Upright versus upside-down faces: How interface attractiveness varies with orientation

Transcription

1 Perception & Psychophysics /994,56 (2), / Upright verss pside-down faces: How interface attractiveness varies with orientation KARL-HEINZ BAML UniversiUit Regensbrg, Regensbrg, Germany A choice experiment is reported in which all pairs and triples of faces from a set of eight moderately attractive faces were presented, both pright and pside down, to 103sbjects. In each orientation, the sbjects had to select the face that appeared more (pairs) or most (triples) attractive to them. For each orientation, the preference probabilities that arose from the pair and triple comparisons cold be described by the BTL rle (Lce, 1959). Ths, each face was represented by two scores, one reflecting its attractivenessin the pright orientationand the otherreflectingits attractiveness in the inverted orientation. Orientation affected the preference probabilities. Qalitatively, score ratios betweenfaces decreased from pright to invertedorientation, sggesting thatthe faces became less discriminable by inversion. Qantitatively, the effect of inversion cold be described by a simple rle that assmes a face's two attractiveness scores to be affinely related across orientations. This reslt indicates that inversion affected all faces abot eqally. The presentfindings are discssed with respect to faces' first- and second-order relational properties, a distinction emphasized in crrent theories of face perception. They sggest that the processing of first- and secondorder relational properties is impaired by inversion to roghly the same degree. Recognition offaces is disrpted by inversion to a far greater extent than is recognition of other classes of visal objects, sch as hoses, airplanes, or landscapes. In the literatre, this phenomenon has been referred to as the face-inversion effect. It was first demonstratedby Yin (1969). Sbseqent stdies have replicated this effect nder qite different experimental conditions, and in this way have demonstrated the very general and robst natre of the phenomenon (see Valentine, 1988, for a review). Originally, the inversion effect was interpreted as evidence for specialized face-recognition processes in visal information processing (Yin, 1969, 1970). Then other approaches were proposed in order to accont for the inversion effectwithot presmingface-specific processes (Diamond & Carey, 1986; Goldstein & Chance, 1980; Rock, 1973). The most inflential of these approaches is that ofdiamondand Carey. These researchers proposed a distinction between two types of spatial information that nderlie the processing of visal objects: first-order relational and second-order relational properties. While first-order relational properties refer to information abot the spatial relationships among parts of an object, second-order relational properties refer to information abot the spatial configration between the The athor's thanks are extended to Hede Helfrich-Holter, Jim Tanaka, and Tim Valentine for their comments on an earlier draft of this manscript, and also to Armin Hartinger and Maria Schnelzer for their help in the experimental work. The athor's mailing address is Institt fr Psychologie, Universitat Regensbrg, Regensbrg, Germany ( heinz@rpss3.psychologie.ni-regensbrg.de). parts ofan object, on the one hand, and the prototypical spatial configration of its parts, on the other (Diamond & Carey, 1986; Tanaka & Farah, 1991). Diamond and Carey hypothesized that inversion was particlarly sensitive to the processing of second-order relational properties while it did not affect the processing offirst-order relational properties. They attribtedthe inversion effect to the se of second-order relational properties that are.important for, althogh not niqe to, face recognition. 1 Indeed, the reslts ofseveral stdies spported the view that inversion affected the processing of second-order relational properties (Diamond & Carey, 1986; Maryama & Endo, 1984; Sergent, 1984; Thompson, 1980; Yong, Hellawell, & Hay, 1987). On the other hand, none ofthese stdies demonstratedthat inversion did not affect first-orderrelational properties to the same degree (Rhodes, Brake, & Atkinson, 1993; Tanaka & Farah, 1991; Valentine, 1988). Recently, two stdies reported experimental reslts that were interpreted as evidence against the hypothesis that second-order relational properties were responsible for the inversion effect. In a mental rotation experiment, Valentine and Brce (1988) fond response-time patterns for rotated faces which were similar to those that Shepard and Metzler (1971) fond for rotated three-dimensional block drawings. The response time of same-different jdgments increased linearly as a fnction of rotation angle when the second of a pair of faces was rotated away from the vertical. Valentine and Brce arged that this reslt did not spport the idea that matching two pright faces involved a process that was qalitatively dif- 163 Copyright 1994 Psychonomic Society

2 164 BAML ferent from that sed to match one pright and one inverted face. Instead, not only first-order relational bt also second-orderrelational properties shold have been extracted from a face when it was presented pside down. Tanaka and Farah (1991) examined the hypothesis that second-order relational properties were disrpted by inversion while inversion did no harm to first-order relational properties in a dot-pattern experiment. They compared the effects ofinversion on the identification ofdot patterns that differed in the extent to which they reqired the encoding ofsecond-order relational properties. They fond that both first-order relational and second-order relational properties were affected by inversion. Specifically, the identification of dot patterns that reqired more encoding of second-order relational properties was not more vlnerable to inversion than was identification of dot patterns that mainly reqired encoding of first-order relational properties. Althogh the reslts of Valentine and Brce (1988) and Tanaka and Farah (1991) are consistent with the view that inversion affects faces' first-order relational and second-order relational properties to the same degree, their reslts provide only loose spport for this view. First, Valentine and Brce's (1988) linearity finding sggests that the rotation ofa face in the vertical indces a very reglar change in the processing of its relational properties. This finding, however,cannot exclde the possibility that it may be mainly the processing of secondorder relational properties that is affected by rotation. For instance, linearity may reslt from the fact that an increasing rotation in the vertical does not affect firstorder relational properties bt has a linear impact on the processing of second-order relational properties. Second, Tanaka and Farah's (1991) reslt stems from experiments in which dot patterns were sed. In order to generalize their reslt to sitations involving faces, the processing offirst- and second-order relational properties shold remain the same for highly artificial dot patterns and photographs of real faces. However, the empirical sondness ofthis prespposition, thogh crcial, is not garanteed. The processing of relational properties indced by artificial stimls sets may well be different from that indced by realistic facial stimli (e.g., Valentine, 1988). The present stdy presents an alternative approach to testing the hypothesis of whether the information processed from pright and inverted faces is the same. This approach is different from the ones sed by Valentine and Brce, or Tanaka and Farah, and may be more conclsive on the isse. The approach relies on the application ofa choice paradigm where preference probabilities are separately measredon sets offaces in the pright and the inverted orientations. The point to be emphasized is that the way preferences between faces are affected by orientation may tell s something abot the way in which the processing of facial properties changes from pright to pside-down orientation. Consider a set of faces with given preferences on them in terms of, say, their perceived attractiveness. Sppose that the perceived attractiveness of a face depends on both its first-order relational and its second-order relational properties (see below). Sppose now that we can accont for the effect of orientation on the preferences by assming that inversion affects all faces-that is, their perceived attractiveness-in the same way. Since faces will vary in terms of the contribtion of first- and second-order relational properties to their overall attractiveness, this accont sggested that inversion has the same effects on faces' first- and second-order relational properties. For instance, if inversion affected only faces' second-order relational properties, inversion wold have a stronger impact on faces with a higher contribtion of secondorder relational properties on attractiveness than on faces with a lower contribtion ofsecond-orderrelational properties; inversion wold affect the faces differently. Based on this argment, the degree ofa face-orientation interaction may indicate the extent to which inversion indces a change in the processing of face properties. Finding a sbstantial face-orientation interaction therefore sggests that information processed from pright and inverted faces is different; finding no sbstantial interaction sggests that the information processed is the same. To bring this approach to some direct experimental application, two isses mst be specified. First, a facial attribte mst be chosen that relies both on faces' firstorder relational properties and on faces' second-order relational properties. Second, an adeqate method mst be chosen to reveal the degree of the face-orientation interaction. There is good empirical evidence indicating that the perceived attractiveness of a face is based on both its first-order relational and its second-order relational properties. Several researchers (Langlois & Roggman, 1990; Perrett, May, & Yoshikawa, 1994) hypothesized that the attractiveness ofa face is a fnction ofits closeness, or deviation, from a facial prototype and ths is a fnction ofits second-orderrelational properties. In fact, these researchers reported experimental reslts in spport oftheir hypotheses. These hypotheses, however, can serve as only rogh approximations (Alley & Cnningham, 1991) that leave room for an additional role of first-order relational properties to affect a face's attractiveness. Indeed, reslts from the stdies of Meerdink, Garbin, and Leger (1990) and Cnningham, Barbee, and Pike (1990) indicate that not only second-order relational properties bt also first-order relational properties have an impact on the perceived attractiveness ofa face. To reveal the degree ofa face-orientation interaction, I dedced a theoretical scale from the preference probabilities, separately for the two orientations. Indeed, if preference probabilities in the two orientations flfill certain reglarities (see Sppes, Krantz, Lce, & Tversky, 1989), they can be sed to infer two theoretical scales ofperceived attractiveness. In this case, each face can be represented by two scores, one score reflecting the perceived attractiveness ofa face in the pright orientation and the other reflecting its perceived attractiveness in the inverted orientation. On the basis of these two scales, I

3 UPRIGHT VERSUS UPSIDE-DOWN FACES 165 compared a face's two scores across orientations. The present stdy searches for a face-independent rle that relates the faces' two scores for the two orientations. It is arged that the extent to which sch a rle can accont for the change in the faces' scores reveals the extent to which inversion affected faces' first-order relational and second-order relational properties to the same degree in the present experiment. METHOD Sbjects The sbjects were 103 psychology stdents at the University of Regensbrg. They were tested individally and were given credit for flfilling degree reqirements. Materials Eight frontal-view photographs of stimls faces of moderate attractiveness were sed. The pictres were all of males between the ages of 21 and 26 years, and all were very similar in terms of hair length and shadows. They wore no eyeglasses, beards, or mstaches, and the expression on their faces was netral. All faces were nknown to the sbjects. The pictres were copied individally onto monochrome slides, and copies were made at two different orientations, pright and pside-down. Apparats The slides were presented sing a Kodak Carosel S-RA 2500 projector that was controlled by a compter. The slides were presented on a blank wall, in front of which the sbject sat at a distance of abot 2 m. Each face sbtended abot 8 ofvisal angle. The sbject indicated his or her response by sing psh bttons, and the response was recorded by the compter. Designand Procedre All 8!/(2!6!) = 28 different pairs and a1l8!/(3!5!) = 56 different triples offaces were presentedto each sbject, in both pright and inverted orientations. The pairs and triples were either presented first in the pright orientation and then, abot 1 week later, in the inverted orientation, or vice versa. Fifty-two sbjects started with the pright orientation, and the remaining 51 sbjects started with the inverted orientation. For each orientation, the presentation of the pairs preceded the presentation ofthe triples. An experimental session consisted of two parts, the presentation ofthe pairs and the presentation ofthe triples. At the beginning of each experimental session, all pairs and all triples were mixed randomly, as were a pair's two faces and a triple's three faces. The two (three) slides of a pair (triple) were presented sccessively with a presentation time of 2 sec for each slide. The presentation of the single slides was interrptedby a l-sec blank field. After the presentation of a pair or triple of faces, the sbject immediately selected which ofthe two or three faces he or she preferred in terms ofattractiveness by pressing one of two (pair comparison) or one ofthree (triple comparison) bttons. Following a 2-sec blank field, the next pair or triple offaces was presented. DataAnalysis For each of the two orientations, a sbject's pair comparisons gave rise to a 28 X2 matrix, a sbject's triple comparisons gave rise to a 56X3 matrix. Each cell of the pair (triple) comparisonmatrix was coded as 1 or 0, depending on whether a face was preferred (I) or not (0) in a pair (triple) comparison. These two matrices represented a sbject's preferences with respect to the presented faces. The pair (triple) comparison matrices were smmed over sbjects to reslt in one pair (triple) comparison matrix for each orientation. These matrices represented the data ofthe experiment. To formlate a statistical model of these preferences, the seqence of choices ofa pair's first or second face was viewed as a seqence of Bernolli trials with nderlying parameter Pab, representing the probability that, for a pair (a,b), face a is preferred over face b. The relative freqency with which face a is chosen from the pair (a,b) is taken as an estimate of Pab' Similarly, the seqence of choices of a triple's first, second, and third face was viewed as a seqence of trials with nderlying parameters Pa;bc and Pb;ac, representing the probabilities that, for a triple of faces (a,b,c), face a is preferred over faces b and c, and face b is preferred over faces a and c. Again, relative freqencies serve as estimatesfor the probabilities (see Sppes et ai., 1989). Based on this statistical model, an orientation's preferences were described by 140 (28X 1+56X2) free parameters. In order to infer a simple theoretical scale of attractiveness which will give a parsimonios accont for an orientation's whole set ofpreference probabilities, some restrictions on the preference probabilities mst be flfilled. The Bradley-Terry-Lce rle (Bradley & Terry, 1952; Lce, 1959; Sppes et ai., 1989; in the following referred to as the BTL rle) was fitted to an orientation's two matrices. This rle sets strong restrictions on the relationship between pair and triple preferences. It demands the following property, called the constant-ratio property, to be tre for all triples offaces (a,b,c): Pab Pba Pa;bc Pb;ac where P ab stands for the probability that given the pair (a,b) face a is preferred over face b, and Pa;bc stands for the probability that given the triple (a,b,c) face a is preferred over faces b and c. This property, in effect, asserts that the strength of preference of the triple's face a over the triple's face b is naffected by the other available alternative, the triple's face c. Ifthe constant-ratio property is satisfied, each face a can be associated with a nmerical vale v(a), so that the preference probabilities for a pair of faces (a,b) are determined by the rle Pab = v(a)/[v( a) + v(b)] and the preference probabilities for a triple of faces (a,b,c) are determined by the rle Pa;bc = v(a)/[v(a)+v(b) +v(c)]. These nmerical vales are niqe p to scalar transformations. Ths, only the nmerical vale of one face can be chosen freely. For the eight face stimli employed in the present experiment, the BTL rle therefore reslts in only seven free parameters to describe an orientation's 140 independent data observations. A likelihood-ratio test (cf. Lindgren, 1976) was sed to determine whether the rle fitted an orientation's data. Given the data, the likelihood of the BTL rle (LBTL) was compared with the likelihood ofthe statistical model (Ls) by sing the property that the term - 2lnL BTdLs is approximately chi-sqare distribted. The parameters ofthe BTL rle were estimated by sing the iterative search procedre PRAXIS (Gegenfrtner, 1992). RESULTS AnalysisofPreference Probabilities Table 1 shows the estimated preference probabilities (relative freqencies) for the pair comparisons, both for the pright and the inverted orientations. The Appendix shows the estimated preference probabilities for the triple comparisons. The preference probabilities varied over a large range ofvales, sggesting considerable differences between faces with respect to their perceived attractiveness. This pattern of reslts held for both ori-

4 166 BAML Table 1 Pair Comparison Matrices for the Two Orientations, Upright and Inverted F, F 2 F 3 F 4 Fs F 6 F 7 Fs Upright Orientation F, F F F Fs F F Fs Inverted Orientation F, F F F Fs F F Fs Note-Each'cell represents the estimated probability that one ofa pair offaces is preferred over the other (F; = face nmber n). entations. To test whether an orientation's preferences differed reliably from indifferent choices, I compared, for each orientation, the (perfect) fit of the statistical model that describes the 140 independent preference probabilities (see Method section) with the fit of a statistical model that restricts all preference probabilities to 1/2 (pairs) or 1/3 (triples). The likelihood-ratio tests were condcted with 140 df The X 2(140) vales of 3, (p <.0001) for the pright orientation and 2, (p <.0001) for the inverted orientation demonstrate that preferences differed reliably from indifferent choices. A first visal comparison ofthe preference probabilities across orientations (see Table 1 and the Appendix) sggests that the probabilities were fairly stable across orientations. However, some tendency showed p for the preferences to be closer to indifferent choices in the inverted orientation than in the pright orientation. For instance, 19 ofthe 28 independent preference probabilities for pairs were less close to indifference (p = 1/2) when the faces were presented pright than when they were presented pside-down. To test whether the preference probabilities cold be assmed to be constant across orientations, I compared the (perfect) fit ofa joint statistical model that described the two orientations' 280 (2 X 140) independent preference probabilities (see Method section) with the fit ofa statistical model that restricted the probabilities to being constant across orientations. The X 2 (140) vale of (p =.007) demonstrates that preferences varied reliably across orientations. Ths, both faces and orientation had a reliable effect on the preferences. Qality of Fit of the BTL Rle To get a more detailed insight into how faces and orientation affect preferences, a more parsimonios accont for an orientation's whole data set is sefl. For each orientation, I examined whether the preferences cold be fitted by the BTL rle. Ifthis rle fitted the data well, each face cold be represented by one positive real-valed score that reflected its perceived attractiveness for the particlar orientation. The BTL rle was simltaneosly fitted to an orientation's pair and triple comparison matrices. For each orientation, the BTL rle has seven free parameters to describe the 140 independent preference probabilities. The x 2 (133) vales of (p =.999) for the pright orientation and (p =.974) for the inverted orientation demonstrate an excellent fit of the rle to the data. Ths, for each orientation, the effect ofthe faces on the preference probabilities can be described by the BTL rle. In the top panel offigre 1, the preference probabilities measred in the two orientations with the predictions of these probabilities when sing the BTL rle ("two individal BTL rles") are compared. The data are merged over the two orientations. If the BTL rle held perfectly, all 280 data points wold fall on the diagonal line. As sggested by the above likelihood-ratio tests, the BTL rle fits the data well. As a reslt, each face is represented by its two BTL scores as described below. Analysis of BTL Scores I compared the scores ofthe faces across orientations. Figre 2 provides bar charts of the faces' scores, separately for each orientation. Becase the BTL rle fitted the data well, an effect oforientation on the BTL scores that was qalitatively similar to the one fond for the preference probabilities wold be expected. Indeed, the scores of the faces show a considerable stability across orientations. The order ofthe faces' scores is hardly affected by inversion, and frthermore the score ratios between faces do not change in a major way across orien-

5 UPRIGHT VERSUS UPSIDE-DOWN FACES 167 TWO INDIVIDUAL BTL-RULES P- pred ONE COMMON BTL-RULE P- pred BTL-RULE + AFFINE RULE 0.75 p- pred ,...---,---,---, P-meas r , P-meas r------,------,-----, P-meas Figre 1.Scatterplotof measredpreferenceprobabilities (p-meas) verss predicted preference probabilities (P-pred) when sing ditterently restricted rles to fit tbe data. The top panel shows tbe qality offitof a jointbtl rle where a face's two parametersarefree to vary across orientations (7+7 = 14 parameters).themiddlepanelshows tbe qality offitof a jointbtl rle where a face's two parameters are restrictedto not varywitb orientation(7 parameters).thebottom panel, f'mally,shows the qality of the fit of a jointbtl rle wherea face's two parameters are restricted to being atrmely related across orientations (7+1 = 8 parameters). The data are merged over tbe two orientations (pright orientation., inverted orientation 0).Iftbe rles accontedperfectlyfor tbe variationin tbe preferenceprobabilities,each panel's 280 datapoints wold fall on the diagonal line. tations. Despite this stability, there is some tendency for the score ratios between faces to be redced by inversion. I examined whether orientation had a reliable effect on the score ratios, that is, whether the assmption that vp(a)/vpcb) = Vinv(a)/VinvCb) holds for all pairs (a,b) had to be rejected statistically. Becase the two scales v p and Vinv are niqe only p to scalar transformations (see Method section), this qestion is eqivalent to testing whether the two scores ofa face can be assmed to be constant across orientations. I fitted a joint BTL rle to the preference matrices ofthe two orientations where the two scores of a face are restricted to not varying with orientation. The fit of this rle to data was compared with the fit ofa joint BTL rle where the two scores of a face were free to vary with orientation. The X 2(7) vale of (p <.001) rejects the hypothesis that score ratios are constant across orientations. Ths, the effect of orientation fond in the preference probabilities is also reflected in the BTL scores. The middle panel of Figre 1 depicts the comparison ofthe preference probabilities measred in the two orientations with the predictions ofthese probabilities when sing a joint BTL rle where the two scores ofa face do not vary with orientation ("one common BTL rle"). Again the data are merged over orientations. As sggested by the above likelihood-ratio test, the fit of this joint BTL rle is somewhat worse than the fit of a joint BTL rle where the two scores ofa face are free to vary with orientation ("two individal BTL rles," top panel). Althoghthe difference in fit is not large, the deterioration is significant. As a reslt, a face cannot be represented by the same BTL score for the two orientations. A rle is needed to describe how a face's two scores are related across orientations. Accordingly, I tested a simple rle to describe the effect of orientation on the faces' scores. This rle assmes that the two scores of a face are affinely related across orientations, that is, Vinv(a) = vp(a)+k for each face a, where k is a real-valed parameter that does not depend on the faces. Eqivalently, across orientations the score ratios between faces are related by the eqation vinvca)/vinv(b) = [vp(a)+k]/[vp(b)+k]. This rle has two interesting properties. First, the affine relationship incldes the assmption that inversion affects all faces eqally. Ths, no interaction is spposed to occr between faces and orientation. The whole effect of orientation is redced to parameterk. Second, the sign ofk determines how preferences are affected by inversion. If k is positive, preferences are less close to indifferent choices in the pright orientationthan they are in the inverted orientation, a reslt sggested above by visal analyses ofthe preference probabilities and the score ratios. On the other hand, if k is negative, preferences are more close to indifferent choices in the pright orientation than they are in the inverted orientation. k=o is eqivalent to an invariance of score ratios and preferences across orientations, a hypothesis already rejected above. To examine whether this rle describes the effect of inversion for the present data, I tested whether the pref-

6 168 BAML l- o V> (/I (/I c: > c l- e V> (/I (/I c: >.;:; I-... -c 3-r , N N Upright o,.. U Rl U. M. M. v. Upside-Down,, Figre 2. Bar charts showing the attractiveness scores of faces for the twoorientations.thescores were estimatedby fitting the BTLrle individaily to an orientation's preference matrices. For each orientation. the face scores were fixed by assigning Face 1 a scale vale of 1. erence matrices of the two orientations cold be fitted by a joint BTL rle where the two scores ofa face were restricted to being affinely related throgh parameter k. Similarly, I compared the fit of this joint BTL rle with the fit ofa joint BTL rle where the two scores ofa face were free to vary with orientation. The X 2 (6) vale of (p =.299) spports the hypothesis that a face's two scores are affinely related across orientations. Ths, the introdction ofthe rotation parameter kredces the chisqare vale by more than 100 points, demonstrating that k plays an essential role in fitting the data. This holds tre even thogh I fond the vale ofk to be fairly small (k=.135). The fact that k is positive indicates that preferences are indeed less close to indifferent choices L/') o. L/') U U. Rl U. U> U Rl U. 3, , o U>,.. co co. in the pright orientation than they are in the inverted orientation. The bottom panel offigre 1 depicts the comparison of the preference probabilities measred in the two orientations with the predictions of these probabilities when sing a joint BTL rle where the two scores of a face are affinely related across orientations ("BTL rle + affine rle"). Again, the data are merged over orientations. As sggested by the above likelihood-ratio test, the data are well fitted. In fact, a visal comparisonwith the fit ofa joint BTL rle where the two scores ofa face are free to vary with orientation ("two individal BTL rles," top panel) demonstrates that the more restrictive, 8-parameter model provides a fit to the data that is eqal to that ofthe less restrictive, 14-parameter model. In this sense, the data are consistent with the hypothesis that the faces' two parameters are affinely related across orientations. In addition, the figre visalizes that the 8 parameter model provides a somewhat better fit to the data than does the 7-parameter model where the two scores of a face are restricted to not vary with orientation ("one common BTL rle," middle panel). These analyses show that the reliable effects of faces and inversion on the preference probabilities can be described in a simple way by assming, first, that each orientation's preference probabilities follow a BTL rle, and second, on the basis of this idea, that a face's two BTL scores are affinely related across orientations. By sing these two rles, the 280 measredpreferenceprobabilities can be described by only eight parameters. Seven ofthese eight parameters reflect the effect offaces; the eighth parameter reflects the effect oforientation in the present data sets. DISCUSSION Preferences between faces were measred with regard to the perceived attractiveness of the faces. This was done with the faces presented pright and with the same faces presented pside down. Inversion affected the preferences. Qalitatively, the preferences between faces became more close to indifferent choices when the faces were presented pside down than when they were presented in the pright orientation. This indicates that the faces became less discriminable when inverted, sggesting an impairment in the processing offacial properties for this orientation. Qantitatively, the effectofinversion cold be described by a simple rle that assmes the two attractiveness scores ofa face to be affinely relatedacross orientations. At the core of this rle is the assmption that inversion affects all faces eqally with no sbstantial interaction between the faces' perceived attractiveness and their orientation. It was arged above that the degree of a face-orientation interaction reflects the extent to which information processed from pright and psidedown faces differs. The reslt that the effectofinversion in this stdy can be describedeqally for the single faces demonstrates a negligible amont offace-orientation in-

7 UPRIGHT VERSUS UPSIDE-DOWN FACES 169 teraction, sggesting that abot the same information is processed from pright and pside-down faces. It is a widely held view in the literatre that inversion is particlarly disrptive to processing faces' secondorder relational properties bt that it hardly affects the processing of faces' first-order relational properties (Diamond & Carey, 1986). Althogh the reslts from a nmber ofstdies were interpreted in favor ofthis view, none ofthose stdies provided neqivocal evidence for it (e.g., Rhodes et al., 1993; Valentine, 1988). Moreover, the data from two more recent stdies were interpreted as evidence for an eqal impairment in the processing of first- and second-order relational properties (Tanaka & Farah, 1991; Valentine & Brce, 1988). While the relationship in these two stdies between reslts and conclsions may still have beeri tentative in natre (see introdction), the reslts from the present stdy provide more direct evidence for this alternative view. Since the same information seems to have been processed from pright and pside-down faces, it is sggested that first- and second-order relational properties are affected by inversion to roghly the same degree. This finding establishes a major challenge to Diamond and Carey's (1986) proposition. It indicates that the distinction between faces' first- and second-order relational properties cannot explain why recognition of faces is disrpted by inversion to a far greater extent than is recognition ofother classes ofvisal objects. Most recently, Rhodes et al. (1993) reported an experiment in which they compared the effects of face inversion for detecting changes that span the continm from first- to second-order relational properties. Using this kind offace maniplation, they fond evidence that second-order relational properties are more sensitive to inversion than first-order relational properties. While their method has the desirable featre that it addresses the qestion of interest very directly, their method depends crcially on the assmption that faces' first- and second-order relational properties can be maniplated independently. As also otlined by Rhodes et al. (p. 50), some of their reslts sggest that this assmption does not hold in general: first- and second-order relational properties appear to be inherently confonded in faces. The degree to which their findings can challenge the view spportedby the presentstdy, that first- and secondorder relational properties are eqally affected by inversion, therefore remains nclear. Attractiveness and Fit ofbtl Rle An orientation's preference matrices cold be well described by the BTL rle, indcing a one-dimensional representation offaces with regard to their perceived attractiveness. At first, this finding might appear to conflict with some crrent theories of attractiveness which sggest several qite different factors as affecting a face's perceived attractiveness, inclding both first- and second-order relational properties (Cnningham, 1986; Cnningham et al., 1990; Meerdink et al., 1990). How- ever, the two findings do not conflict with each other. Instead, the BTL representation implies only that when two faces are compared with regard to their attractiveness, all factors affecting the attractiveness of a face combine into one global score that represents the face's overall attractiveness. Indeed, this score can be interpreted as the sm ofthe attractiveness vales ofthe single factors that affect the attractiveness of a face (Sppes et al., 1989; Tversky, 1972). The present reslts reveal a srprisingly good fit of the BTL rle to orientation data. There are not many data sets in the literatre where choice behavior cold be well described by the BTL rle (Lce, 1977). Face attractiveness as investigated in the present experiment seems to be one ofthose. In fact, I replicated the experiment for pright faces for two other sets of faces, 10 male and 10 female. Again, I sed perceived attractiveness as the facial attribte. For both the males and the females, the BTL rle led to fits to the data that compared well with those fond in the present stdy. Comparing Faces' Attractiveness Scores Across Orientations This stdy focses on interface relationships and the qestion ofhow attractiveness ratios between faces vary with orientation. On the basis ofpair and triple comparisons of faces of eqal orientation, this qestion cold be addressed in a well-fonded way. In order to inclde meaningfl comparisons ofthe two attractiveness scores of a single face across orientations, however, measrements beyond those reported in the present stdy wold be needed. Specifically, preferences between faces of different orientations wold have to be measred. On the "basis ofthese additional measrements, a new scale that simltaneosly qantified the perceived attractiveness of the faces in their pright and their inverted orientations wold have to be developed. Corresponding measrements were not condcted in the present stdy. More formally, the fact that the two scores estimated for a face in this stdy cannot be compared in a meaningfl way across orientations is a simple conseqence of the fact that the two scales for pright and inverted orientation (v p, Vinv) are niqe only p to scalar transformations (see Method section), and that the nits of the two scales can be fixed independently from each other. Ths, whenever the score ofa face is higher in the pright orientation than in the inverted orientationafter two nits for the two scales have been chosen-an appropriate change in the nit of one of the two scales can reverse the order. The development ofone common scale for both orientations, if sccessfl, wold eliminate this freedom. From Inversion to a General Rotation inthe Vertical Inversion ofa face provides only a special case (180 ) of a more general rotation of a face in the vertical. Indeed, rotation angles other than 180 are also known to affect face recognition (Rock, 1973), response times

8 170 BAML (Valentine & Brce, 1988), and/or encoding (Baml, 1992). As a reslt, the attractiveness relationships between faces can be expected to be affected not only by inversion bt also by other rotation angles. The choice paradigm employed in the present stdy can be sed in a straightforward way to stdy how other rotation angles affect the relationships between faces. Specifically, if the affine rle fond in the present stdy to describe the effect of inversion held for any rotation angle, the effect of each rotation angle on interface relationships cold be represented by jst one parameter, k. In this case, the way that parameter k depended on rotation angle wold reveal sefl information on how the rotation angle affects the relationships between faces. On the basis ofvalentine and Brce's (1988), or Rock's (1973), or Baml's (1992) reslts, a monotonic relationship between rotation angle and parameter k may be expected. More specific expectations abot the fnctional form of the relationship between rotation angle and parameter k, however, are hard to derive from previos stdies. In this sense, the qestion ofhow rotation angles other than inversion affect the attractiveness relationships between faces is open to ftre stdies. ThreeFinal Remarks First: By sing attractiveness as a facial attribte, the present stdy provides evidence for the feasibility of sing a simple rle to describe the effect ofinversion on the relationships between faces. To the extent that this rle captres the whole effect of face inversion, its adeqacy shold not depend on the facial attribte employed. Indeed, when sing other facial attribtes, sch as, for instance, distinctiveness or age, the pattern ofreslts shold be similar to that described above with regard to attractiveness. This expectation constittes a strong prediction to be tested in frther experiments. Second: Recall that this stdy sed photographs of eight moderately attractive males of abot eqal age. Althogh these experimental conditions might have favored the reslts obtained in this stdy, the faces employed provided a reasonable test ofthe hypothesis that the same information is processed from pright and inverted faces. Indeed, a considerable range ofperceived attractiveness was spanned by the single faces. The attractiveness ratios for pairs of faces varied from less than 1.1:1 p to more than 40: 1 (cf. Figre 2). Frther experiments mst show whether the rles fond in the present stdy will also apply to qite different sets offacial stimli, inclding those of males of very low or very high perceived attractiveness, older males, and females. Third: The simple rle proposed in this stdy to accont for the inversion effect led to a reasonable description of the data sets. De to its strong restrictions, however, this rle is a simplification. Presmably, some faces do "lose" some of their properties when they are inverted, a phenomenon inconsistent with the idea that inversion affects all faces in the same way. Ftre stdies mst show whether freqency and size of these "losses" are low enogh to accept the view sggestedin this stdy, at least as a first-order model. REFERENCES ALLEY, T. R., & CUNNINGHAM, M. R (1991). Averaged faces are attractive, bt very attractive faces are not average. Psychological Science, 2, BAUML, K.-H. (1992). Diskriminationslernen bei rotierten Gesichtern: Eine Markov-Analyse von Kodierngs- nd Assoziationsprozessen. Zeitschriftfiir Experimentelle nd Angewandte Psychologie, 39, BRADLEY, R A., & TERRY, M. E. (1952). Rank analysis of incomplete block designs. 1. The method of pair comparisons. Biometrika, 39, CUNNINGHAM, M. R (1986). Measring the physical in physical attractiveness: Qasi-experiments on the socio-biology offemale facial beaty. Jornal ofpersonality & Social Psychology, 50, CUNNINGHAM, M. R, BARBEE, A. P., & PiKE, C. L. (1990). What do women want? Facialmetric assessment of mltiple motives in the perception of male facial physical attractiveness. Jornal ofpersonality & Social Psychology, 59, DIAMOND, R., & CAREY, S. (1986). Why faces are and are not special: An effect of expertise. Jornal ofexperimental Psychology: General, 115, GEGENFURTNER, K. (1992). PRAXIS: Brent's algorithm for fnction minimization. Behavior Research Methods, Instrments. & Compters, 24, GOLDSTEIN, A. G., & CHANCE, J. E. (1980). Memory for faces and schema theory. Jornal ofpsychology, 105, LANGLOIS, J. H., & ROGGMAN, L. A. (1990). Attractive faces are only average. Psychological Science, 1, LINDGREN, B. W. (1976). Statistical theory (3rd ed.). New York: Macmillan. LUCE, R D. (1959). Individal choice behavior: A theoretical analysis. New York: Wiley. LUCE, R D. (1977). The choice axiom after twenty years. Jornal of Mathematical Psychology, 15, MARUYAMA, K., & ENDO, M. (1984). Illsory face dislocation effect and configral integration in the inverted face. Tohok Psychologica Folia, 43, MEERDINK, J. E., GARBIN, C. P., & LEGER, D. W. (1990). Cross-gender perceptions of facial attribtes and their relation to attractiveness: Do we see them differently than they see s? Perception & Psychophysics, 48, PERRETT, D. 1.,MAY, K. A., & YOSHIKAWA, S. (1994). Facial shape and jdgements offemale attractiveness. Natre, 368, RHODES, G., BRAKE, S., & ATKINSON, A. P. (1993). What's lost in inverted faces? Cognition, 47, ROCK, 1. (1973). Orientation and form. New York: Academic Press. SERGENT, J. (1984). An investigation into component and configral processes nderlying face perception. British Jornal ofpsychology, 75, SHEPARD, R N., & METZLER, J. (1971). Mental rotation of threedimensional objects. Science, 171, SUPPES, P., KRANTZ, D. H., LUCE, RD., & TVERSKY, A. (1989). Fondations ofmeasrement (Vol. 2, pp ). San Diego: Academic Press. TANAKA, J. w., & FARAH, M. J. (1991). Second-order relational properties and the inversion effect: Testing a theory of face perception. Perception & Psychophysics, 50, THOMPSON, P.(1980). Margaret Thatcher: A new illsion. Perception, 9, TVERSKY, A. (1972). Elimination by aspects: A theory of choice. Psychological Review, 79, VALENTINE, T.(1988). Upside-down faces: A review ofthe effect ofinversion pon face recognition. British Jornal ofpsychology, 79, VALENTINE, T., & BRUCE, V. (1988). Mental rotation of faces. Memory & Cognition, 16, YIN, R K. (1969). Looking at pside-down faces. Jornal ofexperimental Psychology, 81,

9 UPRIGHT VERSUS UPSIDE-DOWN FACES 171 erties, bt made an additional distinction between these two kinds of relational properties on the one hand and more isolated featres on the other. The crcial point in Diamond and Carey's hypothesis, how- ever, dealt with the special role of the second-order relational properties in object inversion; their hypothesis did not incorporate any distinction between first-order relational properties and the more isolated featres. Ths, for sake of brevity, I sefirst-order relational YIN, R. K. (1970). Face recognition by brain-injred patients: A dissociable ability? Neropsychologia, 8, YOUNG, A. w., HELLAWELL, D., & HAY, D. C. (1987). Configrational information in face perception. Perception, 16, NOTE I. Actally, Diamond and Carey (1986) not only distingished be- properties in this paper as a smmary term to refer both to first-order tween objects' first-order relational and second-order relational prop- relational properties and to more isolated featres. APPENDIX The triple comparison matrices for the two orientations, with each cell representing the estimated probability that one of a triple of faces is preferred over the other two. Face Triple Upright Inverted (a,b,c) a;bct b;ac c;ab a;bc b;ac c;ab (FbFz,FJ)* (Fb F2,F4) (Fl,F2,Fs) (FbF2,F6) (FbF2,F7) (Fb F2,Fg) (FbFJ,F4) (F],FJ,Fs) (Fl,FJ,F6) (FbFJ,F7) (FbFJ,Fg) (Fb F4,Fs) (FbF4,F6) (FbF4,F7) (F],F4,Fg) (F1,Fs,F6) (Fb FS,F7) (FbFS,Fg) Q78 (F],F6,F7) (Fb F6,Fg) (FbF7,Fg) (F2,FJ,F4) (F2,FJ,Fs) (F2,F3,F6) : (F2,F3,F7) (F2,F3,Fg) (F2,F4,Fs) (F2,F4,F6) (F2,F4,F7) (F2,F4,Fg) Q78 (F2,Fs,F6) (F2,Fs,F7) (F2,Fs,Fg) (F2,F6,F7) (F2,F6,Fg) (F2,F7,Fg) (F3,F4,Fs) (F3,F4,F6) (F3,F4,F7) (F3,F4,Fg) (F3,Fs,F6) (F3,Fs,F7) (F3,FS,Fg) (F3,F6,F7) (F3,F6,Fg) (F3,F7,Fg) (F4,Fs,F6) (F4,Fs,F7) (F4,Fs,Fg) (F4,F6,F7) (F4,F6,Fg)

10 172 BAML Appendix (Contined) Face Triple Upright Inverted (a,b,c) a;bct b;ac c;ab a;bc b;ac c;ab (F 4,F7,Fs) (Fs,F 6,F7 ) (Fs,F6,Fs) (F s,f7,fs) (F 6,F7,Fs) *F. = face nmber n. ta;bc = face a is preferred over faces band c, and so forth. (Manscript received Jly 1, 1993; revision accepted for pblication Janary 18, 1994.)