The Role of Beliefs in Driving Gender Discrimination Katherine B. Coffman Christine L. Exley Muriel Niederle

Working Paper 18-054

Working Paper 18-054

Copyright © 2017, 2018, 2019, 2020 by Katherine B. Coffman, Christine L. Exley, and Muriel Niederle.

Working papers are in draft form. This working paper is distributed for purposes of comment and discussion only. It may not be reproduced without permission of the copyright holder. Copies of working papers are available from the author.

Funding for this research was provided in part by Harvard Business School.

The Role of Beliefs in Driving Gender Discrimination

Katherine B. Coffman Harvard Business School

Christine L. Exley Harvard Business School

Muriel Niederle Stanford University

The Role of Beliefs in Driving Gender Discrimination

Katherine B. Coffman,

Christine L. Exley,

Muriel Niederle ∗†

December 11, 2019

Abstract

While there is ample evidence of discrimination against women in the workplace, it can

be difficult to understand what factors contribute to discriminatory behavior. We use an

experiment to both document discrimination and unpack its sources. First, we show that,

on average, employers prefer to hire male over female workers for a male-typed task even

when the two workers have identical resumes. Second, and most critically, we use a novel

control condition to identify that this discrimination is not specific to gender. Employers

are simply less willing to hire a worker from a group that performs worse on average, even

when this group is instead defined by a non-stereotypical characteristic. In this way, beliefs

about average group differences are the key driver of discrimination against women in our

setting. We also document some evidence for in-group preferences that contribute to the

gender discrimination observed. Finally, our design allows us to understand and quantify

the extent to which image concerns mitigate discriminatory behavior.

1 IntroductionUnderstanding the drivers of gender differences in labor market outcomes has been an im-

portant topic of study among labor economists, with research identifying sizable roles for oc-

cupational segregation, differences in human capital accumulation, demand for flexibility, and

differences in preferences (Goldin, 2014; Card, Cardoso and Kline, 2016; Olivetti and Petrongolo,

2016).1

∗Coffman: kcoffman@hbs.edu, Harvard Business School; Exley: clexley@hbs.edu, Harvard Business School;Niederle: niederle@stanford.edu, Stanford University, SIEPR and NBER.†Thank you to John Beshears, Alison Wood Brooks, Lucas Coffman, Benjamin Edelman, Leslie John, Johanna

Mollerstrom, Joshua Schwartzstein, Andrei Shleifer, and seminar audiences at Cornell University, the RotmanSchool of Management, and the 2016 ESA North American Meetings for their many useful comments.

1For reviews on gender differences, see Croson and Gneezy (2009), Bertrand (2011), Azmat and Petrongolo(2014) and Niederle (2016).

1

Discrimination also contributes to these gender gaps in earnings and advancement. A large

body of empirical work has documented the existence of gender discrimination in labor markets,

often through clever field experiments or with quasi-experimental data (for reviews, see Riach

and Rich (2002) and Blau and Kahn (2017)). Discrimination against women has been found in

bargaining contexts (Ayres and Siegelman, 1995; Castillo et al., 2013; Bowles, Babcock and Lai,

2007), in hiring, employment, and referral contexts (Neumark, Bank and Nort, 1996; Goldin and

Rouse, 2000; Black and Strahan, 2001; Bertrand and Mullainathan, 2004; Moss-Racusin et al.,

2012; Reuben, Sapienza and Zingales, 2014; Baert, De Pauw and Deschacht, 2016; Bohnet, van

Geen and Bazerman, 2016; Sarsons, 2017a), and in academic contexts (Milkman, Akinola and

Chugh, 2012, 2015; Sarsons, 2017b).2

Despite this large literature, important open questions about gender discrimination remain.

One challenge is identifying the primary driver of observed discrimination. Our study attempts

to answer this key question. Traditionally, economists have viewed discrimination through two

distinct lenses: taste-based or statistical. The taste-based argument posits that discrimination is

rooted in preferences, driven by animus or prejudice (Becker, 1957). Statistical discrimination,

on the other hand, is rooted in rational beliefs about average gender differences in abilities or

skills (Phelps, 1972; Arrow, 1973).3 Recently, researchers have proposed more general forms

of belief-based discrimination, with a particular focus on discrimination based upon inaccurate

beliefs. Bordalo et al. (2016) document a role for representativeness-based stereotypes, where

beliefs about groups are not accurate but instead are biased by differences in low probability but

highly representative tails. These inaccurate beliefs can then give rise to self-stereotyping and

discriminatory behavior (Coffman, 2014; Bordalo et al., 2019). In this spirit, Bohren, Imas and

Rosenberg (2017) document discrimination against women on an online math platform and show

that it is most consistent with discrimination based upon biased beliefs.

However, cleanly disentangling these different drivers of discrimination — particularly ac-

curate beliefs, biased beliefs, and tastes — is challenging. Past literature has mainly followed

three approaches. One approach is to test for a role for beliefs by increasing the amount of

information available on individual workers. The argument is that if more informative perfor-

mance information reduces discrimination, beliefs must play an important role. However, to the

extent that discrimination remains, it is unclear whether it reflects taste-based considerations

or residual belief-based motivations. A second approach is to simply remove information that

identifies the (gender, ethnic, socio-economic, etc.) group to which an individual belongs, for

2Many scholars have identified discrimination against other groups as well, such as the beauty premium(Mobius and Rosenblat, 2006), discrimination based on socioeconomic status (Rao, 2018), and racial or ethnicdiscrimination (for examples, see Fershtman and Gneezy (2001), List (2004), Charles and Guryan (2008), Castilloand Petrie (2010), Gneezy (2012), Doleac and Stein (2013), Lei and Babcock (2013), and Bartos et al. (2016),and, for a review, see Charles and Guryan (2011)). For a review, see also Bertrand and Duflo (2017).

3See Altonji and Blank (1999) and Guryan and Charles (2013) for review and discussion of literature ontaste-based versus statistical discrimination.

2

instance by taking a name off of a resume, ensuring that any discrimination does not reflect

taste-based motivations.4 However, in contexts where there are group differences in performance

on average, removing information that identifies the group to which an individual belongs not

only rules out taste-based motivations, but also inhibits belief-based discrimination. A final

approach, often used in field data, is to ask whether the observed discrimination is consistent

with profit-maximization, suggesting statistical motivations. However, this has typically required

assuming that the discrimination is driven by accurate, rather than inaccurate, beliefs.

Ideally, what one would do to separate any role of beliefs from tastes would be to recreate the

exact hiring environment where we see discrimination against women, divorced from any gender-

specific considerations but not ability-specific considerations. This is our goal. Of course, this

goal would be nearly impossible to achieve in naturally occuring field data. And so, we design

a controlled experiment. We employ a novel control treatment that allows us to benchmark

the level of discrimination we observe against women in a male-typed hiring environment to

discrimination against workers of unknown gender known to be drawn from the same ability

distribution as the women. We do this in a setting where we can provide accurate information

both on individual workers and group differences and where we can measure employer beliefs. It

is in this way that we attempt to separate discrimination based upon tastes and discrimination

based upon beliefs (independent of whether those beliefs are accurate or not).

In a preliminary study, we collect performance information from participants on easy and

hard math and sports quizzes. We use these participants as available “workers” for hire in our

main hiring experiment. In this main hiring experiment, we ask “employers” to make incentivized

choices over available workers. Employers receive information about the easy quiz performances

of available workers. When they choose to hire a worker, they are paid based upon the hard

round performance of that worker, which is unknown to them at the time of the decision.

The hiring experiment involves two treatments that vary only in their labeling of the avail-

able workers: a Gender treatment and a Birth Month treatment. In the Gender treatment,

participants make hiring decisions between female and male workers, labeled as such, while in

the Birth Month treatment, participants make decisions over workers born in an even month or

workers born in an odd month. Importantly, the actual set of available workers is held constant

across the two treatments; we simply vary how the workers are labeled. To implement this, we

use two subsamples to generate our pool of available workers: women born in even months and

men born in odd months. Thus, while the labels used to describe the available workers vary by

treatment, the performances of the available workers do not. In both treatments, participants

4A distinct but related approach is to simulate a context where taste-based considerations are absent, forinstance by studying anonymous workers with randomly-assigned productivities, allowing for clean documentationof statistical discrimination (for instance, Dickinson and Oaxaca (2009) and Dickinson and Oaxaca (2014)). Ofcourse, absent a comparable context where taste-based considerations are relevant, this approach does not allowfor one to back out how much of the observed discrimination is attributable to beliefs versus tastes.

3

receive accurate, detailed information about the distribution of performances across the groups

prior to making their hiring decisions, generating similar ex-post beliefs about the average ability

gap (male/female gap and odd/even gap) across the two treatments. In this way, we abstract

away from the belief formation process.5 Importantly, what we care most about is not whether

beliefs are accurate or not (although our data can speak to this), but instead on how a given set

of beliefs drives behavior in each treatment.

Our design creates a very similar set of beliefs about average group differences across our

two treatments (the Gender treatment and the Birth Month treatment). Then, we can ask

whether women are more or less likely to be hired when they are labeled as women, in the

Gender treatment, than when they are not, in the Birth Month treatment. Because beliefs of

not only individual performances but also average group differences are the same across the two

treatments, differential discrimination across the two treatments can be attributable to gender-

specific tastes, rather than beliefs.

Our new methodology has a number of advantages. First, we elicit employer beliefs of gender

differences, allowing us to directly link beliefs to behavior. Second, our Birth Month treat-

ment better isolates the gender-specific taste component of discrimination by holding beliefs,

both about individuals and groups, constant. It does so by removing information about worker

gender while simultaneously maintaining identical information about individual and group-level

performances. In this way, it provides a useful benchmark to which we can compare the extent

of discrimination against women in the Gender treatment. Holding fixed beliefs about individual

performances and average performance differences across the groups, we can ask whether women

are any less likely to be hired when they are labeled as women rather than as even-month workers.

Third, by replacing the gender labels in the Gender treatment with birth month labels in the

Birth Month treatment, we are able to speak to the relevance of in-group preferences in hiring

decisions, and whether such in-group preferences, are gender-specific.

We find that, on average, women in the Gender treatment are less likely to be hired than

equally able men. That is, when presented with two workers with identical easy quiz perfor-

mances, one of whom is a man and one of whom is a woman, employers are significantly less

likely to hire the woman than the man on average. Perhaps surprisingly, however, this result

is not specific to gender. When we turn attention to the Birth Month treatment, we observe

similar levels of discrimination against even-month workers. This suggests that behavior is not

driven by animus toward women and instead is more consistent with belief-based theories of

discrimination. In particular, beliefs of average group differences are key.

We also find evidence of in-group preferences. In the Gender treatment, female employers hire

women more often than male employers do. But, again, this result does not appear to be driven

5This of course limits our ability to test the stereotype formation channel of Bordalo et al. (2016). We focusinstead on decision-making given a certain set of beliefs.

4

specifically by gender. When we consider the Birth Month treatment, we find a similar in-group

preferences: employers born in even months hire even-month employees more than odd-month

employers do.

One important question is whether our results are driven by image concerns, such as not

wanting to appear sexist. We take this concern seriously in our design, using additional questions

to get at exactly this issue. We explore whether our results persist when there is a veil on

employers’ intentions, reducing image concerns. We operationalize this veil in a manner that

is similar in spirit to the examination of risk as a veil in Exley (2015).6 We continue to find

that women are hired no less often when labeled as women than when labeled as even-month,

suggesting that our failure to document clear evidence of taste-based discrimination against

women is not driven simply by a fear of appearing sexist, or image concerns more generally. The

introduction of a veil of intentions does, however, eliminate in-group preferences, suggesting a

significant role for image concerns in driving these findings.

Our findings point to a key source of gender discrimination: beliefs. In this way, our work

represents an advance within the empirical social science literature on identifying, measuring, and

understanding discriminatory behavior. Our work also has important implications for practition-

ers interested in limiting discrimination within the workplace. First, it is important to recognize

that we all hold beliefs about others based upon the groups to which these others belong (their

gender, their race, their socioeconomic status, etc.). In many cases, these beliefs are inaccurate.

Inaccurate beliefs are particularly likely when information is more scarce, and in contexts where

stereotypes are prevalent. Second, as our experiment shows, these beliefs play a critical role in

individual decision-making: beliefs are likely to inform who we hire, who we promote, and who

we turn to for leadership. When making decisions between a man and a woman, beliefs about

average group-level differences across men and women are likely to fuel discriminatory behavior

even when the individual-level performance information on the man and woman is identical.

2 DesignOur studies are conducted on Amazon Mechanical Turk (MTurk). MTurk has a number of

features that make it an appropriate and useful platform for our experiment. Most centrally,

on Mturk, it is feasible to collect the large amount of data needed to be adequately powered

for the statistical tests we run.7 Importantly, previous work has shown that behavior among

participants on MTurk and participants in laboratory studies is quite similar, including in the

extent to which they pay attention to and follow instructions (Paolacci, Chandler and Ipeirotis,

6See also Dana, Weber and Kuang (2007), Haisley and Weber (2010), Danilov and Saccardo (2017), and forreviews, Kunda (1990) and Gino, Norton and Weber (2016).

7With 400 employers in each our two treatments, we have a statistical power of approximately 80% to detecta 10pp difference in hiring rates of women across our two treatments at the 5% level. Because we run severalversions of our main hiring experiment, detailed below, in total we recruited 3,695 individuals to participate,making the study nearly impossible to run in a laboratory.

5

2010; Halberda et al., 2012). Many classic laboratory studies have been successfully replicated

on MTurk (Rand, 2012).

In a preliminary study, we collect performance information on four types of quizzes from

participants: an easy and a hard math quiz, and an easy and a hard sports quiz.8 We recruit

100 individuals for an advertised 40-minute academic study with a guaranteed completion fee of

$6. Each of the four quizzes contained 10 multiple-choice questions. Workers had three minutes

per quiz to answer as many questions as possible. To incentivize performance, workers received,

as bonus payment within one week, 10 cents for each question they answered correctly, for one

randomly-selected quiz. The workers were also aware that their performances would potentially

be shown to other participants in follow-up experiments.

The purpose of this preliminary study is to produce a pool of available “workers” for hire

in our hiring experiment. For each worker from this preliminary study, we have performance

information that we can reveal to potential “employers,” enabling us to ask employers to make

incentivized hiring decisions over available workers.

In our main hiring experiment, we ask employers to make incentivized choices over available

workers. We recruit 800 new MTurk participants for a 25-minute academic study with a guar-

anteed completion fee of $4. Employers receive information about the easy quiz performances

of available workers. Then, they make a series of hiring decisions from different sets of available

workers. Importantly, across every hiring decision they make, the employers have access to (at

least some) information about the easy quiz performance of the available workers. To incentivize

each hiring decision, the employer earns money based upon the hard quiz performance of the

hired worker in each decision and knows that they will receive, as bonus payment within one

week, any associated payoffs from one randomly-selected decision.

In this sense, our design mimics many real world hiring paradigms in three important respects.

First, the employer receives some information about performance on an initial quality assessment

(the easy quiz) for each candidate. Second, performance on this initial assessment (the easy quiz)

serves as a signal of the candidate’s capacity to perform well in the job (the hard quiz). Third,

the employer stands to gain more by hiring workers who ultimately perform better on the job

(the hard quiz).

We describe our experimental design, and treatment variations, in detail below. Appendix C

provides full experiment instructions and screenshots of the experiment.

2.1 Treatment variations

Our key treatment variation is whether the available workers are described to employers in

terms of their gender, or not. Importantly, our goal is to hold everything else as fixed as possible

8We actually collect data on six total quizzes. The two extra quizzes are an additional easy math quiz and anadditional easy sports quiz. We do not use performance on these two other quizzes in the hiring experiment, sowe choose to omit them from our description of the design in order to reduce confusion.

6

among the workers across the two treatments, including employer beliefs about the performances

of the groups of workers. To do this, we construct an admittedly strange, but very useful set of

available workers. Among the 100 workers who completed our preliminary performance study,

we restrict our attention to two groups of workers: male workers born in odd months (who we

will refer to as “male-odd-month” workers) and female workers born in even months (who we

will refer to as “female-even-month” workers). These two groups of workers will be our available

workers across both treatments of the study. All that will vary is how these workers are described

to employers.9

While employers always make decisions between male-odd-month workers and female-even-

month workers, employers view different labels for these groups of workers. Employers in the

Birth Month treatment are not provided with the gender labels and see all information and

decisions as between odd-month workers and even-month workers. Employers in the Gender

treatment are not provided with the birth month labels and see all information and decisions as

between male workers and female workers. The treatments are otherwise identical.

In addition, employers are randomly assigned to either the math or sports version of the

study. In the math version of the study, employers see information about worker performance on

the easy math quiz and are incentivized according to worker performance on the hard math quiz.

In the sports version of the study, employers see information about worker performance on the

easy sports quiz and are incentivized according to worker performance on the hard sports quiz.

We conducted two versions of the study to better understand how the extent of discrimination

in a male-typed domain might vary depending upon the seriousness or prestige of the task at

hand. That is, we hypothesized that employers might feel that it is more acceptable or socially

appropriate to discriminate against women in a sillier task, like sports, compared to a more

educationally and career relevant task, like math. In most of the analysis below, we pool the two

versions of the study together and discuss them jointly. In Section 3.3, we analyze the differences

between the two versions.

2.2 Elicitation of Prior Beliefs

In the very first stage of the hiring experiment, we elicit employers’ prior beliefs about the

performance gap between male-odd-month workers and female-even-month workers. We explain

that we have recruited a past set of MTurk workers to complete easy and hard 10-question quizzes

in math (sports). Employers are asked to provide their beliefs about average group differences

in performance on these quizzes. Employers in the Birth Month (Gender) treatment are asked

9This design involves the withholding of information from employers (worker gender in the Birth Monthtreatment, and worker birth month in the Gender treatment), but no mis-representation of information. Thus, wefeel that it does not constitute deception. Notably, gender information (not to mention birth month information)is very often withheld from participants in experiments in which participants interact with each other — includingin dictator games, trust games, bargaining games, and tournament competitions. Thus, while the withholding ofthis gender information is salient to the reader given our design, it is not particularly unusual in experimentalparadigms.

7

about their prior beliefs about the average performance gap between odd-month (male) and even-

month (female) workers. In particular, employers predict the difference in average scores between

these groups of workers on a 10-question easy quiz and a 10-question hard quiz. Employers could

indicate any feasible average difference (from -10 to 10 problems solved correctly), with positive

numbers indicating a performance gap in favor of odd-month (male) workers in the Birth Month

(Gender) treatment. These beliefs are not incentivized. We collect this information as a baseline

before employers proceed to the Information Stage.

2.3 Information Stage

After providing prior beliefs, employers proceed to the Information Stage. The Information

Stage is a critical component of our design. In order for the Birth Month treatment to serve as

the right comparison treatment for the Gender treatment, employers must have similar beliefs

about the performances of the two groups of workers in both treatments. That is, our goal

is to have the perceived gender gap in performance be roughly equal to the perceived birth-

month gap in performance when we compare across the two treatments. This is obviously

unlikely in terms of prior beliefs. So, we achieve this by providing accurate and comprehensive

information about performances on the easy quiz in our experiment. We show each employer the

full distribution of performances on the easy quiz for female-even-month workers and male-odd-

month workers. However, rather than just show this information once (i.e. present one histogram

of performances), we reinforce this information by showing different subsamples of this data to

the employer.

In particular, we draw various subsets of performances on the easy quiz of female-even-

month workers and male-odd-month workers from our experiment. To give a concrete example,

an employer sees the following subset of performances: a histogram of the performances of all

female-even-month workers born between the between the 1st - 15th of the month compared

to a histogram of the performances of all male-odd-month workers born between the 1st - 15th

of the month. This allows the employer to compare performances among men and women (or

among even-month and odd-month workers) among this particular subset. Then, they see a new

subset of performances, selected by drawing workers from another range of birth dates. In total,

employers see 12 different sets of distributions, each drawing from a different range of possible

birth dates. The first 11 sets of distributions that employers see are formed by restricting the

subset of workers to those born during different date ranges (see Appendix Table A.1 for more

details on how these distributions are formed). The 11 distributions are unique but overlapping

(that is, an employer only sees the screen with the distributions for workers born between the

1st - 15th of the month once, but she also sees a screen with the distributions for workers born

between the 1st - 10th days of the month). The order in which these 11 sets of distributions

are shown is randomized at the participant level, and each is presented on a separate screen.

For all participants, the 12th and final set of distributions contains the full distributions, the

8

histograms of the performances of all male-odd-month workers and female-even-month workers

in our experiment.10 Figure 1 illustrates how this information was displayed to participants,

using the final set of full distributions.

Figure 1: Worker Performance on the Easy quiz

(a) Math Version (b) Sports Version

Note: Panel A shows the performance distributions in the easy math quiz for female-even-monthworkers, who have an average performance of 3.27, and male-odd-month workers, who have anaverage performance of 3.96. Panel B shows the performance distributions in the easy sports quizfor female-even-month workers, who have an average performance of 4.50, and male-odd-monthworkers, who have an average performance of 5.50. The labels “Female” and “Male” shown hereare those that would appear in the Gender treatment. In the Birth Month treatment, the labelswould instead read “Even” and “Odd”, respectively. Each employer saw these distributions asthe final screen of the Information Stage.

In addition, we seek to increase the employers’ engagement with these data by requiring

them to make incentivized decisions based upon this information. For each set of the 12 sets

of distributions they see, we ask employers to make an incentivized decision. For each set of

distributions, the employers are asked to select one distribution from which they would like to

hire. In the Birth Month treatment, the options are “The Even Month Group”, “The Odd Month

Group” and “Let Chance Determine the Group.” In the Gender treatment, the options are “The

Female Group”, “The Male Group” and “Let Chance Determine the Group.” If an employer

selects “Let Chance Determine the Group,” the computer flips a coin to determine which group

is hired from. We provide the “Let Chance Determine the Group” option in order to better

identify indifference among our population; with this option available, it seems more likely that

10We do not believe there is anything special about this particular construction of subsets of performanceinformation. We felt it would be better for the employers to see this information in many stages, reinforcingthe information, without simply repeating it verbatim. Dividing by day of birth was simply a convenient andavailable formulation for our design.

9

a choice of one of the other two groups reflects a strict preference.

If one of these decisions is randomly selected as the decision that counts for payment, one

worker from the selected group is chosen at random to be hired. The hired worker receives an

additional 25 cents as bonus payment. Employers receive 10 cents for each question answered

correctly by the hired worked on the hard quiz.

The main goal of these choices is to encourage employers to engage with this information,

incorporating it into their beliefs about the average performance gap. While we can consider

choices made during this stage of the experiment when evaluating the degree of discrimination,

they are not as informative about the drivers of discrimination. This is because beliefs are likely

not fixed at this stage in the experiment, and may evolve differently across the two treatments.

Nonetheless, we discuss these Information Stage hiring decisions, and what we can learn from

them, in Section 3.5.

2.4 Elicitation of Posterior Beliefs

After completing the Information Stage, employers are re-asked all of the beliefs questions

from the Elicitation of Prior Beliefs stage. That is, they see the same beliefs questions again so

that we can collect posterior beliefs. Again, these beliefs are not incentivized. The key data for

us are employers’ posterior beliefs of average differences in performance across the two groups.

Our goal is to ensure similar beliefs of average performance differences across the two treatments.

That way, we can isolate the role for gender-specific taste considerations, holding beliefs fixed,

by comparing across the Gender and Birth Month treatment.

2.5 Hiring Stage

After the Information Stage and Elicitation of Posterior Beliefs, we ask employers to make

a series of hiring decisions between specific pairs of workers. In each hiring decision, employers

in the Birth Month treatment are asked whether they want to hire “The Even-Month Worker,”

the “Odd-Month Worker,” or “Let Chance Determine Who is Hired.” Employers in the Gender

treatment are asked whether they want to hire the “The Female Worker,” “The Male Worker,” or

“Let Chance Determine Who is Hired.” Again, the chance option allows employers to explicitly

express indifference. For each hiring decision, employers learn the two workers’ exact perfor-

mances on the easy quiz. We vary the performances of the available workers and the payoffs to

hiring each across a series of 54 decisions, split across six screens (within-subject).

On each screen in the Hiring Stage, employers make nine hiring decisions. For each, they are

told the easy quiz performances of the two workers in the pair. In three of the hiring decisions,

the available workers have the same performance on the easy quiz. The number of questions

answered correctly (out of 10) by the male-odd-month worker versus the female-even-month

worker is (i) 4 versus 4, (ii) 6 versus 6, and (iii) 8 versus 8. In the other six hiring decisions on

the screen, the male-odd-month worker has a weaker performance than the female-even-month

worker: (iv) 4 versus 5, (v) 4 versus 6, (vi) 4 versus 7, (vii) 4 versus 8, (viii) 6 versus 7, and

10

(ix) 6 versus 8. We choose to focus on decisions in which the female-even-month worker weakly

outperforms the male-odd-month worker so that we can easily classify all decisions not to hire

the female worker as discrimination against women.11

Across the screens in the Hiring Stage, we vary the payoffs to hiring each worker. On the

first Hiring Stage screen, the payoff to hiring either worker is 10 cents for each question answered

correctly by their hired worker on the hard quiz.

In these stark, side-by-side hiring decisions, employers may feel that choosing not to hire the

female-even-month worker (who always has a weakly better performance) is harmful to their self-

image or social-image. This might be particularly true in the Gender treatment due to concerns

about perceived sexism. A key question we wish to ask is whether social desirability concerns, or

image concerns more generally, limits the extent of discrimination against the female-even-month

worker. To do this, we introduce additional sets of hiring decisions where we provide a clear and

easily justifiable “excuse” for the employer not to hire the female-even-month worker, similar in

spirit to Exley (2015).12

Paralleling Exley (2015), the potential excuse we provide takes the form of risk. We present

again the hiring decisions from the first screen of the hiring stage. But now, we alter each decision

so that the decision to hire a female-even-month worker becomes risky. In these decisions, the

payoff remains 10 cents per correct answer on the hard quiz if the employer hires a male-odd-

month worker. But, if the employer hires a female-even-month worker, there is some risk that the

employer will receive nothing. In particular, if an employer hires a female-even-month worker,

she receives 10 cents for each question correctly answered by that worker on the hard quiz with

P% chance but no payment for that particular hiring decision with (100-P)% chance. Across

the second through sixth Hiring Stage screens, the risk of hiring the female-even-month worker

increases, as we decrease P from 99, 95, 90, 75 to 50.

From an image perspective, the introduction of this risk can help to serve as a veil on employer

intentions. If an employer does not hire a woman when the payoffs to hiring either the man or

the woman are the same, this could easily be interpreted as discriminatory or sexist behavior.

But, if that same employer does not hire the woman under the risky decisions, an alternative

interpretation is more readily available: the employer is just profit-maximizing. In this way,

11An interesting question is whether the presentation of this particular set of decisions impacts employerdecisions on any given decision. For instance, does the fact that the female worker always weakly outperforms themale worker on any particular Hiring Stage screen impact the propensity to hire women? In order to investigatethis decision, we replicate our main hiring experiment, but allow employers to make only a single choice duringthe Hiring Stage, between two workers each with an easy quiz score of 4. The results of this experiment are largelyconsistent with the results from our main experiment. We describe this experiment and its results in detail inAppendix B.

12Danilov and Saccardo (2017) use a related approach in their paper on ethnic discrimination. While they findno evidence of discrimination by Germans toward Turks in a basic dictator game, when they require that thesender tell a lie in order to provide the receiver with the generous outcome, Germans are less likely to do this fora Turk than a German.

11

employers who would like to discriminate against women in riskless decisions — but feel reluctant

to do so due to image concerns — may feel more able to do so in the risky decisions. Of course,

this explanation seems most relevant in the Gender treatment (where image concerns about

sexism are clear). So, our Birth Month control treatment is helpful in isolating this mechanism.

The introduction of risk decreases the expected payment from hiring a female-even-month

worker in both treatments. So, we expect less hiring of female-even-month workers in the risky

decisions than in the riskless decisions under both treatments. But, if in addition, we see a larger

employer reaction to the introduction of risk in the Gender treatment than in the Birth Month

treatment (that is, a bigger drop in the rate of hiring female-even-month workers), this suggests

that some of the reaction in the Gender treatment is not just about the expected decrease in

payoff. Instead, it may also suggest an enabling of discrimination that had previously been

mitigated by image concerns in the riskless decisions.

If one of these 54 hiring decisions is randomly selected to count for payment, employers receive

the amount of money earned in the selected decision as an additional bonus payment. Their hired

worker from the selected decision also receives 25 cents as additional bonus payment.

This design is summarized in Figures 2 and 3.

Figure 2: Preliminary Study to Create Pool of Workers

Collect performance data from 100 workers• Performance on easy quiz• Performance on hard quiz

Create two pools of workers

25 Women born in Even Months

25 Men born in Odd Months

Present these two pools to employers in the hiring experiment

• Described as women and men in GENDER treatment• Described as even-month and odd-month in BIRTH

MONTH treatment

12

Figure 3: Overview of Hiring Experiment

Elicitation of Prior BeliefsReport believed performance

differences by gender OR birth month.

Information StageReceive accurate information on average performance

differences on easy quiz by gender OR birth month.

Elicitation of Posterior BeliefsReport updated believed performance differences by gender OR birth month.

No Risk Hiring Decisions (Screen 1)Decide whether to hire male/female worker OR

odd month/even month worker.Receive $0.10 per problem solved correctly on hard quiz

by hired worker.

Risky Hiring Decisions (Screens 2 – 6)Decide whether to hire male/female worker OR

odd month/even month worker.Receive $0.10 per problem solved on hard quiz by hired

worker . But, receive payoff with only probability

p = {0.99,0.95,0.90,0.75,0.50} if hired worker is female OR even month.

3 ResultsWe begin by examining employers’ beliefs in Section 3.1. We then investigate hiring decisions

in Section 3.2, revealing an important role for beliefs. To provide more insight, we explore how

the extent of discrimination varies by domain (math versus sports) in Section 3.3, and how the

extent of discrimination varies by type of employer (male versus female) in Section 3.4.

We start first by considering decisions from the Hiring Stage where employers can choose

between a female-even-month worker and a male-odd-month worker with identical easy round

performances. We focus on these decisions for two reasons. First, because the workers have iden-

tical easy round performances, it is a clear-cut environment in which to look for discrimination:

absent discrimination, we expect female-even-month and male-odd-month workers to be hired at

identical rates. Second, since these decisions occur after the Information Stage where all group-

level information is provided, we can more reasonably assume that beliefs have been fixed and are

13

similar across treatments. This allows us not just to document the extent of discrimination, but

also to speak to its sources. After this main analysis, we explore all other decisions that employ-

ers make during our study. In particular, Section 3.5 details the hiring decisions that employers

make during the Information Stage as well as the hiring decisions that employers make when

the female-even-month worker has a better easy round performance than the male-odd-month

worker in the Hiring Stage. Finally, we document the robustness of our findings by showing that

employer decisions look quite similar in a highly simplified follow-up hiring experiment with a

new sample of employers.

3.1 What beliefs do employers hold?

Figure 4 displays the kernel densities of employers’ prior beliefs about the performance gap

on the hard and easy quizzes between male-odd-month workers and female-even-month workers.

As we expect, employers initially believe that male workers outperform female workers by a

significantly larger amount than odd-month workers outperform even-month workers (see Panels

A and C; the p-value for across-treatment differences in prior beliefs is <0.01 for both the easy

and hard quizzes, using either a K-S test or a t-test). After employers complete the Information

Stage, however, posterior beliefs about these gaps are no longer significantly different across

treatments (see Panel B and D; the p-value for across-treatment differences in posterior beliefs

is >0.10 for both the easy and hard quizzes, using either a K-S test or a t-test).13

Our Information Stage “works” in the sense that, afterwards, the extent to which employers

in the Gender treatment believe male workers outperform female workers is the same as the

extent to which employers in the Birth Month treatment believe odd-month workers outperform

even-month workers. Given that, on average, employers believe there is a significant gap in aver-

age performance, this provides a clear motive to statistically discriminate. And, since beliefs are

similar across treatments, we should expect a similar degree of this type of belief-based discrim-

ination across treatments. Thus, if we see different levels of discrimination across treatment, we

can attribute this to gender-specific tastes, rather than beliefs.

Although not central to out investigation, we can also speak to the accuracy of these beliefs.

We note that posterior beliefs get the direction of the difference right on average, but exaggerate

it — believing male-odd-month workers outperform female-even-month workers on both the

hard and easy quizzes by more than they actually do.14 This is in line with the model of biased

13Appendix Table A.2 confirms that this belief convergence is on average statistically significant and showsthat it is similar when separately considering workers’ performance on sports or math.

14In the sports version, the average posterior belief of this gap is 2.31 on the hard quiz and 2.21 on the easyquiz, while the actual average gap in performance is 1.1 on the hard quiz and 1.0 on the easy quiz. If we testwhether the believed gap is significantly different than the observed gap in performance, we reject the null ofequality with p<0.01 for both the hard and easy quiz. Similarly, in the math version, the average posterior beliefof the gap is 1.20 on the hard quiz and 1.15 on the easy quiz, while the actual average gap is 0.451 on the hardquiz and 0.692 on the easy quiz. Again, believed gaps are significantly different than observed gaps, with p<0.01on the hard quiz and p = 0.053 on the easy quiz.

14

beliefs and stereotyping by Bordalo et al. (2016) and the evidence presented in Bordalo et al.

(2019). It is also worth noting that male-odd-month workers directionally outperform female-

even-month workers on the hard quiz even when controlling for easy quiz performance.15 Thus,

an employer who statistically discriminates — i.e. practices belief-based discrimination based

upon accurate beliefs — would hire male-odd-month workers over female-even-month workers

when they have the same easy quiz performance. Of course, a decision-maker who holds biased

beliefs — exaggerating the gap — would similarly make this hiring decision. In our investigation,

we will focus on the predictive power of beliefs generally (both accurate and inaccurate), and

focus on distinguishing any belief-based form of discrimination from taste-based considerations.

Figure 4: Beliefs about performance gap between male-odd-month workers and female-even-months workers

0.1

.2.3

.4.5

Den

sity

-10 -5 0 5 10Performance Gap

Birth Month Treatment Gender Treatment

Two-sided Kolmogorov–Smirnov: p = 0.00

(a) Prior Beliefs (Hard Quiz)

0.0

5.1

.15

Den

sity

-10 -5 0 5 10Performance Gap

Birth Month Treatment Gender Treatment

Two-sided Kolmogorov–Smirnov: p = 0.96

(b) Posterior Beliefs (Hard Quiz)

0.1

.2.3

.4.5

Den

sity

-10 -5 0 5 10Performance Gap

Birth Month Treatment Gender Treatment

Two-sided Kolmogorov–Smirnov: p = 0.00

(c) Prior Beliefs (Easy Quiz)

0.0

5.1

.15

Den

sity

-10 -5 0 5 10Performance Gap

Birth Month Treatment Gender Treatment

Two-sided Kolmogorov–Smirnov: p = 0.19

(d) Posterior Beliefs (Easy Quiz)

3.2 Is there discrimination against women?

We begin by considering hiring decisions between a male-odd-month worker and a female-

even-month worker when all else is equal.16 Employers learn that the payoff rule from hiring

15In particular, a regression of the hard quiz performance on an indicator for female-even-month workers andindicators for each easy round performance level indicates that male-even-month workers, relative to female-even-month workers, answer an average of 0.422 more questions correctly on the hard sports quiz (p = 0.464) and anaverage of 0.538 more questions correctly on the hard math quiz (p = 0.089).

16For raw data and summary information on employer choices over all decisions in the experiment, see AppendixTables A.3, A.4. and A.5.

15

either worker is the same: they receive 10 cents per correct answer on the hard quiz by the hired

worker. Employers also receive identical easy quiz performance information on both workers:

both workers are known to answer 4, 6, or 8 (out of 10) questions correctly on the easy quiz.17

If employers neither engage in belief-based nor taste-based discrimination, we expect that

they should hire female-even-month workers 50% of the time.18 This is not the case. Employers

in the Gender treatment only hire female workers in 43% of decisions, significantly below the

50% benchmark.

This discrimination is not specific to gender, however. In the Birth Month treatment, an

even-month worker is hired in just 37% of decisions, also significantly below the 50% benchmark.

Employers in both treatments discriminate against workers associated with the lower-performing

group, consistent with beliefs driving discrimination against women, rather than tastes.

Table 1 provides corresponding regression results, when clustering standard errors at the

employer-level, from specifications of the following form:

P(female-even-month worker is hired) = β1Gender Treatment

[+β2Posterior(easy gap)]

[+β3Posterior(hard-easy gap)]

+2∑

i=1

3∑j=1

vi × dj + ε (1)

where P(female-even-month worker is hired) is the probability with which a female-even-month

worker is hired, β1 captures the impact of being in the Gender Treatment, β2 and β3 capture the

coefficient estimates on controls for beliefs (see the table note of Table 1 for details), and vi are

dummies for the math versus sports version while dj are dummies for the performance levels of

the workers.

Column 1 of Table 1 Panel A confirms that taste-based considerations, if anything, work

in favor of women: female-even-month workers are 6 percentage points more likely to be hired

when they are labeled as female workers in the Gender treatment than when they are labeled as

even-month workers in the Birth Month treatment.

In Column 2 of Table 1 Panel A, we directly explore the role of beliefs. We add to the

regression two belief measures: the employer’s posterior belief of the average performance gap

in the easy quiz, and her imputed belief of “differential improvement” from the easy to the hard

17In the analysis below, we pool these decisions (4 v 4, 6 v 6, and 8 v 8). Results are quite similar for each ofthe three decisions independently; see Appendix Table A.6.

18We code the probability of a female-even-month worker being hired as 1 if she is hired with certainty, 0.5if chance determines who is hired, and 0 if a male-odd-month worker is instead hired with certainty. AppendixTable A.7 shows that our main results replicate if we consider other methods of classifying decisions across thesethree options.

16

quiz.19 Both measures are signed so that a positive gap indicates a belief of a male-odd-month

advantage. Column 2 shows that beliefs are highly predictive of decisions: employers who believe

the performance gap is larger are less likely to hire the worker from the lower-performing group.20

Columns 3 - 4 of Table 1 Panel A present results from when a “veil” for employers’ intentions

is provided. We do this via the addition of risk into the payoff from hiring a female-even-month

worker but not a male-odd-month worker. The key result is that female-even-month workers are

now hired at equal rates across the two treatments. The fact that, under risk, women are now

no more likely to be hired in the Gender treatment than in the Birth Month treatment suggests

that some of the female preference we observe in the stark, riskless environment may be driven

by image concerns.21 Indeed, the odd columns in Appendix Table A.10 show we do not observe

evidence for female preference (greater hiring of women in the Gender treatment than in the

Birth Month treatment) when separately examining our results for each level of risk. This is

true even when the risk is only a 1% chance of no payment from hiring the female-even-month

worker.

But, it is worth emphasizing here that women are still no less likely to be hired in the Gender

treatment than in the Birth Month treatment, even under risk. Thus, gender-based animus

against women does not arise even when a veil for employers’ intentions exists. It does not

appear to be the case that the absence of taste-based discrimination against women in riskless

decisions was only an artifact of image concerns or social desirability bias. Even when intentions

are veiled, women are no worse off being labeled as women than being labeled as even-month.

Taken together, our results may be summarized as follows. When women are the lower-

performing group, they are discriminated against even when their individual performances are as

good as their counterparts. This type of discrimination, however, is not specific to their gender.

Rather, it appears driven by beliefs. Female workers are treated no worse than even-month

workers, even when there is a veil for employers’ intentions.

After seeing these results, we were curious about the extent to which the observed discrimina-

tion against women was driven by the fact that participants receive ample, detailed information

about the average gender gap in performance in the easy quiz. We hypothesized that this informa-

tion may have helped participants feel justified in their decision to discriminate (either through

19Differential improvement is the idea that participants may believe that the easy quiz performance is dif-ferentially predictive of the hard quiz performance across the two groups. To measure this, we difference theparticipant’s belief of the hard quiz performance gap and the easy quiz performance gap, and include thisdifference-in-differences as a control. By including both differential improvement and posterior belief of theeasy quiz gap, we allow our specifications to depend on beliefs about both the easy and hard quiz. Our resultsare not sensitive to this particular construction. Results available upon request.

20Column 1 of Appendix Table A.8 shows that beliefs are not more predictive in the Gender treatment. So,while we cannot directly rule out that beliefs in the Gender treatment are more or less strongly held than beliefsin the Birth Month treatment, it is not the case that one set of beliefs is more predictive than the other.

21While the preference for hiring women relative to even-month is insignificant under risk, we cannot rejectthat the effects are the same in the risk and riskless decisions. See Columns 1 and 2 of Appendix Table A.9.

17

Table 1: Hiring decisions between workers with the same performance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4)

Panel A: Gender and Birth Month TreatmentsGender Treatment 0.061∗∗∗ 0.057∗∗∗ 0.026 0.024

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

Posterior(easy gap) -0.023∗∗∗ -0.018∗∗∗

(0.003) (0.003)

Posterior(hard-easy gap) -0.011∗∗ -0.011∗∗

(0.005) (0.004)Decision FEs yes yes yes yesObservations 2400 2400 12000 12000

Panel B: Gender and Gender-No-Information TreatmentsNo Information 0.002 0.004 0.005 0.008

(0.020) (0.019) (0.019) (0.019)

Posterior(easy gap) -0.021∗∗∗ -0.015∗∗∗

(0.004) (0.004)

Posterior(hard-easy gap) -0.014∗∗∗ -0.015∗∗∗

(0.005) (0.004)Decision FEs yes yes yes yesObservations 2409 2409 12045 12045

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions of the probability with which a female-even-month worker is hired over a male-odd-month worker, among employers making hiring decisions betweenpairs of workers who have the same performances. Gender Treatment is an indicator for employers in theGender treatment. Posterior(easy gap) is an employer’s posterior belief about the average performancegap in the easy quiz. Posterior(hard − easy gap) is an employer’s posterior belief about the averageperformance gap in the hard quiz minus the employer’s posterior belief about the average performance gapin the easy quiz. No Information is an indicator for employers in the No Information Stage version.Decisions FEs include 6 fixed effects (with one excluded) that capture whether hiring decisions are betweenworkers with equal performances on the sports quiz (4 v 4, 6 v 6, or 8 v 8) or between workers with equalperformances on the math quiz (4 v 4, 6 v 6, or 8 v 8). Columns 1 - 2 involve hiring decisions where thepayoff rule is always the same, while Columns 3 - 4 involve hiring decisions where the payoff rule is riskywhen hiring female-even-month workers and riskless when hiring male-odd-month workers. Panel A involveshiring decisions from the Birth Month and Gender treatments, while Panel B involves hiring decisions fromthe Gender and Gender-No-Information treatments.

experimenter demand, or because their beliefs are so well-informed). Would participants be

willing to discriminate without any information on the distributions of easy quiz performances?

To test this hypothesis, we ran a new No Information Stage version of the experiment that

simply removed the Information Stage for employers. All other aspects of the design were kept the

same, and we again recruited 800 individuals from MTurk. These results are presented in Panel

B of Table 1, both without risk (Columns 1 - 2) and with risk (Columns 3 - 4). We ask whether

18

female workers are hired as often in the Gender treatment of the No Information Stage version

as in our baseline Gender treatment. Column 1 shows that the level of discrimination against

women is identical without the Information Stage; this remains true when we control for beliefs

in Column 2, and when we consider the risky decisions in Columns 3 and 4. Providing ample

information about differences by gender neither promotes nor inhibits discrimination against

women above and beyond what we find when employers only act on their priors. Employers are

willing to discriminate even based on arguably less informed beliefs.22

3.3 Does the domain matter?

Employers in our experiment are randomized into one of two domain conditions: a math

condition or a sports condition. Both math and sports have been perceived as domains of male

advantage by past experimental participants who completed similar quizzes (Coffman, 2014;

Bordalo et al., 2019). In this sense, we would expect belief-based discrimination against women

in both domains. However, consistent with preliminary investigations by Bordalo et al. (2019),

we hypothesized that employers may feel more at ease discriminating in a domain like sports

as compared to math, as it is arguably less academically relevant, less prestigious, and less ego

relevant. Put differently, the image concerns associated with discriminating against women in

math might be much stronger than those associated with discriminating against women in sports.

We pre-tested this hypothesis empirically among the workers who completed our preliminary

study. After completing all of the quizzes, participants in the preliminary study were asked a

question about the social appropriateness of openly expressing beliefs about performance dif-

ferences between men and women, in either math or in sports. In particular, they are told:

imagine that some individual held the following belief: on average, it is likely that men per-

formed better than women in the [math/sports] questions. They are then asked: how socially

appropriate would it be for that individual to share that belief with others? They could select,

very socially inappropriate (1), somewhat socially inappropriate (2), somewhat socially appro-

priate (3), or very socially appropriate (4). Following Krupka and Weber (2013), we incentive

these beliefs by providing a bonus payment if participants provide the modal answer provided

by other participants. Thus, these answers can be interpreted as participants’ beliefs about the

norms surrounding the appropriateness of beliefs of gender differences in these domains.

Among our sample of 100 performance pool participants, we see significant differences in these

assessments for math versus sports. The average rating for sports is 3.15, somewhere between

somewhat socially appropriate and very socially appropriate. The average rating for math, on

the other hand, is 2.35, closer to somewhat socially inappropriate (p<0.001 for the across-domain

22We present a replication of our main beliefs figure, Figure 4, for the No Information Stage version as AppendixFigure A.2. While employers in the Gender treatment still believe there is a significant male advantage inperformance absent information, not surprisingly, there is no significant believed difference in performance in theBirth Month treatment absent our information. Quite understandably then, discrimination against even-monthworkers does not persist in the No Information Stage version (see Appendix Figure A.1).

19

difference). This is consistent with our hypothesis that image concerns related to discriminating

against women in math may be more severe than those associated with discriminating in sports.

Taking this hypothesis to our hiring experiment, we predict that individuals will be more

reluctant to discriminate against women in math than in sports in our Gender treatment, par-

ticularly without the veil of risk. We expect no differences across math and sports in the Birth

Month treatment. In Figure 5, we show the average rate at which the female-even-month worker

is hired in riskless decisions, when both workers have the same performance. We split the data

both by treatment (Gender versus Birth Month) and also by domain (math versus sports).

There is, on average, discrimination against the female-even-month worker across both domains

and treatments. But, consistent with our hypothesis, we see significantly more discrimination

against the female-even-month worker in sports than in math within the Gender treatment. On

the other hand, and again consistent with our hypothesis, the level of discrimination against the

female-even-month worker in the Birth Month treatment is very similar across sports and math.

Figure 5: Math vs Sports Version: Hiring decisions between workers with same performance andpayoff rule

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

GenderTreatment

Birth MonthTreatment

(a) Math Version

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

GenderTreatment

Birth MonthTreatment

(b) Sports Version

We formalize this analysis in Table 2, where we reproduce Table 1 separately for math and

sports. Columns (1) and (2) consider riskless decisions. In math, we see a strong impact of the

Gender treatment, with the female-even-month worker being significantly more likely to be hired

when she is labeled as female rather than even-month. This points to considerable taste-based

preferences in favor of women in the math domain, even conditional on stated beliefs (Column 2).

In sports, we see that it is still the case that women are, if anything, more likely to be hired under

the Gender treatment than under the Birth Month treatment, but the effect is much smaller and

is statistically insignificant. This suggests that there is much less “affirmative action” in sports

20

than in math.

Table 2: Comparing the math and sports versions: hiring decisions between workers with thesame performance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4)

Panel A: Math VersionGender Treatment 0.102∗∗∗ 0.084∗∗∗ 0.071∗∗∗ 0.061∗∗

(0.027) (0.026) (0.027) (0.027)

Posterior(easy gap) -0.028∗∗∗ -0.020∗∗∗

(0.005) (0.005)

Posterior(hard-easy gap) -0.015∗ -0.016∗∗

(0.008) (0.008)Constant 0.379∗∗∗ 0.433∗∗∗ 0.173∗∗∗ 0.213∗∗∗

(0.021) (0.021) (0.017) (0.020)Decision FEs yes yes yes yesObservations 1200 1200 6000 6000

Panel B: Sports VersionGender Treatment 0.021 0.026 -0.019 -0.014

(0.026) (0.026) (0.025) (0.025)

Posterior(easy gap) -0.018∗∗∗ -0.016∗∗∗

(0.004) (0.005)

Posterior(hard-easy gap) -0.008 -0.008(0.005) (0.005)

Constant 0.361∗∗∗ 0.404∗∗∗ 0.208∗∗∗ 0.246∗∗∗

(0.020) (0.020) (0.018) (0.021)Decision FEs yes yes yes yesObservations 1200 1200 6000 6000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions of the probability with which a female-even-month worker is hired over a male-odd-month worker, among employers making hiring decisions betweenpairs of workers who have the same performances. Gender Treatment is an indicator for employers in theGender treatment. Posterior(easy gap) is an employer’s posterior belief about the average performancegap in the easy quiz. Posterior(hard − easy gap) is an employer’s posterior belief about the averageperformance gap in the hard quiz minus the employer’s posterior belief about the average performance gapin the easy quiz. Decisions FEs include 3 fixed effects (with one excluded) that capture whether hiringdecisions are between workers with equal performances of 4 v 4, 6 v 6, or 8 v 8. Columns 1 - 2 involvehiring decisions where the payoff rule is always the same, while Columns 3 - 4 involve hiring decisions wherethe payoff rule is risky when hiring female-even-month workers and riskless when hiring male-odd-monthworkers. Panel A involves hiring decisions from the math version, while Panel B involves hiring decisionsfrom the sports version.

21

When looking at our pooled analysis, we concluded that female-even-month workers were

significantly less likely to be hired than male-odd-month workers on average, conditional on

having identical performances. And, this discrimination was driven by beliefs, not gender-specific

tastes: female-even-month workers were no less likely to be hired when labeled as female than

when labeled as even-month. Splitting the analysis by math and sports reinforces these findings,

as these results replicate within both individual domains. The across-domain differences that

do emerge provide additional insights into our results. The taste-based preference that seems

to work in favor of women (revealed by the Gender versus Birht Month comparison) is much

more prominent in math than in sports, consistent with the idea that employers may be more

reluctant to discriminate against women in areas where there are stronger social norms against

it. Just as employers’ behavior seems to reveal that they think it is more “okay” to discriminate

against even-month workers than female workers, their behavior also seems to reveal that is it

more “okay” to discriminate against women in sports than in math.

In Columns (3) and (4), we consider the risky decisions, again parelleling the analysis of Table

1. Across both math and sports, the introduction of risk directionally reduces the size of the

Gender treatment effect, but not by a significant amount. In math, even with the introduction

of risk, women are still significantly more likely to be hired in the Gender treatment than in

the Birth Month treatment. In sports, there continues to be no significant differences across the

Gender and Birth Month treatments.

3.4 Does the gender of the employer matter?

Figure 6 shows that there is substantial heterogeneity in the extent of discrimination when

considering the four types of employers in our experiment as defined by their gender and their

birth month. In the Gender treatment, male employers engage in substantial discrimination

against female workers, hiring them less than 40% of the time. By contrast, and consistent

instead with no discrimination, female employers hire female workers approximately 50% of the

time.23 Thus, in our setting, men discriminate against women more than women do when gender

is known.

But, what drives this pattern? Is this something specific to gender? Our Birth Month

treatment reveals that this result appears to be driven by in-group preferences. Just as men

hire women less often than women do, odd-month employers hire even-month workers less often

than even-month employers do. While odd-month employers hire even-month workers 32% of

the time, even-month employers hire even-month workers 41% of the time. Thus, across both

treatments, there is less discrimination by in-group employers (female employers in the Gender

treatment, even-month employers in the Birth Month treatment).

23Past evidence on whether female employers are in general more likely to hire female workers is mixed, see forinstance Bagues and Esteve-Volart (2010) and De Paola and Scoppa (2015). In recent work, Cappelen, Falch andTungodden (2019) find that women are more likely to discriminate against low-performing males than men are.

22

Figure 6: Hiring decisions between workers with same performance and payoff rule0

.1.2

.3.4

.5Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Gender Treatment

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Birth Month Treatment

In other words, it is not simply that female employers discriminate less often than male

employers against women. Instead, employers discriminate less often against “in-group” members

more broadly. This is consistent with previous work on in-group preferences that documents that

shared social identities can impact economic outcomes, including studies using both minimal

group paradigms and natural identities (see Tajfel et al. (1971), Brewer (1979), Chen and Li

(2009), Chen and Chen (2011), and Chen et al. (2014)).

We explore this pattern in Table 3. We follow exactly the specifications of Table 1, but start

by replacing the treatment dummy with an in-group dummy (even-month employer in the Birth

Month treatment, or female employer in the Gender treatment) in Column 1. In Columns 2

and 3, we add back in our main treatment dummy and interact it with the in-group dummy,

to ask whether the extent of in-group preferences varies by treatment. Column 1 documents

statistically significant in-group preferences, while Column 2 shows that the extent of in-group

preferences does not vary across treatment. In-group preferences are no stronger in the Gender

treatment than in the Birth Month treatment.

One could ask whether these in-group preferences are well-explained by differences in beliefs

by type of employer. In-group status does appear to influence posterior beliefs. In particular,

within the Gender treatment, male employers on average believe the male advantage in perfor-

mance on the easy quiz is 2.39 points on average, while women believe it is just 1.16 points on

average. Similarly, within the Birth Month treatment, odd-month employers believe the odd-

month advantage in performance on the easy quiz is 2.48 points on average, while even-month

employers believe it is 1.91 points on average. However, these differences in beliefs do not explain

the extent of in-group preferences we observe in our hiring decisions. This can be seen in Column

23

3, where we add the full set of employer beliefs to the regression model. While beliefs are once

again strongly predictive of employer decisions, we see significant in-group preferences on top of

stated beliefs.

Table 3: Hiring decisions between workers with the same performance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4) (5) (6)

In-Group 0.098∗∗∗ 0.090∗∗∗ 0.081∗∗∗ 0.012 0.008 0.001(0.019) (0.026) (0.025) (0.019) (0.025) (0.024)

Gender Treatment 0.055∗∗ 0.057∗∗ 0.022 0.025(0.025) (0.025) (0.025) (0.025)

Gender Treatment*In-Group 0.020 0.007 0.009 -0.002(0.037) (0.036) (0.038) (0.037)

Posterior(easy gap) -0.021∗∗∗ -0.018∗∗∗

(0.003) (0.003)

Posterior(hard-easy gap) -0.011∗∗ -0.011∗∗

(0.005) (0.004)Decision FEs yes yes yes yes yes yesObservations 2400 2400 2400 12000 12000 12000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shownin parentheses. The results are from ordinary least squares regressions of the probability with whicha female-even-month worker is hired over a male-odd-month worker, among employers making hiringdecisions between pairs of workers who have the same performances. In-Group is an indicator foremployers who know they share some demographic characteristic with the female-even-month workers(i.e., even-month employers in the Birth Month treatment and female employers in the Gender treatment).Gender Treatment is an indicator for employers in the Gender treatment. Posterior(easy gap) isan employer’s posterior belief about the average performance gap in the easy quiz. Posterior(hard −easy gap) is an employer’s posterior belief about the average performance gap in the hard quiz minusthe employer’s posterior belief about the average performance gap in the easy quiz. Decisions FEs include6 fixed effects (with one excluded) that capture whether hiring decisions are between workers with equalperformances on the sports quiz (4 v 4, 6 v 6, or 8 v 8) or between workers with equal performances onthe math quiz (4 v 4, 6 v 6, or 8 v 8). Columns 1 - 3 involve hiring decisions where the payoff rule isalways the same, while Columns 4 - 6 involve hiring decisions where the payoff rule is risky when hiringfemale-even-month workers and riskless when hiring male-odd-month workers. All hiring decisions arefrom the Birth Month and Gender treatments.

Interestingly, however, in our setting, in-group preferences appear quite malleable. When

provided with a veil on their intentions via the introduction of risk, employers do not demonstrate

in-group preferences (see Columns 4 - 6 of Table 2, and the odd columns in Appendix Table

A.10 to examine the results separately for each level of risk).24 This suggests that the in-group

24See Columns 3 and 4 of Appendix Table A.9 for an interacted model.

24

preferences in the stark environment of the riskless decisions is likely reflective of image concerns,

including potentially experimenter demand. Alternatively, it may be that in an environment

with limited reasons for selecting one worker over another (i.e. the two workers have identical

easy-round performances), in-group preferences serve as a “tie-breaker” of sorts. But, with the

introduction of risk, this tie-breaker is less relevant, significantly reducing the role of in-group

preferences.25

3.5 Do our results extend to other decisions?

In Sections 3.2, 3.3, and 3.4, we focus on hiring decisions between two workers with identical

easy round performances. It is in these decisions that we see clear evidence of discrimination:

differential treatment by employers of two employees with identical performance information. In

this section, we examine whether our results extend to additional hiring decisions where: (i) the

individual performances of the two workers is uncertain, or (ii) the female-even-month worker

has a better easy round performance than the male-odd-month worker does.

The first set of additional hiring decisions involves the 12 decisions that employers make

as part of the Information Stage. In each of these decisions, the distribution of easy round

performances for some subset of female-even-month workers and the distribution of easy round

performances for some subset of male-odd-month workers are displayed. Employers then choose

between a female-even-month worker and male-odd-month worker, knowing that their worker will

be randomly drawn from the subset of workers comprising the relevant, displayed distribution.

As shown in Appendix Table A.1, these displayed distributions, on average, largely favor male-

odd-month workers. This is not surprising given the true underlying average performance gap.

Out of the 24 distributions that are displayed in the math version or in the sports version, there

are only 3 sets of distributions that favor the female-even-month workers and 1 set of distributions

that neither favors the male-odd-month workers nor the female-even-month workers in terms of

average performance difference within the subset.

Here, we focus on the decisions over the 20 distributions for which the average performance

gap favors the male-odd-month workers.26 The corresponding decisions can be summarized as

follows (and see Appendix Figure A.3 and Appendix Table A.12 for reference). First, employ-

ers are less likely to hire from the lower-performing distribution: female workers in the Gender

treatment are hired 33% of the time when the average performance gap of the displayed distri-

butions favors male workers. Second, this lower hiring rate is not specific to gender: even-month

workers in the Birth Month treatment are hired 29% of the time, significantly lower than the

25If we breakdown the data by domain, we document significant in-group preferences in both math and sports,but again, only in the riskless decisions. Within each domain, the introduction of risk completely eliminatesin-group preferences. See Appendix Table A.11.

26As one would expect, the results are much noisier when only considering the 3 distributions for which theaverage performance gap favors the female-even-month worker (see Appendix Figure A.4 and Appendix TableA.13) and the 1 distribution for which there is no average performance gap (see Appendix Figure A.5 andAppendix Table A.14).

25

33% of the time that female workers are hired in the Gender treatment. This suggests that our

main result replicates in this environment: there is less hiring of women than men, but it is not

driven by gender-specific animus. Third, as in our riskless decisions over workers with identical

performances, evidence for in-group preferences is substantial and significant. Female employers

are 16 percentage points more likely to hire female workers in the Gender treatment than male

employers are, and even-month employers are 6 percentage points more likely to hire even-month

workers in the Birth Month treatment than odd-month employers are. Indeed, the slight prefer-

ence for hiring female, as compared to even-month, workers appears entirely driven by stronger

in-group preferences within the Gender treatment for these Information Stage decisions.

The second set of additional hiring decisions involves employers in the Hiring Stage choos-

ing between pairs of workers in which the female-even-month worker has a better easy round

performance than the male-odd-month worker. The corresponding decisions can be summarized

as follows (and see Appendix Figure A.6 as well as Appendix Table A.15 for reference). First,

female-even-month workers are hired nearly 90% of the time in both of our treatments when the

decision does not involve risk, suggesting that any belief-based discrimination against female-

even-month workers is not so strong as to induce employers not to hire a worker with a strictly

better easy round performance. Second, we again do not observe any evidence consistent with

taste-based animus against female workers: female workers in the Gender treatment are not less

likely to be hired than even-month workers in the Birth Month treatment. Third, while rates

of hiring the female-even-month worker fall with the introduction of risk (just as in the case

of equal performances), the female-even-month worker is still hired no less often in the Gender

treatment than in the Birth Month treatment even under the risky decisions. And, finally, we do

not observe any significant differences according to the demographics of the employer, suggesting

that in-group preferences are not so strong as to overcome a desire to hire the worker with the

better easy round performance.

We also document the robustness of our findings by running an additional, simplified exper-

iment. We describe this experiment and its results in detail in Appendix B. In this simplified

experiment, employers only make a single hiring decision in the Hiring Stage, and we introduce

risk as an “excuse” not to hire across-employer, rather than within-employer. Our results are

quite consistent across the simplified and original hiring experiments. In particular, female-even-

month workers are more likely to be hired under the Gender treatment than under the Birth

Month treatment. And, this continues to be true when we consider the risky decisions, when

there is a veil on employer intentions. Finally, we find at best weak evidence of in-group pref-

erences, even within the riskless decisions in the simplified experiment, again suggesting that

in-group preferences are much less robust than our other findings.

Thus, our conclusions drawn from across the many decisions of the original experiment and

from the new simplified experiment are remarkably consistent. Discrimination against female

26

workers appears to be driven by beliefs, as quite consistently, women are not significantly less

likely to be hired when labeled as women than when labeled as even-month workers. This is true

even when image concerns are largely reduced, through the introduction of a veil on employers’

intentions. Our results suggest that understanding and reducing discrimination requires a clear

focus on addressing beliefs.

4 ConclusionWe leverage a controlled environment in order to disentangle the motivations behind discrim-

ination against women in a male-typed employment environment. We find ample evidence of

discrimination against women, as employers, on average, are significantly less likely to hire a

woman compared to an equally able man.

This discrimination, however, is not specific to gender. When workers are identified as mem-

bers of one of two groups, we observe similar levels of discrimination against the lower-performing

group regardless of whether they are identified according to gender or birth month. This ev-

idence points to an important role for beliefs about performance in explaining discriminatory

practices. Theories of stereotyping and belief formation, like those of Bordalo et al. (2016), may

be particularly important for understanding and reducing discrimination.

We also provide evidence on in-group preferences, finding that male employers, more so than

female employers, discriminate against female workers in the Gender treatment. Again however,

this finding is not specific to gender. Odd-month employers, more so than even-month employers,

discriminate against even-month workers. This evidence highlights how discrimination can be

mitigated by in-group preferences or activated by out-group animus. We caution, however, that

when employers are provided with a veil on their intentions, the addition of risk to the decision, we

no longer observe a significant role for in-group preferences. The evidence for in-group preferences

is quite mixed across the many decisions we consider.

The extent to which these drivers of discrimination are relevant in other settings is almost

certainly context-dependent. In thinking about how our results are likely to generalize to other

settings, it is critical to think about the centrality of beliefs. If we move to an environment or

paradigm in which an employer is likely to believe that women outperform men on average, our

results and framework would predict discrimination against men, again driven by beliefs.

The value in our paper is showing both how and why it is important to construct careful

control comparisons to narrow in on the drivers of discrimination. More specifically, there are

three main takeaways from our approach that could help to inform future investigations on

the drivers of gender discrimination. First, providing ample information on the distribution of

performances for a group — but varying whether the gender of that group is known — is a

powerful way to examine the role of beliefs related to average group differences. Average group

differences are central to many theories of discrimination; collecting data that speaks directly to

their importance provides clearer evidence on these theories. Second, labeling a control group by

27

an arbitrary characteristic — rather than simply removing a gender label — allows one to narrow

in on gender-specific channels. It prevents one from attributing taste-based considerations that

are driven by in-group preferences to gender. Third, introducing a veil for intentions by building

off of the approach in Exley (2015) — given that decisions outside of the laboratory frequently

involve such noise or plausible veils — allows one to speak to the robustness of findings even

in rather abstract environments (e.g., by minimizing the role of image concerns, including those

related to social desirability or experimenter demand concerns, in driving results).

We conclude with four additional points that may be informative to future work as well. First,

while using birth month to label groups may be an unusual approach, it serves our purposes well.

While association with birth month is not a commonly used feature outside of the laboratory, we

note that other group associations based upon school affiliation, section or classroom assignment,

workplace group, office location, or team are quite common. Our paradigm suggests that these

associations could play a non-trivial role in how students and workers from these groups are

evaluated, particularly when other information and potential veils are unavailable.

Second, our findings show that beliefs and tastes may work in opposite directions in driving

discrimination. When considering the overall results in our context, beliefs push employers to

hire women less often, while tastes, if anything, seem to push employers to hire women more

often. This is particularly true in the arguably more career relevant domain of math as compared

to sports. If this pattern persists in other settings, it would make it hard to draw conclusions

about the drivers of discrimination without detailed data on beliefs.

Third, our findings underscore the challenges faced by members of a group that is believed

to be lower-performing on average. Mere membership in a lower-performing group — even when

this membership is outside of the control of the worker and based on an arbitrary characteristic

— is sufficient for discrimination to follow. This challenge, moreover, may be exacerbated for

members of a group that is not well-represented among decision-makers. For instance, had all

employers been males in our study, the overall level of discrimination against female workers

would have been greater.

Finally, it is worth emphasizing that we find evidence for discrimination against women

even in a stark environment. In our riskless decisions, employers make a side-by-side decision

between a male worker and a female worker with identical resumes, there is little ambiguity to

use to justify discrimination, and it is quite cheap to not appear sexist. And, yet, we find that

employers, on average, are willing to discriminate against women. This suggests that belief-

based discrimination is common. While there has been increasing social consensus about the

inappropriateness of animus-driven discrimination, belief-based discrimination has received far

less social scrutiny.

Finding effective remedies remains an important challenge for social scientists, policy-makers,

and organizational leaders. While future work is clearly needed on this issue, we offer a few sug-

28

gestions here. Our results suggest that beliefs about average group differences are an important

driver of discriminatory behavior. Thus, designing processes that reduce reliance on beliefs about

candidates that are informed largely by group membership, and not individual characteristics,

should help to limit the extent to which belief-based discrimination occurs in hiring and other

workplace decision-making. This could mean collecting more, and more objective, information

from each candidate. Acknowledgement of and reflection on beliefs may also be helpful: ask,

what beliefs do I hold about the capabilities of this candidate, what informs those beliefs, and

would I hold those same beliefs if this candidate were of another gender (race, socioeconomic

status, etc.)?

29

ReferencesAltonji, Joseph G., and Rebecca M. Blank. 1999. “Chapter 48 Race and gender in the

labor market.” In . Vol. 3 of Handbook of Labor Economics, 3143 – 3259. Elsevier.

Arrow, Kenneth. 1973. “Discrimination in Labor Markets.” , ed. Orley Ashenfelter and Albert

Rees, Chapter The Theory of Discrimination. Princeton, NJ:Princeton University Press.

Ayres, Ian, and Peter Siegelman. 1995. “Race and gender discrimination in bargaining for

a new car.” The American Economic Review, 304–321.

Azmat, Ghazala, and Barbara Petrongolo. 2014. “Gender and the labor market: What

have we learned from field and lab experiments?” Labour Economics, 30: 32–40.

Baert, Stijn, Ann-Sophie De Pauw, and Nick Deschacht. 2016. “Do employer preferences

contribute to sticky floors?” ILR Review, 69(3): 714–736.

Bagues, Manuel F, and Berta Esteve-Volart. 2010. “Can gender parity break the glass

ceiling? Evidence from a repeated randomized experiment.” The Review of Economic Studies,

77(4): 1301–1328.

Bartos, Vojtech, Michal Bauer, Julie Chytilova, and Filip Matejka. 2016. “Attention

Discrimination: Theory and Field Experiments with Monitoring Information Acquisition.”

American Economic Review, 106(6): 1437–1475.

Becker, Gary S. 1957. The Economics of Discrimination. Chicago, IL:University of Chicago

Press.

Bertrand, Marianne. 2011. “New perspectives on Gender.” Handbook of Labor Economics,

4: 1543–1590.

Bertrand, Marianne, and Esther Duflo. 2017. “Field experiments on discrimination.” Hand-

book of Economic Field Experiments, 1: 309–393.

Bertrand, Marianne, and Sendhil Mullainathan. 2004. “Are Emily and Greg more em-

ployable than Lakisha and Jamal? A field experiment on labor market discrimination.” The

American Economic Review, 94(4): 991–1013.

Black, Sandra E, and Philip E Strahan. 2001. “The division of spoils: rent-sharing and

discrimination in a regulated industry.” The American Economic Review, 91(4): 814–831.

Blau, Francine D, and Lawrence M Kahn. 2017. “The gender wage gap: Extent, trends,

and explanations.” Journal of Economic Literature, 3: 789–865.

30

Bohnet, Iris, Alexandra van Geen, and Max Bazerman. 2016. “When Performance

Trumps Gender Bias: Joint Versus Separate Evaluation.” Management Science, 62(5): 1225–

1234.

Bohren, J Aislinn, Alex Imas, and Michael Rosenberg. 2017. “The Dynamics of Discrim-

ination: Theory and Evidence.” Working Paper.

Bordalo, Pedro, Katherine B. Coffman, Nicola Gennaioli, and Andrei Shleifer. 2019.

“Beliefs about Gender.” American Economic Review, 109(3): 739–773.

Bordalo, Pedro, Katherine Coffman, Nicola Gennaioli, and Andrei Shleifer. 2016.

“Stereotypes.” The Quarterly Journal of Economics, 131(4): 1753–1794.

Bowles, Hannah Riley, Linda Babcock, and Lei Lai. 2007. “Social incentives for gender

differences in the propensity to initiate negotiations: Sometimes it does hurt to ask.” Organi-

zational Behavior and Human Decision Processes, 103(1): 84–103.

Brewer, Marilynn B. 1979. “In-group bias in the minimal intergroup situation: A cognitive-

motivational analysis.” Psychological bulletin, 86(2).

Cappelen, Alexander W., Ranveig Falch, and Bertil Tungodden. 2019. “The Boy Crisis:

Experimental Evidence on the Acceptance of Males Falling Behind.” Working Paper.

Card, David, Ana Rute Cardoso, and Patrick Kline. 2016. “Bargaining, sorting, and the

gender wage gap: Quantifying the impact of firms on the relative pay of women.” Quarterly

Journal of Economics, 131(2): 633–686.

Castillo, Marco, and Ragan Petrie. 2010. “Discrimination in the lab: Does information

trump appearance?” Games and Economic Behavior, 68.

Castillo, Marco, Ragan Petrie, Maximo Torero, and Lise Vesterlund. 2013. “Gender

differences in bargaining outcomes: A field experiment on discrimination.” Journal of Public

Economics, 99: 35–48.

Charles, Kerwin Kofi, and Jonathan Guryan. 2008. “Prejudice and Wages: An Empiri-

cal Assessment of Becker’s The Economics of Discrimination.” Journal of Political Economy,

116(5): 773–809.

Charles, Kerwin Kofi, and Jonathan Guryan. 2011. “Studying discrimination: Fundamen-

tal challenges and recent progress.” Annual Review of Economics, 3(1): 479–511.

Chen, Roy, and Yan Chen. 2011. “The potential of social identity for equilibrium selection.”

The American Economic Review, 101(6): 2562–2589.

31

Chen, Yan, and Sherry Xin Li. 2009. “Group identity and social preferences.” The American

Economic Review, 99(1): 431–457.

Chen, Yan, Sherry Xin Li, Tracy Xiao Liu, and Margaret Shih. 2014. “Which hat to

wear? Impact of natural identities on coordination and cooperation.” Games and Economic

Behavior, 84: 58–86.

Coffman, Katherine Baldiga. 2014. “Evidence on Self-Stereotyping and the Contribution of

Ideas.” The Quarterly Journal of Economics, 129(4): 1625–1660.

Croson, Rachel, and Uri Gneezy. 2009. “Gender Differences in Preferences.” Journal of

Economic Literature, 47(2): 448–474.

Dana, Jason, Roberto A. Weber, and Jason Xi Kuang. 2007. “Exploiting moral wig-

gle room: experiments demonstrating an illusory preference for fairness.” Economic Theory,

33: 67–80.

Danilov, Anastasia, and Silvia Saccardo. 2017. “Disguised Discrimination.” Working Paper.

De Paola, Maria, and Vincenzo Scoppa. 2015. “Gender discrimination and evaluators’

gender: evidence from Italian academia.” Economica, 82(325).

Dickinson, David L., and Ronald L. Oaxaca. 2009. “Statistical Discrimination in Labor

Markets: An Experimental Analysis.” Southern Economic Journal, 76(1): 16–31.

Dickinson, David L, and Ronald L. Oaxaca. 2014. “Wages, Employment, and Statistical

Discrimination Evidence from the laboratory.” Economic Inquiry, 52(4): 1380–1391.

Doleac, Jennifer L., and Luke C.D. Stein. 2013. “The Visible Hand: Race and Online

Market Outcomes.” The Economic Journal, 123.

Exley, Christine L. 2015. “Excusing Selfishness in Charitable Giving: The Role of Risk.”

Review of Economic Studies, 83(2): 587–628.

Fershtman, Chaim, and Uri Gneezy. 2001. “Discrimination in a segmented society: An

experimental approach.” The Quarterly Journal of Economics, 116(1): 351–377.

Gino, Francesca, Michael I. Norton, and Roberto A. Weber. 2016. “Motivated Bayesians:

Feeling Moral While Acting Egoistically.” Journal of Economic Perspectives, 30(3): 189–212.

Gneezy, Uri John List, Michael K. Price. 2012. “Toward an Understanding of Why People

Discriminate: Evidence from a Series of Natural Field Experiments.” NBER Working Paper.

32

Goldin, Claudia. 2014. “A Grand Gender Convergence: Its Last Chapter.” American Economic

Review, 104(4): 1091–1119.

Goldin, Claudia, and Cecilia Rouse. 2000. “Orchestrating Impartiality: The Impact of”

Blind” Auditions on Female Musicians.” The American Economic Review, 90(4): 715–741.

Guryan, Jonathan, and Kerwin Kofi Charles. 2013. “Taste-based or Statistical Discrim-

ination: The Economics of Discrimination Returns to its Roots.” The Economic Journal,

123(572).

Haisley, Emily C., and Roberto A. Weber. 2010. “Self-serving interpretations of ambiguity

in other-regarding behavior.” Games and Economic Behavior, 68: 614–625.

Halberda, Justin, Ryan Ly, Jeremy B Wilmer, Daniel Q Naiman, and Laura Ger-

mine. 2012. “Number sense across the lifespan as revealed by a massive Internet-based sam-

ple.” Proceedings of the National Academy of Sciences, 109(28): 11116–11120.

Krupka, Erin L., and Roberto A. Weber. 2013. “Identifying social norms using coordi-

nation games: Why does dictator game sharing vary?” Journal of the European Economic

Association, 11(3): 495–524.

Kunda, Ziva. 1990. “The Case for Motivated Reasoning.” Psychological Bulletin, 108(3): 480–

498.

Lei, Lai, and Linda C. Babcock. 2013. “Asian Americans and workplace discrimination: The

interplay between sex of evaluators and the perception of social skills.” Journal of Organiza-

tional Behavior, 34(3): 310–326.

List, John A. 2004. “The nature and extent of discrimination in the marketplace: Evidence

from the field.” The Quarterly Journal of Economics, 119(1): 48–89.

Milkman, Katherine L., Modupe Akinola, and Dolly Chugh. 2012. “Temporal Distance

and Discrimination: An Audit Study in Academia.” Psychological Science, 23(7): 710–717.

Milkman, Katherine L, Modupe Akinola, and Dolly Chugh. 2015. “What happens be-

fore? A field experiment exploring how pay and representation differentially shape bias on the

pathway into organizations.” Journal of Applied Psychology, 100(6).

Mobius, Markus M, and Tanya S Rosenblat. 2006. “Why beauty matters.” American

Economic Review, 96(1): 222–235.

Moss-Racusin, Corinne A, John F Dovidio, Victoria L Brescoll, Mark J Graham,

and Jo Handelsman. 2012. “Science faculty’s subtle gender biases favor male students.”

Proceedings of the National Academy of Sciences, 109(41): 16474–16479.

33

Neumark, David, Roy J. Bank, and Kyle D. Van Nort. 1996. “Sex Discrimination in

Restaurant Hiring: An Audit Study.” Quarterly Journal of Economics, 111(3): 915–941.

Niederle, Muriel. 2016. “Gender.” In Handbook of Experimental Economics. Vol. 2, , ed. John

Kagel and Alvin E. Roth, 481–553. Princeton University Press.

Olivetti, Claudia, and Barbara Petrongolo. 2016. “The Evolution of Gender Gaps in In-

dustrialized Countries.” Annual Review of Economics, 8(1): 405–434.

Paolacci, Gabriele, Jesse Chandler, and Panagiotis G. Ipeirotis. 2010. “Running exper-

iments on amazon mechanical turk.” Judgment and Decision Making, 5(5): 411–419.

Phelps, Edmund S. 1972. “The statistical theory of racism and sexism.” The American Eco-

nomic Review, 62(4): 659–661.

Rand, David G. 2012. “The promise of Mechanical Turk: How online labor markets can help

theorists run behavioral experiments.” Journal of theoretical biology, 299: 172–179.

Rao, Gautam. 2018. “Familiarity Does Not Breed Contempt: Generosity, Discrimination and

Diversity in Delhi.” Working Paper.

Reuben, Ernesto, Paola Sapienza, and Luigi Zingales. 2014. “How stereotypes impair

women’s careers in science.” Proceedings of the National Academy of Sciences, 111(12): 4403–

4408.

Riach, P. A., and J. Rich. 2002. “Field Experiments of Discrimination in the Market Place.”

The Economic Journal, 112(483).

Sarsons, Heather. 2017a. “Interpreting Signals in the Labor Market: Evidence from Medical

Referrals.” Working Paper.

Sarsons, Heather. 2017b. “Recognition for Group Work: Gender Differences in Academia†.”American Economic Review: Papers & Proceedings, 107(5): 141–145.

Tajfel, Henri, Michael G Billig, Robert P Bundy, and Claude Flament. 1971. “Social

categorization and intergroup behaviour.” European Journal of Social Psychology, 1(2): 149–

178.

34

Appendices

35

A Additional Results

Table A.1: Descriptive statistics on performance distributions seen in Part 1

Performance distribution in easy round of math quiz1 2 3 4 5 6 7 8 9 10 11 12

Avg Male Advantage 2.10 2.07 1.62 1.27 0.91 0.65 0.50 0.00 -0.38 -0.43 -0.80 0.69Men FOSD Women yes yes yes yes no no no no no no no noWorkers born on days 1-5 21-25 21-31 26-31 16-31 1-10 1-15 11-15 11-20 16-20 6-10 1-31Number of workers 9 8 19 11 30 16 20 4 15 11 7 50

Performance distribution in easy round of sports quiz1 2 3 4 5 6 7 8 9 10 11 12

Avg Male Advantage 4.08 2.50 1.34 2.25 2.08 1.85 0.73 0.69 0.72 0.50 0.24 1.00Men FOSD Women yes yes yes no no no no no no no no noWorkers born on days 21-31 11-15 16-20 21-31 6-10 16-31 1-10 1-15 11-20 21-27 1-5 1-31Number of workers 7 7 6 10 7 16 34 27 13 5 20 50

Avg Male Advantage is the average performance of male workers minus the average performance of female workers. MenFOSD Women indicates whether the performance of male workers first order stochastically dominates the performance of femaleworkers. Workers born on days indicates the days during which workers in a particular distribution were born. Number ofworkers indicates how many workers are involved in a particular distribution (note that employers are not aware of the samplesize for each distribution however). Results are sorted by FOSD, then Avg Male Advantage.

36

Table A.2: Beliefs about the performance gap between male-odd-months workers and female-even-month workers

Hard quiz performance gap Easy quiz performance gapPrior Posterior Prior − Prior Posterior Prior −

Posterior Posterior(1) (2) (3) (4) (5) (6)

Math and Sports VersionGender Treatment 2.344∗∗∗ 0.031 -2.313∗∗∗ 1.845∗∗∗ -0.368 -2.213∗∗∗

(0.177) (0.220) (0.256) (0.159) (0.225) (0.243)

Constant 0.387∗∗∗ 2.121∗∗∗ 1.734∗∗∗ 0.327∗∗∗ 2.209∗∗∗ 1.882∗∗∗

(0.095) (0.153) (0.183) (0.090) (0.167) (0.181)Observations 800 800 800 800 800 800

Math VersionGender Treatment 0.997∗∗∗ -0.354 -1.352∗∗∗ 0.597∗∗∗ -0.906∗∗∗ -1.502∗∗∗

(0.229) (0.294) (0.358) (0.194) (0.296) (0.316)

Constant 0.389∗∗∗ 1.899∗∗∗ 1.510∗∗∗ 0.374∗∗∗ 1.980∗∗∗ 1.606∗∗∗

(0.136) (0.217) (0.273) (0.124) (0.222) (0.243)Observations 400 400 400 400 400 400Sports VersionGender Treatment 3.705∗∗∗ 0.425 -3.280∗∗∗ 3.105∗∗∗ 0.180 -2.925∗∗∗

(0.234) (0.322) (0.359) (0.222) (0.331) (0.365)

Constant 0.385∗∗∗ 2.340∗∗∗ 1.955∗∗∗ 0.280∗∗ 2.435∗∗∗ 2.155∗∗∗

(0.134) (0.215) (0.245) (0.131) (0.248) (0.268)Observations 400 400 400 400 400 400

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from an ordinary least squares model of the beliefs about the extent of theperformance gap between male-odd-month workers and female-even-month workers. Gender Treatment isan indicator for when workers are labeled according to their gender as opposed to birth month. Results arefrom the Birth Month and Gender treatments.

37

Table A.3: By performance and risk levels: probability of hiring female-even-month worker

When The Level of Risk Is0% 1% 5% 10% 25% 50%

Gender Treatment4 v 4 performances 0.44 0.30 0.26 0.23 0.17 0.154 v 5 performances 0.85 0.69 0.62 0.51 0.31 0.234 v 6 performances 0.90 0.83 0.78 0.71 0.45 0.284 v 7 performances 0.91 0.88 0.83 0.80 0.62 0.374 v 8 performances 0.92 0.88 0.87 0.85 0.71 0.496 v 6 performances 0.41 0.29 0.27 0.23 0.16 0.146 v 7 performances 0.81 0.67 0.60 0.52 0.29 0.246 v 8 performances 0.92 0.83 0.80 0.72 0.51 0.318 v 8 performances 0.44 0.30 0.25 0.23 0.17 0.14Birth Month Treatment4 v 4 performances 0.37 0.26 0.21 0.20 0.15 0.114 v 5 performances 0.85 0.68 0.58 0.48 0.27 0.204 v 6 performances 0.87 0.82 0.73 0.64 0.41 0.234 v 7 performances 0.90 0.86 0.81 0.79 0.60 0.314 v 8 performances 0.90 0.88 0.85 0.83 0.70 0.446 v 6 performances 0.36 0.24 0.23 0.20 0.15 0.136 v 7 performances 0.80 0.66 0.54 0.48 0.27 0.186 v 8 performances 0.87 0.85 0.75 0.68 0.45 0.268 v 8 performances 0.38 0.27 0.23 0.23 0.15 0.13

i v j performances indicate that the female-even-month worker has aperformance of i and the male-odd-month worker has a performanceof j.

38

Table A.4: By performance and risk levels: summary of hiring decisions over workers

When The Level of Risk Is0% 1% 5% 10% 25% 50%

Gender Treatment - Given Equal PerformancesProbability of hiring female-even-month worker 0.43 0.30 0.26 0.23 0.17 0.14Fraction hiring female-even-month worker for sure 0.15 0.16 0.14 0.13 0.09 0.07Fraction letting chance decide who is hired 0.56 0.28 0.24 0.19 0.16 0.15Fraction hiring male-odd-month worker for sure 0.29 0.56 0.62 0.68 0.75 0.79N 1206 1206 1206 1206 1206 1206Birth Month Treatment - Given Equal PerformancesProbability of hiring female-even-month worker 0.37 0.26 0.22 0.21 0.15 0.12Fraction hiring female-even-month worker for sure 0.10 0.14 0.12 0.11 0.07 0.04Fraction letting chance decide who is hired 0.54 0.23 0.21 0.19 0.15 0.16Fraction hiring male-odd-month worker for sure 0.36 0.63 0.67 0.69 0.77 0.80N 1194 1194 1194 1194 1194 1194Gender Treatment - Given Women Perform BetterProbability of hiring female-even-month worker 0.89 0.80 0.75 0.68 0.48 0.32Fraction hiring female-even-month worker for sure 0.84 0.76 0.71 0.65 0.44 0.28Fraction letting chance decide who is hired 0.08 0.08 0.07 0.07 0.07 0.08Fraction hiring male-odd-month worker for sure 0.07 0.16 0.21 0.28 0.48 0.64N 2412 2412 2412 2412 2412 2412Birth Month Treatment - Given Women Perform BetterProbability of hiring female-even-month worker 0.87 0.79 0.71 0.65 0.45 0.27Fraction hiring female-even-month worker for sure 0.82 0.75 0.68 0.62 0.41 0.22Fraction letting chance decide who is hired 0.08 0.07 0.07 0.07 0.08 0.10Fraction hiring male-odd-month worker for sure 0.09 0.17 0.26 0.31 0.51 0.68N 2388 2388 2388 2388 2388 2388

39

Table A.5: By employer type and risk: summary of hiring decisions over workers with equalperformances

When The Employer ismale male female femaleodd even odd even

month month month month

Gender Treatment - Riskless DecisionsProbability of hiring female-even-month worker 0.37 0.39 0.51 0.48Fraction hiring female-even-month worker for sure 0.11 0.13 0.18 0.19Fraction letting chance decide who is hired 0.51 0.52 0.67 0.57Fraction hiring male-odd-month worker for sure 0.38 0.35 0.16 0.24N 342 327 261 276Gender Treatment - Risky DecisionsProbability of hiring female-even-month worker 0.19 0.23 0.26 0.20Fraction hiring female-even-month worker for sure 0.10 0.14 0.12 0.10Fraction letting chance decide who is hired 0.18 0.17 0.28 0.21Fraction hiring male-odd-month worker for sure 0.72 0.69 0.60 0.69N 1710 1635 1305 1380Birth Month Treatment - Riskless DecisionsProbability of hiring female-even-month worker 0.31 0.39 0.34 0.44Fraction hiring female-even-month worker for sure 0.06 0.12 0.08 0.13Fraction letting chance decide who is hired 0.50 0.53 0.51 0.61Fraction hiring male-odd-month worker for sure 0.44 0.34 0.40 0.26N 315 267 315 297Birth Month Treatment - Risky DecisionsProbability of hiring female-even-month worker 0.17 0.18 0.20 0.22Fraction hiring female-even-month worker for sure 0.08 0.08 0.12 0.11Fraction letting chance decide who is hired 0.19 0.18 0.17 0.21Fraction hiring male-odd-month worker for sure 0.73 0.73 0.71 0.68N 1575 1335 1575 1485

40

Table A.6: By each performance level: hiring decisions between workers with the same perfor-mance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4)

Panel A: Performance Level of 4Gender Treatment 0.069∗∗∗ 0.061∗∗∗ 0.034∗ 0.031

(0.022) (0.022) (0.019) (0.019)

Posterior(easy gap) -0.031∗∗∗ -0.020∗∗∗

(0.004) (0.004)

Posterior(hard-easy gap) -0.007 -0.011∗∗

(0.006) (0.005)

Constant 0.337∗∗∗ 0.419∗∗∗ 0.182∗∗∗ 0.234∗∗∗

(0.019) (0.021) (0.016) (0.019)Decision FEs yes yes yes yesObservations 800 800 4000 4000

Panel B: Performance Level of 6Gender Treatment 0.053∗∗ 0.052∗∗ 0.027 0.026

(0.022) (0.022) (0.019) (0.019)

Posterior(easy gap) -0.022∗∗∗ -0.017∗∗∗

(0.004) (0.004)

Posterior(hard-easy gap) -0.016∗∗∗ -0.013∗∗∗

(0.006) (0.005)

Constant 0.345∗∗∗ 0.403∗∗∗ 0.178∗∗∗ 0.221∗∗∗

(0.019) (0.021) (0.016) (0.018)Decision FEs yes yes yes yesObservations 800 800 4000 4000

Panel C: Performance Level of 8Gender Treatment 0.061∗∗∗ 0.059∗∗ 0.017 0.014

(0.023) (0.023) (0.019) (0.019)

Posterior(easy gap) -0.016∗∗∗ -0.018∗∗∗

(0.004) (0.004)

Posterior(hard-easy gap) -0.009 -0.011∗∗

(0.006) (0.004)

Constant 0.351∗∗∗ 0.391∗∗∗ 0.194∗∗∗ 0.242∗∗∗

(0.020) (0.022) (0.017) (0.019)Decision FEs yes yes yes yesObservations 800 800 4000 4000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions of the same specifications as in Panel Aof Table 1 but simply broken out according to the performance levels of the workers.

41

Table A.7: Hiring decisions between workers with the same performance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4)

Panel A: DV = 1 if Female Worker Hired for Sure, 0 otherwiseGender Treatment 0.050∗∗∗ 0.048∗∗ 0.018 0.017

(0.019) (0.019) (0.016) (0.016)

Posterior(easy gap) -0.015∗∗∗ -0.008∗∗

(0.004) (0.003)

Posterior(hard-easy gap) -0.008∗ -0.005(0.005) (0.004)

Constant 0.070∗∗∗ 0.109∗∗∗ 0.089∗∗∗ 0.111∗∗∗

(0.017) (0.020) (0.014) (0.016)Decision FEs yes yes yes yesObservations 2400 2400 12000 12000

Panel B: DV = 1 if Let Chance Decide, 0 otherwiseGender Treatment 0.022 0.018 0.016 0.014

(0.033) (0.033) (0.023) (0.022)

Posterior(easy gap) -0.016∗∗∗ -0.020∗∗∗

(0.006) (0.004)

Constant 0.542∗∗∗ 0.583∗∗∗ 0.193∗∗∗ 0.245∗∗∗

(0.030) (0.033) (0.020) (0.023)Decision FEs yes yes yes yesObservations 2400 2400 12000 12000Panel C: DV = 1 if Male Worker Hired For Sure, 0 otherwiseGender Treatment -0.072∗∗ -0.066∗∗ -0.034 -0.031

(0.030) (0.029) (0.026) (0.026)

Posterior(easy gap) 0.031∗∗∗ 0.028∗∗∗

(0.005) (0.005)

Posterior(hard-easy gap) 0.013∗ 0.018∗∗∗

(0.007) (0.006)

Constant 0.389∗∗∗ 0.308∗∗∗ 0.717∗∗∗ 0.644∗∗∗

(0.028) (0.028) (0.023) (0.026)Decision FEs yes yes yes yesObservations 2400 2400 12000 12000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions of the same specifications as in Panel Aof Table 1 except for the dependent variables being altered as noted

42

Figure A.1: In the No Information Stage version, hiring decisions between workers with sameperformance and payoff rule

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Gender Treatment0

.1.2

.3.4

.5Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Birth Month Treatment

Figure A.2: No Information Stage: Beliefs about performance gap between male-odd-monthworkers and female-even-months workers

0.1

.2.3

.4.5

Den

sity

-10 -5 0 5 10Performance Gap

Birth Month Treatment Gender Treatment

Two-sided Kolmogorov–Smirnov: p = 0.00

(a) Prior Beliefs (Hard Quiz)

0.2

.4.6

Den

sity

-10 -5 0 5 10Performance Gap

Birth Month Treatment Gender Treatment

Two-sided Kolmogorov–Smirnov: p = 0.00

(b) Prior Beliefs (Easy Quiz)

43

Table A.8: With belief interactions: hiring decisions between workers with the same performance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2)

Panel A: Gender and Birth Month TreatmentsGender Treatment 0.044∗∗ 0.028

(0.022) (0.024)

Posterior(easy gap) -0.026∗∗∗ -0.017∗∗∗

(0.004) (0.004)

Posterior(hard-easy gap) -0.014∗∗ -0.007(0.007) (0.005)

Gender Treatment*Posterior(easy gap) 0.006 -0.002(0.006) (0.007)

Gender Treatment*Posterior(hard-easy gap) 0.005 -0.008(0.009) (0.009)

Decision FEs yes yesObservations 2400 12000

Panel B: Gender and Gender-No-Information TreatmentsNo Information 0.011 -0.009

(0.023) (0.026)

Posterior(easy gap) -0.020∗∗∗ -0.019∗∗∗

(0.005) (0.005)

Posterior(hard-easy gap) -0.009 -0.015∗∗

(0.006) (0.007)

No Information*Posterior(easy gap) -0.002 0.009(0.007) (0.007)

No Information*Posterior(easy gap) -0.009 -0.000(0.009) (0.008)

Decision FEs yes yesObservations 2409 12045

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions of the probability with which a female-even-month worker is hired over a male-odd-month worker, among employers making hiring decisions betweenpairs of workers who have the same performance and payoff rules. Gender Treatment is an indicator foremployers in the Gender treatment. Posterior(easy gap) is an employer’s posterior belief about the averageperformance gap in the easy quiz. Posterior(hard − easy gap) is an employer’s posterior belief about theaverage performance gap in the hard quiz minus the employer’s posterior belief about the average performancegap in the easy quiz. No Information is an indicator for employers in the No Information Stage version.Decisions FEs include 6 fixed effects (with one excluded) that capture whether hiring decisions are betweenworkers with equal performances on the sports quiz (4 v 4, 6 v 6, or 8 v 8) or between workers with equalperformances on the math quiz (4 v 4, 6 v 6, or 8 v 8). Column 1 involves hiring decisions where thepayoff rule is always the same, while Column 2 involves hiring decisions where the payoff rule is risky whenhiring female-even-month workers and riskless when hiring male-odd-month workers. Panel A involves hiringdecisions from the Birth Month and Gender treatments, while Panel B involves hiring decisions from theGender and Gender-No-Information treatments.

44

Table A.9: With risk interactions: hiring decisions between workers with the same performance

(1) (2) (3) (4)

Risky Payoff -0.174∗∗∗ -0.174∗∗∗ -0.152∗∗∗ -0.152∗∗∗

(0.016) (0.016) (0.015) (0.015)

Gender Treatment 0.061∗∗∗ 0.059∗∗∗

(0.019) (0.018)

Gender Treatment*Risky Payoff -0.035 -0.035(0.023) (0.023)

In-Group 0.100∗∗∗ 0.086∗∗∗

(0.019) (0.018)

In-Group*Risky Payoff -0.088∗∗∗ -0.088∗∗∗

(0.023) (0.023)

Posterior(easy gap) -0.019∗∗∗ -0.019∗∗∗

(0.003) (0.003)

Posterior(hard-easy gap) -0.011∗∗∗ -0.011∗∗∗

(0.004) (0.004)Decision FEs yes yes yes yesObservations 14400 14400 14400 14400

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-leveland shown in parentheses. The results are from ordinary least squares regressions ofthe probability with which a female-even-month worker is hired over a male-odd-monthworker, among employers making hiring decisions between pairs of workers who have thesame performance and payoff rules. Risky Payoff is an indicator for hiring decisionswhere the payoff rule is risky when hiring female-even-month workers and riskless whenhiring male-odd-month workers. Gender Treatment is an indicator for employers inthe Gender treatment. In-Group is an indicator for employers who know they sharesome demographic characteristic with the female-even-month workers (i.e., even-monthemployers in the Birth Month treatment and female employers in the Gender treatment).Posterior(easy gap) is an employer’s posterior belief about the average performancegap in the easy quiz. Posterior(hard − easy gap) is an employer’s posterior beliefabout the average performance gap in the hard quiz minus the employer’s posterior beliefabout the average performance gap in the easy quiz. Decisions FEs include 6 fixed effects(with one excluded) that capture whether hiring decisions are between workers with equalperformances on the sports quiz (4 v 4, 6 v 6, or 8 v 8) or between workers with equalperformances on the math quiz (4 v 4, 6 v 6, or 8 v 8). All hiring decisions are from theBirth Month and Gender treatments.

45

Table A.10: When separately considering by the risk level: hiring decisions between workers with the same performance

Level of risk in payoff rule if a female-even-month worker is hired1% 5% 10% 25% 50%

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

In-Group 0.027 -0.000 0.034 0.036 0.012 0.004 -0.019 -0.018 0.006 0.019(0.024) (0.033) (0.023) (0.032) (0.023) (0.031) (0.020) (0.027) (0.018) (0.024)

Gender Treatment 0.017 0.036 0.012 0.016 0.029(0.033) (0.031) (0.031) (0.028) (0.024)

Gender Treatment*In-Group 0.057 -0.000 0.017 -0.001 -0.026(0.048) (0.047) (0.046) (0.040) (0.036)

Decision FEs yes yes yes yes yes yesObservations 2400 2400 2400 2400 2400 2400 2400 2400 2400 2400

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown in parentheses. The results are from ordinaryleast squares regressions of the probability with which a female-even-month worker is hired over a male-odd-month worker, among employersmaking hiring decisions between pairs of workers who have the same performance but different payoff rules. In-Group is an indicator foremployers who know they share some demographic characteristic with the female-even-month workers (i.e., even-month employers in the BirthMonth treatment and female employers in the Gender treatment). Gender Treatment is an indicator for employers in the Gender treatment.Decisions FEs include 6 fixed effects (with one excluded) that capture whether hiring decisions are between workers with equal performanceson the sports quiz (4 v 4, 6 v 6, or 8 v 8) or between workers with equal performances on the math quiz (4 v 4, 6 v 6, or 8 v 8). Columns 1 -2, 3 -4, 5 - 6, 7 - 8, and 9 - 10, involve hiring decisions where the payoff rule involves 1%, 5%, 10%, 25%, and 50% of no payment from hiringfemale-even-month workers, respectively. All hiring decisions are from the Birth Month and Gender treatments.

46

Table A.11: Comparing the math and sports versions: hiring decisions between workers with thesame performance

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4) (5) (6)

Panel A: Math VersionIn-Group 0.077∗∗∗ 0.081∗∗ 0.080∗∗ -0.009 0.004 0.003

(0.027) (0.037) (0.035) (0.028) (0.034) (0.033)

Gender Treatment 0.107∗∗∗ 0.099∗∗∗ 0.085∗∗ 0.082∗∗

(0.038) (0.037) (0.039) (0.039)

Gender Treatment*In-Group -0.015 -0.032 -0.029 -0.042(0.054) (0.052) (0.055) (0.054)

Posterior(easy gap) -0.027∗∗∗ -0.021∗∗∗

(0.005) (0.005)

Posterior(hard-easy gap) -0.015∗ -0.016∗∗

(0.008) (0.007)

Constant 0.392∗∗∗ 0.340∗∗∗ 0.393∗∗∗ 0.214∗∗∗ 0.172∗∗∗ 0.212∗∗∗

(0.021) (0.027) (0.026) (0.020) (0.022) (0.025)Decision FEs yes yes yes yes yes yesObservations 1200 1200 1200 6000 6000 6000Panel B: Sports VersionIn-Group 0.120∗∗∗ 0.099∗∗∗ 0.087∗∗ 0.034 0.013 0.002

(0.026) (0.036) (0.035) (0.025) (0.036) (0.036)

Gender Treatment 0.010 0.016 -0.033 -0.026(0.033) (0.033) (0.033) (0.033)

Gender Treatment*In-Group 0.049 0.042 0.039 0.032(0.052) (0.051) (0.051) (0.051)

Posterior(easy gap) -0.015∗∗∗ -0.015∗∗∗

(0.004) (0.005)

Posterior(hard-easy gap) -0.006 -0.007(0.005) (0.005)

Constant 0.320∗∗∗ 0.314∗∗∗ 0.356∗∗∗ 0.185∗∗∗ 0.202∗∗∗ 0.243∗∗∗

(0.018) (0.025) (0.026) (0.017) (0.025) (0.027)Decision FEs yes yes yes yes yes yesObservations 1200 1200 1200 6000 6000 6000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions involving the same specifications as inTable 3. Panel A involves hiring decisions from the math version, while Panel B involves hiring decisionsfrom the sports version.

47

Figure A.3: When the displayed female-even-month worker distribution has an average easyround performance that is lower: hiring decisions between workers in the Information Stage

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Gender Treatment0

.1.2

.3.4

.5Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Birth Month Treatment

48

Table A.12: When the displayed female-even-month worker distribution has an average easyround performance that is lower: hiring decisions between workers in the Information Stage

(1) (2) (3) (4)

Gender Treatment 0.044∗∗ -0.000 0.007(0.019) (0.025) (0.021)

In-Group 0.107∗∗∗ 0.056∗∗ 0.032(0.019) (0.027) (0.023)

Gender Treatment*In-Group 0.107∗∗∗ 0.077∗∗

(0.038) (0.032)

Posterior(easy gap) -0.050∗∗∗

(0.003)

Posterior(hard-easy gap) -0.024∗∗∗

(0.004)Decision FEs yes yes yes yesObservations 8000 8000 8000 8000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at theemployer-level and shown in parentheses. The results are from ordinary leastsquares regressions of the probability with which a female-even-month worker ishired over a male-odd-month worker, among employers making decisions in theInformation Stage involving displayed distributions that indicate that the averageeasy round performance is lower for the female-even-month workers than it is formale-odd-month workers. In-Group is an indicator for employers who knowthey share some demographic characteristic with the female-even-month workers(i.e., even-month employers in the Birth Month treatment and female employersin the Gender treatment). Gender Treatment is an indicator for employers inthe Gender treatment. Posterior(easy gap) is an employer’s posterior beliefabout the average performance gap in the easy quiz. Posterior(hard − easygap) is an employer’s posterior belief about the average performance gap in thehard quiz minus the employer’s posterior belief about the average performancegap in the easy quiz. Decisions FEs include 20 fixed effects (with one excluded)that capture each of the 20 relevant sets of sports or math distributions. Allhiring decisions are from the Birth Month and Gender treatments.

49

Figure A.4: When the displayed female-even-month worker distribution has an average easyround performance that is higher: hiring decisions between workers in the Information Stage

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Gender Treatment0

.1.2

.3.4

.5Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Birth Month Treatment

50

Table A.13: When the displayed female-even-month worker distribution has an average easyround performance that is higher: hiring decisions between workers in the Information Stage

(1) (2) (3)

In-Group 0.107∗∗∗ 0.056∗∗ 0.032(0.019) (0.027) (0.023)

In-Group 0.049∗∗ 0.052 0.050(0.024) (0.035) (0.035)

Gender Treatment 0.027 0.023(0.036) (0.036)

Gender Treatment*In-Group -0.007 -0.014(0.048) (0.048)

Posterior(easy gap) -0.010∗∗

(0.004)

Posterior(hard-easy gap) -0.002(0.007)

Decision FEs yes yes yesObservations 1200 1200 1200

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clus-tered at the employer-level and shown in parentheses. The resultsare from ordinary least squares regressions of the probability withwhich a female-even-month worker is hired over a male-odd-monthworker, among employers making decisions in the Information Stageinvolving displayed distributions that indicate that the average easyround performance is higher for the female-even-month workers thanit is for male-odd-month workers. In-Group is an indicator for em-ployers who know they share some demographic characteristic withthe female-even-month workers (i.e., even-month employers in theBirth Month treatment and female employers in the Gender treat-ment). Gender Treatment is an indicator for employers in theGender treatment. Posterior(easy gap) is an employer’s poste-rior belief about the average performance gap in the easy quiz. Pos-terior(hard − easy gap) is an employer’s posterior belief aboutthe average performance gap in the hard quiz minus the employer’sposterior belief about the average performance gap in the easy quiz.Decisions FEs include 3 fixed effects (with one excluded) that cap-ture each of the 3 relevant sets of sports or math distributions. Allhiring decisions are from the Birth Month and Gender treatments.

51

Figure A.5: When there is no average difference in the average easy round performances betweenthe displayed distributions: hiring decisions between workers in the Information Stage

0.1

.2.3

.4.5

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Gender Treatment0

.1.2

.3.4

.5Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Birth Month Treatment

52

Table A.14: When there is no average difference in the average easy round performances betweenthe displayed distributions: hiring decisions between workers in the Information Stage

(1) (2) (3)In-Group -0.032 -0.037 -0.036

(0.043) (0.061) (0.061)

Gender Treatment 0.001 0.004(0.060) (0.061)

Gender Treatment*In-Group 0.009 0.010(0.086) (0.086)

Posterior(easy gap) 0.001(0.008)

Posterior(hard-easy gap) -0.003(0.012)

Constant 0.404∗∗∗ 0.404∗∗∗ 0.400∗∗∗

(0.030) (0.041) (0.044)Decision FEs no no noObservations 400 400 400

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clus-tered at the employer-level and shown in parentheses. The resultsare from ordinary least squares regressions of the probability withwhich a female-even-month worker is hired over a male-odd-monthworker, among employers making decisions in the Information Stageinvolving displayed distributions that indicate that there is no aver-age difference in easy round performances between the two distribu-tions. In-Group is an indicator for employers who know they sharesome demographic characteristic with the female-even-month work-ers (i.e., even-month employers in the Birth Month treatment andfemale employers in the Gender treatment). Gender Treatment isan indicator for employers in the Gender treatment. Posterior(easygap) is an employer’s posterior belief about the average performancegap in the easy quiz. Posterior(hard − easy gap) is an employer’sposterior belief about the average performance gap in the hard quizminus the employer’s posterior belief about the average performancegap in the easy quiz. All hiring decisions are from the Birth Monthand Gender treatments.

53

Figure A.6: Hiring decisions between workers with same payoff rule but different performances(i.e., the female-even-month worker has a better easy round performance than the male-odd-month worker)

0.1

.2.3

.4.5

.6.7

.8.9

1Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Gender Treatment0

.1.2

.3.4

.5.6

.7.8

.91

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Birth Month Treatment

54

Table A.15: Hiring decisions between workers with different performances (i.e., the female-even-month worker has a better easy round performance than the male-odd-month worker)

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4) (5) (6)

In-Group 0.015 0.026 0.023 -0.021 -0.029 -0.033(0.016) (0.025) (0.025) (0.020) (0.027) (0.027)

Gender Treatment 0.031 0.029 0.024 0.026(0.023) (0.023) (0.025) (0.025)

Gender Treatment*In-Group -0.021 -0.022 0.016 0.010(0.032) (0.032) (0.039) (0.039)

Posterior(easy gap) -0.003 -0.011∗∗∗

(0.003) (0.004)

Posterior(hard-easy gap) 0.005 -0.007(0.004) (0.005)

Decision FEs yes yes yes yes yes yesObservations 4800 4800 4800 24000 24000 24000

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shownin parentheses. The results are from ordinary least squares regressions of the probability with whicha female-even-month worker is hired over a male-odd-month worker, among employers making hiringdecisions between pairs of workers in which the even-month-female worker has a better performancethan the male-odd-month worker. In-Group is an indicator for employers who know they share somedemographic characteristic with the female-even-month workers (i.e., even-month employers in theBirth Month treatment and female employers in the Gender treatment). Gender Treatment is anindicator for employers in the Gender treatment. Posterior(easy gap) is an employer’s posteriorbelief about the average performance gap in the easy quiz. Posterior(hard − easy gap) is anemployer’s posterior belief about the average performance gap in the hard quiz minus the employer’sposterior belief about the average performance gap in the easy quiz. Decisions FEs include 12 fixedeffects (with one excluded) that capture whether hiring decisions involve a male-odd-month workerversus a female-even-month worker where the female-even-month worker has a better performance onthe sports quiz (4 v 5, 4 v 6, 4 v 7, 4 v 8, 6 v 7, 6 v 8) or on the math quiz (4 v 5, 4 v 6, 4 v 7, 4 v 8,6 v 7, 6 v 8) . Columns 1 - 3 involve hiring decisions where the payoff rule is always the same, whileColumns 4 - 6 involve hiring decisions where the payoff rule is risky when hiring female-even-monthworkers and riskless when hiring male-odd-month workers. All hiring decisions are from the BirthMonth and Gender treatments.

B Replication in a simplified environmentWe also document the robustness of our findings in a second, simplified experiment. In this

simplified experiment, employers make just a single hiring decision in the Hiring Stage, and we

introduce risk as an “excuse” not to hire across-employer, rather than within-employer.

The design is as follows. First, employers complete the Elicitation of Prior Beliefs Stage.

Second, they proceed to a very simplified version of the Information Stage. In this simplified

55

version of the Information Stage, they are shown only the final screen of the Information Stage

from the original experiment: the complete histogram of performances of all female-even-month

workers and male-odd-month workers. We do not ask them to make a choice about whom to

hire from these distributions. They simply see this single histogram of all performances. Third,

employers complete the Elicitation of Posterior Beliefs Stage. Fourth, and most importantly,

the Hiring Stage is shortened so that employers only make a single hiring decision. They must

choose to hire either a female-even-month worker or a male-odd-month worker, each of whom

earned a score of 4 on the easy quiz.

Just as in our original experiment, we randomly assign employers to either the Gender treat-

ment or the Birth Month treatment. To accommodate the one-hiring decision nature of the sim-

plified experiment, we also randomly assign employers to either make this decision in the risky

or riskless paradigm. All employers in the riskless paradigm make a hiring decision in which the

payoff rule for hiring either worker is the same, while all employers in the risky paradigm make a

hiring decision in which the payoff rule is altered so that there is only a 99% chance of receiving

the payoff if the female-even-month worker is hired. Thus, in this simplified experiment, the

introduction of risk is an across-subject treatment variation, rather than within-subject. Finally,

all employers are assigned to the math version of the experiment, reducing the number of em-

ployers we need in order to be adequately powered to identify differences across the original and

simplified versions of the experiment.

We ran this study on Amazon Mechanical Turk in November 2019 with 1995 employers; it was

advertised as a 10-minute academic study with a completion payment of $1.50, plus additional

payment based upon the hiring decision made.27 With the exception of the details described

above, the rest of the experimental design exactly replicates the original hiring experiment,

including our use of the same available pool of workers.

In the results below, we compare the decisions made by employers in our simplified experi-

ment to the decisions made by employers in our original experiment, restricted to the common

decision. That is, we ask whether the hiring decisions made between two workers, both with

a performance of 4 on the easy quiz in math, vary across the simplified and original versions

of the experiment. Figure B.1 compares the riskless decisions across the simplified experiment

and the original experiment. The results are remarkably consistent. Most importantly, our main

result strongly replicates: within math, women are no less likely to be hired under the Gender

treatment than under the Birth Month treatment. In fact, while there is significant discrimina-

tion against the female-even-month worker in the Birth Month treatment, women are hired at a

rate indistinguishable from 50% within the Gender treatment in both the simplified and original

experiments for this particular decision (math, riskless, 4 v 4).

27We recruited 2000 Mturkers, but 5 employers either did not submit the proper evidence of HIT completionor had participated in the original version of this experiment despite our efforts to exclude them. We drop those5 Mturkers from our sample.

56

Figure B.1: Comparing the simplified versus original math version: Hiring decisions betweenworkers with same performance levels of 4 and same payoff rule

0.1

.2.3

.4.5

.6Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

GenderTreatment

Birth MonthTreatment

(a) Simplified Math Version0

.1.2

.3.4

.5.6

Prob

abilit

y of

hiri

ng fe

mal

e-ev

en-m

onth

wor

ker

GenderTreatment

Birth MonthTreatment

(b) Original Math Version

Table B.1 formalizes this analysis by replicating Table 1 for both the simplified and original

versions of the experiment, again restricted to the decisions common to both experiments. Our

focus is on the comparison between Panels A and B. We see a large degree of consistency across

the two experiments. In both riskless and risky decisions, the Gender treatment has a significant,

positive impact on the probability of hiring women in both the simplified and original experiments

(Columns 1 and 2 for riskless, Columns 3 and 4 for risky).

We found that in our original experiment, there was some evidence of in-group preferences

among employers, with less discrimination against female-even-month workers by women in the

Gender treatment, and by even-month employers in the Birth Month treatment. But, these

in-group preferences were not robust to the introduction of risk, nor did they persist in decisions

where the female-even-month worker performed better than the male-odd-month worker on the

easy quiz. We can further probe the robustness of these in-group preferences by exploring them

in the simplified experiment. As Figure B.2 and Table B.2 shows, the simplified experiment

suggests that in-group preferences are not particularly robust. In particular, we find much

weaker evidence of in-group preferences even within the riskless decisions. In fact, there are no

significant in-group preferences conditional on stated beliefs in the simplified experiment, either

in the riskless or the risky decisions.

57

Table B.1: Comparing the simplified versus original math version: hiring decisions betweenworkers with the same performance level of 4

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4)

Panel A: Simplified Math VersionGender Treatment 0.127∗∗∗ 0.105∗∗∗ 0.151∗∗∗ 0.148∗∗∗

(0.025) (0.023) (0.026) (0.025)

Posterior(easy gap) -0.046∗∗∗ -0.022∗∗∗

(0.003) (0.004)

Posterior(hard-easy gap) -0.025∗∗∗ -0.018∗∗∗

(0.004) (0.005)

Constant 0.371∗∗∗ 0.468∗∗∗ 0.222∗∗∗ 0.267∗∗∗

(0.017) (0.018) (0.017) (0.019)Observations 1001 1001 994 994

Panel B: Original Math VersionGender Treatment 0.116∗∗∗ 0.090∗∗∗ 0.094∗∗ 0.084∗∗

(0.032) (0.032) (0.037) (0.037)

Posterior(easy gap) -0.034∗∗∗ -0.022∗∗∗

(0.006) (0.007)

Posterior(hard-easy gap) -0.009 -0.016(0.009) (0.010)

Constant 0.371∗∗∗ 0.438∗∗∗ 0.242∗∗∗ 0.284∗∗∗

(0.022) (0.024) (0.025) (0.027)Observations 400 400 400 400

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shownin parentheses. The results are from ordinary least squares regressions of the probability with which afemale-even-month worker is hired over a male-odd-month worker, among employers making hiring decisionsbetween pairs of workers who have the same performances of 4. Gender Treatment is an indicator foremployers in the Gender treatment. Posterior(easy gap) is an employer’s posterior belief about theaverage performance gap in the easy quiz. Posterior(hard − easy gap) is an employer’s posterior beliefabout the average performance gap in the hard quiz minus the employer’s posterior belief about the averageperformance gap in the easy quiz. Columns 1 - 2 involve hiring decisions where the payoff rule is alwaysthe same, while Columns 3 - 4 involve hiring decisions where the payoff rule is risky when hiring female-even-month workers and riskless when hiring male-odd-month workers. Panel A involves hiring decisionsfrom the simplified math version (which only collected data about hiring decisions between two workers whoboth have a performance level of 4 and when there was no risk of risk equal to 1%) , while Panel B involveshiring decisions from the original math version (restricting to to the data about hiring decisions betweentwo workers who both have a performance level of 4 and when there was no risk of risk equal to 1%).

58

Figure B.2: Comparing the simplified versus original math version: hiring decisions betweenworkers with same performance levels of 4 and same payoff rule

0.1

.2.3

.4.5

.6Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(a) Simplified Math Version: Gender Treatment

0.1

.2.3

.4.5

.6Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(b) Simplified Math Version: Birth Month Treatment

0.1

.2.3

.4.5

.6Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(c) Original Math Version: Gender Treatment

0.1

.2.3

.4.5

.6Pr

obab

ility

of h

iring

fem

ale-

even

-mon

th w

orke

r

Overall OddMonthMale

EvenMonthMale

OddMonthFemale

EvenMonthFemale

(d) Original Math Version: Birth Month Treatment

59

Table B.2: Comparing the simplified versus original math version: hiring decisions betweenworkers with the same performance level of 4

Same payoff rule Risk in payoff rule only forfemale-even-month workers

(1) (2) (3) (4) (5) (6)

Panel A: Simplified Math VersionIn-Group 0.048∗ -0.009 -0.004 0.062∗∗ 0.012 0.014

(0.026) (0.035) (0.033) (0.026) (0.033) (0.033)

Gender Treatment 0.071∗∗ 0.070∗∗ 0.104∗∗∗ 0.108∗∗∗

(0.034) (0.031) (0.034) (0.033)

Gender Treatment*In-Group 0.125∗∗ 0.080∗ 0.102∗∗ 0.088∗

(0.051) (0.047) (0.051) (0.051)

Posterior(easy gap) -0.046∗∗∗ -0.021∗∗∗

(0.003) (0.004)

Posterior(hard-easy gap) -0.024∗∗∗ -0.018∗∗∗

(0.004) (0.005)

Constant 0.412∗∗∗ 0.375∗∗∗ 0.468∗∗∗ 0.269∗∗∗ 0.216∗∗∗ 0.258∗∗∗

(0.017) (0.024) (0.023) (0.017) (0.022) (0.024)Observations 1001 1001 1001 994 994 994Panel B: Original Math VersionIn-Group 0.142∗∗∗ 0.174∗∗∗ 0.170∗∗∗ 0.002 0.004 0.004

(0.032) (0.043) (0.040) (0.037) (0.050) (0.050)

Gender Treatment 0.147∗∗∗ 0.133∗∗∗ 0.100∗ 0.096∗

(0.044) (0.042) (0.053) (0.052)

Gender Treatment*In-Group -0.070 -0.092 -0.011 -0.025(0.063) (0.061) (0.074) (0.074)

Posterior(easy gap) -0.033∗∗∗ -0.022∗∗∗

(0.006) (0.007)

Posterior(hard-easy gap) -0.010 -0.016(0.009) (0.010)

Constant 0.360∗∗∗ 0.288∗∗∗ 0.355∗∗∗ 0.289∗∗∗ 0.240∗∗∗ 0.282∗∗∗

(0.022) (0.029) (0.029) (0.026) (0.033) (0.036)Observations 400 400 400 400 400 400

∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Standard errors are clustered at the employer-level and shown inparentheses. The results are from ordinary least squares regressions involving the same specifications asin Table 3. Panel A involves hiring decisions from the simplified math version (which only collected dataabout hiring decisions between two workers who both have a performance level of 4 and when there was norisk of risk equal to 1%) , while Panel B involves hiring decisions from the original math version (restrictingto to the data about hiring decisions between two workers who both have a performance level of 4 andwhen there was no risk of risk equal to 1%). 60