Beauty and the Labor Market: Accounting for the Additional Effects of Personality and Grooming

Beauty and the Labor Market: Accounting for theAdditional Effects of Personality and Groominglabr_511 228..251

Philip K. Robins — Jenny F. Homer — Michael T. French

Abstract. This paper examines the influence of three non-cognitive personal traits — beauty, person-ality, and grooming — on the labor market earnings of young adults. It extends the analyses ofHamermesh and Biddle [1994, American Economic Review 84(5): 1174–1194] and others who focusprimarily on the effects of beauty on labor market earnings. We find that personality and groomingsignificantly affect wages, and their inclusion in a model of wage determination reduces somewhat theeffects of beauty. We also find some evidence of employer discrimination based on these traits in thesetting of wages.

1. Introduction

In recent years, economists have expanded the study of labor market discrimination toinclude the effects of physical appearance (or beauty) on earnings. The seminal paper byHamermesh and Biddle (1994) finds a ‘plainness penalty’ of 5–10 per cent and a slightly lower‘beauty premium’ for both men and women in the workplace. Hamermesh and Biddle’sanalysis suggests that some degree of employer discrimination is present, but the premiumsand penalties may also reflect occupational crowding, customer discrimination, or productiv-ity differences. Since the original study, Hamermesh and his colleagues have investigatedsimilar topics ranging from the impact of lawyers’ appearance on their salaries to the likeli-hood of attractive politicians being elected (Biddle and Hamermesh, 1998; Hamermesh, 2006;Hamermesh and Parker, 2005; Hamermesh et al., 2002; Pfann et al., 2000). Other recentstudies include French (2002), who analyses self-reported appearance data and finds a beautypremium for female workers but not male workers, Mobius and Rosenblat (2006), who

Philip K. Robins — Jenny F. Homer — Michael T. French (author for correspondence), HealthEconomics Research Group, Department of Sociology, Department of Epidemiology and PublicHealth, and Department of Economics, University of Miami, 5202 University Drive, Merrick Building,Room 121F, P.O. Box 248162, Coral Gables, FL 33124-2030, USA. E-mail: [email protected].

The authors are grateful for research assistance from Robin Prize and editorial/administrative assis-tance from Carmen Martinez and William Russell. This research uses data from Add Health, a programproject directed by Kathleen Mullan Harris and designed by J. Richard Udry, Peter S. Bearman, andKathleen Mullan Harris at the University of North Carolina at Chapel Hill, and funded by grantP01-HD31921 from the Eunice Kennedy Shriver National Institute of Child Health and Human Devel-opment, with cooperative funding from 23 other federal agencies and foundations. Special acknowledge-ment is due to Ronald R. Rindfuss and Barbara Entwisle for assistance in the original design. Informationon how to obtain the Add Health data files is available on the Add Health website (http://www.cpc.unc.edu/addhealth). No direct support was received from grant P01-HD31921 for this analysis. Theauthors are entirely responsible for the research and results reported in this paper and their position oropinions do not necessarily represent those of the Carolina Population Center or the University of Miami.

LABOUR 25 (2) 228–251 (2011) DOI: 10.1111/j.1467-9914.2010.00511.xJEL J24, J31, J71© 2011 CEIS, Fondazione Giacomo Brodolini and Blackwell Publishing Ltd,

9600 Garsington Rd., Oxford OX4 2DQ, UK and 350 Main St., Malden, MA 02148, USA.

investigate the possible causes of a beauty premium within an experimental labor market, andFletcher (2009), who examines whether the beauty premium varies with measured ability.Finally, Doran and Hersch (2009) perform a comprehensive analysis of the robustness of theestimated beauty effect found by Hamermesh and Biddle and conclude that the effect isunstable and generally not statistically significant.

The purpose of this paper is to extend the beauty literature by considering the effects of twoadditional (and possibly correlated) non-cognitive personal traits on the labor market earn-ings of a sample of young adults. The importance of non-cognitive personal traits in influ-encing labor market earnings was discussed over three decades ago in Bowles and Gintis(1976) and more recently was summarized in Bowles et al. (2001). Borghans et al. (2008)present a more extensive discussion of this topic and link it to the literature in the field ofpsychology. They argue that certain non-cognitive personal traits are productivity enhancingand can therefore influence labor market earnings for some occupations.

Personality attractiveness and grooming are the two additional non-cognitive traits consid-ered in our analysis. A worker’s personality attractiveness and grooming can enhance his or herproductivity and generate higher earnings by improving interactions and communications withsupervisors, co-workers, and customers. For example, positive personality traits amongworkers can lead to a more friendly and relaxed workplace, improving overall productivity. Injobs involving contact with customers, well-groomed individuals may make a customer feelmore secure and trusting, which could lead to greater success in selling products and obtainingcontracts. Even in jobs that do not involve direct customer contact, well-groomed individualsmay appear more professional and confident to their co-workers and supervisors. Poorgrooming habits and/or objectionable personalities could be associated with lower earnings ifthese traits contribute to workplace conflict and poor customer relations. Above averagepersonality attractiveness and grooming do not necessarily imply above average productivity,however, which raises the possibility of employer discrimination based on these personal traits.

In most of the related literature, personality attractiveness and grooming are not taken intoaccount (Ritts et al., 1992), primarily because data on such traits are generally unavailable.1

This is a potentially important limitation because studies that only examine beauty maymistakenly attribute significant wage effects to this trait alone. A few studies have attemptedto overcome this limitation through the inclusion of variables related to other characteristicsin their earnings models (Hamermesh and Parker, 2005; Hamermesh et al., 2002; Süssmuth,2006). For example, Hamermesh et al. (2002) include women’s spending on clothing andcosmetics as a proxy for grooming in addition to interviewer appearance ratings. Althoughthese authors find that beauty increases earnings, spending on beauty enhancements producesonly a small incremental benefit in terms of higher wages. Other studies focus on weight andheight in place of a direct measure of physical attractiveness (Crosnoe and Muller, 2004; Loh,1993), whereas Frieze et al. (1991) and Harper (2000) use weight and height in addition to aphysical attractiveness measure.

Beauty could lead to higher or lower earnings, depending on the employee and the nature ofthe job. If physically attractive individuals invest relatively less time and other resources inactivities that develop their human capital, then physical attractiveness may lead to lowerearnings.2 Beauty could also have a positive effect on earnings if there is discrimination in thelabor market favoring attractive individuals or if beauty is productivity enhancing in occupa-tions involving extensive contact with the public (e.g. fashion models, entertainers, newscast-ers). If beauty is correlated with personality attractiveness and grooming, the beauty premiumor plainness penalty estimated in previous studies may be partially picking up the effects of theseomitted factors. French et al. (2009) find that when examined alone, beauty is positively related

229Beauty and the Labor Market

© 2011 CEIS, Fondazione Giacomo Brodolini and Blackwell Publishing Ltd

to high school grade point average among a sample of young adults. The effects of beauty turnnegative or lose statistical significance, however, when personality attractiveness and groomingare added to the model. A statistically significant grade premium is present for well-groomedmale and female students, and an even larger penalty exists for poorly groomed male students.Female students with pleasant personalities also receive a grade premium.

Given the range of possibilities when considering multiple aspects of non-cognitive personaltraits, this paper investigates whether controlling for personality attractiveness and groomingalters the relationships found in previous studies between beauty and labor market earnings. Wealso attempt to determine whether the estimated effects for beauty, personality attractiveness,and grooming reflect discriminatory practices on the part of employers, factors that enhanceworker productivity (including customer discrimination), or some combination of these factors.

The next section describes our empirical model of earnings and personal traits.3 Section 3discusses the Add Health data, the construction of the variables used in the analysis, and thedata limitations. Section 4 presents the results from the basic model. Section 5 exploreswhether the core findings can be traced to employer discrimination. Section 6 examines thesensitivity of the findings to alternative empirical specifications. Finally, Section 7 presents oursummary and conclusions.

2. The empirical framework

The primary analysis presented in this paper consists of estimating a series of earningsequations of the following form:

Lnwage Beauty Personality Grooming Xi i i i i= + + + + +β β β β β μ0 B P G X [1]

where Lnwage is the natural logarithm of individual i’s hourly earnings, Beauty is a categoricalmeasure of an individual’s physical attractiveness, Personality is a categorical measure of anindividual’s personality attractiveness, Grooming is a categorical measure of an individual’spersonal appearance, X is a set of control variables (including traditional human capitalmeasures), and m is a random error term. First, equation [1] is estimated assuming bP and bG

are zero (i.e. excluding Personality and Grooming from the model). In this basic model, wefollow the initial specification of Hamermesh and Biddle (1994) as closely as possible to permitdirect comparisons with their findings. Separate equations are estimated for men and women.Next, Personality and Grooming are added sequentially to determine their independent effectsand whether their inclusion alters the inferences for the beauty measures. To test the sensitivityof our basic findings, we examine additional specifications in which we alter the controlvariables, re-estimate the models using different samples, account for possible sample selectionbias, and use alternative measures of the personal traits.

In addition to our primary empirical analysis, we attempt to identify whether employerdiscrimination is present. First, we interact our personal trait measures with an index reflectingthe degree to which a person’s occupation is likely to be productivity enhancing. In the secondapproach, we follow Hamermesh and Biddle’s (1994) strategy of estimating separate equa-tions within certain occupations where the personal traits are likely to be productivity enhanc-ing and comparing the results with the estimated effects in occupations where the personaltraits are unlikely to be productivity enhancing. In defining the appropriate occupations forcomparison, we use information from the recently developed Occupational InformationNetwork online database (O*NET Online), which replaced the Dictionary of OccupationalTitles (DOT) as the primary source of occupational information in the USA.4

230 Philip K. Robins — Jenny F. Homer — Michael T. French


3. Data and variables

3.1 The National Longitudinal Study of Adolescent Health (Add Health)

Our analysis uses two waves of data from Add Health, a school-based, longitudinal studyof adolescent health-related behaviors and their consequences in young adulthood. Wave 1was administered during 1994–95 and included 20,745 adolescents sampled from 80 highschools and 52 middle schools. In 2001–2, 15,170 respondents were re-interviewed in Wave 3when they were 18–28 years old.5 Numerous questions were asked about employment, includ-ing employment status, hours of work, wages and earnings, and occupation. The sampleanalysed in this paper consists of 6,074 individuals (3,136 men and 2,938 women) who (1) werenot in school at the time of the Wave 3 survey; (2) were working at least 20 hours per week intheir current job; and (3) had valid employment and occupational data (see Table 1).

The Add Health data have many appealing features for our purposes including interviewerassessments of three key personal traits and a rich set of economic and demographic charac-teristics for each respondent.6 To our knowledge, this is the first study to use the Add Healthdata to examine the relationships between non-cognitive personal traits and earnings.7

Despite numerous advantages, the Add Health data have at least three important draw-backs that might limit the precision and generalizability of our findings. First, the respondentsare relatively young and their careers are probably still in transition. The average age of themen and women in our sample is 22 years. Thus, the individuals in our sample are relativelynew participants in the labor market — they average about 1.5 years at their current job andabout 6 years since they first had a regular job.8

Second, personal traits may be more volatile among this age group than among olderworkers. If so, then the recorded characteristics may not fully reflect the traits that willultimately determine long-run earnings. French et al. (2009) reported that the changes inappearance ratings for Add Health respondents across Waves 1 and 3 were somewhat more

Table 1. Construction of analysis sample

N = 15,170 respondents in Waves 1 and 3 of Add Health↓

N = 14,390 after excluding respondents who were disabled, had poor self-reported health, ever had aphysical or nervous condition that kept them from working, or in the military

↓N = 14,372 after excluding respondents who were missing personal appearance measures

↓N = 13,558 after excluding respondents who never had a job or never worked at least 9 weeks at a jobthat was at least 10 hours per week

↓N = 8,435 after excluding respondents who were attending school full or part time

↓N = 6,432 after excluding respondents who were currently working less than 10 hours per week orwho were working less than 20 hours per week at their main job

↓N = 6,172 after excluding respondents who were missing hourly wages at their current job or whoreported working but said their wages were zero

↓N = 6,074 after excluding respondents who were missing O*NET data

↓ ↓ ↓The analysis sample of 6,074 includes 3,136 men and 2,938 women



pronounced than those reported in the sample of adults analysed by Hamermesh and Biddle(1994).

Third, physical attractiveness, personality attractiveness, and grooming are measured con-currently (in Wave 3) with the labor market variables and other non-cognitive traits that couldalso influence earnings. Consequently, the estimated effects for personality attractiveness andgrooming (and possibly beauty as well) may be biased due to reverse causality or importantomitted variables that are correlated with the traits we consider. To account for these possi-bilities, we explored using an instrumental variables procedure to derive our estimates.Unfortunately, existing measures in the data set (e.g. other personal characteristics of therespondents and their families) were weak instruments, thereby producing imprecise esti-mates, and the merging of external instruments (e.g. community characteristics) is not possiblebecause Add Health does not provide geographic identifiers. We also considered using per-sonal trait measures from Wave 1 as instruments for the Wave 3 measures, but the estimateswere imprecise.

3.2 The personal trait measures

The Add Health interviewers responded to separate questions on three key personal traitspertaining to the respondents’ physical attractiveness, personality attractiveness, and groom-ing (see Appendix A for the specific survey questions).9 For each of these traits, five responsesto a single question were allowed — two for above average, one for average, and two for belowaverage. Because the interviewers answered a series of questions about the relative traits ofrespondents, one would expect to see a symmetrical distribution resembling a normal curvewith ‘average’ as the modal response and a fairly equal number of responses on either side ofthe ‘average’. However, inspection of the data reveals a preference among interviewers for topranking the respondents, with about twice the number being designated above average thanbelow average. To address the generosity of the interviewers, we coded the two highestcategories as separate groups, the third category as ‘average’, and the bottom two categoriesas a combined ‘below average’ group. This recoding yielded a four-category distribution withat least 4 per cent of the sample in every group.10 Table 2 presents the cross-tabulations for thethree sets of measures.11

In our empirical models, we use the personal traits measured at Wave 3 and estimate theireffects on current earnings. Wave 3 interviews were conducted with 15,170 of the 20,745 Wave1 respondents. To examine the possibility of attrition bias, we compared the personal traitsmeasures at Wave 1 for respondents in our analysis sample with those who were onlyinterviewed at Wave 1. For most of the measures, there were no significant differences betweenthe two groups with the exception of two personality attractiveness measures. Those who wereonly surveyed in Wave 1 were less likely to be rated as having an attractive personality(p < 0.05) and more likely to be rated as having a less than average personality attractiveness(p < 0.05) compared with respondents in our analysis sample.

3.3 Hourly wage measure

To construct an hourly wage measure for the Add Health respondents, we use informationon the rate of pay, weekly hours of work, and the pay period from their main job where theywork the most hours. About 75 per cent of the respondents who are currently employed reportan hourly wage. For the others, we use information on their rate of pay for each pay periodand the number of hours typically worked during a week to construct an hourly wage



Tab

le2.

Cro

ss-t

abul

atio

nof

pers

onal

trai

tm

easu

res

(N=

6,07

4)

Phy

sica

latt

ract

iven

ess

Per

sona

lity

attr

acti

vene

ss

Tot

alV

ery

attr

acti

veA

ttra

ctiv

eA

vera

geL

ess

than

aver

age

Ver

yat

trac

tive

Att

ract

ive

Ave

rage

Les

sth

anav

erag

e

Per

sona

lity

attr

acti

vene

ssV

ery

attr

acti

vepe

rson

alit

y40

938

714

722

[6.7

3][6

.37]

[2.4

2][0

.36]

Att

ract

ive

pers

onal

ity

203

1,27

982

169

[3.3

4][2

1.06

][1

3.52

][1

.14]

Ave

rage

pers

onal

ity

4647

71,

758

159

[0.7

6][7

.85]

[28.

94]

[2.6

2]L

ess

than

aver

age

pers

onal

ity

538

115

139

[0.0

8][0

.63]

[1.8

9][2

.29]

Gro

omin

gV

ery

wel

lgro

omed

297

199

5730

326

171

6026

583

[4.8

9][3

.28]

[0.9

4][0

.49]

[5.3

7][2

.82]

[0.9

9][0

.43]

[9.6

0]W

ellg

room

ed29

01,

169

607

5441

61,

186

467

512,

120

[4.7

7][1

9.25

][9

.99]

[0.8

9][6

.85]

[19.

53]

[7.6

9][0

.84]

[34.

90]

Ave

rage

groo

min

g75

788

2,01

316

721

296

61,

729

136

3,04

3[1

.23]

[12.

97]

[33.

14]

[2.7

5][3

.49]

[15.

90]

[28.

47]

[2.2

4][5

0.10

]L

ess

than

aver

age

groo

min

g1

2516

413

811

4918

484

328

[0.0

2][0

.41]

[2.7

0][2

.27]

[0.1

8][0

.81]

[3.0

3][1

.38]

[5.3

0]T

otal

663

2,18

12,

841

389

965

2,37

22,

440

297

6,07

4[1

0.92

][3

5.91

][4

6.77

][6

.40]

[15.

89]

[39.

05]

[40.

17]

[4.8

9][1

00.0

0]

Cel

lper

cent

ages

repo

rted

inbr

acke

tsbe

low

the

cell

freq

uenc

ies.



measure. As is common with most measures of hourly earnings calculated in this manner,there are unusual values at both the high and low end. The hourly wage in our sample rangesfrom less than $0.001 to over $1,000. To minimize the effect of outliers, we estimate all modelsusing robust regression, which is a compromise between dropping selective outliers andseriously violating the assumptions of ordinary least squares (OLS) regression. Robust regres-sion is a form of weighted least squares and assigns very low weights to extreme values of thedistribution based on information from the residuals.12 To check the stability of our results, weestimate alternative models in which we follow the more standard procedure of excludingextreme outliers (see Section 6 below).

3.4 Control variables

As indicated earlier, we include control variables that essentially follow Hamermesh andBiddle’s (1994) specification to account for demographic, economic, and geographic charac-teristics. The full set of control variables is presented in Table 3. In Section 6, we report resultsusing alternative sets of control variables.

3.5 Descriptive statistics

Descriptive statistics for the hourly wage, personal traits, and control variables are pre-sented in Table 3. Unless otherwise noted, all variables are measured at Wave 3. The meanwage in our sample (2001/2002 dollars) is $12.82 ($12.91 for men and $12.72 for women).Between 10 and 20 per cent of the individuals are rated in the highest category for each of thethree personal trait measures (the percentages are higher for women in all areas), between 35and 40 per cent are in the second highest category (again, these percentages are higher forwomen), and between 4 and 7 per cent of the cases are rated in the less than average category(these percentages are about the same for men and women, implying a lower proportion ofwomen are in the ‘average’ category for all three measures).

The average age in the sample is 22 years. Roughly one-fifth of the sample is currently married(the percentage is higher for women). Forty-eight per cent are high school graduates with nofurther education, 25.9 per cent have some college, and 15.7 per cent have a college degree.Whites comprise over half of the sample and Hispanics and Blacks are about 18 per cent each.

4. Results from the basic model

Results for the personal trait variables from the basic model are presented in Table 4.13

When only the beauty measures are included, the results are qualitatively similar to thosefound in previous studies (e.g. Hamermesh and Biddle, 1994). There is an estimated beautypremium of approximately 12 per cent for very physically attractive men, 7 per cent for veryphysically attractive women, and 4 per cent for physically attractive men and women. Forthose rated as less than average in physical attractiveness, we estimate a plainness penalty of1.8 per cent for men and 2.5 per cent for women, but neither penalty is statistically significant.

When we include the personality attractiveness measures in the model, the beauty premiumsessentially remain unchanged for men, but are lower for women. Although not statisticallysignificant, the plainness penalties remain unchanged for both groups. None of the personalityattractiveness variables is statistically significant for men, but there is between a 4 and 5 percent premium for above average personalities for women.



Table 3. Sample statistics

Men(N = 3,136)

Women(N = 2,938)

Full sample(N = 6,074)

Mean (SD) Mean (SD) Mean (SD) Min. Max.

Wage measuresHourly wage at current job*** 12.911 12.725 12.821 0.004 1,327.5

(34.397) (47.745) (41.391)Natural logarithm of hourly wage at current job*** 2.281 2.139 2.212 -5.561 7.191

(0.808) (0.933) (0.873)

Personal appearance measuresVery physically attractive*** 0.079 0.141 0.109 0 1Physically attractive*** 0.342 0.378 0.359 0 1Less than average physical attractiveness 0.062 0.067 0.064 0 1Very attractive personality*** 0.121 0.199 0.159 0 1Attractive personality 0.381 0.401 0.391 0 1Less than average personality* 0.054 0.044 0.049 0 1Very well groomed*** 0.073 0.120 0.096 0 1Well groomed*** 0.315 0.386 0.349 0 1Less than average grooming*** 0.062 0.046 0.054 0 1

Control variablesHispanic* 0.186 0.166 0.177 0 1Missing Hispanic 0.002 0.002 0.002 0 1Black** 0.172 0.195 0.183 0 1Missing Black 0.004 0.002 0.003 0 1American Indian 0.041 0.035 0.038 0 1Missing American Indian 0.005 0.002 0.004 0 1Asian 0.065 0.068 0.066 0 1Missing Asian 0.005 0.003 0.004 0 1English spoken at home 0.882 0.895 0.888 0 1Age*** 22.492 22.375 22.436 18 28

(1.676) (1.625) (1.652)Currently married*** 0.194 0.246 0.219 0 1High school degree*** 0.503 0.456 0.480 0 1Some college 0.255 0.264 0.259 0 1College degree*** 0.115 0.202 0.157 0 1Excellent self-reported health*** 0.325 0.285 0.306 0 1Very good self-reported health 0.427 0.418 0.423 0 1Good self-reported health*** 0.213 0.249 0.230 0 1Years since first regular joba,*** 6.201 5.755 5.986 0 15

(2.603) (2.342) (2.490)Years in current job*** 1.607 1.386 1.500 0 15

(1.845) (1.557) (1.715)Had resident father at Wave 1*** 0.728 0.696 0.712 0 1Had resident mother at Wave 1 0.935 0.940 0.937 0 1Professional resident father at Wave 1 0.058 0.070 0.064 0 1Missing professional resident father at Wave 1 0.005 0.003 0.004 0 1White collar resident father at Wave 1** 0.158 0.141 0.150 0 1Missing white collar resident father at Wave 1 0.005 0.003 0.004 0 1Resident mother or father at Wave 1 attended college 0.433 0.422 0.428 0 1Missing resident mother or father at Wave 1 attended college*** 0.027 0.016 0.022 0 1Resident mother at Wave 1 born in the USA** 0.765 0.789 0.777 0 1Missing resident mother at Wave 1 born in the USA 0.002 0.002 0.002 0 1Resident father at Wave 1 born in the USA 0.588 0.574 0.581 0 1Missing resident father at Wave 1 born in the USA** 0.003 0.001 0.002 0 1

O*NET measuresAverage O*NET score*** 50.075 55.631 52.762 14.417 76.300

(11.042) (9.105) (10.524)Job skill level*** 2.570 2.739 2.652 1 5

(0.833) (0.908) (0.874)

Standard deviations reported in parentheses for continuous variables. All variables from Wave 3 unless otherwise noted.a Regular job is defined as a job for at least 10 hours per week for at least 9 weeks.* Statistically significant differences between men and women, p < 0.10. ** Statistically significant differences between men and women, p < 0.05.*** Statistically significant differences between men and women, p < 0.01. (Kruskal–Wallis tests for equality of populations.)



When both personality attractiveness and grooming are included in the model, the beautypremium is reduced slightly for men and is no longer statistically significant for women. Beingvery well groomed and well groomed generate premiums of about 4–5 per cent for men, butonly the premium for being well groomed (0.034) is statistically significant for women.14

Personality attractiveness continues to be important for women and seems to matter morethan being physically attractive. For men, being physically attractive generates a largerpremium than having an attractive personality or being well groomed.

The results from the full model suggest that some of the estimated effects of beauty foundby Hamermesh and Biddle (1994) could actually be due to other correlated personal traits.These results, however, should be interpreted with caution because our sample is younger andthe personal trait measures we use are based entirely on interviewer assessments.15 Nonethe-less, for this young sample, we find that beauty is not the only personal trait that matters in thelabor market. Beauty appears to be the most important of the three personal traits consideredhere for men (although grooming also plays an important role), but it is the least important forwomen after accounting for the effects of personality attractiveness and grooming.

4.1 Do the results reflect employer discrimination or productivity differences?

As we explained earlier, a beauty premium may reflect employer discrimination ifemployers reward physically attractive workers or penalize unattractive workers solely

Table 4. Effects of personal traits on natural logarithm of hourly wages at current job

Men (N = 3,136) Women (N = 2,938)

b/se b/se b/se b/se b/se b/se

Very physically attractive 0.122*** 0.111*** 0.089*** 0.065*** 0.038* 0.023(0.024) (0.027) (0.028) (0.019) (0.022) (0.023)

Physically attractive 0.044*** 0.040*** 0.027* 0.041*** 0.024* 0.013(0.014) (0.015) (0.016) (0.014) (0.014) (0.015)

Less than average physicalattractiveness

-0.018 -0.026 -0.019 -0.025 -0.026 -0.015(0.027) (0.028) (0.029) (0.025) (0.027) (0.028)

Very attractive personality 0.022 0.008 0.051** 0.042**(0.023) (0.024) (0.020) (0.020)

Attractive personality 0.011 0.002 0.041*** 0.032**(0.015) (0.015) (0.015) (0.015)

Less than average personality 0.029 0.031 0.020 0.018(0.030) (0.030) (0.033) (0.033)

Very well groomed 0.051* 0.027(0.028) (0.023)

Well groomed 0.043*** 0.034**(0.016) (0.015)

Less than average grooming -0.023 -0.038(0.028) (0.031)

R2 0.204 0.204 0.206 0.288 0.290 0.292

* p < 0.10, ** p < 0.05, *** p < 0.01.Coefficient and standard errors (in parentheses) reported. ‘Average’ is the excluded category for all the personal traitmeasures. All models are estimated using robust regression and control for age, age squared, marital status, educa-tion, self-reported health status, years since first regular job, years in current job, race, Hispanic ethnicity, whetherEnglish is spoken at home, employment status of residential father at Wave 1, presence of residential father at Wave1, presence of residential mother at Wave 1, whether residential mother and father were born in the USA, and whethera parent attended college. Full regression results are presented in Appendix B.



because of their appearance rather than their actual productivity. A similar argument can bemade when considering whether there is discrimination in the workplace against employeeswith certain grooming habits and personalities. Because personal traits may also be highlycorrelated with productivity, it is difficult to empirically determine which if any effects are dueto discrimination.

Discrimination can potentially be identified by evaluating the effects of personal traits inoccupations where these traits are important and comparing them with the effects of personaltraits in occupations where these traits are less important (Hamermesh and Biddle, 1994).16

For example, the personal traits are probably associated with higher productivity in occupa-tions involving significant contact with the public (e.g. modeling, entertaining, teaching, sales,news reporting), but less so or not at all in those occupations where contact with the public isminimal (e.g. computer programming, bookkeeping, journalism).17 Hamermesh and Biddle(1994) use the DOT classification scheme, opinions of eight individuals, and information froma survey of employers to identify occupations where physical attractiveness is potentiallyproductivity enhancing. They specify a partially interactive earnings model with the beautyvariables, a dummy for being in an occupation where looks are likely to be productivityenhancing, and interactions between the beauty variables and the occupation dummy. Theirresults suggest that beauty is productivity enhancing in certain occupations, but also generateshigher earnings in occupations where beauty is not likely to be an important factor. Theyconclude that the beauty premium in the labor market reflects a combination of higherproductivity and employer discrimination.

To test for the possible existence of employer discrimination, we adopt two approachesusing the Occupational Information Network (O*NET Online), which is an updated versionof the DOT database used by Hamermesh and Biddle (1994). The O*NET database containssix descriptors of various characteristics associated with job performance (e.g. ‘work activi-ties’) and several more finely defined elements within each descriptor (e.g. ‘working directlywith the public’, ‘selling or influencing people’). For each DOT occupation, an importancescore ranging from 0 to 100 is recorded for each element based on surveys of workers withineach occupation (higher scores indicate that the element is more important). We selected 12elements that indicated to us that personal traits would be important. Then, for the mainoccupation reported by each Add Health sample member, we recorded the O*NET scores forthe 12 elements and derived a summary score by computing the unweighted average of the 12individual element scores. Appendix C describes in more detail how the O*NET database wasused to classify occupations.

In our first approach, we interact each of the personal trait measures with the standardized(Z-score) O*NET measure. If higher wages are due mainly to productivity-enhancing factors,the interaction terms should be statistically significant and positive for the desirable traits(better than average looks, personality attractiveness, and grooming) and statistically signifi-cant and negative for the undesirable traits. In our models, none of the interaction terms isstatistically significant at even the 10 per cent level for either men or women.18 Thus, we cannotreject the hypothesis that the wage differentials are independent of occupational classification,implying some degree of employer discrimination in setting wages.

Second, we divide the sample into those above and below the median O*NET score (i.e.those in which personal traits are more likely to be productivity enhancing and those in whichpersonal traits are less likely to be productivity enhancing) and estimate separate earningsequations for each group. Job skill level (i.e. job zone), a measure in O*NET ranging from 1to 5 to reflect the skill level associated with each occupation, is included in the models tocontrol for the general level of preparation (i.e. training and education) required in the



occupation.19 If employer discrimination is present, we would expect to find significant effectsof the personal traits in occupations below the median O*NET score (i.e. occupations in whichthe personal traits are less important for job performance). Furthermore, the absence ofsignificant effects of the personal traits in occupations above the median O*NET score couldsignal a less overt form of employer discrimination.

Tables 5A (men) and 5B (women) present the effects of the personal traits on hourly wagesfor occupations above and below the median O*NET scores.20 For both men and women, wefail to reject (p < 0.10) the null hypothesis that the coefficients are equal for the two occupa-tional groups (columns C and F) in the model with all three physical appearance variables.This finding is consistent with the results from the interaction model. For men (Table 5A), thebeauty premium appears to span all occupations and is somewhat larger in occupations wherebeauty is less important. There also appears to be a penalty for less than average grooming inoccupations with O*NET scores below the median. The results for physical attractiveness andless than average grooming suggest that some employer discrimination may be present. On theother hand, being well groomed generates a significant premium in occupations where thepersonal traits are expected to be more important, suggesting that being well groomed isproductivity enhancing. For women (Table 5B), we do not find strong evidence of employerdiscrimination, as no significant premiums or penalties exist in occupations where the personaltraits are less important for job performance. In those occupations where the personal traitsare expected to be important (column F), women with attractive or very attractive personali-ties receive a statistically significant premium, whereas poorly groomed women receive asignificant penalty

5. Sensitivity tests

We now turn to an extensive set of sensitivity checks to examine the robustness of our corefindings to alternative specifications and to address several issues that could possibly introducebias. Tables 6A (men) and 6B (women) present results for six additional specifications of thefully augmented models reported in Table 4.

First, we include interviewer fixed effects to account for possible heterogeneity acrossinterviewers (column A). The analysis sample included surveys conducted by 431 interviewers.Given the large number of interviewers, we constructed one indicator variable for all inter-viewers who conducted five or fewer interviews (approximately 17 per cent of the sample) andperson-specific indicators for each interviewer who conducted more than five interviews.

We then control for several additional variables that could affect earnings and also becorrelated with the personal traits (column B). We include the O*NET score to control forpotential occupational crowding,21 job skill level to control for productivity differences withinoccupation, the Peabody Picture Vocabulary Test (PVT) to account for cognitive skills, threeinterviewer dummy variables (whether the interviewer and the respondent were of the samegender, whether they were of the same race, and whether the interviewer attended college) tocontrol for interviewer heterogeneity, and body mass index (BMI) to further control forphysical characteristics (Morris, 2006).

The model in column C expands the sample to include workers who are also full-time orpart-time students. Next, although we use robust regression to systematically down-weightoutliers, previous studies of wages have excluded outliers prior to estimation (e.g. Hamermeshand Biddle, 1994). In column D, we re-estimate the main specifications using OLS regressionand excluding observations with hourly wages below $2.13 (the federal minimum for tipped



Tab

le5A

.E

ffec

tsof

pers

onal

trai

tson

natu

rall

ogar

ithm

ofho

urly

wag

eat

curr

ent

job,

bym

edia

nO

*NE

Tsc

ore

(men

)

Mod

els

Bel

owth

em

edia

n(N

=1,

571)

Abo

veth

em

edia

n(N

=1,

565)

AB

CD

EF

Ver

yph

ysic

ally

attr

acti

ve0.

123*

**0.

135*

**0.

114*

**0.

100*

**0.

091*

*0.

068*

(0.0

32)

(0.0

36)

(0.0

37)

(0.0

33)

(0.0

37)

(0.0

39)

Phy

sica

llyat

trac

tive

0.03

6**

0.04

1**

0.03

30.

061*

**0.

049*

*0.

031

(0.0

18)

(0.0

20)

(0.0

20)

(0.0

19)

(0.0

21)

(0.0

22)

Les

sth

anav

erag

eph

ysic

alat

trac

tive

ness

-0.0

18-0

.037

-0.0

160.

005

0.00

8-0

.009

(0.0

35)

(0.0

37)

(0.0

38)

(0.0

38)

(0.0

41)

(0.0

42)

Ver

yat

trac

tive

pers

onal

ity

-0.0

17-0

.027

0.01

50.

001

(0.0

32)

(0.0

32)

(0.0

32)

(0.0

33)

Att

ract

ive

pers

onal

ity

-0.0

03-0

.008

0.03

10.

021

(0.0

20)

(0.0

20)

(0.0

21)

(0.0

22)

Les

sth

anav

erag

epe

rson

alit

y0.

057

0.06

3-0

.005

-0.0

13(0

.040

)(0

.040

)(0

.042

)(0

.043

)V

ery

wel

lgro

omed

0.06

20.

052

(0.0

41)

(0.0

37)

Wel

lgro

omed

0.01

50.

066*

**(0

.021

)(0

.022

)L

ess

than

aver

age

groo

min

g-0

.074

**0.

067

(0.0

35)

(0.0

43)

Job

skill

leve

l0.

183*

**0.

183*

**0.

181*

**0.

085*

**0.

085*

**0.

084*

**(0

.014

)(0

.014

)(0

.014

)(0

.010

)(0

.010

)(0

.010

)R

20.

326

0.32

70.

329

0.23

00.

231

0.23

6T

est

stat

isti

cfo

rw

heth

erth

epe

rson

altr

aits

coef

ficie

nts

inM

odel

Car

eeq

ualt

oth

ose

inM

odel

F(p

valu

e)1.

58(0

.115

)

*p

<0.

10,*

*p

<0.

05,*

**p

<0.

01.

Coe

ffici

ent

and

stan

dard

erro

rs(i

npa

rent

hese

s)re

port

ed.‘

Ave

rage

’is

the

excl

uded

cate

gory

for

allt

hepe

rson

altr

ait

mea

sure

s.A

llm

odel

sar

ees

tim

ated

usin

gro

bust

regr

essi

onan

dco

ntro

lfor

job

skill

leve

l,ag

e,ag

esq

uare

d,m

arit

alst

atus

,edu

cati

on,s

elf-

repo

rted

heal

thst

atus

,yea

rssi

nce

first

regu

lar

job,

year

sin

curr

ent

job,

race

,H

ispa

nic

ethn

icit

y,w

heth

erE

nglis

his

spok

enat

hom

e,em

ploy

men

tsta

tus

ofre

side

ntia

lfat

her

atW

ave

1,pr

esen

ceof

resi

dent

ialf

athe

rat

Wav

e1,

pres

ence

ofre

side

ntia

lm

othe

rat

Wav

e1,

whe

ther

resi

dent

ialm

othe

ran

dfa

ther

wer

ebo

rnin

the

USA

,and

whe

ther

apa

rent

atte

nded

colle

ge.



Tab

le5B

.E

ffec

tsof

pers

onal

trai

tson

natu

rall

ogar

ithm

ofho

urly

wag

eat

curr

ent

job,

bym

edia

nO

*NE

Tsc

ore

(wom

en)

Mod

els

Bel

owth

em

edia

n(N

=1,

514)

Abo

veth

em

edia

n(N

=1,

424)

AB

CD

EF

Ver

yph

ysic

ally

attr

acti

ve0.

048*

*0.

039

0.03

80.

042

0.00

7-0

.008

(0.0

23)

(0.0

26)

(0.0

29)

(0.0

29)

(0.0

34)

(0.0

36)

Phy

sica

llyat

trac

tive

0.01

80.

010

0.00

90.

041*

0.02

00.

011

(0.0

16)

(0.0

17)

(0.0

18)

(0.0

22)

(0.0

24)

(0.0

24)

Les

sth

anav

erag

eph

ysic

alat

trac

tive

ness

-0.0

13-0

.018

-0.0

23-0

.014

-0.0

110.

007

(0.0

29)

(0.0

31)

(0.0

33)

(0.0

40)

(0.0

44)

(0.0

45)

Ver

yat

trac

tive

pers

onal

ity

0.01

80.

017

0.06

6**

0.05

5*(0

.023

)(0

.024

)(0

.032

)(0

.033

)A

ttra

ctiv

epe

rson

alit

y0.

022

0.02

20.

050*

*0.

042*

(0.0

17)

(0.0

18)

(0.0

24)

(0.0

24)

Les

sth

anav

erag

epe

rson

alit

y0.

027

0.02

6-0

.005

0.00

3(0

.039

)(0

.039

)(0

.052

)(0

.052

)V

ery

wel

lgro

omed

0.00

20.

029

(0.0

29)

(0.0

36)

Wel

lgro

omed

0.00

60.

018

(0.0

17)

(0.0

24)

Les

sth

anav

erag

egr

oom

ing

0.01

7-0

.110

**(0

.038

)(0

.049

)Jo

bsk

illle

vel

0.16

5***

0.16

4***

0.16

4***

0.15

2***

0.15

2***

0.15

1***

(0.0

11)

(0.0

11)

(0.0

11)

(0.0

12)

(0.0

12)

(0.0

12)

N1,

514

1,51

41,

514

1,42

41,

424

1,42

4R

20.

424

0.42

40.

424

0.33

60.

340

0.34

3T

est

stat

isti

cfo

rw

heth

erth

epe

rson

altr

aits

coef

ficie

nts

inM

odel

Car

eeq

ualt

oth

ose

inM

odel

F(p

valu

e)0.

98(0

.453

)

*p

<0.

10,*

*p

<0.

05,*

**p

<0.

01.

Coe

ffici

ent

and

stan

dard

erro

rs(i

npa

rent

hese

s)re

port

ed.‘

Ave

rage

’is

the

excl

uded

cate

gory

for

allt

hepe

rson

altr

ait

mea

sure

s.A

llm

odel

sar

ees

tim

ated

usin

gro

bust

regr

essi

onan

dco

ntro

lfor

job

skill

leve

l,ag

e,ag

esq

uare

d,m

arit

alst

atus

,edu

cati

on,s

elf-

repo

rted

heal

thst

atus

,yea

rssi

nce

first

regu

lar

job,

year

sin

curr

ent

job,

race

,H

ispa

nic

ethn

icit

y,w

heth

erE

nglis

his

spok

enat

hom

e,em

ploy

men

tsta

tus

ofre

side

ntia

lfat

her

atW

ave

1,pr

esen

ceof

resi

dent

ialf

athe

rat

Wav

e1,

pres

ence

ofre

side

ntia

lm

othe

rat

Wav

e1,

whe

ther

resi

dent

ialm

othe

ran

dfa

ther

wer

ebo

rnin

the

USA

,and

whe

ther

apa

rent

atte

nded

colle

ge.



Tab

le6A

.Se

nsit

ivit

yte

sts

ofba

selin

em

odel

inT

able

4(m

en)

Bas

elin

e(T

able

4)A

BC

DE

F1:

Bel

owm

edia

nF

2:A

bove

med

ian

Mod

els

b/se

b/se

b/se

b/se

b/se

b/se

b/se

b/se

Ver

yph

ysic

ally

attr

acti

ve0.

089*

**0.

099*

**0.

084*

**0.

064*

**0.

125*

**0.

128*

**0.

089*

*0.

080*

*(0

.028

)(0

.029

)(0

.029

)(0

.024

)(0

.032

)(0

.032

)(0

.037

)(0

.038

)P

hysi

cally

attr

acti

ve0.

027*

0.02

7*0.

029*

0.02

8**

0.02

40.

021

0.01

50.

026

(0.0

16)

(0.0

16)

(0.0

16)

(0.0

13)

(0.0

18)

(0.0

18)

(0.0

20)

(0.0

21)

Les

sth

anav

erag

eph

ysic

alat

trac

tive

ness

-0.0

19-0

.020

-0.0

05-0

.035

-0.0

21-0

.029

-0.0

170.

027

(0.0

29)

(0.0

30)

(0.0

31)

(0.0

25)

(0.0

33)

(0.0

33)

(0.0

40)

(0.0

39)

Ver

yat

trac

tive

pers

onal

ity

0.00

80.

011

-0.0

090.

001

-0.0

09-0

.009

-0.0

07-0

.013

(0.0

24)

(0.0

25)

(0.0

24)

(0.0

20)

(0.0

27)

(0.0

27)

(0.0

30)

(0.0

34)

Att

ract

ive

pers

onal

ity

0.00

20.

002

0.01

5-0

.003

-0.0

03-0

.003

0.00

90.

014

(0.0

15)

(0.0

16)

(0.0

16)

(0.0

13)

(0.0

17)

(0.0

17)

(0.0

20)

(0.0

21)

Les

sth

anav

erag

epe

rson

alit

y0.

031

0.04

80.

041

0.04

10.

073*

*0.

082*

*0.

062

-0.0

07(0

.030

)(0

.031

)(0

.032

)(0

.026

)(0

.035

)(0

.035

)(0

.040

)(0

.041

)V

ery

wel

lgro

omed

0.05

1*0.

063*

*0.

045

0.04

2*0.

057*

0.05

5*0.

005

0.09

0**

(0.0

28)

(0.0

29)

(0.0

29)

(0.0

24)

(0.0

32)

(0.0

32)

(0.0

37)

(0.0

38)

Wel

lgro

omed

0.04

3***

0.04

8***

0.03

8**

0.03

2**

0.05

6***

0.05

6***

0.02

10.

065*

**(0

.016

)(0

.016

)(0

.016

)(0

.013

)(0

.018

)(0

.018

)(0

.021

)(0

.021

)L

ess

than

aver

age

groo

min

g-0

.023

-0.0

38-0

.017

-0.0

35-0

.016

-0.0

16-0

.037

0.00

3(0

.028

)(0

.029

)(0

.030

)(0

.025

)(0

.032

)(0

.032

)(0

.037

)(0

.039

)A

vera

geO

*NE

Tsc

ore

-0.0

09**

*(0

.001

)Jo

bsk

illle

vel

0.15

2***

0.07

9***

0.09

5***

(0.0

09)

(0.0

14)

(0.0

09)

PV

Tsc

ore

0.00

1(0

.000

)B

MI

0.00

0(0

.001

)G

ende

rm

atch

0.01

5(0

.016

)R

ace

mat

ch-0

.000

(0.0

16)

Inte

rvie

wer

atte

nded

colle

ge-0

.001

(0.0

13)

N3,

136

3,13

62,

708

4,35

03,

086

3,08

11,

617

1,51

9R

20.

206

0.29

30.

272

0.19

00.

153

0.29

00.

248

*p

<0.

10,*

*p

<0.

05,*

**p

<0.

01.

Coe

ffici

enta

ndst

anda

rder

rors

(in

pare

nthe

ses)

repo

rted

.‘A

vera

ge’i

sth

eex

clud

edca

tego

ryfo

ral

lthe

pers

onal

trai

tmea

sure

s.A

llm

odel

sex

cept

Dar

ees

tim

ated

usin

gro

bust

regr

essi

onan

dco

ntro

lfor

age,

age

squa

red,

mar

ital

stat

us,e

duca

tion

,sel

f-re

port

edhe

alth

stat

us,y

ears

sinc

efir

stre

gula

rjo

b,ye

ars

incu

rren

tjob

,rac

e,H

ispa

nic

ethn

icit

y,w

heth

erE

nglis

his

spok

enat

hom

e,em

ploy

men

tsta

tus

ofre

side

ntia

lfat

her

atW

ave

1,pr

esen

ceof

resi

dent

ial

fath

erat

Wav

e1,

pres

ence

ofre

side

ntia

lmot

her

atW

ave

1,w

heth

erre

side

ntia

lmot

her

and

fath

erw

ere

born

inth

eU

SA,a

ndw

heth

era

pare

ntat

tend

edco

llege

.Mod

el‘A

’inc

lude

sin

terv

iew

erfix

edef

fect

s.M

odel

‘B’c

ontr

ols

for

allt

heva

riab

les

inth

eba

selin

em

odel

inad

diti

onto

aver

age

O*N

ET

scor

e,jo

bsk

illle

vel,

PV

Tsc

ore,

BM

I,an

din

terv

iew

erch

arac

teri

stic

s.M

odel

‘C’i

nclu

des

full-

and

part

-tim

est

uden

ts.M

odel

‘D’u

ses

OL

San

dex

clud

esre

spon

dent

sea

rnin

gle

ssth

an$2

.13

per

hour

(min

imum

wag

efo

rti

pped

empl

oyee

sin

2001

/200

2)an

dm

ore

than

$200

per

hour

.Mod

el‘E

’is

atw

o-st

epH

eckm

anse

lect

ion

mod

elus

ing

one

iden

tify

ing

vari

able

(num

ber

ofch

ildre

nun

der

age

12in

the

hous

ehol

d)to

acco

unt

for

poss

ible

sele

ctio

nbi

as.T

heto

taln

umbe

rof

obse

rvat

ions

inth

efir

stst

age

is4,

506.

The

coef

ficie

nt(s

tand

ard

erro

r)fo

rnu

mbe

rof

child

ren

is0.

095*

(0.0

58).

The

coef

ficie

nton

the

inve

rse

mill

sra

tio

term

is-0

.024

(0.0

59).

Mod

els

‘F1’

and

‘F2’

use

aver

age

impo

rtan

cesc

ore

for

two

O*N

ET

mea

sure

s(p

erfo

rmin

gfo

ror

wor

king

dire

ctly

wit

hth

epu

blic

and

selli

ngor

influ

enci

ngot

hers

)be

low

the

med

ian

(F1)

and

abov

eth

em

edia

n(F

2).



Tab

le6B

.Se

nsit

ivit

yte

sts

ofba

selin

em

odel

inT

able

4(w

omen

)

Bas

elin

e(T

able

4)A

BC

DE

F1:

Bel

owm

edia

nF

2:A

bove

med

ian

Mod

els

b/se

b/se

b/se

b/se

b/se

b/se

b/se

b/se

Ver

yph

ysic

ally

attr

acti

ve0.

023

0.02

50.

017

0.03

1*-0

.000

2-0

.008

0.03

60.

011

(0.0

23)

(0.0

23)

(0.0

24)

(0.0

18)

(0.0

27)

(0.0

28)

(0.0

30)

(0.0

34)

Phy

sica

llyat

trac

tive

0.01

30.

014

0.00

50.

017

0.01

30.

014

0.02

50.

016

(0.0

15)

(0.0

15)

(0.0

16)

(0.0

12)

(0.0

18)

(0.0

18)

(0.0

19)

(0.0

22)

Les

sth

anav

erag

eph

ysic

alat

trac

tive

ness

-0.0

15-0

.017

-0.0

020.

006

-0.0

15-0

.018

-0.0

190.

015

(0.0

28)

(0.0

27)

(0.0

28)

(0.0

23)

(0.0

33)

(0.0

33)

(0.0

36)

(0.0

40)

Ver

yat

trac

tive

pers

onal

ity

0.04

2**

0.07

1***

0.03

5*0.

030*

0.02

80.

021

0.01

80.

058*

(0.0

20)

(0.0

20)

(0.0

20)

(0.0

16)

(0.0

24)

(0.0

24)

(0.0

25)

(0.0

30)

Att

ract

ive

pers

onal

ity

0.03

2**

0.04

9***

0.03

6**

0.00

70.

036*

*0.

033*

0.02

80.

035

(0.0

15)

(0.0

15)

(0.0

15)

(0.0

12)

(0.0

18)

(0.0

18)

(0.0

19)

(0.0

22)

Les

sth

anav

erag

epe

rson

alit

y0.

018

0.05

30.

030

0.02

9-0

.002

-0.0

020.

031

0.04

5(0

.033

)(0

.032

)(0

.033

)(0

.026

)(0

.038

)(0

.039

)(0

.043

)(0

.046

)V

ery

wel

lgro

omed

0.02

70.

032

0.01

60.

022

0.05

0*0.

052*

-0.0

190.

036

(0.0

23)

(0.0

24)

(0.0

23)

(0.0

18)

(0.0

27)

(0.0

28)

(0.0

29)

(0.0

35)

Wel

lgro

omed

0.03

4**

0.01

80.

024

0.03

0**

0.05

1***

0.05

0***

0.01

40.

022

(0.0

15)

(0.0

15)

(0.0

15)

(0.0

12)

(0.0

17)

(0.0

17)

(0.0

18)

(0.0

21)

Les

sth

anav

erag

egr

oom

ing

-0.0

38-0

.061

**-0

.016

-0.0

33-0

.040

-0.0

380.

018

-0.1

22**

*(0

.031

)(0

.031

)(0

.032

)(0

.027

)(0

.037

)(0

.037

)(0

.041

)(0

.044

)A

vera

geO

*NE

Tsc

ore

-0.0

07**

*(0

.001

)Jo

bsk

illle

vel

0.14

2***

0.06

6***

0.14

2***

(0.0

08)

(0.0

11)

(0.0

09)

PV

Tsc

ore

0.00

1**

(0.0

00)

BM

I-0

.004

***

(0.0

01)

Gen

der

mat

ch-0

.002

(0.0

17)

Rac

em

atch

0.00

2(0

.015

)In

terv

iew

erat

tend

edco

llege

0.00

8(0

.012

)

N2,

938

2,93

82,

550

4,55

32,

871

2,86

71,

543

1,39

4R

20.

292

0.61

50.

385

0.25

90.

235

0.32

90.

394

*p

<0.

10,*

*p

<0.

05,*

**p

<0.

01.

Coe

ffici

enta

ndst

anda

rder

rors

(in

pare

nthe

ses)

repo

rted

.‘A

vera

ge’i

sth

eex

clud

edca

tego

ryfo

ral

lthe

pers

onal

trai

tmea

sure

s.A

llm

odel

sex

cept

Dar

ees

tim

ated

usin

gro

bust

regr

essi

onan

dco

ntro

lfor

age,

age

squa

red,

mar

ital

stat

us,e

duca

tion

,sel

f-re

port

edhe

alth

stat

us,y

ears

sinc

efir

stre

gula

rjo

b,ye

ars

incu

rren

tjob

,rac

e,H

ispa

nic

ethn

icit

y,w

heth

erE

nglis

his

spok

enat

hom

e,em

ploy

men

tsta

tus

ofre

side

ntia

lfat

her

atW

ave

1,pr

esen

ceof

resi

dent

ial

fath

erat

Wav

e1,

pres

ence

ofre

side

ntia

lmot

her

atW

ave

1,w

heth

erre

side

ntia

lmot

her

and

fath

erw

ere

born

inth

eU

SA,a

ndw

heth

era

pare

ntat

tend

edco

llege

.Mod

el‘A

’inc

lude

sin

terv

iew

erfix

edef

fect

s.M

odel

‘B’c

ontr

ols

for

allt

heva

riab

les

inth

eba

selin

em

odel

inad

diti

onto

aver

age

O*N

ET

scor

e,jo

bsk

illle

vel,

PV

Tsc

ore,

BM

I,an

din

terv

iew

erch

arac

teri

stic

s.M

odel

‘C’i

nclu

des

full-

and

part

-tim

est

uden

ts.M

odel

‘D’u

ses

OL

San

dex

clud

esre

spon

dent

sea

rnin

gle

ssth

an$2

.13

per

hour

(min

imum

wag

efo

rti

pped

empl

oyee

sin

2001

/200

2)an

dm

ore

than

$200

per

hour

.Mod

el‘E

’is

atw

o-st

epH

eckm

anse

lect

ion

mod

elus

ing

one

iden

tify

ing

vari

able

(num

ber

ofch

ildre

nun

der

age

12in

the

hous

ehol

d)to

acco

unt

for

poss

ible

sele

ctio

nbi

as.T

heto

taln

umbe

rof

obse

rvat

ions

inth

efir

stst

age

is4,

629.

The

coef

ficie

nt(s

tand

ard

erro

r)fo

rnu

mbe

rof

child

ren

is-0

.175

***

(0.0

56).

The

coef

ficie

ntof

the

inve

rse

mill

sra

tio

term

is-0

.181

***

(0.0

51).

Mod

els

‘F1’

and

‘F2’

use

aver

age

impo

rtan

cesc

ore

for

two

O*N

ET

mea

sure

s(p

erfo

rmin

gfo

ror

wor

king

dire

ctly

wit

hth

epu

blic

and

selli

ngor

influ

enci

ngot

hers

)be

low

the

med

ian

(F1)

and

abov

eth

em

edia

n(F

2).



employees in 2001/2002) and hourly wages above $200.22 Column E reports the results using atwo-stage Heckman procedure to account for possible sample selection bias due to restrictingthe estimation sample to persons with observed wages.23 Although not strictly necessary due tonon-linearities of the selectivity correction term, we include one identifying variable in theselection equation that is excluded from the wage equation: the number of children under age12 in the household. (Results for both stages of the Heckman model are available upon requestfrom the authors.) As in column D, persons with outlier wages are excluded from the estimatedwage equation. Finally, column F reports the results of an alternative attempt to determine thepossible existence of employer discrimination, in which the average O*NET scores are calcu-lated using only two of the 12 job elements in Table 5 — ‘performing or working directly withthe public’ and ‘selling or influencing others’. We believe these two elements best describeoccupations where the personal traits would be especially productivity enhancing.

For men (Table 6A), the results for physical attractiveness, personality attractiveness, andgrooming are virtually unaltered after accounting for interviewer fixed effects (column A). Thecoefficients in the fixed-effects model are slightly larger in magnitude than the coefficients inthe baseline model and the coefficient for being very well groomed is more precisely estimated.Column B presents results for the model with additional control variables. The averageO*NET score has a statistically significant negative coefficient, which is inconsistent with thecrowding hypothesis and simply indicates that wages tend to be lower in occupations wherethe personal traits are more important.24 As expected, the coefficient for the PVT score ispositive and statistically significant. Systematic interviewer bias does not appear to be a majorconcern, as none of the interviewer characteristics is statistically significant.

The results for men are also fairly insensitive to the addition of about 1,200 full- andpart-time students to the analysis sample (column C). When we manually exclude outliersfrom the analysis sample and use OLS instead of robust regression (column D), the wagepremium for being very physically attractive increases from 0.089 to 0.130, and an unexpectedpremium now emerges for having a less than average personality. This may simply be astatistical artifact, but it could also suggest that men with unappealing personalities actuallyreceive higher wages, a finding supported in work by Mueller and Plug (2006).

When accounting for possible sample selection bias (column E), results are very similar tothose in column D, where sample selection is ignored. The identifying variable, number ofchildren under age 12, is positive and marginally significant (p < 0.10), indicating that menwith children are more likely to be employed. The coefficient associated with the selectivitycorrection term is not statistically significant for men, however, indicating the absence ofselection bias.

The results in columns F1 and F2 for men are similar to those in Table 5A in that theysuggest the presence of employer discrimination in favor of physically attractive individuals.Employer discrimination with respect to grooming is not supported by the data, as significantpremiums for being very well groomed or well groomed are only found in occupations wherethe personal traits are likely to be productivity enhancing.

Table 6B reports the same alternative specifications for women. Results with interviewerfixed effects (column A) are similar to those from the baseline model. The coefficients forpersonality attractiveness are somewhat greater in magnitude and more significant than in thebaseline model, being well groomed is no longer significant, and less than average groomingis now significant at the 5 per cent level. In column B, few changes occur in the effects of thepersonal trait measures with the exception of being well groomed, which is no longer statis-tically significant. When working students are added to the sample (column C), the resultsremain qualitatively similar to those from the baseline model, although a small yet statistically



significant beauty premium emerges for the very physically attractive and the effect of havingan attractive personality is no longer statistically significant.

When we manually drop outlier wages instead of performing robust regression (column D),the positive effects for having an attractive personality and being well groomed and very wellgroomed get a bit larger, with the latter becoming statistically significant at the 10 per centlevel. The wage premium for having a very attractive personality decreases in magnitude andbecomes statistically non-significant. However, none of the changes in the coefficients is verylarge and none exceeds their standard error.

As is the case for men, the estimates in column E (sample selection correction) are verysimilar to those in column D (no sample selection correction). Unlike the results for men,however, the selectivity correction term is statistically significant (p < 0.01). Furthermore, theidentifying variable (the number of children under age 12) is negative and statistically signifi-cant in the women’s selection equation, indicating that women with children are less likely tobe employed in our sample.

The results for women when using O*NET scores for only two of the job elements (columnsF1 and F2) are similar to those presented in Table 5B with all 12 elements. Although physicalattractiveness is not a significant determinant of wages for young women, the wage premiumfor women with very attractive personalities and the wage penalty for less than averagegrooming appear to be concentrated in occupations where attractive personalities and goodgrooming are likely to be productivity enhancing.

The final model we considered (not reported in the tables but available on request from theauthors) is one in which we construct a personal traits ‘index’ equal to the cumulative score forthe three personal trait measures. Using 1 for the lowest category and 5 for the highestcategory, the constructed personal traits index ranges from 3 to 15. The intent here is to assesswhether beauty, personality, and grooming collectively exert a cumulative effect on wages. Asexpected, based on the earlier results, the personal traits index has a positive and highlystatistically significant effect on hourly wages for both men and women. The coefficient of theindex variable is 0.02 for both men and women, with a standard error of 0.003.25 This resultsuggests that, holding other individual characteristics constant, workers with relatively lowbeauty ratings may possess other desirable personal traits (e.g. good grooming and/or apleasant personality) that compensate for this ‘shortcoming’ and enable the workers to avoida possible wage penalty.

6. Summary and conclusions

This paper examined the influence of three non-cognitive personal traits — beauty, person-ality attractiveness, and grooming — on labor market earnings. The analysis represents anextension of previous studies that focus mainly on beauty. It is important to highlight that oursample is younger than the samples examined in most of the previous studies and, hence, ourresults may not be directly comparable with them. Nonetheless, one aspect of our findings isconsistent with the earlier studies in that a statistically significant beauty premium is presentwhen beauty is the only personal trait included in the model. Contrary to much of theliterature, we do not find evidence of a plainness penalty for either gender.

The key outcome of our study is that personality attractiveness and grooming also affectwages, and their inclusion in a model of wage determination reduces the effects of beauty. Inthe case of young men, the beauty premium is reduced by between 25 and 40 per cent whereassignificant wage premiums emerge for being well groomed and for being very well groomed.For young women, the beauty premium loses statistical significance (although it is still



positive) when personality attractiveness and grooming are added to the model. Being wellgroomed and having a pleasant personality generate wage premiums for women. In certainspecifications, we also find evidence of a wage penalty for poor grooming among both genders.We conclude therefore that part of the beauty premium estimated in previous studies may bereflecting better grooming for men and women, and a pleasant personality for women.

It appears that the beauty premium for men arises from a combination of higher produc-tivity (possibly due to customer discrimination) and employer discrimination. In furtheranalyses, we find that the beauty premium is statistically significant and larger in occupationswhere the personal traits are less likely to be productivity enhancing than in occupations wherethe personal traits are more likely to be productivity enhancing, suggesting some employerdiscrimination in setting wages. A similar result was found in Hamermesh and Biddle (1994)using different data. On the other hand, we do not find evidence of employer discriminationon the basis of good grooming even though a statistically significant penalty is present forpoor grooming in occupations where grooming is not likely to be productivity enhancing. Forwomen, there is little evidence of employer discrimination on the basis of grooming andpersonality. The premium for an attractive personality and the penalty for poor groomingemerge in those occupations where these traits are expected to influence productivity.

Numerous sensitivity tests are performed, and the results generally remain stable withchanges in model specification, correction for sample selection, and adjustment in samplecomposition. Overall, the results of this paper largely confirm earlier theoretical and empiricalstudies on the importance of physical attractiveness in wage determination. However, unlikeprevious studies, our results suggest that other non-cognitive personal traits are also impor-tant determinants of labor market success. An appropriate agenda for future research there-fore is to further examine (e.g. using different data sets and measures) the relative roles playedby various cognitive and non-cognitive traits in the labor market.

Appendix A: Personal traits questions

Q1. How physically attractive is the respondent?1 Very unattractive2 Unattractive3 About average4 Attractive5 Very attractive

Q2. How attractive is the respondent’s personality?1 Very unattractive2 Unattractive3 About average4 Attractive5 Very attractive

Q3. How well groomed was the respondent?1 Very poorly groomed2 Poorly groomed3 About average4 Well groomed5 Very well groomed

Notes: These questions were part of the interviewer remarks. The interviewer wasasked to describe the respondent, the neighborhood, and the circumstancesand surroundings of the interview as part of a separate section that couldonly be accessed by the interviewer using a password. Respondents wereunable to review the interviewers’ questions or responses, and these ques-tions were to be completed as soon as possible after leaving the respondent.



Appendix B: Full regression results for Table 4

Men Women

Personal traits measuresVery physically attractive 0.122*** 0.111*** 0.089*** 0.065*** 0.038* 0.023

(0.024) (0.027) (0.028) (0.019) (0.022) (0.023)Physically attractive 0.044*** 0.040*** 0.027* 0.041*** 0.024* 0.013

(0.014) (0.015) (0.016) (0.014) (0.014) (0.015)Less than average physical attractiveness -0.018 -0.026 -0.019 -0.025 -0.026 -0.015

(0.027) (0.028) (0.029) (0.025) (0.027) (0.028)Very attractive personality 0.022 0.008 0.051** 0.042**

(0.023) (0.024) (0.020) (0.020)Attractive personality 0.011 0.002 0.041*** 0.032**

(0.015) (0.015) (0.015) (0.015)Less than average personality 0.029 0.031 0.020 0.018

(0.030) (0.030) (0.033) (0.033)Very well groomed 0.051* 0.027

(0.028) (0.023)Well groomed 0.043*** 0.034**

(0.016) (0.015)Less than average grooming -0.023 -0.038

(0.028) (0.031)

Control variablesHispanic 0.030 0.029 0.026 0.068*** 0.065*** 0.067***

(0.021) (0.022) (0.022) (0.021) (0.021) (0.021)Missing Hispanic -0.362** -0.363** -0.372*** -0.137 -0.137 -0.137

(0.142) (0.142) (0.142) (0.123) (0.123) (0.123)Black -0.072*** -0.072*** -0.073*** -0.046*** -0.046*** -0.048***

(0.018) (0.018) (0.018) (0.016) (0.016) (0.016)Missing Black 0.083 0.092 0.094 -0.055 -0.032 -0.033

(0.221) (0.221) (0.221) (0.397) (0.396) (0.396)American Indian -0.031 -0.029 -0.035 0.041 0.041 0.040

(0.033) (0.033) (0.033) (0.034) (0.034) (0.034)Missing American Indian 0.156 0.151 0.137 -0.004 -0.022 -0.003

(0.180) (0.180) (0.180) (0.323) (0.323) (0.323)Asian 0.040 0.039 0.037 0.102*** 0.098*** 0.098***

(0.030) (0.030) (0.030) (0.029) (0.029) (0.029)Missing Asian -0.274* -0.275* -0.260 -0.080 -0.081 -0.086

(0.160) (0.161) (0.161) (0.187) (0.187) (0.187)English spoken at home 0.057** 0.055* 0.054* -0.016 -0.016 -0.013

(0.029) (0.029) (0.029) (0.028) (0.028) (0.028)Age 0.318*** 0.317*** 0.317*** 0.263*** 0.259*** 0.264***

(0.083) (0.084) (0.083) (0.087) (0.087) (0.087)Age squared -0.007*** -0.007*** -0.007*** -0.005*** -0.005*** -0.005***

(0.002) (0.002) (0.002) (0.002) (0.002) (0.002)Currently married 0.065*** 0.065*** 0.064*** -0.003 -0.005 -0.006

(0.016) (0.016) (0.016) (0.014) (0.014) (0.014)High school degree 0.070*** 0.070*** 0.068*** 0.130*** 0.131*** 0.124***

(0.020) (0.020) (0.020) (0.022) (0.022) (0.022)Some college 0.125*** 0.124*** 0.119*** 0.212*** 0.209*** 0.202***

(0.022) (0.022) (0.022) (0.024) (0.024) (0.024)College degree 0.350*** 0.348*** 0.335*** 0.448*** 0.444*** 0.436***

(0.028) (0.028) (0.028) (0.026) (0.026) (0.027)Excellent self-reported health 0.104*** 0.105*** 0.100*** 0.061** 0.056* 0.055*

(0.035) (0.035) (0.035) (0.030) (0.030) (0.030)Very good self-reported health 0.068* 0.070** 0.067* 0.045 0.040 0.040

(0.035) (0.035) (0.035) (0.029) (0.029) (0.029)Good self-reported health 0.053 0.055 0.052 0.021 0.015 0.016

(0.036) (0.036) (0.036) (0.030) (0.030) (0.030)Years since first regular job 0.019*** 0.019*** 0.018*** 0.014*** 0.013*** 0.013***

(0.003) (0.003) (0.003) (0.003) (0.003) (0.003)Years in current job 0.031*** 0.031*** 0.030*** 0.031*** 0.031*** 0.031***

(0.004) (0.004) (0.004) (0.004) (0.004) (0.004)Had resident father at Wave 1 0.058** 0.059** 0.058** 0.029 0.029 0.029

(0.028) (0.028) (0.028) (0.028) (0.028) (0.028)Professional resident father at Wave 1 -0.015 -0.015 -0.017 -0.001 -0.001 -0.001

(0.029) (0.029) (0.029) (0.026) (0.026) (0.026)Missing professional resident father at Wave 1 -0.031 -0.030 -0.031 -0.101 -0.111 -0.111

(0.095) (0.096) (0.095) (0.122) (0.122) (0.122)White collar resident father at Wave 1 0.003 0.003 0.002 0.038** 0.040** 0.039**

(0.018) (0.018) (0.018) (0.019) (0.019) (0.019)



Appendix B: Continued

Men Women

Missing white collar resident father at Wave 1 0.000 0.000 0.000 0.000 0.000 0.000(0.000) (0.000) (0.000) (0.000) (0.000) (0.000)

Had resident mother at Wave 1 0.003 0.003 0.004 0.014 0.016 0.015(0.034) (0.034) (0.034) (0.034) (0.034) (0.034)

Resident mother or father at Wave 1attended college

0.013 0.013 0.012 0.041*** 0.039*** 0.037***(0.014) (0.014) (0.014) (0.014) (0.014) (0.014)

Missing resident mother or father atWave 1’s education

-0.027 -0.027 -0.028 -0.002 -0.000 0.001(0.040) (0.040) (0.040) (0.048) (0.048) (0.048)

Resident mother at Wave 1 born in USA -0.041 -0.040 -0.042 -0.010 -0.012 -0.012(0.027) (0.027) (0.027) (0.027) (0.027) (0.027)

Missing resident mother at Wave 1 bornin the USA

-0.033 -0.030 -0.029 0.052 0.066 0.066(0.168) (0.168) (0.168) (0.138) (0.138) (0.138)

Resident father at Wave 1 born in the USA -0.013 -0.012 -0.010 -0.016 -0.017 -0.017(0.028) (0.028) (0.028) (0.028) (0.028) (0.028)

Missing resident father at Wave 1 bornin the USA

-0.075 -0.072 -0.072 0.019 0.037 0.026(0.134) (0.134) (0.134) (0.268) (0.268) (0.268)

Constant -1.769* -1.764* -1.749* -1.353 -1.312 -1.363(0.928) (0.928) (0.927) (0.955) (0.954) (0.955)

N 3,136 3,136 3,136 2,938 2,938 2,938R2 0.204 0.204 0.206 0.288 0.290 0.292

* p < 0.10, ** p < 0.05, *** p < 0.01. Coefficient and standard errors (in parentheses) reported. ‘Average’ is the excluded category for all the personal traitmeasures. All specifications were estimated using robust regression.

Appendix C: Derivation of O*NET scores

The Occupational Information Network (O*NET) is a recent occupational network data-base that contains information on hundreds of occupations based on surveys with workersfrom each occupation. The development of O*NET is funded by a grant from the USDepartment of Labor/Employment and Training to the North Carolina Employment SecurityCommission.

O*NET Online (http://www.onetcenter.org) allows users to search for specific occupations.In addition, occupations are classified by O*NET descriptors, categories of occupationalinformation collected for O*NET — Standard Occupational Classification (SOC) occupa-tions. The O*NET database includes six descriptors (knowledge, skills, abilities, work activi-ties, interests, and work values), and each contains more specific elements. Occupationsreceive an importance score ranging from 0 to 100 for each element; higher scores indicate thatan element is more important in an occupation. We selected the following 12 elements:

• Work Activities — Interacting with Others — Assisting and Caring for Others• Work Activities — Interacting with Others — Communicating with Persons Outside

Organization• Work Activities — Interacting with Others — Establishing and Maintaining Interpersonal

Relationships• Work Activities — Interacting with Others — Performing for or Working Directly with the

Public• Work Activities — Interacting with Others — Selling or Influencing Others• Skills — Social Skills — Negotiation• Skills — Social Skills — Persuasion• Skills — Social Skills — Service Orientation• Skills — Social Skills — Social Perceptiveness• Skills — Basic Skills — Speaking



• Knowledge — Sales and Marketing• Knowledge — Customer and Personal Service

For this analysis, the O*NET data needed to be merged into the Add Health data set. One ofthe questions in Add Health asked respondents to select what they do in their job from fourlists of occupation codes based on the 1998 Standard Occupational Classification. Level Icontained a list with the broadest categories, and subsequent levels contained more specificdescriptions. We used Level II occupation codes, and replaced missing Level II codes withLevel I codes for 43 observations. We then searched O*NET for the Add Health Level IIcategories and recorded the O*NET importance scores for each element. Many of the Level IIAdd Health occupation descriptions were broader than the categories in O*NET, as theO*NET categories contain information about a set of occupations. To adjust for this, wecollected data in O*NET for each occupation within the group and averaged the scorestogether to calculate a category score that corresponded with Add Health’s occupation codes.For example, Add Health’s occupation codes included a category for ‘Advertising, Marketing,Promotions, Public Relations, and Sales Managers’. To estimate the O*NET scores for thiscategory, we took the average of the importance scores of four different O*NET occupations:(1) Advertising and Promotions Managers; (2) Marketing Managers; (3) Sales Managers; and(4) Public Relations Managers. The variable Average O*NET Score in the analysis is theaverage for each occupation category of the importance scores for the 12 elements.

O*NET also classifies each occupation into one of five zones: job zone 1 requires little orno preparation, job zone 2 requires some preparation, job zone 3 requires medium prepa-ration, job zone 4 requires considerable preparation, and job zone 5 requires extensivepreparation. The job zone score, ranging from 1 to 5, is assigned to each occupation inO*NET. An approach similar to what was described above was used to match occupationsfrom Add Health to the O*NET database and construct the variable job skill level based onjob zone.

Notes

1 An exception is Das and De Loach (2010), who analyse the effects of grooming on earnings.However, they lack measures of beauty and personality attractiveness.

2 This could occur because physically attractive individuals might feel that their appearance will leadto future rewards in the labor market.

3 Hereafter, we use the term ‘personal traits’ to refer collectively to the three non-cognitive traits weconsider: physical attractiveness (i.e. beauty), personality attractiveness, and grooming.

4 O*NET was developed for the US Department of Labor by the National Center for O*NETDevelopment (see http://www.onetcenter.org/overview.html).

5 The age range for 93.5 per cent of the sample used in the current analysis was 20–25.6 It is not clear that interviewer assessments of these personal traits are more accurate than the

self-assessments used in most previous studies, although one could make a case that interviewer assess-ments have less bias. Later, in a sensitivity analysis, we attempt to determine whether there is anyevidence of systematic biases in the interviewer assessments.

7 Mocan and Tekin (2010) use Add Health data to examine the relationship between beauty andcriminal activity. They also estimate wage equations similar to those in Hamermesh and Biddle (1994)and briefly discuss the results without showing them in a table. However, because beauty and the labormarket was not the central focus of their investigation, they do not include measures of personalityattractiveness or grooming. Other recent studies using Add Health data include Tekin and Markowitz(2008), Carmalt et al. (2008), and Norton and Han (2008).



8 Regular job is defined as a job for at least 10 hours per week for at least 9 weeks. The person mayor may not have been continuously employed since that date, so this variable is not necessarily a measureof work experience.

9 The majority of previous research on physical attractiveness uses ratings based on photographs ofthe subjects’ faces (Ritts et al., 1992), which may be less reliable than using ratings based on face-to-faceinterviewer assessments.

10 Other coding strategies yielded similar inferences (results are available on request from the authors).11 Including all three personal trait measures in a single equation could be problematic if the measures

are highly correlated. None of the correlations exceeds 0.5 and, as expected, the similar rankings for eachtrait are positively correlated. For example, the correlation coefficient for being rated as very physicallyattractive and having a very attractive personality is 0.439, whereas the correlation coefficient for beingvery physically attractive and being very well groomed is 0.418. Although sizeable, these correlations arenot likely to create serious multicollinearity problems. To examine this issue further, we calculatedvariance inflation factors for the core models. All of the variance inflation factors of the personal traitmeasures are less than 2, which is well below the usual threshold for potential multicollinearity problems(Kennedy, 2003).

12 We use the rreg command in Stata.13 The full regression results are presented in Appendix B. The estimated effects of the traditional

human capital variables (i.e. age, education, and work experience) are in accordance with previousliterature, but the coefficients are reduced slightly as additional non-cognitive personal traits are addedto the model.

14 Das and De Loach (2010) also find a positive effect of good grooming for men but they find anegative effect for women. However, as mentioned earlier, they are unable to control for beauty orpersonality attractiveness in their models. Additionally, their measure for grooming (time spent groom-ing) differs conceptually from the measure used here (a ranking from very well groomed to very poorlygroomed).

15 In a sensitivity analysis later in the paper, we examine alternative specifications that account forpossible differences across interviewers, including a model with variables measuring interviewer charac-teristics and a model allowing for interviewer fixed effects.

16 This procedure is only valid if there are no unobservable variables influencing the selection ofworkers into these occupations. If such unobservable variables are important, appropriate statisticalmethods (e.g. instrumental variables) would be necessary to obtain consistent estimates of the effects ofbeauty within the selected occupations.

17 It should be noted that in occupations involving frequent contact with the public, wages may varywith the personal traits because of customer discrimination. However, customer discrimination isactually productivity enhancing if it leads to a higher marginal revenue product of the worker. Althoughit is important to acknowledge this possible scenario, we are unable to determine whether variation inwages in occupations involving significant contact with the public is due to customer discrimination ormore traditional productivity-enhancing factors.

18 The results from the O*NET interaction models are available on request from the authors.19 It turns out that the coefficient estimates for the personal trait variables are insensitive to the

inclusion of the job zone variable.20 We also divided the occupations into thirds and quartiles of the O*NET scores for finer distinctions.

The qualitative results (available on request) were virtually identical to those obtained when we dividedthe occupations above and below the median O*NET scores. Also, in the sensitivity section below, wetest another specification using the O*NET scores.

21 Hamermesh and Biddle (1994) argue that pushing unattractive workers to occupations in whichpersonal traits are not productivity enhancing depresses wages in those occupations relative to wages inoccupations where those traits are more important. This implies that the coefficient on O*NET score willbe positive.

22 The wage questions in the Add Health Survey asked respondent to include tips, but it is likely thatincome from tips was underreported or excluded entirely by some respondents.



23 A standard problem in estimating wage equations is that wages are not observed for non-workers.24 Many of the occupations where personal traits are important involve entry-level positions with

limited skill requirements (e.g. receptionist, cashier). This result may also indicate that we have not fullyaccounted for potential crowding.

25 A quantitative interpretation of the coefficient should be avoided because the numerical rankingsmaking up the index are on an ordinal scale and do not have any meaningful cardinal properties.Nonetheless, the results suggest that workers with a higher index value (i.e. a more desirable combinationof traits) are rewarded with higher wages.

References

Biddle J. E. and Hamermesh D. S. (1998) ‘Beauty, Productivity, and Discrimination: Lawyers’ Looksand Lucre’, Journal of Labor Economics 16(1): 172–201.

Borghans L., Duckworth A. L., Heckman J. J. and Weel B. (2008) ‘The Economics and Psychology ofPersonality Traits’, Journal of Human Resources 43(4): 972–1059.

Bowles S. and Gintis H. (1976) Schooling in Capitalist America: Educational Reform and the Contradic-tions of Economic Life, New York: Basic Books.

Bowles S., Gintis H. and Osborne M. (2001) ‘The Determinants of Earnings: A Behavioral Approach’,Journal of Economic Literature 39(4): 1137–1176.

Carmalt J., Cawley J., Joyner K. and Sobal J. (2008) ‘Body Weight and Matching with a PhysicallyAttractive Romantic Partner’, Journal of Marriage and Family 70(5): 1287–1296.

Crosnoe R. and Muller C. (2004) ‘Body Mass Index, Academic Achievement, and School Context:Examining the Educational Experiences of Adolescents at Risk of Obesity’, Journal of Health andSocial Behavior 45(4): 393–407.

Das J. and De Loach S. B. (2010) ‘Mirror, Mirror, on the Wall: The Effect of Time Spent Grooming onEarnings’, Journal of Socio-Economics doi:10.1016/j.socec.2010.06.005.

Doran K. and Hersch J. (2009) ‘The Beauty Premium Is Not Robust’, Working paper. Available at:http://www.nd.edu/~kdoran/BeautyNotRobust.pdf.

Fletcher J. M. (2009) ‘Beauty versus Brains: Early Labor Market Outcomes of High School Graduates’,Economics Letters 105(3): 321–325.

French M. T. (2002) ‘Physical Appearance and Earnings: Further Evidence’, Applied Economics 34(5):569–572.

French M. T., Robins P. K., Homer J. F. and Tapsell L. M. (2009) ‘Effects of Physical Attractiveness,Personality, and Grooming on High School GPA’, Labour Economics 16(4): 373–382.

Frieze I. H., Olson J. E. and Russell J. (1991) ‘Attractiveness and Income for Men and Women inManagement’, Journal of Applied Social Psychology 21(13): 1039–1057.

Hamermesh D. S. (2006) ‘Changing Looks and Changing “Discrimination:” The Beauty of Economists’,Economics Letters 93(3): 405–412.

Hamermesh D. S. and Biddle J. E. (1994) ‘Beauty and the Labor Market’, American Economic Review84(5): 1174–1194.

Hamermesh D. S. and Parker A. (2005) ‘Beauty in the Classroom: Instructors’ Pulchritude and PutativePedagogical Productivity’, Economics of Education Review 24(4): 369–376.

Hamermesh D. S., Meng X. and Zhang J. (2002) ‘Dress for Success — Does Primping Pay?’, LabourEconomics 9(3): 361–373.

Harper B. (2000) ‘Beauty, Stature, and the Labour Market: A British Cohort Study’, Oxford Bulletin ofEconomics and Statistics 62(Suppl. 1): 771–800.

Kennedy P. (2003) A Guide to Econometrics, 5th edn, Cambridge, MA: The MIT Press.Loh E. S. (1993) ‘The Economic Effects of Physical Appearance’, Social Science Quarterly 74(2):

420–438.Mobius M. M. and Rosenblat T. S. (2006) ‘Why Beauty Matters’, American Economic Review 96(1):

222–235.Mocan H. N. and Tekin E. (2010) ‘Ugly Criminals’, Review of Economics and Statistics 92: 15–30.



Morris S. (2006) ‘Body Mass Index and Occupational Attainment’, Journal of Health Economics 25(2):347–364.

Mueller G. and Plug E. (2006) ‘Estimating the Effect of Personality on Male and Female Earnings’,Industrial and Labor Relations Review 60(1): 3–22.

Norton E. C. and Han E. (2008) ‘Genetic Information, Obesity, and Labor Market Outcomes’, HealthEconomics 17(9): 1089–1104.

Pfann G. A., Biddle J. E., Hamermesh D. S. and Bosman C. M. (2000) ‘Business Success and BusinessesBeauty Capital’, Economics Letters 67(2): 201–207.

Ritts V., Patterson M. L. and Tubbs M. E. (1992) ‘Expectations, Impressions, and Judgments ofPhysically Attractive Students: A Review’, Review of Educational Research 62(4): 413–426.

Süssmuth B. (2006) ‘Beauty in the Classroom: Are German Students Less Blinded? Putative PedagogicalProductivity Due to Professors’ Pulchritude: Peculiar or Pervasive?’, Applied Economics 38(2):231–238.

Tekin E. and Markowitz S. (2008) ‘The Effects of Suicidal Behavior on Productive Activities of YoungAdults’, Southern Economic Journal 75(2): 300–331.



Beauty and the Labor Market: Accounting for the Additional Effects of Personality and Grooming

Documents