Article Commentary: On the Difficulty of Averaging Faces: Comments on Langlois and Roggman

In this journal, Langlois and Roggman (1990) argued that faces with morphologically features will be seen as attractive by conspecifics. In humans, this process takes the following form: The perceiver develops mental prototypes of faces by averaging the features of the faces she sees and then uses those prototypes as baselines for judging the attractiveness of faces seen subsequently. Langlois and Roggman also described a novel technique for averaging facial photographs. Simulated photographs produced by this technique were used to test their hypothesis that average, i.e., normal, faces are especially attractive. Although I support a different account of facial attractiveness (Carello, Grosofsky, Shaw, Pittenger, & Mark, 1989), I do not dispute Langlois and Roggman's conclusion about the attractiveness of faces. Indeed, their study is an elegant test of their hypothesis. However, three characteristics of their technique for producing a two-dimensional representation of an average compromise the technique's ability to recover either morphologically faces or mental prototypes of faces. The severity of the resulting problems is contingent partly on how we choose to define normal face and partly on details of the processes by which people might construct mental prototypes of faces. Briefly, their technique involves reducing black-andwhite facial photographs to a 512 x 512 matrix of graylevel values, averaging the values at each point in the matrix over individual faces, and printing the resulting matrix as a simulated photograph. The first problem lies in the fact that averages are computed over gray levels located in a fixed matrix rather than over the spatial locations of anatomically defined points. These two processes produce different faces. For example, suppose that upper eyelids appear as dark lines in photographs and that eyelid sizes are bimodally distributed (i.e., most eyes are either rather large or rather small, few are mid-sized). Averaging of gray levels at fixed points in a matrix would generate two eyelid lines in the composite face, one near each of the modal sizes, rather than a single lid at the mean anatomical position. This type of problem appears to have occurred in Langlois and Roggman's study: On pages 116 and 117, they report the need to apply a smoothing operation to the averaged faces to remove double edges. Although the concept of morphologically is not yet precisely defined, it seems plausible to suppose that it should be based on the shapes and locations of anatomical features. Although mental prototypes conceivably could be based on gray-level representations of faces, it is more plausible to suppose they too would be based on anatomical features.