Limitations of traditional morphometrics in research on the attractiveness of faces

ERIK HOLLAND www.femininebeauty.info The traditional morphometrics approach to shape comparisons involves computing multiple interlandmark distances without taking into account the geometric configuration of the landmarks. A recent example of this approach is a study by Potter and Corneille (2008). They had participants rate the attractiveness of computer-generated European, African, and Asian male faces, and they computed the Euclidean distances between each face and the group prototypes. They found that faces are rated more attractive when they are closer to their group prototype. This letter addresses differing conclusions in the literature, the methodological shortcomings of Potter and Corneille, and another study that explored a similar topic, with a special focus on guiding future researchers around the pitfalls of traditional morphometrics. - Conclusions Different From Potter and Corneilles (2008) In many non-European populations, the attractive face is less ethnic-looking and closer to European norms than the average. This has been documented for KoreanAmerican women evaluated by their co-ethnics (Choe, Sclafani, Litner, Yu, & Romo, 2004) and also for the profile of African-Americans (Farrow, Zarrinnia, & Azizi, 1993; Martin, 1964; Polk et al., 1995; Sushner, 1977; Thomas, 1979; but see Sutter & Turley, 1998, for a null find). Aesthetic facial cosmetic surgeries in East Asians (Ahn, 2006; Dobke, Chung, & Takabe, 2006; Lam, 2005) and African-Americans (Rohrich & Muzaffar, 2003) also tend to cluster in the direction of European norms. Rhodes et al. (2005) found that Eurasian faces obtained by morphing European and East Asian faces were rated more attractive than European or East Asian faces. To my knowledge, this is the only study that has documented a shift toward East Asian norms increasing the perceived attractiveness of European faces, but this study had numerous shortcomings. These authors had the participants rate composite face morphs, rather than individual faces, for attractiveness. Some adjustments for differences in face size need to be performed when morphing faces together: A common practice, also employed by Rhodes et al., is to equalize interpupillary distance. However, a single interlandmark distance is a poor approximation of face size. In one standard implementation for controlling for size, one computes the center of mass of a form with unit mass at each of its landmarks. This is known as the centroid. One obtains the centroid size by summing the squared distances of a forms landmarks from its centroid. Then, scaling all forms to the same centroid size adjusts for size. Another problem with Rhodes et al. (2005) is that all groups of the faces used (European, Asian) should have had similar distributions of attractiveness and femininity with respect to the norms in the respective ethnic groups. This is because the average of attractive faces is rated more attractive than the average of nonattractive faces (Johnston & Oliver-Rodriguez, 1997; Perrett, May, & Yoshikawa, 1994), and the femininity of a womans face is a much more powerful correlate of beauty than its prototypicality (Rhodes, 2006); the prototypical female face is at the 50th percentile of femininity among women. But we have no indications that these requirements are met in Rhodes et al., and they would be difficult to fulfill. Yet another problem with Rhodes et al. (2005) is that when one uses face composites, one cannot readily assess the effect on attractiveness when faces across a range of attractiveness are transformed along ethnic lines. Furthermore, Rhodes et al. assumed face shapes of ethnically mixed offspring to be an average of the parental face shapes, but this is not true for the majority of face-shape variables (Martnez-Abadas et al., 2006). Methodological Issues Faces generated by FaceGen Modeller. Potter and Corneille (2008) generated faces using FaceGen Modeller (www.facegen.com). FaceGen is mainly used by game developers. It is also used by police to generate 3-D sketches of suspects. However, there are concerns about how well FaceGen parallels reality. In comparison with European faces, the nasion is displaced inferiorly in subSaharan Africans (Africans) and East Asians (Hennessy & Stringer, 2002), but FaceGen achieves this effect primarily by raising the eyebrows in Africans and Asians, not by lowering the nasion. The flattest nasal bones are found in sub-Saharan African populations (Hanihara, 2000), but FaceGen makes East Asian nasals flatter than African nasals. Europeans tend to have shorter chins than Africans and East Asians (Bastir, Rosas, & Kuroe, 2004), but not so in FaceGen. These limitations are of little relevance to game developers or police because these groups need only to generate faces that approximate target faces. Also, FaceGen can be used as a very basic educational tool, but psychological research is another matter. If the research addresses basic perceptions of ethnicity, sex differences, or attractiveness and involves no facial measurements, then FaceGen can be used for convenience; but when minutiae of shape variables and measurements are involved, it is best to use actual faces. Variation between and within populations. Most skull shape variation in humans lies within populations (Roseman & Weaver, 2004). Because there is a correlation structure underlying differences between populations, geographical clusters appear with the assessment of multiple interlandmark craniofacial distances (Brace & Hunt, 1990) or the geometric configuration of the landmarks (Hennessy & Stringer, 2002). Nevertheless, there also is clinal variation (Hanihara, 1996, 2000). Hence, with representative sampling of a population, it should be possible to obtain faces that represent variation within this population, as well as faces somewhat shifted toward the norms of other populations on multiple counts. To investigate whether facial attractiveness varies along the discriminant distinguishing ethnic groups, this component must be isolated from total face shape variation. However, Rhodes et al. (2005) could not do this and, hence, did not address the extent to which the higher attractiveness of Eurasian faces resulted from variation along the discriminant distinguishing European from Asian faces, rather than from other shape components. Potter and Corneille (2008) used the rand lock feature in FaceGen, which keeps dimensions that are key to a group constant, to generate random faces for each population. It is questionable how well FaceGen achieves this. If it does so well, the authors did not have faces within a group that varied among themselves along the discriminant distinguishing this group from other groups and, hence, could not answer how attractiveness varies as a function of distance from other-group prototypes. Even if FaceGen Modeller had produced random faces within a group, so that a few faces were, overall, somewhat shifted toward other-group norms (...truncated)