Robert Brooks, Evolution & Ecology Research Centre and School of Biological, Earth and Environmental Sciences, The University of New South Wales, Sydney, NSW 2052, Australia. Tel.: +61 2 9385 2587; fax: +61 2 9385 1558; e-mail: firstname.lastname@example.org
Studies of the attractiveness of female bodies have focussed strongly on the waist, hips and bust, but sexual selection operates on whole phenotypes rather than the relative proportions of just two or three body parts. Here, we use body scanners to extract computer-generated images of 96 Chinese women’s bodies with all traits unrelated to body shape removed. We first show that Chinese and Australian men and women rate the attractiveness of these bodies the same. We then statistically explore the roles of age, body weight and a range of length and girth measures on ratings of attractiveness. Last, we use nonlinear selection analysis, a statistical approach developed by evolutionary biologists to explore the interacting effects of suites of traits on fitness, to study how body traits interact to determine attractiveness. Established proxies of adiposity and reproductive value, including age, body mass index and waist-to-hip ratio, were all correlated with attractiveness. Nonlinear response surface methods using the original traits consistently outperform all of these indices and ratios, suggesting that indices of youth and abdominal adiposity tell only part of the story of body attractiveness. In particular, our findings draw attention to the importance of integration between abdominal measures, including the bust, and the length and girth of limbs. Our results provide the most comprehensive analysis to date of the effect of body shape and fat deposition on female attractiveness.
Physical attractiveness is an important determinant of evolutionary, social and economic (Hamermesh & Biddle, 1994; Mobius & Rosenblat, 2006; Lynn, 2009) success. The dimensions of an individual’s body can provide observers with information about that individual’s suitability as a potential mate (i.e. their biological sex and self-identified gender, Johnson & Tassinary, 2007), value as a mate or long-term partner (e.g. by signalling age, fertility, resource-holding potential or developmental history, Gangestad & Scheyd, 2005; Rilling et al., 2009; Singh, 1993b) and the threat they present as a sexual competitor (Buunk & Dijkstra, 2005). A rapidly growing body of work has, in the last two decades, placed the study of attractiveness solidly within the paradigm of Darwinian selection. Our mating, pairing and social preferences have been fashioned by the benefits of choosing the best possible mates, partners and group mates, wherever circumstances allow. These preferences find their expression not only in the mating and social decisions we make today, but also in novel modern contexts such as hiring and remuneration (Hamermesh & Biddle, 1994; Mobius & Rosenblat, 2006; Swami et al., 2008), purchasing and marketing (Saad, 2004; Miller, 2009) and the value of payments such as tips to staff (Lynn, 2009).
WHR increases with abdominal adiposity, and it has been suggested that correlates of WHR such as body mass (Tassinary & Hansen, 1998; Tovee et al., 1998), body volume (Fan et al., 2004), abdominal depth (Rilling et al., 2009) and cues of femininity may be more relevant cues of attractiveness. Studies have shown that body mass index (BMI = weight/height2), its two-dimensional area proxy, perimeter–area ratio (PAR) (Toveéet al., 1999) and another related measure volume–height index (VHI, volume/chin height2) (Fan et al., 2004) are associated with female attractiveness. Overall, studies of WHR and BMI in industrialized countries show that WHR in the lower typical range of approximately 0.7 and perceived BMI in the lower normal range of 18.5–20 are both correlated with high attractiveness (Weeden & Sabini, 2005; but see Swami et al., 2010; Dixson et al., 2010; Furnham et al., 2005). A recent study modelling patterns of fat deposition indicates that the relationship between attractiveness judgements and WHR is because of the way in which fat is deposited during weight gain rather than particular hormonal effects on relative fat deposition on the waist and hips (Cornelissen et al., 2009b). WHR may, therefore, only be a convenient proxy for attractiveness rather than conveying any information beyond that conveyed by BMI, PAR or VHI.
Bodies are complex phenotypes that vary in a large number of properties, many of which are correlated with one another. It remains unclear whether there is more to attractiveness than the few indices of adiposity, size and torso shape that have been the focus of the overwhelming share of body studies to date. One fruitful approach to dealing with the highly correlated body measures is to estimate a more manageable number of parameters that together describe shape, and this has been applied to the study of human body attractiveness (Toveéet al., 2002; Fan et al., 2007; Smith et al., 2007). A complementary approach is based on tools that revolutionized the study of natural selection in wild animals: multiple regression-based methods for estimating and visualizing the strength of linear and nonlinear selection on suites of correlated traits (Lande & Arnold, 1983; Phillips & Arnold, 1989). With these methods, it is possible to estimate not only the effects of each trait on a fitness measure such as survival or mating success, but also the complex ways in which traits interact. Rather than the experimenter testing whether a given index or suite of indices predicts attractiveness, it is possible to build and assess the performance of a statistical model that identifies the most important predictors of attractiveness.
Selection analyses work best when the experimenter can be sure that they have considered all of the variables that could be relevant. It is then more likely that the traits that directly affect fitness will be included in the model, rather than spurious correlates of those traits. To do this for human bodies, we need a large number of measures of a suite of bodies and a means to test the attractiveness of those bodies. Body scanning technology has been used by the clothing and textile industry to provide full scans of human bodies that can then be exported to three-dimensional modelling software. The scanning software can also provide millimetre-accurate estimates of a large number (hundreds) of measures (Fan et al., 2004; 2005; 2007). This has also allowed for detailed anthropometric studies of external body dimensions and how they change with weight, age, nutrition and disease risk factors (Wells et al., 2007; 2008a,b). Body scanning technology also allows experimenters interested in attractiveness to show raters a set of still or rotating computer-generated images and to measure their ratings of attractiveness. Body scanning has one further advantage for cross-culturally relevant studies. Differences in stature, build and adiposity between racial, cultural and socioeconomic groups are often correlated with differences between these groups in other traits such as facial features, hair colour and texture, and clothing. With body scans, the experimenter can present images that are stripped off these traits or, potentially, in which these traits are manipulated. The extent to which the features that make bodies attractive are shared across cultures remains an important and largely unresolved question in the study of human mate choice.
Here, we measure the attractiveness of scans of 96 bodies of Chinese women to a sample of Australian adults, most of European descent. We first compare the attractiveness ratings with those provided by a group of Chinese raters in Hong Kong to establish some measure of cross-cultural robustness to these ratings. We then statistically explore the roles of age, body weight and a range of length and girth measures on ratings of attractiveness. We start with simple indices like BMI and WHR and build towards an eight-trait nonlinear response surface analysis that captures a very high proportion of the variation in average attractiveness ratings. By combining one of the largest samples of real women’s bodies with a large number of ratings and a comprehensive and holistic approach to model fitting, we provide some of the most complete evidence yet on how traits interact to determine the attractiveness of bodies.
Ninety-six adult women were scanned, 51 of whom were scanned using a [TC]2 body scanner at Shanghai and 45 of whom were scanned using a Lectra body scanner at Tianjin, resulting in a suite of body measurements and a 3D computer-generated image of each body (see Fan et al., 2004 for more details about scanning, see Table S1 for descriptive statistics). Models were a mix of university students and volunteers from the local communities, ranging in age from 20 to 49 years. Models wore tight-fitting body underwear that did not alter the shape of the body. For four of the scans, one or more measures were missing, resulting in sample sizes of between 92 and 96 bodies, depending on the traits used in each analysis. Waist girth was measured as the minimum circumference around the waist area and hip girth was measured as the circumference at the greater trochanter.
The 3D object models were each imported into Autodesk® Maya® (Autodesk Inc., San Francisco, CA, USA) 3D modelling software and standardized as blue wireframe models (against a grey background) to remove skin tone and texture stimuli and obscure facial features. Standardized 1280 × 720 resolution video files were created for each model. Each video showed the model, positioned centrally and scaled proportionally to actual size, rotating 360 degrees clockwise (starting and ending front onto the viewer) over 12 s.
We recruited 92 participants from UNSW campus (66) and Manly hospital (26) in Sydney, Australia, to rate the models. Raters comprised 40 men and 52 women aged 18–58 (mean age 28.3 and 30.3 years for men and women, respectively). Each participant was given a brief background on the study and instructions on how to complete the study. We presented stimuli and recorded responses using MediaLab (v2008; Empirisoft Corp., New York, NY, USA) psychology research software running on a personal computer with a 1280 × 1024 resolution screen. We recorded each rater’s age, gender (male or female) and sexual orientation (heterosexual, homosexual, bisexual, none/other) before presenting any stimuli. We then showed each rater 5-s clips of 12 videos (chosen to represent a range of body types to acclimatize them to the range of models. Raters were then shown 117 model videos (the 96 original videos plus 20 in which the size of the model had been reduced by 20%, and which were not included in the analyses presented here) in random order and asked to rate each in turn. Raters were asked to rate each model as quickly as possible on a seven-point Likert scale that asks raters to specify the extent to which they find the model attractive. Raters took 5.35 s, on average, to rate each model. Once a rater entered a rating, the video stopped and the next model was presented.
We first analysed data from male and female rates separately. We calculated the average Likert scores for male and female raters for each model. We then divided each average score by the average for all raters of a given sex to give relative fitness as required for selection analysis (Lande & Arnold, 1983). Models with relative attractiveness scores below 1 are less attractive and those with scores above 1 are more attractive than average. To compare our respondent’s ratings with the ratings that a group of Hong Kong Chinese raters had given a subset of 51 of the models in previous studies (Fan et al., 2007), we converted the nine-point Likert scale used in those studies to a seven-point range and then standardized average scores in the same way. Because the attractiveness ratings of Australian men and women were so tightly correlated (see Results), all further analyses were carried out on a mean of male and female raters’ scores.
We explored the determinants of attractiveness using linear and nonlinear regression. For each model, we knew the scanned woman’s age and weight. After scanning, body measurements were obtained using the standard scanner software. We start by analysing the effects of age, height, weight and waist and hip-related dimensions on attractiveness using multiple regressions. Then, to gain a more comprehensive understanding of the basis of body attractiveness, we employ multivariate nonlinear selection analysis (Lande & Arnold, 1983; Phillips & Arnold, 1989) and visualization (Blows & Brooks, 2003).
A large number of the measures provided by this software measure only slightly different aspects of the same thing, leading to considerable multicollinearity if they are all included in the analyses. We therefore conducted our analyses in two stages. First, we explored a subset of 23 measurements that measured different aspects of the body first using univariate correlation and then by exploring forward and backward stepwise multiple regressions (including linear, quadratic and cross-product terms). From this process, we identified eight variables that were important predictors of attractiveness in at least one of the above analyses: height, the vertical distance between waist and hip (dvertW–H), arm length, upper arm girth, bust girth, waist girth, thigh girth and ankle girth. These eight variables, when included in a linear regression together show only mild autocorrelation [variance inflation factor (VIF) < 5], but the addition of any more variables leads to excessive autocorrelation (VIF > 10). We therefore eliminated the other 18 variables from our response surface analyses: vertical distance from bust to chin, vertical distance from waist to bust, waist height, hip height, vertical distance from the side of the neck to bust, neck column girth, back shoulder width, across back width, bust to bust, neck base girth, wrist girth, bust arc width, under bust girth, hip girth and knee girth.
We followed standard methods for linear and nonlinear selection analysis (Lande & Arnold, 1983; Phillips & Arnold, 1989). We standardized all variables to have a mean of 0 and unit standard deviation. This ensures that the estimated regression coefficients can be interpreted as standardized selection gradients: the predicted change from one generation of selection if heritability equals 1 (Lande & Arnold, 1983). We estimated linear (i.e. directional) selection gradients (βi) by fitting a multiple regression including an intercept and the linear terms for the eight traits we had chosen to include in the response surface. We then fitted a multiple regression that included the linear terms, their squares, and cross-products of each pair of linear terms. Quadratic selection gradients (γii) are equal to double the regression coefficients estimated for the squared terms from this regression (Stinchcombe et al., 2008), and correlation selection gradients (γij) are given by the coefficients for the cross-product terms.
The multivariate form of nonlinear selection may be misinterpreted by simple inspection of the nonlinear selection gradients. In particular, the strength and significance of nonlinear selection may be underestimated (Phillips & Arnold, 1989; Blows & Brooks, 2003). We therefore performed a canonical rotation of the gamma matrix to find the major axes of the nonlinear response surface (Phillips & Arnold, 1989; see method in app. 1 of Blows & Brooks, 2003). This canonical rotation resulted in a matrix, M, comprising eight eigenvectors, mi, each describing a major axis of the response surface. The strength of nonlinear selection along each eigenvector is given by its eigenvalue, λi. The strength of directional selection (θi) along each eigenvector and the significance of both directional and nonlinear selection along each eigenvector were obtained by including linear and quadratic forms of all eigenvectors in a new multiple regression model [i.e. the double linear regression (DLR) method] described by Bisgaard & Ankenman (1996) These results were visualized using thin-plate splines in the fields package in R.
There was strong agreement among the four sets of raters: Hong Kong men and women and Sydney men and women (pairwise correlation coefficients all r > 0.955, Hong Kong ratings for 51 models and Sydney ratings for 96 models exist). Correlation between Sydney men and women r = 0.988, P < 0.001.
Scans of younger women’s bodies were significantly more attractive (Fig. 1). The regression equation describing the linear and quadratic effects of age on attractiveness explained 46.4% of the variance in attractiveness ( = 0.464, F2,94 = 42.6, P < 0.0001; β= −0.043 ± 0.006 SE, t94 = −7.58, P < 0.0001; γ= 0.0017 ± 0.0005 SE, t94 = 2.82, P = 0.0001). The significant quadratic term is because of the greatest drop in attractiveness occurring between 20 and 30 years of age, and the drop slowing substantially thereafter.
Height, weight and BMI
We explored all possible univariate and multiple regression models involving height and weight. The best model, in terms of Mallow’s Cp (Draper & Smith, 1981), included linear terms for height and weight and did not include quadratic or cross-product terms (height β= 0.163 ± 0.029 SE, t94 = 5.46, P < 0.0001; weight β= −0.319 ± 0.029, t94 = −10.66, P < 0.0001). This model explained 55% of the variance in attractiveness of body scans (overall model = 0.548; Model F2,94 = 59.2, P < 0.0001). As illustrated in Fig. 2a, body scans of taller and lighter women were most attractive. Height and weight are the two constituents of body mass index (BMI = weight/height2), and this single composite measure performed no worse or better as a predictor of attractiveness than did the multiple regression that included height and weight ( = 0.551, BMI β= −0.102 ± 0.009 SE, t = −10.90, P < 0.0001; Fig. 2b).
Waist and hips
Of all the variables we measured, waist girth was the strongest correlate of relative attractiveness and always the first term added to any forward stepwise multiple regression. To test the usefulness of waist and hip dimensions, we explored multiple regressions involving height, waist girth, hip girth, the vertical distance between waist and hips and their quadratic and cross-product terms. The best model included linear terms for all four measures, plus waist2 and (vdist waist–hip)2 (Table 1). This model explained 92% of the variance in relative attractiveness.
Table 1. Best multiple regression model involving linear, quadratic and cross-product terms involving standardized height, waist girth, hip girth and vertical distance between waist and hips.
= 0.914, Mallows’ Cp = 7.26.
β are standardized selection gradients.
All excluded terms had stepwise P < 0.13.
By contrast, with this approach, waist–hip ratio alone explained only 70% of the variance in relative attractiveness (F1,95 = 223.3, P < 0.0001, b= −4.64 ± 0.039 SE), and this was not improved by fitting the quadratic form of WHR (a significant negative quadratic term would be indicative of an intermediate optimum WHR) or fitting height or waist–hip vertical distance in any combination. The relationship between waist, hip and attractiveness is shown in Fig. 3. It is important to note that a smaller waist is more attractive for any given hip girth.
Full response surface
Throughout the process of exploring both the effects of all linear and girth measures (i.e. all traits estimated from the body scan, no weight, volume or composite index measures), four traits had significant linear effects: height and arm length (positive effects) and waist girth and upper arm girth (negative effects). Using forward and backward stepwise model-fitting approaches, the best model of linear, quadratic and correlational effects also included contributions from ankle girth, thigh girth, bust girth and the vertical distance between waist and hips. We therefore fitted a full linear selection analysis using these eight variables and a full nonlinear response surface (Table 2).
Table 2. The vector of standardized directional selection gradients (β) and the matrix (γ) of standardized quadratic (on diagonal) and correlational (below diagonal) selection gradients on the suite of eight traits implicated in selection.
U. arm g.
Asterisks indicate gradients where P < 0.05.
Upper arm girth
Canonical rotation of the nonlinear response surface indicated the presence of four axes of positive quadratic and four axes of negative quadratic selection. Unusually, for a multivariate response surface, there was significant quadratic selection along every one of the eight eigenvectors of the response surface (Table 3). Interpreting every element of eight eigenvectors would be pointless, so we discuss only the major features. First, the presence of strong stabilizing selection on m7 and m8 suggests that certain combinations of intermediate trait values are more attractive than more extreme combinations. The large contributions made by upper arm, bust, thigh and ankle girth to m7 and by thigh, waist and bust girths to m8 are driven by the strong nonlinear selection on these traits (Table 2), particularly but not limited to thigh2, bust × upper arm girth and bust × ankle girth.
Table 3. The major dimensions [eigenvectors, (mi)] of nonlinear selection identified by canonical rotation of γ.
U. arm g.
Asterisks indicate gradients where P < 0.05.
The eigenvalues (λi) of each eigenvector are equivalent to the quadratic gradients of selection along each eigenvector.
The two axes of strongest concave selection, m1 and m2, were both also under significant directional selection, suggesting an area of low attractiveness at small values of m1 and large values of m2. These women had small busts, large waists, large ankles, short arms and long distances between waist and hips. The latter trait is associated with larger waists, as abdominal adiposity tends to cause the waist to appear higher up than it would otherwise do.
We found that within the natural distribution of women’s body shapes, established proxies of adiposity and reproductive value such as BMI and WHR were all correlated with attractiveness. Both BMI and WHR are correlated with age, but both measures explained substantially more variance in attractiveness than did age alone, with WHR explaining 70%, BMI explaining 55% and age explaining only 46% of the variance in attractiveness. Despite the fact that BMI and especially WHR are convenient indices of attractiveness, nonlinear response surface methods using the original traits, rather than ratios, consistently outperform both of these indices, suggesting that these indices of abdominal adiposity tell only part of the story of body attractiveness.
Age, BMI and WHR
BMI was as good a predictor of attractiveness as any multiple regression model that included height, weight, their squares and cross-products. This is unsurprising as BMI is really just a combination of height and weight. On the other hand, regressions involving height, waist girth, hip girth and waist–hip vertical distance plus their squares and cross-products provided much greater explanatory power than WHR alone (90% vs. 70% of variance explained –). This suggests that although WHR does capture some of the key information underlying attractiveness, it is limited in its power to explain the effects of torso shape and adiposity on attractiveness. The sample of body scans that we used included only three women with WHR below 0.7. It therefore remains an open question whether attractiveness would have stabilized around or even beyond this value. Extrapolating from our analyses, it is more likely that values for WHR below 0.7 would have been even more attractive, as Dixson et al. (2007) found for Chinese women. Nonetheless, the simplest interpretation of our results is that waist girth per se is probably more important as a determinant of attractiveness than the ratio of waist to hip girth.
Throughout our study, waist girth was clearly an important univariate predictor of attractiveness. Women with narrow waists, especially relative to their height, were considered much more attractive. The importance of narrow waist girth is consistent with a recent study that showed both waist girth and abdominal depth are negatively correlated with attractiveness of women’s bodies (Rilling et al., 2009). By manipulating photographs presented to raters, Rozmus-Wrzesinska & Pawlowski (2005) showed that manipulations of waist, rather than hip width, have greatest effect on attractiveness ratings. The preference for narrow-waisted women in our study and others’ (Rozmus-Wrzesinska & Pawlowski, 2005; Rilling et al., 2009) may have the same evolutionary underpinnings as have been argued for small WHR: the fact that narrow waists are often a reliable indicator of youth, health and fecundity, and therefore of reproductive value (Rimm et al., 1988; Singh, 1993a; Zaadstra et al., 1993; Weeden & Sabini, 2005). Male preferences for narrow waists may well have evolved because of the direct benefits of mating with women of high reproductive value.
Beyond ratios: eight-trait response surface
Our eight-trait response surface identifies a number of important interactions among traits in determining the attractiveness of models. The two canonical axes of strongest stabilizing selection m7 and m8 are heavily weighted to the five measures of girth: upper arm, bust, thigh, ankle and waist girth, suggesting an intermediate optimum for girth measures. Much of this selection would have been missed by looking only at the quadratic effects of each of these traits (thigh girth being an exception), as it is the correlational terms, involving the interactions among measures, that most strongly influence the significant nonlinear selection on these axes. The selection on m7 and m8 suggests that the various girth measures may indeed be tightly integrated, with departures from an intermediate optimum in these trait combinations resulting in reduced attractiveness. The strong directional selection coupled with negative quadratic selection on m5 favours slender waists, thighs, upper arms and busts, combined with short waist–hip distances, indicative of low abdominal adiposity, but such slenderness has a limit within the normal range of women’s bodies.
Girth measures and the vertical distance between waist and hips are all influenced by adiposity, whereas arm length and overall height are not. It seems reasonable to interpret the overall pattern of linear and nonlinear selection, and particularly the selection on m8, m7, m5, m2 and m1 as a strong indication that the six measures affected by adiposity interact in complex ways in determining the attractiveness of a female body. It is important to note that not all fat, and thus not all measures of adiposity, have the same effects on fitness or well-being (Rimm et al., 1988; Zaadstra et al., 1993; Lassek & Gaulin, 2006). Based on the results that we present here, the same can be said about attractiveness; it is likely that the girth of limbs, bust and waist relative to one another and relative to overall height is more important than simple measures of overall adiposity (e.g. BMI) or of waist adiposity (e.g. WHR). The increased complexity of our empirical approach does make interpretation much less simple than, for instance, accounts of the role of WHR. However, we do believe the greater explanatory power offered by multivariate studies together with the removal of the inferential complications that are introduced by the use of ratios and indices (Jasienski & Bazzaz, 1999; Voracek, 2009) make this a promising approach for the future.
The bust, in context
The two axes of strongest concave selection, m1 and m2, were both also under significant directional selection. This pattern is often seen in sexual signals where a large part of the range of trait values is uniformly unattractive and then one particular combination of traits (in this case high m1 and low m2) is favoured, resulting in a rising peak on the edge of a fitness surface (LeBas et al., 2003; Bentsen et al., 2006; Gerhardt & Brooks, 2009). These models had large busts, small waists, narrow ankles, long arms and short distances between waist and hips. The latter trait is associated with narrow waists, as abdominal adiposity tends to cause the waist to appear higher up than it would otherwise do. Once again these results point to the fact that attractive bodies have a suite of attractive traits in the right combinations, some of which are associated with waist, hips and thighs, and some of which are elsewhere.
The strong contrast between larger busts and narrower waist, ankles and thighs in m1 merits further interpretation. This axis is under the strongest directional selection of any of the eight canonical axes, as well as the second strongest nonlinear selection – in this case a pattern of concave selection that serves to strengthen the directionality of the linear selection, a common pattern in sexual signals (Bentsen et al., 2006; Gerhardt & Brooks, 2009). Large values of m1 are favoured over small and intermediate values, and women with these values would have had large busts, small waists and narrow thighs and ankles. The pattern of contrasting selection for larger busts and narrower waists may be a general one. Previous studies have shown that larger breasts only elevate attractiveness in combination with narrower waists (Singh & Young, 1995; Furnham et al., 2006; Grundl et al., 2009). In one of the earliest studies of how WHR and breast size interact to determine attractiveness, Singh & Young (1995) showed that line drawings of women with large breasts and narrow waists are most attractive. In a study using manipulated computer-generated images, a large bust-to-waist ratio (BWR) was as good a correlate of attractiveness as a small WHR (r = 0.63 vs. −0.61, respectively, Grundl et al., 2009). In our study, BWR was a significant predictor of attractiveness in univariate regressions (b = 3.10 ± 0.25 SE; P < 0.001), but not quite as good as WHR (b = 4.65 ± 0.31 SE). In recent eye-tracking studies, men looked at and fixed first on either the breast or the waist area compared with the face, legs or pubic region (Dixson et al., 2009). Men also looked for longer at breasts, but attractiveness scores were more likely to be based on waist scores alone than the breast scores. Another eye-tracking study showed that raters, especially men, pay most attention to the breast and waist area and very little attention to the hip, leg, pubic or face regions when judging attractiveness (Cornelissen et al., 2009a).
The evolution of human female breast size and shape and particularly the role of sexual selection are the subject of long-running controversy (Cant, 1981; Mascia-Lees et al., 1986; Singh & Young, 1995; Pawlowski, 1999; Jasienska et al., 2004; Furnham et al., 2005, 2006; Dixson et al., 2009). It appears from these studies and from our findings that bust size, waist girth and, potentially, other traits are tightly integrated such that the effects of breast size cannot usually be detected in univariate analyses (where large busts are less attractive because of the correlation with overall adiposity) or by manipulating a small number of trait values. In particular, our study contributes to the evidence that large breasts (or large bust girth) are only attractive when combined with a narrow waist.
A study of Polish women has shown that women with large breasts combined with narrow waists have estradiol and progesterone profiles associated with high fecundity compared with women who have only one or neither of these attributes (Jasienska et al., 2004). It has been suggested that the evolution of permanently enlarged breasts in humans may have started as a side effect of the evolution of gynoid patterns of subcutaneous fat deposition under oestrogenic control (Pawlowski, 1999) and that large breasts may indicate good nutritional status (Cant, 1981).
Measures of length
Scanned bodies of taller women who had longer arms were strongly preferred. Much less attention has been paid in the literature to measures of body or limb length (but see Rilling et al., 2009; Bogin & Varela-Silva, 2010; Fan et al., 2004) than to measures of girth and adiposity. In a study using rotating video images of 42 women wearing form-fitting skin-coloured lycra leotards, Rilling et al. (2009) found that both stature and leg length had significant positive effects on attractiveness ratings. Likewise, in a study of 31 body scans (which formed a subset of our sample) presented in the same way as we presented scans here, Fan et al. (2004) showed that height (relative to body volume) and leg length have positive effects on attractiveness. Neither of these studies report effects of arm length, although Rilling et al. report preference for bodies with more slender arms (mid-arm circumference), consistent with our findings on upper arm girth. A study using experimentally modified computer-generated images showed that intermediate leg-to-body ratios are most attractive (Frederick et al., 2010), indicating a preference for legs that were in proportion to the body. Our results indicate that long, slender arms are highly attractive and that slender legs and tall models are highly attractive. In our variable selection stage, leg length did not contribute significantly to attractiveness in univariate or multivariate analyses. The extent to which any real contribution from leg length was obscured by the statistical contribution of other effects, such as thigh and ankle girth and total height, remains unknown, other than to say that because of the reported effects of leg length on attractiveness, we explored a possible contribution of leg length quite extensively during selection of variables for our eight-trait model and did not find evidence that it should be included.
Our results indicate that attractiveness ratings by men and women and by Hong Kong Chinese and Australian raters were strongly correlated, suggesting considerable cross-cultural consistency. The two samples of raters should not, however, be interpreted as entirely culturally independent because they are both drawn from mostly young adults in urban, developed economies with many shared media experiences. Nonetheless, the cross-national and cross-gender consistency that we document suggests that when models are stripped of their most obvious racial and cultural features, including hair and skin colour, facial detail and clothing, the features that make bodies attractive tend to be shared by men and women from China and Australia. This insight also broadens the cultural relevance of earlier work (Fan et al., 2004; 2005; Liu et al., 2006) that used some of the same models with only Hong Kong Chinese raters. It is also consistent with a recent study showing that men from Xi’an in central China tend to show preferences for images of women with smaller WHR (Dixson et al., 2007) because such preferences have been well documented in men and women from a large number of populations including Caucasian Americans (Singh, 1993a; Singh & Luis, 1994), African Americans, Indonesians and some remote tribal populations (Marlowe et al., 2005; Dixson et al., 2007). Other studies have shown that preferences for particular body types, including particular WHR, BMI or body mass values tend to vary consistently among world regions as well as with socio-economic and cultural differences within countries (e.g. Swami et al., 2010). The extent to which complex multivariate relationships that determine attractiveness, like those we document in this study, vary among countries, cultures and with socio-economic status remains an open question. However, because established methods exist for statistically comparing multivariate fitness surfaces (Chenoweth & Blows, 2005; Rundle et al., 2008), it is an empirically tractable one. The body scanning technique that we use here provides an excellent empirical approach to partitioning the effects of body dimensions per se from more obvious correlates of nationality, ethnic group and cultural groups. It is likely to become an important tool for research on the interactions between Darwinian evolution, culture and social psychology in shaping judgements and norms of attractiveness.
Thanks to the many volunteers who rated bodies for this project. RB was supported by a fellowship from the Australian Research Council. Thanks are also due to the Hong Kong Polytechnic University for its funding support to Prof. JT Fan through an Inter-faculty project (Project No: G-YG13) and a Niche Area Project (J-BB6T).