“Beauty” premium for social scientists but “unattractiveness” premium for natural scientists in the public speaking market

Public speaker data and speaking fees

The data on 734 public speakers, collected between December 2013 and January 2014 (Chan et al., 2014a, 2014b), were taken from speaker profiles on the websites of eight North-America based speaker agencies (BigSpeak Speakers Bureau, Keppler Speakers, Leading Authorities Speakers Bureau, Premiere Motivational Speakers Bureau, Speakerpedia, Speakers Platform, Sweeney Agency, and Washington Speakers Bureau). The final sample comprises 217 full-time academics (over half the career spent at an academic institution), 151 part-time academics (over half the career spent at a private or government institution while affiliated with academia for some of this period), and 366 nonacademics (no work or affiliation with any academic institution). The lower end of the fee range represents the minimum speaking fee (unless no range is specified), with the highest of the lower-end range values used as the fee for speakers listed on more than one agency website (Chan et al., 2014a, 2014b).

Facial attractiveness data

Facial attractiveness is based on photographs (frontal portraits) of the speakers obtained from public domains such as personal, professional, and institutional websites, or via a Google image search. Our searches successfully located a frontal portrait for 726 of the 734 speakers in the sample. We use Anaface.com, the web-based software for photo analysis that evaluates factors such as horizontal symmetry, nose to ear length ratio, nose width to face width ratio, mouth width to nose width ratio, face width to face height ratio, and eye width to innerocular distance ratio (Halford and Hsu, 2014; Hoegele et al., 2016; Sung et al., 2016; Babin et al., 2019; Baugh et al., 2019), to proxy for facial attractiveness. The software assigns each photo a beauty score from 1 to 10 based on facial geometry differentiated by gender. Because facial averageness or symmetry is a prime candidate not only for sexual selection and reproduction but also for overall phenotypic quality and developmental health (Alcock and Thornhill, 2014; Grammer and Thornhill, 1994; Thornhill and Gangestad, 1993; Trivers, 1985), facial averageness and symmetry are highly correlated with human beauty ratings (Komori et al., 2009; Rhodes et al., 2001). Using a facial structure-based proxy is equally important and avoids any familiarity bias, which could otherwise be an issue given our subjects’ considerable public exposure (e.g., through TED talks or best-selling books). However, such a project cannot be classified as an objective measure, because individual (Hönekopp, 2006) and cross-cultural differences exist with respect to the features of an attractive face (Little et al., 2011), a point that Darwin (1871) emphasized over 200 years ago. Measurements to assess attractiveness can vary culturally (Yu and Shepard, 1990) and are subject to westernization (Yu and Shepard, 1998), as many studies have been conducted in North America, Britain, or Australia (Little et al., 2011). However, some meta-analysis evidence indicates the existence not only of within-culture agreements, but also of strong cross-ethnic and cross-cultural agreements among raters about who is and who is not attractive (reliability between r = 0.88 and r = 0.94). This may indicate some universal standards by which attractiveness is judged (Langlois et al., 2000; Rhodes 2006). Such cross-cultural or cross-ethnic consistency in ratings are particularly evident when evaluating female faces (Johnston and Franklin, 1993; Little et al., 2011) or when evaluating unattractive faces (Sorokowski et al., 2013). Apicella et al. (2007) also find a preference for averageness within an isolated hunter-gatherer society, namely the Hadza of Northern Tanzania). Agreement is also higher among relatives than strangers (Bronstad and Russell, 2007). For example, Cunningham et al. (1995) find that Asian, Hispanic, and White judges gave higher ratings to faces with neonate large eyes, greater distance between eyes, and small noses.

Because this facial geometry is predicated on 17 facial landmarks placed manually on the speakers’ images (see Fig. S1 in the SI Appendix), between June 5 and June 20, three research assistants made independent evaluations of each sample photo to minimize potential measurement error from manual placement. The average of the three scores was then used in the subsequent analyses (see Table S1). In the Supplementary Information, we offer detailed documentation regarding the complete procedure for measuring facial attractiveness of our sample speakers and the process of validating the attractiveness scores obtained from Anaface.com with facial geometry-based beauty measures suggested by prior literature. Notably, a recent study by Babin et al. (2019) found that Anaface scores are weakly (but significantly) positively correlated with human attractiveness ratings, thus providing support for the internal validity of using Anaface score as a proxy measure for facial attractiveness. Nevertheless, due to research funding and capacity limitations, we are unable to conduct the ideal validation exercise with an independent sample of models and human raters.

Performance and influence data

Internally within academia, performance can be estimated by the total number of publications and citations, citations per paper, and h-index from Google Scholar and Scopus for speakers with an academic background. The number of web pages mentioning the speaker’s full name, which were indexed by the Google search results, served to quantify external influence and prominence. Search queries (excluding pages from an.edu domain) were automated on April 14, 2014 using the Google search application programming interface (API) (Chan et al., 2014a, 2014b). An exclusion rule was applied manually to avoid spurious matches, and a search count was deemed invalid if five or more pages were not attributed to the speaker (N = 154) (Chan et al., 2014a, 2014b). External impacts were additionally measured by TED talk experience (number of times invited to present at TED conferences before 2013), number of books published (titles listed with the US Library of Congress), and recognition (number of New York Times best-selling books, number of weeks on the NYT list, and any of 21 major nonfiction book awards) (Chan et al., 2014a, 2014b).

Speaker biographical information

Based on the automated word searches, we categorize speakers by field of expertize into three groups: natural science (e.g., physics, biology and medicine, N = 187), social science [business] (e.g., economics, finance, management, and marketing, N = 358), and social science [other] (e.g., sociology, psychology, and political science, N = 189) (Chan et al., 2014a). We differentiate those two social science groups based on the sample size (see Supplementary Table S2), including under business-related speakers all those fields that are usually part of a Business School. As is evident from the descriptive statistics these scholars are more active in the speaking market. Supplementary Table S2 also details the field of studies for the sample speakers. The information on gender, ethnicity, and professional age (years since highest degree) is taken from speaker biographies, CVs, LinkedIn entries, and Wikipedia images. We also use the within-field average rankings from the 2004–2018 QS World University rankings in Natural Sciences and Social Sciences and Management (UniversityRankings.ch) to quantify the quality of education at the university that awarded each speaker’s highest degree. Because the rankings only cover the top 500 universities, we assign a rank of 500 to any universities that do not appear on the rankings.

Data analysis

All analyses use ordinary least-squares (OLS) regressions with heteroscedasticity consistent (robust) standard errors^{Footnote 2}. Because academic achievements and external influence are the main determinants of the scholars’ speaking fees (Chan et al., 2014a, 2014b), we first examine the relation between these two variables as a check for conflation in the main association between facial beauty and speaking fees. In analyzing facial beauty’s effect on scholarly achievements and external influence, we control for the scholars’ biographical characteristics: gender (male = 0, female = 1), ethnicity (dummy variables for African and Asian, with Caucasian as the reference group), professional age (years since highest degree), professional age squared^{Footnote 3}, QS ranking of the university of highest degree, whether a Nobel Prize winner (laureate = 1, others = 0), whether US based (outside US = 0, US = 1), academic engagement (less than half the entire career spent in academia = 0; more than half the entire career spent in academia = 1), and academic discipline (natural sciences, social sciences, business). Based on (Chan et al., 2014a, 2014b), we proxy the scholars’ academic achievement by number of publications (productivity) and citations (academic impact) and their external influence by number of webpages (in noneducational network domains) with the scholar’s full name indexed on Google, all assumed to be the main predictors of speaking fees charged. To assess whether the discipline is mediating the relation between facial beauty and speaking market value, we use interaction terms to estimate the difference in slope coefficients between facial beauty and discipline. As the distributions are skewed, we also log-transform the number of publications, number of citations, Google search results, and minimum speaking fees. Descriptive statistics are reported in Supplementary Table S3. When reporting the results of our analyses, we include four statistical significance levels, 0.001, 0.01, 0.05, and 0.1.