“Just another pretty face”: A multidimensional scaling approach to face attractiveness and variability

TIMOTHY POTTER OLIVIER CORNEILLE KIRSTEN I. RUYS Universit Catholique de Louvain Louvain-la-Neuve Belgium Findings on both attractiveness and memory for faces suggest that people should perceive more similarity among attractive than among unattractive faces. A multidimensional scaling approach was used to test this hypothesis in two studies. In Study 1, we derived a psychological face space from similarity ratings of attractive and unattractive Caucasian female faces. In Study 2, we derived a face space for attractive and unattractive male faces of Caucasians and non-Caucasians. Both studies confirm that attractive faces are indeed more tightly clustered than unattractive faces in people's psychological face spaces. These studies provide direct and original support for theoretical assumptions previously made in the face space and face memory literatures. - A widely discussed hypothesis in the face literature suggests that attractiveness relates positively to averageness (for recent reviews, see Rhodes, 2006; Rhodes & Zebrowitz, 2002). A recent meta-analytic review has also shown that averageness and high levels of sexual dimorphism are attractive in both male and female faces (Rhodes, 2006; see also Perrett, May, & Yoshikawa, 1994). Both findings suggest that attractive faces should be perceived to be more alike than unattractive faces, because attractive faces must conform either to the population average or to an optimal direction of deviation from it. Unattractive faces, in contrast, may deviate from the population average in a number of ways. The hypothesis that people perceive attractive faces to be more alike than unattractive faces may have important consequences for face memory, and it seems consistent with effects reported in the face memory literature. Specifically, attractive faces elicit more false memory than do unattractive faces (Corneille, Monin, & Pleyers, 2005; Monin, 2003). If attractive faces are more alike, they should be more densely clustered in face space, and so should have a higher probability of being mistaken for a previously seen face already encoded in memory (see also Lewis, 2004; Light, Kayra-Stuart, & Hollander, 1979; Mueller, Heesacker, & Ross, 1984; Vokey & Read, 1992). Not only are false alarm rates higher for attractive than for unattractive faces, but in addition hit rates are higher for unattractive faces. The latter finding would be explained if unattractive faces are less alike, and therefore can be encoded in a more precise and discriminating fashion (see also Light, Hollander, & Kayra-Stuart, 1981; Vokey & Read, 1992). Surprisingly enough, however, no study to date has directly examined whether attractive faces are indeed more clustered in face space. We set out to do so here, using multidimensional scaling (MDS) to derive psychological face spaces for sets that contain attractive and unattractive female Caucasian faces (Study 1) and attractive and unattractive male faces of various ethnicities (Study 2). We tested whether attractive faces are perceived to be more alike than unattractive faces by comparing the variability of attractive and unattractive faces in these derived face spaces. Method Participants. Fifty-nine psychology undergraduates from the Universit Catholique de Louvain (54 females, 5 males) participated in exchange for partial course credit. Materials. We downloaded portraits of 80 young Caucasian females from a casting database (www.interfaces.nl). The pictures were all in black and white, and the faces had similar facial expressions. Some pictures had different poses, and we mirror-reversed some of these to ensure that each pose had a similar orientation and that the right shoulder was in the foreground. Twenty voluntary participants contacted on campus (half males, half females) rated the faces for attractiveness on a scale from 1 (very unattractive) to 9 (very attractive). On the basis of these ratings (Cronbachs .96), we selected two sets of faces: 15 attractive female faces (M 6.9, SD 0.4, range 6.27.6) and 15 unattractive female faces ( M unattractive faces, p .402), and a Levenes test, which showed equality of variances between these distributions ( p .987). Thus, data points were distributed in a balanced way around the mean. These analyses suggest that the larger mean distances of the unattractive faces are not due to a few highly atypical outliers. Finally, we examined the relationship between the attractiveness ratings (bimodally distributed) and the (nonnormally distributed) distance from each face to the cen2.6, SD 0.4, range 1.93.3). Attractive and unattractive faces did not differ significantly in age (respectively, M 22.3 years, SD 4.0, and M 25.2 years, SD 6.8) [ t(28) 1.4, p .174], and Levenes test also showed equality of variances ( p .34). There was minimal pose variation between the pictures selected. If anything, poses varied slightly more for the attractive faces (because of two pictures that were framed relatively more closely than the other faces, and one face that was not as well framed as all the others), which ran counter to our hypothesis. Procedure. The task was presented on a computer (PC compatible). All instructions were presented onscreen. Participants were asked to make similarity judgments between pairs of faces on a Likert scale from 1 (extremely dissimilar) to 7 (extremely similar). We specifically instructed participants to concentrate on the faces. We told them, You are going to see a pair of faces in each new screen. You will have to, for each pair of faces, give a global and spontaneous judgment. There is no wrong or right answer. Simply look at the traits of these faces, and spontaneously judge their global degree of similarity. Before proceeding to the similarity judgment task, participants were presented with the 30 faces, displayed successively in a random order at the center of the screen for 1,000 msec apiece, so that they could see the range of variation in the set. The participants then rated the similarity of 435 pairs of faces (each of the 30 faces presented once with each of the other 29 faces). One of the faces was positioned on the upper left side of the screen and the other on the upper right side. Text reminding the participants of the scale was positioned in the bottom center of the screen. The pairs of faces remained visible until participants responded by pressing a key on the keyboard (from 1 to 7). A rest screen appeared for 30 sec every 100 presentations. Results We first converted the ratings of similarity into dissimilarity rating matrixes for all participants. We used the INDSCAL technique (a form of weighted MDS) to process these data, which enabled us to compute the model best representing the one used by the group of participants as a whole. This technique is to be favored (Martens & Zacharov, 2000) because it accounts for differences in the importance assigned by each participant to each dimension (the weight). The ALSCAL procedu (...truncated)