PSYCHOLOGICAL SCIF.NCE

Research Article

ARE AVERAGE FACIAL CONFIGURATIONS ATTRACTIVEONLY BECAUSE OF THEIR SYMMETRY?

Gillian Rhodes,'- Alex Sumichr and Graham Byatt'--' University of Western Australia. Perth. Australia, and -University of Canterbury. Christchurch. New Zealand

Abstract—Several commentators have suggested that the attractive-ness of averai-e facial configurations could be due solely to associatedchanges in symmetry. If this symmetry hypothesis is correct, then aver-ageness should not account Jor significant variance in attractivenessratings when the effect of symmetry is partialed out. Furthermore,citanges in attractiveness produced by manipulating the averageness ofindividual faces should disappear when all the images are made per-fectly symmetric. The e.xperiments reported support neither prediction.Symmetiy and averageness {or distinctiveness. the converse of aver-ageness} made independent contributions to attractiveness (Experi-ments I and 2), and changes in attractiveness resulting fmm changes inaverageness renuiined when the images were made peifectly symmetric(Experiment 2). Tlie.se results allow us to reject the symmetry hypothe-sis, and strengthen the evidence that facial averageness is attractive.

Images of surprising beauty can be produced by averaging facestogether (Galton. 1878: Langloi.'i & Roggmmi. 1990; Langloi,s.Roggman. & Mtisselman, 1994; Rhodes & Tremewan, I996J. Tbisobservation has generated considerable debate about why such imagessboiitd be attractive. Are they attractive because they are average, orare other factors respon.sible?

Pait ofthe appeal of uveriiged composites no doubt stems from tbesmootb complexions and soft-focus look that result when faces areblended (Benson & PerTett, 1992). Blending artifacts cannot be thewboie story, however, because averageness is still attractive whenihesc artifacts are eliminated (e.g., when line drawings of faces areused; Rhodes & Tremewan. 1996).

Another striking feature of averaged facial composites is their highdegree of bilateral symmetry, which results from the elimination offluctuating asymmetries (small, random deviations from perfect sym-metry) wben laces are averaged together.' Thi,s observation has ledseveral commentators to conjecture that symmetry may explain theattractiveness of averaged composites or of averageness generally(Alley & Cunningham. 1991; Grammer & Thornbill. 1994; Ridley.1992: Thornhill & Gangestad. 1993).- We refer to this conjecture astbe symnwny hvpoihesis.

Langloiset al. (1994) have rejected the symmetry hypothesis, argu-ing tbat facial symmetry is not attractive and therefore cannot accountfor tbe attractiveness of averaged composites. They sbowed tbat sym-metry did not correlate wiih attractiveness, and that perfectly symniet-

Address correspondence to Gillian Rhodes. Department of Psychology,Univensity of Western Au.stralia. Nedlami.s. Penh. WA 6907, Australia; e-mail:gi II @psy,uwa.edu.au.

1, Averaged faces may nol be pcrleclly symmetric because direclionala.symmetrics, which have a coiisisieni bias in the population, are not eliminatedby averaging.

2, Thornhill and Gangostad (1993) also noted tliat symmetry and average-ness could both contribute to ihe itltractiveness of averaged compo.site.s.

ric versions of face.s were less attractive than the originals. Otber stud-ies have also found ihat perfectly symmetric versions of faces are lessattractive than the originals (Kowner, 1996; Samuels. Buttcrworth,Roberts, Graupner, & Hole, 1994; Swaddle & Cutbill, 1995). Howev-er, these studies are not without flaws. The symmetric images (exceptin Swaddle & Cuthill. 1995) were constructed by reflecting eachhemiface about the midline. With this technique, asymmetries iit thelocation of midline features, or in the size of the hemifaces (Previc,1991), can result in odd-looking symmetric images (e.g,. abnormal eyespacing, nose width, or aspect ratio). It is difficult, therefore, to deter-mine whetber the un attractiveness of these images is due to their sym-metry or their structural abnormalities. In Swaddle and Cutbill's(1995) study, the generalizability of the results was limited by failureto display the whole face (only iniemal facial features were shown).

Recently, we (Rhodes, Pn)ffitt, Grady. & Sumich. in press) haveshown that perfectly sytumetric images, made by blending eacb facewith tls mirror Image, are more attractive tban the original faces, andthat the attractiveness of individual faces can be altered by manipulat-ing their level of symmetry.- Signillcant correlations between facialsymmetry and attractiveness have also been reported in several studies(Gramtner & Thornbill. 1994; Jones & Hill. 1993, for some ethnicgroups; Rhodes ct al.. in press; Zebrowitz. Voinescu. & Collins, 1996).These findings lit well with evidence that bodily symmetry is attrac-tive to humans and many otber species (Brooks & Pomiankowski.1994; Concar. 1995; Mpller & Pomiankow.ski, 1993; Thornhill &Gangestad, 1994: Watson & Thornhiil. 1994),

The symmetry hypothesis cannot, therefore, be ruled out a priori onthe grounds tbat symmetry is not attractive. We report here the resultsof two studies designed to test tbe symmetry hypothesis. If it is cor-rect, then symmetry should explain a significant proportion ofthe vari-ance in attractiveness judgments when the effect of averagencss ispartialed out. and averageness should not account for a significant pro-portion of variance when the effect of symmeti-y is partialed out. Wetested these predictions using regression analyses in Experiments Iand 2, We also reasoned ihat if changes in attractiveness produced bymanipulating averageness are due solely to associated changes indegree of symmetry, then changes in attractiveness should disappearwhen all the images are made perfectly symmetric. We tested this pre-diction in Experiment 2.

EXPERIMENT 1

Subjects rated tbe lUiractiveness of male and female averaged com-posites (eacb created by mathematically averaging 24 same-sex faces),and low-, normal-, and high-averageness versions of 48 faces. The

3, The perfecily symmetric images were more uiiructive than other blends,suggesting that the result is not a blending ariifact.

52 Copyright © 1999 American Psychological Sociely VOL. 10. NO. I.JANUARY 1999

PSYCHOLOGICAL SCIENCE

Gillian Rhodes. Alex Sumich. and Graham Byatt

high- and low-averageness versions were created by warping pho-tographs of faces (normal averageness) halfway toward (high) or awayfrom (low) the same-,sex average (,see Method). This procedureextends Rhodes and Tremewan's (1996) line-drawing inanipulation ofaverageness to photographic images, and we expected to replicate thefinding that attractiveness increases with manipulated averagene,ss.Such a result would provide strong additional evidence that average-ness is attractive in the absence of blending artifacts.

We also obtained symmetry, averageness, and expression ratings forall the images. Our main aim was to determine whether variations inaverageness account for significant variance in attractiveness ratingswhen the effects of symmetry are partialed out. Alternatively, changesin symmetry and averageness could both contribute to variance inattractiveness ratings. A secondary question was whether posiliviiy ofexpression also contributes to the attractiveness of averageness. Aver-aged composites have pleasant expressions (see Fig, 1), and low-aver-ageness faces created by caricaturing sometiiiies have slightlyunpleasant expressions. Positive expressions are attractive (e.g.. Cun-nhigham, Roberts, Barbee, Druen. & Wu, 1995), so we hypothesizedthat expression might contribute lo the attractiveness of averagenes.s."'

Method

SubjectsThirty-six Caticasian undergraduates (18 males. 18 females)

received SIO each for participating in a 1-hr session.

StimuliBlack-and-white photographs were taken of 48 young adults (24

Tiiale, 24 female) from senior high school classes and the universitycommuTiity. All faced the camera directly and displayed neutralexpressions. Gryphon's Morph'^" software was used to manipulate theaverageness of the faces. First, a fixed set of 120 landmark points wasfound (by eye) on each face, aTid their locations iTidicaied with amouse-click. These points captured the spatial layout of the internalfeatures, face outliTie, inner and outer hairline, ears, and any prominentfacial lines (cf, Rhodes & Tremewan, 1996). The program automati-cally joined subsets of these points with smooth curves to create facialcontours, and interpolated three additional points between each pair ofpoints on a contour. The resulting 656 points were u,seJ to manipulatethe images. Second, the program created an average male compositeand an average female composite, by computing the mean location ofeach of these 656 points for same-sex faces, warping individual same-sex faces onto this average configuration, and then averaging the gray-level values in corresponding regions of the faces (see Beate & Keil,1995, for more details). This averaging technique should produceaverages superior to those produced by pixel-based techniques that donot align facial features or anatomical locations before averaging (cfLanglois & Roggman, 1990). Finally, the high- and low-averagenessversions of each face were created by moving each point (along thedifference vector connecting corresponding points on the same-sexaverage and the target face) either halfway toward (high) or the samedistance away from (low) the average, and remapping the original pat-tern of gray levels, or "texture." onto this new configuration. Eachimage measured approximately 12.5 x 10 cm.

4. Any such contribution mighl be small given that the original faces al!liad nciiual expressions.

Fig. 1. Male and female averages from Experiment I, Each is theaverage of 24 faces.

The stimuli consisted of the male and female avei-ages (averagedcomposites) together with a high-averageness version, a low-average-ness version, and the original (normal-averageness) version of eachface. A second set of images was created by placing each face insidean oval mask, which hid most of the hair but displayed the face out-line, chin, and inner hairline. Masked versions of the stimuli are shownin Figure 1 (male and female averages) and the top row of Figure 2(low-, normal', and high-averageness versions).

VOL. 10, NO. 1. JANUARY 53

F.SYCH01,0GICAL

Facial Attractiveness

Fig. 2, Examples of low-, nonnal-. and high-averageness versions of aface. The examples in the top row were tised in both experiments, andthose in the bottom row are perfectly symmetric version.s of the sameimages, used in Experiment 2 only.

ProcedureOne version of eacb face was assigned to each of three blocks, bal-

anced for averageness level and sex of face. The male and female aver-ages were added to Block I. Blocks were presented in three differentorders (123. 231. 312). witb image order randomized within blocks.SuperLab (CedmsCorp.) controlled presentation of the tmages on aPowerMac 6300 computer and recorded responses. Each image wasdisplayed until subjects made a rating, using tbe keyboard keys Ithrough 7. Faces were rated on 7-point scales of attractiveness, aver-ageness. symmetry, and expression (1 = negative, 7 = positive). All theimages were rated on one scale before tbe next scale was presented.Attractiveness was always rated first, and the order of tbe other threerating tasks was counterbalanced with block order and sex of subject.Half the subjects saw masked and balf saw unmasked faces, counter-balanced across task order, block order, and sex of subject. Subjectswere tested individually.

Results and Discussion

All four .sets of ratings were highly reliable, with Cronbach alphasfrom .94 to .97. A four-way analysis of variance (ANOVA) was car-ried out on the attractiveness ratings, with display condition (mask vs,no mask) and sex of subject as between-subjects factors, and average-ness level (low vs. normal vs. bigb) and sex of face as repeated mea-sures factors.-' Planned comparisons were used to test predicteddifferences between averageness levels.

5. Syminelry, ijveragcncss. and expressicm ratings were collected primari-ly for use in the regression analy.scs. bill ANOVA. ^ on these nitings showed thatall three increased significiintly with increases in (maniptiUued) averagenesslevel. The averaged coinpo,sitcs were aiso rated io^ more symmetric (bolh Ma -(y6l than any individual face, a,* expected if averaging eiiniinaies lluciiiatinga,symme tries.

Mean attractiveness, averageness, symmetry, and expression rat-ings were calculated for each image, averaging across subjects, A mul-tiple regression analysis was carried out on tliese ratings, withattractiveness as the dependem variable and averageness, symmetry.and expression as tbe independent variables.

Attractiveness ANOVAAttractiveness increased witb manipulated averageness, as ex-

pected, ra, 64) = 288.26. /; < .0001 (Ms = 2.4, 3.4, and 3.8 for low,noritial. and bigb averageness, respectively: /s > 6.62. /'s < .0001),Tbis result extends Rhodes and Tremewan's (1996) findings witb linedrawings to more realistic pbotograpbic images. There was a signifi-cant effect of sex of face (female: A = 3.3. male: M = y.\).F(\. 32) =11.96. /; < .002, and a tbree-way interaction between sex of face, aver-ageness level, and display condition. F{1. 64) = 9.84, p < .0002.Female faces showed a bigger drop in attractiveness tban male faceswben distorted away from the average wben tbe hair was visible (pre-sumably because ofthe greater distoning effect on female bair, whichtends to be fuller and longer tban male bair). This interaction does notqualify our main finding that attractiveness increased with average-ness, because tbat finding was observed for both sexes and displayconditions (for low, nonnal. and high averageness, respectively, Ms =2.6. 3,5, and 3.9 for masked female faces: Ms = 2,3, 3.?, and 3,9 torfemale faces with no mask; A/s = 2.4, 3,4, and 3.8 tor masked malefaces; Ms = 2.2. 3.2. and 3,5 for male faces with no mask).

Attractiveness of averaged compositesThe average male face was more attractive (M = 4.6) than any indi-

vidual male face, and the average female face was more attractive (M= 5.2) than all but one female face (M = 5.3).

Predicting attrcictivenessOverall, averageness. symmetiy. and expression accounted for

71'/r of tbe variance in attractiveness ratings. F(3, 142) ^ 121.12,/J <.0001. Table 1 shows the zero-order and partial correlations of eachindependent variable with attractiveness. The main results of interestare the partial cotrelations. whicb indicate the (unique) contributionof each variable when ihe effects of the otber variables ;ire partialedout.

There was no support for the symmetry hypothesis. Although sym-metry explained significant variance in attractiveness ratings (evenwhen tbe effects of averageness and expression were partialed out),averageness still accounted for significant variance when the effect ofsymmetry (and expression) was partialed out. Therefore, even thoughsymmetry and averageness were significantly correlated, li 144) = .70,p < .001 (all versions). K46) = .30, p < . 05 (undistorted faces), theattractiveness of averageness was not due solely to this accompanyingincrease in symmetry. Averageness, symmetry, and expression eachcontributed independently to variance in attractiveness ratings. Thisfinding held when all versions (low, normal, and bigb averageness,plus the averaged composites) of the faces were considered, and whenonly the undistorled faces were considered.

Wben male and female faces were considered separately, tbeseconclusions held for all versions of the faces. However, when theundistorted faces were considered (reducing the range of each variableand tbe power of tbe analyses), only averageness made a (significant)unique contribution to male attractiveness, and only expression madea unique contribution to female attractiveness.

54 VOL. in. NO. I.JANUARY IW9

PSYCHOLOGICAL

Gillian Rhodes. Alex Sumich, and Graham Byatt

Table 1. Partial and :.ero-oriler {in parentheses) correlations of attractivenesswith ayerageness. symmetry, and e.xpres.sion: Experiment 1

Faces

All versionsAll {n = 146)Male (n = 73)Female (/; = 73)

UndistortedAll {n = 48)Male (n = 24)Female {n = 24)

. , , < m . . , , < . O L . .

Correlations withaverageness

40*** (,7t***).63*** (,88***)28* (.60*'*'*)

•^" j* / 4'^***"i

.75*** (.83***)

.01 (.12)

V<.o.),.

Con'elations witlisymmetry

.54*** (.76***)

.40*** (.81***)

.51*** (,7I***)

.40* (.46***)

.07 (.53**),30 (.37)

Conelations wilhexpression

.45*** (.46***)

.33** (.25*)

.54*** (.64***)

,47** (.46***),23 (.19)

EXPERIMENT 2

In Experiment 2, we tested the symmetry hypothesis experimen-tally. If the changes in attractiveness a,ssociated vfith our manipulationof averageness in Experiment 1 were due solely to changes in sym-metry, as asserted by the symmetry hypothesis, then they should dis-appear when all the images are made perfectly symmetric. We testedthis prediction by comparing the attractiveness of the images fromExperiment I and their symmetric counterparts (Fig. 2, bottom row).

We also attempted to replicate our regression results, with tvi'o pro-cedural improvements. First, we replaced averageness ratings withdistinctiveness (the converse of averageness) ratings, in case subjectsinterpreted "averageness" to mean "average-looking" (i.e., not partic-ularly "good-looking"), rather than spatially average, as intended. Asin previous studies, we interprei these distinctiveness ratings as a (con-verse) measure of averageness (Rhodes &. Tremewan. 1996), Second,we obtained each set of ratings from a different group of raters loavoid any carryover effects between rating scales.

In addition, we asked subjects to rate opposite-sex faces on "appealas a ptitential life partner," in a preliminary investigation of whetherperceived averageness or symmelry might influence reproductivechoices, as predicted by biological accounts of such preferences(Kaliek, Zebrowitz. Langlois. & Johnson. 1998: M0ller & Swaddle.1997; Shackelford & Larsen. 1997; Thomhill & Mtiller. 1997),

Method

SubjectsOne hundred and eighty Caucasian undergraduates (90 males. 90

females) received $5 each for participating. None had participated inExperiment I.

StinntliFor each image in Experiment I, including the male and female

averages, we created a perfectly symmetric composite by blendingthat image with its mirror image, using Gryphon's Morph"'"' software(see Rhodes et al., in press, for details). This procedure averages thelocations of corresponding landmark points on the two images, warpseach image onto that average configuration, and average,s the gray-level values in corresponding facial regions. Blemishes and jewelrywere removed from the original images using the cloning tool in

Adobe Photoshop before ihe symmetric versions were made. Six ver-sions of each face were used in the experiment: the low-, normal-, andhigh-averageness versions from Experiment I. and their perfectlysymmetric counterparts (see Fig. 2). Normal and pertectly symmetricversions of the male and female averages were also used. All imageswere displayed in the oval masks used in Experiment 1.

ProcedureThe procedure was similar to the procedure in Experiment 1, wilh

the following differences. Six blocks of trials were constructed byassigning one of the six versions of each face to each block, counier-balancing averageness level, symmetry level, and sex of face acrossblocks. The male and female averages and their perfectly symmetricversions were assigned to Block 1. Six different block orders wereused (123456, 234561. etc.), with image order randomized withinblocks. Block order was counterbalanced with ,sex of subjeci. Five dif-ferent groups of 36 subjects (18 male. 18 female) rated all the imagesfor either attiacliveness, symmetry, distinctiveness. expression, ormale appeal (opposite-sex faces only). Subjects were instructed to ralea face as distinctive if il would be easy to spot in a crowd and to ralemate appeal high il the person looked appealing as a potential lite part-ner (i.e.. someone to setde down and have children with). Subjectswere requested lo use the full range of the scales.

Results and Discussion

All sets of raiings were highly reliable, with Cronbach alphas from.91 lo ,98. A four-way ANOVA was earned mil on the mean attrac-tiveness ratings, with sex of subject as a between-subjects factor, andaverageness level (low vs. normal vs, high), symmetry' level (normalvs. perfect), and sex of face as repeated measures factors. A three-wayANOVA was carried out on mean mate-appeal ratings, with .' ex ofsubjeci as a between-subjecls factor, and averageness level and sym-metry level as repeated measures factors. Planned comparisons wereu.sed to test predicted differences between averageness levels.

Mean altractiveness, mate-appeal, distinclivencss. symmetry, andexpression ratings were calculated for each image, averaging acrosssubjects. Two multiple regression analyses were carried out on Lheseratings, one with attractiveness as the dependent variable and the olherwith mate appeal as the dependent variable. In both cases, distinctive-ness, symmetry, and expression were the independent variables.

VOL, m. NO. I.JANUARY 1999 55

PSYCHOLOGICAL SCIENCE

Facial Atiracliveness

Attractiveness ANOVAThe increase in attractivcnes.s a.ssociated with an increase in aver-

agcnes.s level was tinaffected by making all tlie iinage.s perfectly syni-nietric (see Fig. 3), Attractiveness increased with averageness level.Fa. 68) = 214.57, p < .0001 (A-/s = 2.1, 3.0, and 3.4 for low, nonnal.and high averageness. respectively: /s > 6,65, ps < .0001), and withsyriiinelry level, F( I. M) = 72.22, p < .0001 (Ms = 2,7 and 3.0 for nor-mal iiiu! perfect symmetry, respectively), but there was no interactioitbetween averageness level and symmetry level, F < \. These restdtsare inconsistent with the symmetry hypothesis.

There were several effects involving sex of face, but these did notaffect our coitcltision that averageness is attractive und that its appealis not due solely to changes in symmetry. Female faces were generallymore attractive than male faces, F{]. 34) = 6.05. p < ,02, and anincrease in symmetry increased the attractiveness of female faces (per-fect symmetry: M = 3.1, normal symmetry: M = 2.8) mc»re than malefaces (perfect symmetry: M = 2.H. nonnal symmetry: M = 2,6), F{i,34) = 14.10, p < .0006. There was also a three-way interactionbetweea sex of face, symmetry level, and averageness, F{2,68) = 4.44.p < .02 (see Fig. 3). This interaction has no obvious theoretical signif-icance, and planned comparisons showed that the increase in attrac-tiveness from low- to normal- to high-averageness versions wassignificant for both male and female faces, at both symmetry levels, allts> 10.16, ps< ,0001,

Attractiveness of averaged compositesAs expected, male and female averages were more attractive (male

average: M = 4,4, feuiak' average: M = 4.K) than any (undistottcd)sainc-sex face. Similar results were obtained when the perfectly sym-metric averages were compared with perfectly symmetric individualfaces. The symmetric tnale average was more attractive (M = 4,1) thanany indiviJtial symmetric male face, and the symmetric female aver-age was more attractive (M - 4.9) than all but two symmetric individ-ual female faces {Ms - 5.0 and 5,1), The averages with normal and

atc'•atcV)tn0)c(D

6

5

4-

3-

2-

perfect symmetry - F facesperfect symmetry - M facesnormal symmetry - F facesnormal symmetry - M faces

Low Normal High

Averageness

Fig. 3. Mean attractiveness ratings as a function of averageness level,symmetry level, and sex of face in Experiment 2. For clarity, SE barshave been omitted (maximum S£' = 0.15), F = female, M = male.

perfect symmetry did nol differ significantly in attractivenes.s. foreither male or female images, both /s < I,

Mate-appeal ANOVAThe mate-appeal ratings showed that averageness and symmetry

were both appealing in a potential life partiici", Fl2, 68) = 149.54, /) <,0001 (A/s = 1.8, 2,5. and 3,0 for low. normal, and high averageness)and Fi 1, 34) = 34.40, /; < .0001 (A/s = 2.3 and 2.5 for normal and per-fect symmetry). Averageness interacted with symmetry level. F(2, 68)- 6.01. /)< .004, However, rather than the effect of averageness beingreduced when all the images were perfectly symmetric, as expectetl ifthe appeal of averageness is due to an accotnpanying increase in sym-mett^, the effect was slightly increased (Fig. 4). Moreover, maleappeal increa,sed with averageness for both symmetry levels, all /s >10,82, /)s < .0001. indicating that the mate appeal of more averagefacial configurations cannot be due solely to their greater symmetry.

Mate appeal of averaged compositesThe average male and female faces had greater mate appeal (male:

M = 4.8, female: M = 4.5) than any individual same-sex face. The per-fectly symmetric average male and female faces also had greater mateappeal (male: M = 4.7, female: M = 4,5) than any individual synunet-rk same-se,\ face. Averages with nonnal and perfccl symmetry did notdiffer in mate appeal for either male or female images, hoth fs < I.These results exactly parallel those obtained for attractiveness.

Predicting attractivenessOverall, disiinctiveness, symmetry, and expression explained 67%

of the variance in attractiveness ratings for the full set of images, FO.288) = 199.08, p < ,0001. Table 2 shows the zero-order and piuiial cor-relations of each independenl variable with attractiveness. There wasno support for the symmetry hypothesis. Although symmetryexplained significant variance in attractiveness judgments when theeffects ot distinctiveness (converse of averageness) and expression

c

Rat

"rafit

Ap

pe

raE

ean

6

5

4

3

2

perfect symmetrynormal symmetry

Low Normal

Averageness

High

Fig. 4. Mean tnate-appeal ratings as a function of averageness leveland symmetry level in Experiment 2. SE bars are shown.

56 VOL, 10. NO. I.JANUARY

PSYCHOLOGICAL SCIENCE

Gillian Rhodes. Alex Sumich, and Graham Byatt

Table 2, Fartial and zero-order (in pareniheses) correlations of attractiveness withflistinctiveness. symmetry, and expression:

Faces

All versionsAll (n = 292)Male(H= 146)Female {n = 146)

UndistoiiedAll (» = 48)Male in = 24)Female (;i = 24)

*p < .05. **/3 < .01. ***

Coirelations withdistinctiveness

-.71*** (-.77***)-.88*** (-.90***)-.50*** (-.67***)

_ ^^5* (_ 57*** '_ -j2'^=** 1— 80***'

.23 (-.24)

/x.OOI.

E.\periment 2

Correlations withsymmetry

.30*** (.43***)

.25** (.43***)3'^*** / 44***)

.48*** (.51***)

.32 (.60**)

.46* (.43*)

Correlations withexpression

.37*** (.46***)34*** (29***)45*** ( (-,2***)

45** (.45***).43* (.28).66*** (.65***)

were parlialcd out. distinctiveness still accounted for significant vari-ance in attractiveness after the effect of symmetry (and expression)was partialed out. Therefore, although symmetry and distinctivenesswere significantly negatively correlated. r{290) = -.33. p < .001 (allversions). r(46) - -.34, /) < . 02 (undistorted faces), the unattractive-ness of distitictivcness was not dtie solely to tlic accompanyingdecrease in symmetry.

Distinctiveness, symmetry, and expression each contributed uniquevariance to attractiveness. These results held for hoth male and femalefaces when all versions of the faces were con.sidercd. but for undis-torted faces, the distinctiveness eifect was restricted to male faces (cfaverageness result in Experiment 1). and the synuiietry effect tofemale faces. There was a narrower range of averageness (ExperimentI) and distinctiveness (Experiment 2) ratings lor female (4.0-5.9,averageness: 2.6^.7. distinctiveness) than male faces (3.9-6.1, aver-ageness; 2.6-5.1. distinctiveness), but this seeins unlikely to fullyaccount for the sex difference. Nor could a sex difference in the rangeof ratings account for restriction of the .symmetry effect to femalefaces, because the range was larger for male (2.0-4.7) than female(2.6-5.1) faces. A greater effect of symmetry on female than maleattractiveness was also found in the ANOVA results for attractiveness.These two sex differences were also appiu-ent in the results for all ver-sions of the faces (although in that analysis both factors were signifi-

cant predictors for both sexes). Therefore, it is possible thai facia!symmetry is a strongei' pE-eiiictor of female than male atU'activeness.and that facial averagene.ss (converse of distinctiveness) is a strongerpredictor of male than female attractiveness. Other studies, however,have found either no sex difference (Thomhill & Gangestad. 1994) orthe opposite sex difference (Gangestad. Thomhill. & Yeo, 1994), forbodily symmetry. Moreover, although averageness c<intributed only tomale attractiveness in Grammer and Thomliill's study (1994), aver-ageness was unattractive (cf. our finding that distinctiveness was unat-tractive). These inconsistencies indicate that the sex differences weobtained should be viewed with caution.

Predicting mate appealOverall, the three variables explained 60'! ofthe variance in mate-

appeal ratings for the full set of images, F(3. 288) - 147.20. /; < .OOOI.There was no support for the symmetry hypothesis (see Table 3). Dis-tinctiveness was a significant predictor of mate appeal even when theeffect of symmetry (and expression) was partialed out. The overall pat-tern of results was similar to that obtained for attiactivenes.s, althoughdistinctiveness and symmetry appeared to have stronger effects onattractiveness than on mate appeal, whereas expression had a strongerinfluence on mate appeal. The latter result may reflect a sensible desirefor a pleasant disposition in a life partner.

Table 3. Pariial ctnd zero-order (in parentheses) correlaiions of mate appeal withdistinctiveness, symmetry, and expression: Experiment 2

FacesCorrelations with

distinctivenessCorrelations with

symmetryCorrelations with

expression

All versionsAll (/I = 292) -.61*** (-.6K***)Male (n= 146) -.67*** (-.68***)Female (n = 146) -.37*** (-.59***)

UndistortedAll(n = 48) -.35* (-.50***)Male(/; = 24) -.47* (-.49*)Female (/(= 24) .04 (-.29)

.15t (.34***)3 I * * * ( 3C|***\

.30^^

.12

.41-i

(.37**)(.34)(.37)

.47*** (.54***)44=1:** (.38***)57*** (.59***)

5Q*** / 52**-*')

.47* (.35)72*** / 75***1

t/x.lO. -05. .01.

VOL. 10. NO. 1. JANUARY i999 57

RSYCHOLOGICAL SCIENCE

Facial Attractiveness

More generally, the finding that averageness and symmetry areiippealinjz in a life partner provides preliminary evidence that the.sepreferences might intluence reproduciive behavior. More direct evi-dence ot" an influence on reproductive behavior and fitness is nowneeded, to test predictions that these preferences increase reproductivesuccess (e.g., Grammer& Thomhill, 1994; Symons. 1979).

CONCLUSIONS

Our results settle the dispute about whether the attractiveness ofavenigcness is due solely to associated changes in symmetry. Aver-ageness remained a significant predictor of attractiveness when theeffect of symmetry was partialed out. and changes in attractivenessproduced hy experimentally manipulating the averageness of indi-vidual faces remained when the images were made perfectly sym-metric. These results rule out the symmetry hypothesis. Theattractiveness of average facial configurations cannot be attributedsolely to their symmetry.

Blending artifacts can also be ruled out as a complete account ofIhc attractiveness of average facial configurations. Rhodes and Treme-wan (1996) showed that individual line-drawing faces can be mademore (or less) attractive by distorting them towiU'd (or away from) anaverage eonllguration. TTie use of line drawings allowed blending arti-facts to be eliminated, but at the cosi of losing realism. By extendingthese results (o more realistic photographic images (also without intro-ducing blending lutifacts), we have further strengthened the case thataverageness is attractive.

It now seems clear thai the attractiveness of averageness cannot beexplained by symmetry, blending artifacts, or a variety of other factors(see Langlois et al.. 1994. for a review). Perhaps it is time to shiftattention frotn the question of whether averageness is genuinely attrac-tive to the question of why averageness is attractive. At one level, onemight ask what mechanisms render average faces attractive. Does pro-totype abstraction play a role, as has been hypothesized (Langlois &Roggman. 1990)7 Is this a kind of mere-exposure effect, whereby thesiniilarily ofan unseen average to many seen faces generates a positiveresponse to that average? Does some (unconscious) statistical analysisof a population of faces contribute? Al another level, one can ask whyhumans find averageness attractive. Is the preference culturallyinstilled, or has it been built in by evolutionary processes, perhapsbecause it increases reproductive success? Attention to these questionsshould enrich understanding of human nature as well as the nature ofbeauty.

Acknowledgments—This work was supported by grams from the M;irsJetiFoundation of New Zealand and the Australian Research Council.

REFERENCES.4lti;y. r.R.. & Cunningham, M.R. (IWt). Avcrugeii faces arc attraciivL-. lun very atitaclivc

laces lire not iiverage. I'syiliohinirulScmue. 2. 121-12.'i.Beak-. J.M.. & Kei]. F.C. (t995). Caiegoritat effects in ilic pert-epiiun of facts. Conniiiim.

57. 217-239.Uenson. P.. & Perfett. D. (1992). Face to face with the perfect image. New Scientist. /SOT,

Bruoks. M.. & Pomiankow.ski. A. (1994). .Symmetry is in the eye of the tieholdcr. Trendsin Ernhgy and Eviiluiinii. 9. 201-202.

Concar. D. (19951. Sex mid the syninietriL-;tl liiicty. Ni'w .Stienii.it. 146. 41M4.Cunningham. M.R-. Robens. A.R.. Barbee. A.I', Druen. P.B.. & Wu. C.-H. (19')^). -Their

idcii.s of beauiy are, on lhe whdie, ihc .same its imrs": Cotisistenfy and variabiliiy inthc crcjss-ciilttir.it percepiion of female physical aitracliveness. Jutinuil nf t'er.siiinit-ily and Sitiial Psychology. 68. 261-279.

GalKm, F. (1878). Conipiisite portraits, .tnumal nf the Atuhniitiiliigical /n.siiiule of (ireninrittiin iH Ireland. H. L12-I42.

Gangesuid. S.. Ttiomhili. R.. & Yeo, R. (I'J'M). Facial uttractivene.ss, tievclopmentui slii-bility itnil llucluiiling aBynimetrj. Ethotoay & Sociohiotogy. 15. 73-85.

Grajnmer. K., & Tbtirnhill, R. (1994). Human {Homo siipieii.1) facial attraciivettes.s andsexual selection: The rote of .symmetry and averageness. Jminuil of CiimpurnihrI'syclwhsy. 108. 233-242.

.lones. D.. & Hill. K. (1993). Criteria of facial attractivene.ss in live populations. HiimcinM-mirc.J. 271-2%.

Kalick. S.M.. Zebriiwit/, L.A.. t^nglois. J.H., & Johnson. R.M. 11998), Diies human facialaitrac live ness honestly advenise health? Longitudinnl rfala on an evolutionary ques-tion. P.sychologiail Science. 9. 8-13,

Kowner, R. (19%). h'acial asymmetry and attractiveness judgment in developmental per-spective. Jounml of Experimenial Psycliol/ifi\: Human Fenipiiim and I'erfcr-maiue. 23. (i62-67.'i.

Larglois. J H,. & Rocgman, L.A. (1990), Attnictive faces are only average. l'.\\iiuili>i;''al.SV;«wv, ;, 1I.S-I2I,

I-iinglois. J.H., Roggman. L.A,, & Mu.sselman, L. 11994). What is average and whiil is nolaverage about ailractive faces.' P.\yclnili>i:kiil .Stieiue. 5. 214-220.

Moller, A.P., & Pomiankowski. A. (1993), Fliicluatinp asymmetry and sexual selection.Genetiici. 89. 2(-)7-279.

Mwller, A.P., & Swaiidle, J.P. (1997). A.'iymmelry. rleveliipmt'iilal stability <tnd evcliitum.Oxford. England: Oxford University Press.

Previc. RH. (19911. A general theory eoneerning the prenatal origins of cerebral lateral-i/ation in humans. Psychoio^iva! Review. 98. 299-334,

Rhodes;. G., Proffitt. R. Grady. J,. & Siimich. A. (in press). Facial symmetry and the per-ception of beauiy. P.iychcmtimii- Hulletin & Review,

Rhodes. G.. & Tremewan, T. (1996). Averageness. exaggeration, and hicial attractiveness.Psychological Science. 7. 105-110.

Ridley. M. (1992). Swallows attd scorpionllies find symmeliy is lieiiuliful. Science. 257.327-328.

Saiiiiiels, C.A.. Butierworth. C... Roberts. T. Gratipner. L,. & Hole. G. (1994). Racial aes-thetics: Babies prefer attractiveness lo symmetry, Perceptiim. 2.?, N2.'-831.

Shackclford. T.K.. & Larsen. R.J. (1997). Facial a.iynmieiry as an indicator of psycholog-ical, emotional, and ph)'siological distress. Journal of PersiimditY mid Sociut P.sv-ch>lo,Ky. 72.456-166.

.Swaddle. J.P.. & Cuthill, I.C'. (1995), Asymmetry and human facial aitractivencss: Sym-metr> may nol always be beautiful. Procfedin^a nf ihc Rovat Society of London. B.261. 1 l l - l 16.

Symons, D. (1979). The cvolutiini ofhumtm .'iexiuitiry. Oxford, England: Oxford tlniversi-ly Press.

Thornhili. R.. & Gangestad, S,W. (1993). Hunwn facial heauly, tinman Nature. •/,237-269,

Thomhill. R,. & Gatigestad, S.W. (1994). Human fluctuating asymmetry and sexualbehavior. Psychotogical Science. 5. 297-302.

Thomhill. T.. & Moller, A.P. (1997). Developmental stability, disea.se and medicine, liio-logicat Reviews. 72. 497-548.

Watsim, P.M., SThomhill. R. 11994). Fluctuating a-sytntneiry and sexual selection. Tremttiin F.coliifiy ami Evolution. 9, 21-25,

Zebrowit/, L.A.. Voiitescu, L.. & Collin,s. M.A. (1996). Wide-eyed and crooked-faced—Dclcmiinanls of perceived and real honesty across thc lifespan. Permnaiily andSocial Psychology Bulletin. 22. 125S-1269.

IRLCEIVED 11/18/97; ACCEPTED 3/16/98)

58 VOL. 10. NO. 1. JANUARY