Theories and empirical strategies

to detect discrimination

Ingar K. Haaland

Trial lecture, June 12, 2019

1 / 33
Discrimination: How to define it

• Definition from Bertrand and Duflo (2017): Members of a minority

group being treated differentially than members of a majority group
with otherwise identical characteristics in similar circumstances.

2 / 33
Discrimination: How to define it

• Definition from Bertrand and Duflo (2017): Members of a minority

group being treated differentially than members of a majority group
with otherwise identical characteristics in similar circumstances.

• Focus of today: Labor market discrimination based on gender

and race

2 / 33
Roadmap

Economic theories of discrimination

Empirical strategies to detect discrimination

3 / 33
Discrimination and economic theory

In standard economic models with perfect competition, there is no

labor market discrimination:

• Employers maximize profits

• Workers are paid their marginal product

• No unemployment

4 / 33
Discrimination and economic theory

In standard economic models with perfect competition, there is no

labor market discrimination:

• Employers maximize profits

• Workers are paid their marginal product

• No unemployment

To study discrimination, economists need to adjust the standard

model

4 / 33
Discrimination: Theory

Economic theories differ in why they predict discrimination:

• Taste-based discrimination (Becker, 1957)

• Animus towards minorities

• Statistical discrimination (Arrow, 1973; Phelps, 1972)

• Using race or gender to make inferences about unobserved skills

• Implicit discrimination (Bertrand et al., 2005)

• Discrimination driven by subconscious biases

5 / 33
Discrimination: Theory

Economic theories differ in why they predict discrimination:

• Taste-based discrimination (Becker, 1957)

• Animus towards minorities

• Statistical discrimination (Arrow, 1973; Phelps, 1972)

• Using race or gender to make inferences about unobserved skills

• Implicit discrimination (Bertrand et al., 2005)

• Discrimination driven by subconscious biases

Distinguishing between the theories is important for welfare

analysis and optimal policy

5 / 33
Taste-based discrimination

• Core assumption in standard economics: Employers maximize

profits
• Becker (1957): Employers maximize utility

• Employers want higher profits, but have an animus towards

minority workers
• Employers are willing to forego profits to avoid hiring minority
workers

6 / 33
Taste-based discrimination: Sketch of model

• Employers can hire male and female workers at market wages

wm and wf , respectively
• Male and female workers are equally productive

• A prejudiced employer incurs a psychic cost, de , from hiring a

female worker

• An employer wants to hire a female worker only if wm − wf > de

7 / 33
Taste-based discrimination: Wages in equilibrium

• Scenario 1: Employers are equally prejudiced

• The pay gap equals the psychic cost of hiring a female worker:
wm − wf = de

• Scenario 2: Employers differ in their prejudice

• Prejudiced employers hire only male workers

• Less prejudiced employers hire only female workers

• Pay gap wm − wf determined by the prejudice of the marginal

employer

8 / 33
Taste-based discrimination: Wages in equilibrium

• Scenario 1: Employers are equally prejudiced

• The pay gap equals the psychic cost of hiring a female worker:
wm − wf = de

• Scenario 2: Employers differ in their prejudice

• Prejudiced employers hire only male workers

• Less prejudiced employers hire only female workers

• Pay gap wm − wf determined by the prejudice of the marginal

employer

• Key prediction of model: Discrimination determined by

marginal prejudice, not by average prejudice

8 / 33
Statistical discrimination

• Core assumption in standard economics: Perfect information

• Arrow (1973); Phelps (1972): Employers face uncertainty about

worker productivity
• Employers maximize profits

• Race or gender is used to make inferences about unobservable

skills

9 / 33
Statistical discrimination: An example for illustration

1. The employer has different prior beliefs about the productivity

of males and females
• The employer expects males to produce 15 units on average and
females to produce 10 units on average

2. The employer observes how a worker performs on a test

• The test correctly predicts worker productivity on average, but is
imprecise

3. The employer updates his beliefs about worker productivity

after observing the test result
• If prior beliefs and the test are normal and equally precise, the
employer updates his beliefs by taking the average of the two:
Posterior about females with test score 20: 0.5 ∗ 20 + 0.5 ∗ 10 = 15
Posterior about males with test score 20: 0.5 ∗ 20 + 0.5 ∗ 15 = 17.5

10 / 33
Statistical discrimination: Implications

• Workers with identical observable skills receive different wages

depending on their type
• Workers are paid equal to their expected productivity

• Discrimination arise because firms use gender or race to make

inferences about unobservable skills
• There would be no discrimination if skills were perfectly
observable, or if employers could design a perfectly informative
costless test

• Key prediction of the model: The gender pay gap reflects actual
productivity differences between males and females

11 / 33
Is statistical discrimination a problem?

• Accurate statistical discrimination: fairness–efficiency trade-off

• People are treated differently based on factors outside of their

control

• Inaccurate statistical discrimination: inefficient and unfair

• Natural to expect biases in beliefs in settings where people have

limited experience (Bohren et al., 2019)

• Regardless of accuracy: Problem of self-fulfilling prophecies

• Anticipating future discrimination, minorities might invest less in

their human capital (Coate and Loury, 1993; Lundberg and Startz,
1983)

12 / 33
Implicit discrimination

• A core assumption in standard economics: People know their

preferences and make choices consistent with those preferences

• Tversky and Kahneman (1974): Heuristics and biases affect

choices

• Implicit discrimination: Employers have no animus towards

minorities, but subconscious biases against minorities give rise
to discrimination (Bertrand et al., 2005)

• Key prediction: Employers with larger implicit biases engage in

more discrimination

13 / 33
Roadmap

Economic theories of discrimination

Empirical strategies to detect discrimination

14 / 33
Detecting discrimination: Empirical strategies

Two main classes of studies:

• Documenting the existence of discrimination

• Testing different theories of discrimination

Three main classes of data:

• Non-experimental data

• Experimental data

• Quasi-experimental data

15 / 33
Detecting discrimination from non-experimental data:
Analysis of earnings differences

• Across countries, females earn less than males – the gender wage
gap

• Parts of the gender wage gap can typically be explained by

observable differences in human capital between males and
females
• Measure of discrimination: the unexplained gender wage gap

• I.e., the part of the gender wage gap that remains after we account
for all observable differences between males and females

16 / 33
Detecting discrimination from non-experimental data:
Critiques

• Risk of explaining too much

• One should not account for factors that are themselves the product
of discrimination

• Risk of explaining too little

• Some skills are unobservable to the researcher, but observable to

employers
• Females and males might systematically differ on these
unobservable characteristics

17 / 33
The gold standard of causal inference: randomized
experiments

• In a randomized experiment, the researcher first randomizes

subjects into two equal groups
• Randomization is the only way to ensure that two groups are equal
on both observable and unobservable characteristics

• The researcher then assigns a treatment to one of the groups

• Differences between the two groups can be fully attributed to the

treatment administered by the researcher

18 / 33
Detecting discrimination from randomized experiments

• Studying the “causal” effect of race or gender on labor market

outcomes is impossible
• Researchers cannot randomly assign race or gender to different
groups of people

• Basic idea behind correspondence studies: Manipulate

perceptions of race and gender through names on fictitious
resumes

19 / 33
Detecting discrimination in hiring: A correspondence study

• Bertrand and Mullainathan (2004) sent out more than 4800

fictitious resumes to help-wanted ads in Boston and Chicago
newspapers
• Half of the resumes had typical white-sounding names, such as
“Emily” and “Greg”
• The other half of the resumes had typical black-sounding names,
such as “Lakisha” and “Jamal”

• Differences in callback rates for interviews can only be attributed

to the names on the resumes

20 / 33
Detecting discrimination in hiring: A correspondence study

• Bertrand and Mullainathan (2004) sent out more than 4800

fictitious resumes to help-wanted ads in Boston and Chicago
newspapers
• Half of the resumes had typical white-sounding names, such as
“Emily” and “Greg”
• The other half of the resumes had typical black-sounding names,
such as “Lakisha” and “Jamal”

• Differences in callback rates for interviews can only be attributed

to the names on the resumes

• Main finding: White-sounding names receive 50 percent more

callbacks for interviews than black-sounding names

20 / 33
Critiques of correspondence studies

• The studies survey the average firm, not the marginal firm

• Employers might infer other characteristics from the names (e.g.,

social status)

21 / 33
Detecting discrimination with quasi-experimental data:
Natural experiments

• Natural experiments are naturally occurring events that

resemble a randomized experiment

• Some change of events create a quasi-random variation that

allows researchers to credibly test for the existence of
discrimination

22 / 33
Detecting discrimination in hiring: A natural experiment
on blind auditions

• Orchestras conduct auditions for vacancies where the candidates

perform before a committee

• In the 1970s and 1980s, many orchestras adopted blind auditions

• Goldin and Rouse (2000) ask whether blind auditions

disproportionately benefit female candidates

23 / 33
Detecting discrimination in hiring: A natural experiment
on blind auditions

• The timing of the adoption of blind auditions varied between

different orchestras
• Many candidates auditioned for several orchestras during the
period
• Identification comes from candidates who auditioned both with
and without screens

• Main result of the study: Blind auditions disproportionately

benefited female candidates

24 / 33
Detecting discrimination in hiring: A natural experiment
on blind auditions

• The timing of the adoption of blind auditions varied between

different orchestras
• Many candidates auditioned for several orchestras during the
period
• Identification comes from candidates who auditioned both with
and without screens

• Main result of the study: Blind auditions disproportionately

benefited female candidates
• Caveat: Study cannot rule out that females perform
disproportionately better under blind auditions

24 / 33
Detecting discrimination: Testing theory predictions

• The leading theories of discrimination make different testable

predictions:
• Taste-based discrimination: Discrimination determined by the
marginal prejudice among employers
• Statistical discrimination: Discrimination determined by actual
productivity differences between groups
• Implicit discrimination: Discrimination determined by
subconscious biases of employers

• Confirming a unique theoretical prediction strengthens support

for the theory

25 / 33
Testing theory predictions: Taste-based discrimination

• Key prediction of the Becker model: Discrimination determined

by the prejudice of the marginal employer
• In the US, large wage gap between blacks and whites

• Blacks are a small fraction of the population ⇒ Discrimination

should be determined by the least prejudiced employers

26 / 33
Testing theory predictions: Taste-based discrimination

• Charles and Guryan (2008) test whether racial prejudices among

whites predict the unexplained black–white wage gap
• Measure of prejudice: Survey questions on racial attitudes from a
decades-long representative survey

• Main finding: Only prejudice among the least prejudiced whites

predicts the unexplained black-white wage gap
• Median prejudice or prejudice among the most prejudiced whites
do not predict the unexplained wage gap

• Evidence consistent with taste-based discrimination

27 / 33
Testing theory predictions: Statistical discrimination

• Key prediction of statistical discrimination: Pay differences

between groups should reflect actual productivity differences

• If pay differences do not reflect group differences in productivity,

could reflect taste-based discrimination or inaccurate statistical
discrimination

28 / 33
Testing theory predictions: Statistical discrimination

• Bohren et al. (2019) test for statistical discrimination in a stylized

online experiment
• “Employers” can hire “workers” to work on a math task

• The employers are informed about whether the workers are

American or Indian

• Main findings: Employers offer higher wages to Indian workers,

but American workers slightly outperform Indian workers
• Employers believe that Indian workers will outperform American
workers
• Information about actual group differences reduces discrimination

• Evidence consistent with inaccurate statistical discrimination

29 / 33
Testing theory predictions: Implicit discrimination

• According to theories of implicit discrimination, subconscious

biases lead to discrimination against minorities

• Subconscious biases can be measured using the

implicit-association test (IAT, Greenwald et al., 1998)

30 / 33
Detecting implicit biases: The IAT

Figure: Screenshot from [Link]

31 / 33
Detecting implicit biases: The IAT

Figure: Screenshot from [Link]

31 / 33
Detecting implicit biases: The IAT

Figure: Screenshot from [Link]

31 / 33
Testing theory predictions: Implicit discrimination

• Glover et al. (2017) examine how manager bias affects the job
performance of minority workers
• Cashiers hired on short-term contract are quasi-randomly assigned
to different managers
• Managers differ in their implicit bias towards minorities (as
measured by IAT scores)

• Main result: Minority cashiers perform worse when they are

assigned a biased manager
• No impact on job performance of non-minority cashiers

• Minority cashiers report fewer interactions with biased managers

32 / 33
Strategies to reduce discrimination

The best strategy to reduce discrimination depends the source of

discrimination:
• Taste-based discrimination

• Scope to reduce discrimination through contact with minority

groups (Burns et al., 2018; Lowe, 2018; Rao, 2019)

• Statistical discrimination

• Under inaccurate statistical discrimination, scope to reduce

discrimination by correcting biases in beliefs (Bohren et al., 2019)

• Implicit discrimination

• Scope to reduce discrimination by raising awareness of biases in

implicit judgement (Alesina et al., 2018; Pope et al., 2018)

33 / 33
References I
Alesina, Alberto, Michela Carlana, Eliana La Ferrara, and Paolo
Pinotti, “Revealing Stereotypes: Evidence from Immigrants in
Schools,” Technical Report, NBER Working Paper No. 25333 2018.
Arrow, Kenneth, “The Theory of Discrimination,” in Orley
Ashenfelter and Albert Rees, eds., Discrimination in the Labor
Market, Princeton University Press, 1973, pp. 3–33.
Becker, Gary S, The Economics of Discrimination, University of
Chicago Press, 1957.
Bertrand, Marianne and Esther Duflo, “Field Experiments on
Discrimination,” in Esther Duflo and Abhijit Banerjee, eds.,
Handbook of Field Experiments, Vol. 1, Elsevier, 2017, chapter 8.
and Sendhil Mullainathan, “Are Emily and Greg More
Employable than Lakisha and Jamal? A Field Experiment on Labor
Market Discrimination,” American Economic Review, 2004, 94 (4),
991–1013.
, Dolly Chugh, and Sendhil Mullainathan, “Implicit
Discrimination,” American Economic Review, 2005, pp. 94–98.
1/4
References II
Bohren, J. Aislinn, Kareem Haggag, Alex Imas, and Devin G Pope,
“Inaccurate Statistical Discrimination,” NBER Working Paper
25935 2019.
Burns, Justine, Lucia Corno, and Eliana La Ferrara, “Interaction,
prejudice and performance. Evidence from South Africa,”
Technical Report, IFS Working Paper 2018.
Charles, Kerwin Kofi and Jonathan Guryan, “Prejudice and Wages:
An Empirical Assessment of Becker’s The Economics of
Discrimination,” Journal of Political Economy, 2008, 116 (5), 773–809.
Coate, Stephen and Glenn C Loury, “Will Affirmative-Action
Policies Eliminate Negative Stereotypes?,” American Economic
Review, 1993, 83 (5), 1220–1240.
Glover, Dylan, Amanda Pallais, and William Pariente,
“Discrimination as a Self-Fulfilling Prophecy: Evidence from
French Grocery Stores,” Quarterly Journal of Economics, 2017, 132 (3),
1219–1260.

2/4
References III
Goldin, Claudia and Cecilia Rouse, “Orchestrating Impartiality: The
Impact of “Blind” Auditions on Female Musicians,” American
Economic Review, 2000, 90 (4), 715–741.
Greenwald, Anthony G, Debbie E McGhee, and Jordan LK
Schwartz, “Measuring individual differences in implicit cognition:
the implicit association test.,” Journal of Personality and Social
Psychology, 1998, 74 (6), 1464.
Lowe, Matt, “Types of Contact: A Field Experiment on Collaborative
and Adversarial Caste Integration,” Unpublished Manuscript, 2018.
Lundberg, Shelly J and Richard Startz, “Private discrimination and
social intervention in competitive labor market,” American
Economic Review, 1983, 73 (3), 340–347.
Phelps, Edmund S, “The Statistical Theory of Racism and Sexism,”
American Economic Review, 1972, 62 (4), 659–661.
Pope, Devin G, Joseph Price, and Justin Wolfers, “Awareness
Reduces Racial Bias,” Management Science, 2018, 64 (11), 4967–5460.

3/4
References IV

Rao, Gautam, “Familiarity does not breed contempt: Diversity,

discrimination and generosity in Delhi schools,” American Economic
Review, 2019, 109 (3), 774–809.
Tversky, Amos and Daniel Kahneman, “Judgment Under
Uncertainty: Heuristics and Biases,” Science, 1974, 185 (4157),
1124–1131.

4/4