreportedextraversion better predicted extraversion ratings elicited by indi-vidual faces than did scores on any of theother selfreport personal-ity dimensions. Previous literature indicates that accurateperception of other traits is less consistent across studies. We foundno other evidence of selfstranger correlations that were consistentacross both target sexes, but emotional stability and openness to ex-perience were also accurately perceived in male target faces. Accu-

rate perception of an openness/culture dimension has beenreported in zero acquaintance studies (e.g.,Passini & Norman, 1966)and for some adjective pairs in response to static facial images (e.g.,polishedcrude in Borkenau & Liebler, 1992). Borkenau andLiebler also report accurate perception ofemotionalstabilityinmale,

but not female, targets, as we report here. As in other studies, the re-lationship between selfreported personality and strangerratertraits is stronger for male than for female target faces.

The observed sex differences in selfreported personality matchsex differences in social stereotypes. Women consider themselves to

be lower in emotional stability and higher in agreeableness than domen(Costa, Terracciano,& McCrae, 2001),so it isessentialto analyze

the male and female targets separately. If raters consider womenmore agreeable than men, and targets report the same, significant

but somewhat misleading accuracy may be discovered(Paunonen, 1991).

While many studies have found somewhat accurate judgments oftargets extraversion, there are considerable inconsistencies acrossstudies in the accuracy of judgments of other traits. This may reflectthe use of different instruments to assess target personality. For ex-ample, Norman (1963) used 20-item 5factor questionnaires,whereas Costa, Terracciano, and McCrae (2001) employed the60-item Neurotocism, Extraversion, and Openness (NEO) personal-ity inventory. Additionally, varied tasks have been used to investi-

gate the raters" attributions of personality (e.g., one item trait scalesasusedhereand in, for example,Albrightetal., 1988, versusthe mul-tiple-item trait scales used in some other articles, e.g., Borkenau &Lieber, (1992). In this experiment, we had a relatively large number



12/35


TABLE2a.Crosscorrelationsbetweenselfandotherra

teddimensionsforindividualmalefaces).

SelfReportPersonality

RatedPersonality

Agreeableness

Conscientiousness

Extraversion

EmotionalStability

Opennes

s

Agreeableness

0.007

0.151

0.003

0.013

0.04

4

Conscientious

ness

0.005

0.005

0.168*

0.031

0.01

7

Extraversion

0.069

0.136

0.237**

0.229**

0.01

4

EmotionalSta

bility

0.029

0.161

0.008

0.182*

0.04

4

Openness

0.200

0.114

0.072

0.048

0.22

0**

Note.Congruentratingsunderlined.*p

Documents

Personality judgments from natural and composite facial images: more evidence for a “kernel of truth” in social perception