Why Beauty Matters An Experimental Investigation

1
Why Beauty MattersAn Experimental Investigation

• Markus Mobius (Harvard University)
• Tanya Rosenblat (Wesleyan University)
• April 2004

2
Is Beauty in the Eye of the Beholder?

• Surprisingly psychologists say No
• Strong agreement on what is considered
  beautiful in facial photograph ratings across
  genders and across cultures
• Therefore beauty can be measured (objectively)!

3
Is Beauty in the Eye of the Employer?

• Extensive research on beauty in social psychology
  and human resource management
• In economics, Hamermesh and Biddle (1994) Beauty
  Premium
• Establish that looks matter even after
  controlling for many observable characteristics
  (actual labor market experience, years of tenure
  in a firm, union status, firm size, race,
  geographic location, fathers' occupation,
  childhood background, immigrant status of
  respondents and their parents and grandparents)

4
Psychology Literature
I. How are beautiful people perceived by others?

• Attractiveness or Beauty-Is-Good Stereotype
  viewed superior along several dimensions
  personality traits (sociability, dominance,
  sexual warmth, modesty, character), mental
  health, intelligence and academic ability, and
  social skills

5
Psychology Literature
II. To what extent is this stereotype true?

• Kernel of Truth Hypothesis
• Attractive people are treated better by others
  throughout their life cycle.
• Physical attractiveness rating does not change
  much throughout life cycle.
• A self-fulfilling prophecy? gt Become more
  confident and more persuasive

6
Experimental Literature

• Physical attractiveness in Experiments
• Ultimatum Game (Solnick and Schweitzer (1999))
• Prisoners Dilemma (Mulford, Orbell, Shatto and
  Stockard (1998), Kahn, Hotes and Davis (1971))
• Public Goods (Andreoni and Petrie (2004))
• Trust Games (Eckel and Wilson (2004))
• Dictator Game

7
How does beauty affect wages?
Decompose the effects of beauty

• Becker-type discrimination (employers have a
  taste for good-looking employees)
• Ability Effect - more physically-attractive have
  superior skills at performing a task
• Stereotype, Confidence and Persuasion Effects
  during wage negotiation process

8
How does beauty affect wages?
Wage Negotiation

• Employer forms a belief about workers ability
• Direct Stereotype Channel raises employer
  belief about worker ability directly (because
  beauty is good)
• Indirect Stereotype Channel raises employer
  belief indirectly during verbal interaction
  through characteristics correlated with beauty

9
How does beauty affect wages?
Wage Negotiation

• Worker forms a belief about his own ability
• Confidence Channel raises worker confidence in
  his ability
• Employer decides on the wage based on his prior
  and workers confidence
• Persuasion Channel raises wage by increasing
  weight on workers confidence

10
(No Transcript)
11
Experimental Design
Job Description

• Employees were hired to perform a skilled task
  of solving Yahoo! mazes for 15 minutes.
• Before interviews they had a chance to solve a
  practice maze of level Easy
• During employment period they solved mazes one
  level of difficulty higher

12
Experimental Design
13
Experimental Design
14
Experimental Design
Why Mazes?

• We would not expect beauty to be directly
  productive for this task. We can therefore focus
  on worker/employer interaction alone
• The task requires true skill. Gneezy, Niederle
  and Rustichini (2003) have shown that there is
  considerable variation in skill and speed of
  learning for performing this task.

15
Experimental Design

• Neither worker nor employer have well defined
  focal points for predicting future performance if
  presented with the practice time.
• There is a significant amount of learning
  possible in performing this task during the
  allocated 15 min time period.
• This allows for overconfidence effects and also
  for true persuasion a confident worker might
  truly believe that she can solve many mazes even
  though she did poorly in the practice round, and
  possibly can convince the employer to believe her.

16
Experimental Design

• Playing the main game at the next level of
  difficulty opens room for additional uncertainty
  and thus further over-confidence and persuasion
  effects.

17
Experimental Design
Each worker enters her resume information

• College major, name of the degree granting
  institution, matriculation year, hobbies, team
  sports, age, gender, dream job, the number of
  jobs previously held, the number of job
  interviews they have participated in, and whether
  they have internet connection at home (income
  proxy)
• Time it took to complete the practice round

18
(No Transcript)
19
Experimental Design
In addition,

• Each worker is asked to form an estimate of how
  many mazes she will be able to solve in 15
  minutes
• This information is only provided for the
  experimenter and is not revealed to the
  employers.
• Compensation is structured in an incentive
  compatible manner to induce workers to truthfully
  reveal their estimates.
• Workers and employers complete a control
  questionnaire to make sure they understand how
  payments are calculated.

20
Experimental Design
Each worker participates in 5 treatments in
random order

• Treatment A Resume only without a facial
  photograph.
• Treatment B Resume and facial photograph.
• Treatment C Resume without a photograph and oral
  telephone communication.
• Treatment D Resume with a facial photograph and
  oral telephone communication.
• Treatment E Resume with a facial photograph and
  face-to-face interview.

21
Only matters in treatments C, D, E
Does interaction of beauty and confidence matter
in Treatments C, D, E?
Treatments C, D, E (especially C - speech only!)
Treatments B, D, E (especially B - picture only!)
22
Experimental Design
Timing

• Workers enter their resume information and
  confidence estimates.
• Workers interact with employers (treatments C, D,
  E) or employers see workers files (treatments A,
  B).
• Employers find out whether their estimates of
  worker productivity will be used to compensate
  employees (80 of the time).

23
Experimental Design

• Employers decide on their estimates of worker
  productivity and submit their choices to the
  experimenter after they have been the audience to
  all 5 candidates.
• Note, that all workers are hired, but get
  different compensation.
• Workers participate in 15 minute work period.
• Total compensation is determined for workers and
  employers.

24
Experimental Design
Why is employer wage used only in 80 of the
cases?

• To distinguish between
• Employers choosing to transfer some money to
  workers independent of their skill and
• Compensation for perceived skill
• Use this to check for direct taste-based
  discrimination.

25
Experimental Design
Compensation of Workers

• Workers get a piece rate of 100 points for each
  maze they solve during the work period.
• Workers get a wage determined by each employer.
  This wage is used 80 of the time. 20 of the
  time the wage is set by the experimenter all
  wages are paid by the experimenter.
• 40 points are subtracted from workers
  compensation for each maze they mispredict (above
  or below their estimate). This provides a
  marginal incentive of 60 points per maze to
  continue solving mazes even after they hit their
  estimate.

26
Experimental Design
Compensation of Employers

• Employers get a fixed fee of 4000 points.
• During the interview and resume review they form
  an estimate of how many mazes each candidate can
  solve. This number times 100 points becomes
  employee wage in 80 of cases.
• Regardless of whether employer wage is used or
  not 40 points are subtracted from employers
  compensation for each maze they mispredict (above
  or below their estimate for each employee).

27
Experimental Design
Beauty Ratings

• By a panel of 50 independent evaluators on a
  scale from 1 to 5
• 1 - homely, far below average in attractiveness
  2 - plain, below average in attractiveness 3 -
  of average beauty 4 - above-average and 5 -
  strikingly handsome or beautiful.
• Standard passport-type photographs were presented
  to evaluators in random order via a website.

28
Subjects

• Undergraduate and masters students from Tucuman
  University, Argentina
• instructions in Spanish delivered orally and via
  a computer
• subjects completed a control questionnaire to
  ensure understanding of compensation schemes
• 33 sessions of 5 workers and 5 employers each
  worker being reviewed by 5 employers (825
  observations)

29
Subjects

• Subjects were paid 12 pesos for participation
  additional earnings described above
• Average earnings 25 pesos for an experiment of up
  to one and a half hours in length.
• Made sure subjects did not know each other prior
  to the experiment.

30
Employee Subject Pool Description

• Subjects from 3 university campuses, 85 from UNT
  
• 56 male
• Average age 22.9 more graduate students
• Majors business and economics (21) science,
  medicine, and information technology (46)
  humanities and arts (33)
• 51 have internet access at home (80 from
  private 41 from public)

31
Employee Subject Pool Description

• 61 participated in team sports
• 43 had no previous work experience (out of them
  63 never interviewed for a job)
• Those with work experience worked in education,
  information technology, retail sales, business,
  public sector, arts, food production and service,
  and industry.
• Intensity of interpersonal interaction on a job
• Hobbies in computers, recreation (listening to
  music, reading), creative tasks (writing,
  drawing, composing music), sports

32
Average Performance

• The mean number of mazes solved was 9.5 (10.9 for
  men 7.8 for women)
• The average maze during 15 minute trial took 94
  sec the average practice time was 127 sec
• Subjects systematically underestimated their own
  productivity by 24 on average.
• Employers underestimated workers productivity in
  a similar manner (20 on average).

33
Variable Transformations
Confidence Measure

• Estimated number of rounds (ln)

Ability Measure

• Actual number of rounds (ln)

Prediction based on extrapolation from the
practice round

• Ln (1560/Practice)

Becker Discrimination

• SETWAGE1 if employer estimate was used to
  determine workers wage

34
Beauty Measure
Detrend beauty ratings to get rid of measurement
error

• Measurement error arises because each rater has a
  distinct definition of baseline beauty
• Formally, for each rater we take her average
  beauty rating and subtract it from each raw
  rating for subject in order to define the
  centered rating
• The measure BEAUTY for subject is then defined as
  the mean over all raters centered rating.

35
Procedure for Data Analysis

• 1. Relationship between beauty and ability
• 2. Relationship between beauty and confidence
• 3. Wage regressions without controlling for
  confidence
• 4. Wage regressions with a control for confidence
• 5. Persuasion Effect
• 6. Pooled Regression

36
Beauty and Ability

• Regression of actual ability during 15 min work
  period measured by LNACTUAL on age, sex, family
  wealth (approximated by INTERNET), and physical
  attractiveness (with and w/o decision variables).
• MALE is significant men are 30 better at
  solving mazes in 15 minutes (can be also seen
  from summary statistics 10.9 vs 7.8)
• Beauty is NOT statistically significant!

37
(No Transcript)
38
Beauty is NOT statistically significant
39
Beauty is NOT statistically significant
Practice time doesnt fully predict actual ability
40
Men have better skills
Beauty is NOT statistically significant
Practice time doesnt fully predict actual ability
41
Beauty and Ability

• Regression of predicted ability extrapolated from
  the practice round PREDICT on age, sex, family
  wealth, and physical attractiveness
• Again men do better in the practice round
• In addition older subjects do better (with
  decreasing returns to age)
• Beauty is not significant
• Both regressions show that there is no
  relationship between beauty and ability!
• Note that we run two specifications with and
  without major and hobby choices.

42
(No Transcript)
43
Beauty is NOT statistically significant
44
Men are more skilled
Beauty is NOT statistically significant
45
Older subjects do better, but with decreasing
returns
Men are more skilled
Beauty is NOT statistically significant
46
Confidence

• Regression of confidence on beauty, true ability
  and worker characteristics.

47
(No Transcript)
48
Actual ability weakly raises confidence (but
coef. is not 1)
49
Rely on practice performance more so than is
justified based on ability regression before
50
Men are not more confident if we control for
actual ability
Rely on practice performance
Actual ability weakly raises confidence (but
coef. is not 1)
51
Raising Beauty by 1 standard deviation raises
confidence by 13
52
Beauty raises confidence equally for men and
women.
53
Confidence

• Regression of confidence on beauty, true ability
  and worker characteristics.
• Actual ability weakly raises confidence. A 1
  increase in actual ability increases confidence
  by about 0.15.
• Note that if confidence were truthful and based
  only on self- knowledge about true abilities then
  we would expect a coefficient close to 1 on
  LNACTUAL and all other variables to be not
  significant.
• The biggest boost of confidence is performance in
  the practice round a 1 increase in predicted
  performance raises confidence by at least .4.

54
Confidence

• Male subjects are not more confident once we
  control for their higher average ability in
  solving mazes.
• Physical attractiveness raises confidence equally
  for men and women since coefficient on
  interaction term beautymale is not significant.
• There is a strongly significant (at the 1 percent
  level) effect of physical attractiveness on
  confidence. Raising beauty by one standard
  deviation increases confidence about 13.
• This effect is very large if we define a
  beautiful person to be one standard deviation
  above the mean and a plain person to be one
  standard deviation below then the plain subject
  is about 26 less confident than the beautiful
  subject.

55
Confidence

• Confidence is by no means a truthful reflection
  of actual ability. The large coefficient on
  PREDICT suggests that subjects have a hard time
  evaluating their own ability and tend to
  over-extrapolate from their practice performance.
  
• Interestingly, physical attractiveness has a very
  large confidence-enhancing effect while gender
  has none. Although men in our sample are more
  confident, they actually perform better at the
  task.

56
Wage Regressions (w/o Confidence Controls)

• Fixed effects regressions of wages on workers
  characteristics including BEAUTY but excluding
  CONFIDENCE. (Separate regression for each
  treatment).
• y is wage of worker j set by employer i
• is employer fixed effect
• B is worker jth beauty
• S is SETWAGE1 if employer determines workers
  wage directly
• X vector of CV characteristics

57
Dep. Var LNWage (w/o CV Controls)
Beauty Premium
58
Dep. Var LNWage (w/ CV Controls)
Beauty Premium
59
Dep. Var LNWage (w/o CV Controls)
Practice Performance Matters a lot!
Beauty Premium
No Evidence for direct taste-based discrimination
60
Dep. Var LNWage (w/ CV Controls)
Practice Performance Matters a lot!
Beauty Premium
Not much evidence for direct taste-based
discrimination
61
Wage Regressions (w/o Confidence Controls)

• Regressions of wages on workers characteristics
  including BEAUTY but excluding CONFIDENCE.
  (Separate regression for each treatment).
• First of all, there is a beauty premium in our
  experiment in all treatments except A ranging
  from 9.4 to 12.7 without CV controls and from 12
  to 17 with CV controls.
• SETWAGEBEAUTY is not significant there is no
  evidence for direct taste-based Becker-type
  discrimination
• 1 increase in practice performance increases
  wages by .4 (from coefficient on PREDICT)
• MALE is significant in treatments C and D only.

62
Wage Regressions (w/ Confidence Controls)

• Fixed effects regressions of wages on workers
  characteristics including BEAUTY and CONFIDENCE.
  (Separate regression for each treatment).
• C is worker js confidence

63
Dep. Var LNWage (w/ CV Controls)
Confidence matters only in treatments with verbal
interaction
64
Dep. Var LNWage (w/ CV Controls)
Beauty Premium declines in those treatments
Confidence matters only in treatments with verbal
interaction
65
Dep. Var LNWage (w/ CV Controls)
As before, actual performance doesnt matter and
practice performance does.
Beauty Premium declines in those treatments
Confidence matters only in treatments with verbal
interaction
66
Wage Regressions (w/ Confidence Controls)

• Same as regressions before but with an additional
  control for confidence.
• As expected, confident subjects do better in
  treatments with verbal interaction.
• A 1 increase in confidence raises wages by about
  0.18 to 0.33.
• The beauty effects in treatments B to E are
  smaller by about 2 to 4. This decline is
  consistent because we know that one standard
  deviation in beauty increases confidence by about
  13.

67
Wage Regression w/ Confidence Controls

• The coefficient on MALE is the same as before
• LNACTUAL is still not significant
• SETWAGELNESTIMATED and SETWAGEBEAUTY are also
  not significant

68
Confidence channel
69
Other Covariates

• One percent increase in practice performance
  raises wages by about .4 percent in treatments A
  and B and .3 percent in treatments C, D, and E gt
  Employers put less emphasis on practice
  performance when they can interact verbally with
  the worker
• Actual Ability is NOT statistically significant
  in any treatment.
• Gender effects in treatments C and D only
• Age effects in treatments D and E only
• Team sports and internet are not significant.

70
Testing for Persuasion Channel

• Fixed effects regressions of wages on workers
  characteristics including BEAUTY, CONFIDENCE, and
  BEAUTYCONFIDENCE. (Separate regression for each
  treatment).

71
Testing for Persuasion Channel

• Coefficient on the interaction term is not
  significant gt no evidence for persuasion channel
  

72
Pooled Regression

• AUDIO 1 if worker and employer can talk to each
  other (treatments C, D, and E)
• VISUAL1 if employer can see workers picture
  (treatments B, D, and E)
• FTF1 if there is face-to-face communication
  (treatment E)
• Interact PREDICT, BEAUTY and LNESTIMATED with the
  dummies above and include CV controls.

73
(No Transcript)
74
Direct Stereotype
75
Direct Stereotype
Indirect Stereotype
76
Direct Stereotype
Indirect Stereotype
Confidence
77
Pooled Regression

• Direct Stereotype Channel identified by
  coefficient on BEAUTYVISUAL (7.2 wage gain for
  each standard deviation in beauty)
• Indirect Stereotype Channel is captured by
  coefficient on BEAUTYAUDIO (10.4 wage gain for
  each standard deviation in beauty)
• Confidence Channel raises wage by .3 for each 1
  increase in confidence. This translates into
  3.6 increase in wage for one standard deviation
  increase in beauty

78
3.6 increase in wage for 1 standard deviation
increase in beauty
No Evidence
10.4 gain for 1 standard deviation increase in
beauty
7.2 gain for 1 standard deviation increase in
beauty
79
Policy Implications

• Job interviews are currently the most common
  method of employee selection.
• Direct discrimination can be minimized by
  reducing face-to-face interactions and relying on
  telephone interviews instead or hard data like
  test scores.
• For example, Goldin and Rouse (2000) have found
  that blind auditions reduce gender discrimination
  in hiring women musicians.
• We find that blind interview procedures (like
  telephone interviews) can reduce beauty premium
  by 40 (due to elimination of direct stereotype
  effects).
• Elimination of verbal interaction can eliminate
  beauty premium completely. Too drastic

80
What We Dont Know

• Is taste based discrimination present in repeated
  relationships?
• Do students care more about physical
  attractiveness than older human resource
  officers?
• Are employers over-interpreting visual and audio
  stimuli because those can be productive in most
  other environments?
• Can we design an experiment in which
  self-confidence of workers is payoff-relevant for
  employers?