Compare text

Find the difference between two text files

Real-time diff

Unified diff

Collapse lines

Highlight change

Syntax highlighting

Tools

Diffchecker Desktop The most secure way to run Diffchecker. Get the Diffchecker Desktop app: your diffs never leave your computer!Get Desktop

Comparing sensitive data, confidential files or internal emails?

Most legal and privacy policies prohibit uploading sensitive data online. Diffchecker Desktop ensures your confidential information never leaves your computer. Work offline and compare documents securely.

Download Diffchecker Desktop Learn more

Untitled Diff

Created 5 years agoDiff never expires

427 removals

463 lines

444 additions

464 lines

Gender Differences in Recognition for Group Work

Heather Sarsons

Heather Sarsons, Klarita Gërxhani, Ernesto Reuben, and Arthur Schram

∗

February 4, 2019

September 15, 2019

Abstract

Does gender influence how credit for group work is allocated? Using data from

academic economists’ CVs, I test whether coauthored and solo-authored publications

academic economists’ CVs, we test whether coauthored and solo-authored publica-

matter differently for tenure for men and women. Because coauthors are listed alpha-

tions matter differently for tenure for men and women. Because coauthors are listed

betically in economics, coauthored papers do not provide specific information about

alphabetically in economics, coauthored papers do not provide specific information

each contributor’s skills or ability. Solo-authored papers, on the other hand, provide

about each contributor’s skills or ability. Solo-authored papers, on the other hand,

a relatively clear signal of ability. I find that conditional on publication quality and

provide a relatively clear signal of ability. We find that conditional on publication

other observables, men are tenured at roughly the same rate regardless of whether

quality and other observables, men are tenured at roughly the same rate regardless of

they coauthor or solo-author. Women, however, become less likely to receive tenure

whether they coauthor or solo-author. Women, however, become less likely to receive

the more they coauthor. The result is most pronounced for women coauthoring with

tenure the more they coauthor. The result is most pronounced for women coauthoring

men and less pronounced among women who coauthor with other women. An on-

with men and less pronounced among women who coauthor with other women. Two

line experiment finds similar patterns when women perform male-stereotyped tasks.

experiments suggest that both stereotypes surrounding a task as well as the evalua-

However, women receive equal credit for joint work with men when the task is per-

tors’ gender affect who receives credit. Taken together, our results are best explained

ceived to be gender-neutral. Taken together, the results suggest that gender and stereo-

by gender and stereotypes having a noticeable influence on the allocation of credit for

types influence the allocation of credit for group work.

group work.

∗

I especially thank Roland Fryer, Claudia Goldin, Larry Katz, David Laibson, and Amanda Pallais for

Sarsons, University of Chicago Booth (heather.sarsons@chicagobooth.edu); Gërxhani, European Uni-

their guidance and encouragement. I also thank Mitra Akhtari, Amitabh Chandra, John Coglianese, Oren

versity Institute; Reuben, New York University Abu Dhabi and LISER; Schram, Amsterdam School of Eco-

Danieli, Ellora Derenoncourt, Florian Ederer, Ben Enke, Raissa Fabregas, Nicole Fortin, Peter Ganong, Ed-

nomics and European University Institute. Sarsons especially thanks Roland Fryer, Claudia Goldin, Larry

ward Glaeser, Siri Isaksson, Sara Lowes, Rob McMillan, Eduardo Montero, Gautam Rao, Alex Segura, Nihar

Katz, David Laibson, and Amanda Pallais for their guidance and encouragement. We also thank Mitra

Shah, Peter Tu, Justin Wolfers, and participants at SOLE 2016, the Early Behavioural Economics Conference,

Akhtari, Amitabh Chandra, John Coglianese, Oren Danieli, Ellora Derenoncourt, Florian Ederer, Ben Enke,

the Harvard Business School Gender Initiative, and the Harvard Inequality Seminar for helpful comments

Raissa Fabregas, Nicole Fortin, Nickolas Gagnon, Peter Ganong, Edward Glaeser, Siri Isaksson, Emir Ka-

and suggestions. This paper is intentionally solo-authored.

menica, Sara Lowes, Rob McMillan, Eduardo Montero, Gautam Rao, Alex Segura, Nihar Shah, Peter Tu,

Jeroen van de Ven, Justin Wolfers, and various conference and seminar participants for their helpful com-

ments and suggestions.

1 Introduction

Do employers use gender when allocating credit for group work, particularly when in-

dividual contributions are unobserved? Organizations increasingly rely on group work

for production (Lazear and Shaw, 2007), yet there is little empirical evidence document-

ing how credit for group work is allocated. Unless employers can perfectly observe each

worker’s contribution to the team’s output, they must decide how to allocate credit with-

out having full information as to what each worker did. This could leave room for demo-

graphic characteristics, such as gender, to influence the allocation of credit.

In this paper, I test whether uncertainty over an individual’s contribution to a project

In this paper, we test whether uncertainty over an individual’s contribution to a project

leads to differential attribution of credit that contributes to the gender promotion gap. In

many industries, women are not only hired at lower rates than men are, they are also

promoted at lower rates.

This paper explores whether gender differences in credit for

group work exist and whether they explain part of the promotion gap.

I primarily look at the tenure decisions of academic economists to test whether gender

We primarily look at the tenure decisions of academic economists to test whether gen-

influences the allocation of credit for coauthored papers. Economics is a relevant setting

der influences the allocation of credit for coauthored papers. Economics is a relevant

as there is a large tenure gap between men and women, and because the amount of coau-

setting as there is a large tenure gap between men and women, and because the amount

thoring has risen dramatically in recent years (Ginther and Kahn, 2004; Hammermesh,

of coauthoring has risen dramatically in recent years (Ginther and Kahn, 2004; Hammer-

2013). Using data from economists’ CVs, I track individuals’ career trajectories and com-

mesh, 2013). Using data from economists’ CVs, we track individuals’ career trajectories

pare whether the trajectory is different for individuals who coauthor versus solo-author,

and compare whether the trajectory is different for individuals who coauthor versus solo-

and whether there is a difference by gender.

author, and whether there is a difference by gender.

Within economics, I find that men and women who solo-author most of their work

Within economics, we find that men and women who solo-author most of their work

have similar tenure rates conditional on a proxy for the quality of papers. However, an

additional coauthored paper is correlated with an 8.2% increase in tenure probability for

men but only a 5.6% increase for women. This gap is significantly less pronounced for

women who coauthor with women, suggesting that the attribution of credit is related

to the gender mix of coauthors. Furthermore, a man who coauthors is no less likely to

receive tenure than a comparable man who solo-authors even though there is presumably

more uncertainty as to how much work he did. A counterfactual exercise suggests that

this difference in credit allocation explains 60% of the unconditional gender gap in tenure

rates and 84% of the gap that remains after controlling for average paper quality, citations,

tenure and PhD institution ranks, and field.

To ensure that I am not picking up on ability differences between men and women, I

To ensure that we are not picking up on ability differences between men and women,

control for the quality of papers using both journal rankings and citations, allowing for

we control for the quality of papers using both journal rankings and citations, allowing

Blau and DeVaro (2007), for example, find that across jobs, women are less likely to be promoted than

men conditional on worker’s performance and ability ratings. In the UK, female managers are nearly 40%

less likely to be promoted than male managers (Elmins et al. 2016).

less likely to be promoted than male managers (Elmins et al., 2016).

a comparison of men and women with similar research portfolios. The results are also

for a comparison of men and women with similar research portfolios. The results are also

robust to including other individual-level controls such as length of time to tenure and

robust to including other individual-level controls such as length of time to tenure and the

the seniority of one’s coauthors, as well as tenure year, tenure institution, and primary

seniority of one’s coauthors, as well as tenure year, tenure institution, and primary field

field fixed effects.

fixed effects.

I argue that these results are most consistent with a story of women receiving less credit

We argue that these results are most consistent with a story of women receiving less

for their joint work with men because of bias. To show this, I first use current CV and

credit for their joint work with men because of bias. To show this, we first use current

citation data to compare the productivity of men and women who did and did not receive

CV and citation data to compare the productivity of men and women who did and did

tenure at the institution where they initially went up for tenure. While the estimates are

not receive tenure at the institution where they initially went up for tenure. While the

imprecise, I find suggestive evidence that women who coauthor and are denied tenure

estimates are imprecise, we find suggestive evidence that women who coauthor and are

produce more solo-authored papers that publish in high-ranking journals than men who

denied tenure produce more solo-authored papers that publish in high-ranking journals

are denied tenure. Data on citations show a similar result.

than men who are denied tenure. Data on citations show a similar result.

I then rule out several alternative explanations for the empirical patterns. For example,

We then rule out several alternative explanations for the empirical patterns. For ex-

several papers have demonstrated that selection into coauthorship in economics is not

ample, several papers have demonstrated that selection into coauthorship in economics

random.

is not random.

I test for selection into coauthorship and do not find any evidence that women

We test for selection into coauthorship and do not find any evidence that

coauthor with high ability or more senior men. I also look at the timing of coauthorship

women coauthor with high ability or more senior men. We also look at the timing of coau-

and find no evidence that women begin coauthoring if they have a slower start to their

thorship and find no evidence that women begin coauthoring if they have a slower start

careers. The empirical patterns are also inconsistent with taste-basted discrimination.

to their careers. The empirical patterns are also inconsistent with taste-basted discrimina-

Because the CV data does not allow me to rule out the possibility that women actually

tion.

contribute less to papers that are coauthored with men, I conduct an online experiment

Because the CV data do not allow us to rule out the possibility that women actually

designed to test whether this drives credit allocation. In the experiment, I first hire indi-

contribute less to papers that are coauthored with men, we conduct two experiments de-

viduals to complete quizzes on topics that are either male or female-stereotyped. I then

signed to test whether real or perceived differences in contributions drive credit allocation.

hire participants who act as “predictors” and are randomized into a solo treatment or a

In the first experiment, we first hire individuals to complete quizzes on topics that are ei-

group treatment. Predictors in the solo treatment are shown two individual’s separate

ther male or female-stereotyped. We then hire participants who act as “predictors” and

quiz scores while predictors in the group treatment are shown the combined score of two

are randomized into an individual treatment or a joint treatment. Predictors in the indi-

individuals. They are then asked to predict the performance of each participant on future

vidual treatment are shown two individual’s separate quiz scores while predictors in the

quizzes.

joint treatment are shown the combined score of two individuals. They are then asked to

I find that in the group treatment, women are predicted to perform worse than their

predict the performance of each participant on future quizzes.

male counterparts for male-stereotyped quizzes, suggesting that participants making the

In the joint treatment, women are predicted to perform worse than their male counter-

predictions believe that women contributed less to the combined score (that is, they per-

parts for male-stereotyped quizzes, suggesting that predictors believe that women con-

formed worse). However, if pairs performed a female-stereotyped quiz, women and men

tributed less to the combined score (that is, they performed worse). However, if pairs

are given equal credit. To understand whether these results are driven by participants’

performed a female-stereotyped quiz, women and men are given equal credit. To under-

beliefs about the ability distributions of men and women, I randomly provide some par-

stand whether these results are driven by participants’ beliefs about the ability distribu-

tions of men and women, we randomly provide some participants with the distribution of

scores on the initial quiz by gender. Women appear to be given equal credit in the female-

stereotyped quiz because participants view it as being gender-neutral. That is, they do not

Boschini and Sjögren (2007) test whether coauthorship patterns in economics are gender neutral. They

See, for example, Boschini and Sjögren (2007), Garcia and Sherman (2015), and Bikard et al (2015).

find that women are more likely to solo-author than men and that gender homophily in coauthoring in-

creases with women’s representation in a field. Garcia and Sherman (2015) argue that the alphabetical pub-

lishing norms in economics influence both the types of people authors work with and the types of projects

they will work on.

ticipants with the distribution of scores on the initial quiz by gender. Women appear to

realize that women tend to outperform men. Showing participants the gender distribution

be given equal credit in the female-stereotyped quiz because participants view it as being

of scores corrects this belief and women are then predicted to have a better performance

gender-neutral. That is, they do not realize that women tend to outperform men. Show-

in future female-stereotyped quizzes but it does not affect the predicted performance gap

ing participants the gender distribution of scores corrects this belief and women are then

for women and men performing male-stereotyped tasks.

predicted to have a better performance in future female-stereotyped quizzes. Surpris-

The second experiment is conducted in a more natural setting with human resources

ingly, this treatment does not affect the predicted performance gap for women and men

personnel. Following a similar design, we again test whether women are less likely than

performing male-stereotyped tasks.

men to receive credit for good group performance. We additionally elicit the HR person-

nels’ beliefs about male and female performance and find that differences in the allocation

of credit are largely driven by differences in beliefs. We also find that male HR personnel

are more likely to hire in favor of men, and women in favor of women.

This paper replicates and builds off of the results in Sarsons (2017), which shows the

basic correlational patterns between paper composition and tenure. In this paper, I repli-

basic correlational patterns between paper composition and tenure. In this paper, we

cate the results using more data and then use the C.V. data and an experiment to establish a

replicate the results using more data and then use the C.V. data and two experiments to es-

channel through which gender influences the allocation of credit. The paper also relates to

tablish a channel through which gender influences the allocation of credit. The paper also

a large literature seeking to understand difference in labor market outcomes between men

relates to a large literature seeking to understand difference in labor market outcomes be-

and women. Factors such as productivity, personality and behavioural differences (such

tween men and women. Factors such as productivity, personality and behavioural differ-

as competition aversion), and fertility preferences have been shown to explain some differ-

ences (such as competition aversion), and fertility preferences have been shown to explain

ences in career choice and progression.

some differences in career choice and progression.

In academia in particular, studies have pointed

In academia in particular, studies have

to both supply-side factors, including differences in subject matter interest (Dynan and

pointed to both supply-side factors, including differences in subject matter interest (Dy-

Rouse, 1997) and the availability of role models (Hale and Regev, 2014; Carrell, Page, and

nan and Rouse, 1997) and the availability of role models (Hale and Regev, 2014; Carrell et

West, 2010); demand-side factors, such as implicit bias (Milkman, Akinola, and Chugh,

al., 2010); demand-side factors, such as implicit bias (Milkman et al., 2015; Moss-Racusin et

2015; Moss-Racusin et al., 2012); and institutional factors (Antecol et al., 2018). This paper

al., 2012); and institutional factors (Antecol et al., 2018). This paper directly tests whether

directly tests whether the differential treatment of work output contributes to the gender

the differential treatment of work output contributes to the gender gap.

gap.

The remainder of the paper is organized as follows. Section 2 describes the data and

shows that a tenure gap exists between male and female economists. In Section 3, I show

shows that a tenure gap exists between male and female economists. In Section 3, we

that the tenure gap closes as women produce more solo-authored papers but does not

show that the tenure gap closes as women produce more solo-authored papers but does

close as they produce more coauthored papers. Women have a consistently lower proba-

not close as they produce more coauthored papers. Women have a consistently lower

bility of tenure for each additional coauthored paper than men. I show that the results are

probability of tenure for each additional coauthored paper than men. We show that the

robust to accounting for attrition, and to using different journal rankings and definitions

results are robust to accounting for attrition, and to using different journal rankings and

of tenure. In Section 4, I argue that the results are in line with a story in which women

definitions of tenure. In Section 4, we argue that the results are in line with a story in which

receive less credit for joint work with men. I test alternative explanations of the relation-

women receive less credit for joint work with men. We test alternative explanations of the

ship between coauthorship and tenure and argue that none can fully explain the observed

relationship between coauthorship and tenure and argue that none can fully explain the

empirical patterns. Section 5 discusses the structure and results of the experiment. Section

observed empirical patterns. Section 5 discusses the design and results of the experiments.

6 concludes.

Section 6 concludes.

There is a large literature documenting gender differences in productivity, attitudes toward different

types of work, and family choices. See, for example, Niederle and Vesterlund (2007), Antecol et al. (2016),

types of work, and family choices. See, for example, Niederle and Vesterlund (2007), Buser et al. (2014),

Ceci et al. (2014), and Ginther and Kahn (2004).

Antecol et al. (2018), Ceci et al. (2014), Reuben et al. (2017), and Ginther and Kahn (2004).

2 Data

To examine the relationship between paper composition and tenure, I construct a dataset

To examine the relationship between paper composition and tenure, we construct a dataset

using the CVs of economists who came up for tenure between 1985 and 2014 at one of the

top 35 U.S. PhD-granting universities

. The academic progression documented in the CVs

makes it possible to evaluate the relationship between an individual’s research output and

career progression. I can then compare the degree of collaborative work and reward for

career progression. We can then compare the degree of collaborative work and reward for

that work, and compare these results for men versus women.

2.1 Sample Selection and Data Overview

I include only PhD-granting institutions in the sample as tenure evaluation at these schools

We include only PhD-granting institutions in the sample as tenure evaluation at these

is primarily based on research output, of which I have a clear measure. Other institutions

schools is primarily based on research output, of which we have a clear measure. Other

like liberal arts colleges place greater weight on teaching ability for tenure, something

institutions like liberal arts colleges place greater weight on teaching ability for tenure,

that I cannot measure. I exclude business and public policy schools for similar reasons.

something that we cannot measure. We exclude business and public policy schools for

similar reasons.

It is reasonable to assume that the top 35 economics departments in the U.S. emphasize

It is reasonable to assume that the top 35 economics departments in

research which is measured by the number and quality of papers one produces.

the U.S. emphasize research which is measured by the number and quality of papers one

produces.

One problem in collecting tenure information is that the CVs of individuals who went

up for tenure, were denied it, and left to industry or government are difficult to find, lead-

ing to a sample selection problem. To deal with this issue, I collected historical faculty lists

ing to a sample selection problem. To deal with this issue, we collected historical faculty

from 23 of the 35 schools and locate over 90% of faculty who had ever gone up for tenure

lists from 23 of the 35 schools and locate over 90% of faculty who had ever gone up for

at these 23 institutions. For the remaining 12 schools that did not have historical faculty

tenure at these 23 institutions. For the remaining 12 schools that did not have historical

lists available, I looked at the top 75 U.S. institutions, the top 5 Canadian institutions, and

faculty lists available, we looked at the top 75 U.S. institutions, the top 5 Canadian institu-

the top 5 European institutions to locate anyone who went up for tenure at a top 35 U.S.

tions, and the top 5 European institutions to locate anyone who went up for tenure at a top

school and then moved to another institution. I also checked economists’ CVs at the ma-

35 U.S. school and then moved to another institution. We also checked economists’ CVs at

jor Federal Reserve Boards and other large research institutes, such as Mathematica, in the

the major Federal Reserve Boards and other large research institutes, such as Mathemat-

U.S. While there might still be a sample selection problem, I show in Section 3.2.1 that the

ica, in the U.S. While there might still be a sample selection problem, we show in Section

results are robust to using only the sample for which I have historical faculty lists.

3.2.1 that the results are robust to using only the sample for which we have historical

From individuals’ CVs, I code where and when they received their PhDs, their em-

faculty lists.

From individuals’ CVs, we code where and when they received their PhDs, their em-

ployment and publication history, and their primary and secondary fields. When looking

at the relationship between publications and tenure in the main analysis, I only include

at the relationship between publications and tenure in the main analysis, we only include

papers that were published up to and including the year an individual goes up for tenure.

Book chapters are not included in the paper count. In a robustness check, I include papers

that were published one and two years after tenure.

The list of institutions are taken from the RePEc/IDEAS Economics Department rankings. The list of

schools included can be found in Appendix A

schools included can be found in Appendix C.

Business and policy schools might also value teaching differently and put weight on different types of

journals.

To control for the quality of a person’s publications, I primarily use the “AER equiv-

papers that were published up to and including the year an individual goes up for tenure.

Book chapters are not included in the paper count. In a robustness check, we include

papers that were published one and two years after tenure.

To control for the quality of a person’s publications, we primarily use the “AER equiv-

alent” ranking measure developed by Kalaitzidakis et al. (2003). This measure converts

journal publications into their equivalent number of American Economic Review papers

journal publications into their equivalent number of American Economic Review papers.

Less than 10% of journal articles cannot be converted because the journal does not appear

in the ranking. In these cases I give the publication a ranking of zero.

in the ranking. In these cases we give the publication a ranking of zero.

Using the AER-equivalent measure instead of a list journal rank allows for different

distances between journal ranks and for multiple journals to hold the same rank. For

example, the top field journals can all hold the same rank. Other journal rankings force

a ranking among these even though the journals might count the same amount toward

tenure depending on one’s field. For robustness, I replace this paper quality measure

tenure depending on one’s field. For robustness, we replace this paper quality measure

with the RePEc/IDEAS ranking of economics journals in Section 3.2.2.

Finally, I include citations, measured in 2015, of pre-tenure papers as a control variable.

Finally, we include citations, measured in 2015, of pre-tenure papers as a control vari-

These citations were scraped from Google Scholar.

able. These citations were scraped from Google Scholar.

I supplement this dataset with results from a survey designed to measure individuals’

We supplement this dataset with results from a survey designed to measure individu-

beliefs about the returns to various types of papers. The survey also contains information

als’ beliefs about the returns to various types of papers. The survey also contains informa-

on how frequently individuals present their papers. The exact questions and nature of the

tion on how frequently individuals present their papers. The exact questions and nature

survey are discussed in greater detail in Section 4.

of the survey are discussed in greater detail in Section 4.

2.2 Construction of Tenure

To determine whether someone received tenure, I follow the guidelines on each school’s

To determine whether someone received tenure, we follow the guidelines on each school’s

website (as of 2015) as to when tenure decisions are made. The majority of schools require

faculty to apply for tenure 7 years after their initial appointment. I therefore consider

faculty to apply for tenure 7 years after their initial appointment. We therefore consider

years 6-8 to be the “tenure window” in which someone applies for tenure to account for

people who go up for tenure early or late (because of a leave of absence, for example). I as-

people who go up for tenure early or late (because of a leave of absence, for example).

sume that an individual is denied tenure if s/he moves to a university ranked 5 positions

We assume that an individual is denied tenure if s/he moves to a university ranked 5

below the initial institution during the tenure window. Similarly, I assume that an indi-

positions below the initial institution during the tenure window. Similarly, we assume

vidual is denied tenure if he moves from academia to industry during the tenure window.

that an individual is denied tenure if s/he moves from academia to industry during the

Defining tenure in this way accounts for the fact that some people switch institutions 2-3

tenure window. Defining tenure in this way accounts for the fact that some people switch

years after their initial appointment, not because they were denied tenure but for personal

institutions 2-3 years after their initial appointment, not because they were denied tenure

preferences, and that some people might choose to move to a comparable school around

the time of tenure even though they were offered tenure at their original institution. For

The American Economic Review is regarded as one of the top journals in economics. Most journal

publications are therefore converted to be some fraction of an AER paper.

If someone does not have any solo or coauthored papers, I set the relevant journal ranking to zero and

If someone does not have any solo or coauthored papers, we set the relevant journal ranking to zero and

include a dummy variable indicating that the individual has no solo (or coauthored) papers. This enables

me to keep using the full sample.

us to keep using the full sample.

example, someone who moves from MIT to Harvard after 7 years was presumably offered

but for personal preferences, and that some people might choose to move to a comparable

tenure at MIT but chose to move to Harvard for other reasons.

school around the time of tenure even though they were offered tenure at their original

institution. For example, someone who moves from MIT to Harvard after 7 years was

presumably offered tenure at MIT but chose to move to Harvard for other reasons.

As mentioned, a person who moves 5 or fewer years after his or her initial appointment

is not assumed to have been denied tenure since s/he moved before the tenure window

starts. If someone moves before the tenure window, I use the second institution they were

starts. If someone moves before the tenure window, we use the second institution they

at to determine tenure. For example, if a person’s first job is at University A but s/he

were at to determine tenure. For example, if a person’s first job is at University A but s/he

moves to University B after three years, I use University B as the tenure institution but

moves to University B after three years, we use University B as the tenure institution but

do not start the tenure clock over. I do not restart the clock because the data shows that

do not start the tenure clock over. We do not restart the clock because the data shows that

in over 80% of cases, the individual still appears to go up for tenure within 8 years of his

or her appointment at the first institution. However, I do extend this tenure clock in a

or her appointment at the first institution. However, we do extend this tenure clock in a

robustness check.

Individuals who move from an academic institution into industry before the tenure

window are excluded from the sample.

2.3 Summary Statistics

Table 1 presents summary statistics of the data. Approximately 68% of the full sample

received tenure, but this masks a stark difference between men and women. Only 52% of

women received tenure while 73% of men did.

Total Papers,Solo-authored, andCoauthoredare the number of papers in each group that

an individual had published by the time of tenure. These publication counts do not in-

clude books or book chapters. Papers published in non-economics journals (such as a

political science journal) are included but receive a ranking of 0 (the lowest ranking). The

results are robust to excluding publications in non-economics journals.

There is no statistically significant difference in the number of papers that men and

women produce. Panel B looks at differences in the quality of papers. Men are no more

likely to publish their papers in "Top 5" journals (American Economic Review, Economet-

likely to publish their papers in “Top 5” journals (American Economic Review, Economet-

rica, Journal of Political Economy, Quarterly Journal of Economics, and The Review of

Economic Studies) than women. The only statistically significant productivity difference

is that men tend to publish their coauthored papers in slightly higher-ranking journals.

Specifically, men’s coauthored papers have an average ranking of 0.34 AER-equivalents

while women’s coauthored papers have an average ranking of 0.30 AER-equivalents. I

while women’s coauthored papers have an average ranking of 0.30 AER-equivalents. We

therefore control for the quality of papers, measured using the AER-equivalent ranking as

well as average citations, throughout the analysis.

Panel C displays differences in coauthoring patterns between men and women.Num-

ber Unique CAsis the number of unique coauthors an individual has had by tenure. Men

and women have roughly the same number of coauthors but there are some differences in

the types of people men and women coauthor with. For example, women are less likely to

coauthor with senior faculty and more likely to coauthor with other assistant professors.

This could in part be driven by the fact that they are also more likely to coauthor with

other women, many of whom are also junior professors.

For illustrative purposes, I plot the number of women and men who have various

For illustrative purposes, we plot the number of women and men who have various

combinations of solo and coauthored papers in Figure 1, as well as the average proba-

combinations of solo and coauthored papers in Figure 1, as well as the average probability

bility of receiving tenure for each paper combination in Figure 2. Most men and women

of receiving tenure for each paper combination in Figure 2. Most men and women have

have a similar combination of solo and coauthored papers. Figure 2 illustrates that in-

a similar combination of solo and coauthored papers. Figure 2 illustrates that individu-

dividuals with a large number of either solo or coauthored papers are likely to receive

als with a large number of either solo or coauthored papers are likely to receive tenure.

tenure. However, Panel A suggests that women with a higher fraction of their papers that

However, Panel A suggests that women with a higher fraction of their papers that are

are solo-authored have a better chance of receiving tenure than women with a mix of solo

solo-authored have a better chance of receiving tenure than women with a mix of solo

and coauthored papers. I examine this claim formally in the next section.

and coauthored papers. We examine this claim formally in the next section.

3 Empirical Strategy and Results

3.1 Main Results

I show three main results. I first establish that a significant tenure gap exists between

We show three main results. We first establish that a significant tenure gap exists between

men and women. I then show that the gap becomes more pronounced the more women

men and women. We then show that the gap becomes more pronounced the more women

coauthor, and that women who solo-author all of their papers have comparable tenure

rates to men. Finally, I show that the gender of a woman’s coauthor matters. Women who

rates to men. Finally, we show that the gender of a woman’s coauthor matters. Women

coauthor with other women do not suffer a coauthor penalty.

who coauthor with other women do not suffer a coauthor penalty.

3.1.1 The Tenure Gap

Figure 3 plots the coefficient

from estimating

if st

=β

TotPapers

+β

TotPapers

+γ

′

+θ

if st

(1)

separately for men and women using OLS. The dependent variable,T

if st

, is an indicator

that individualiin fieldfat schoolsreceives tenure in yeart.TotPapers

is the number of

papers (both coauthored and solo-authored) individualihas at the time he or she went up

for tenure. A quadratic in the number of papers is included to capture non-linearities in

how publications matter for tenure. The vector of individual-level controls,Z

, includes

average journal rank (measured as average AER-equivalents), the log of total citations,

the number of years it tookito go up for tenure, and the average number of coauthors

the number of years it tookito go up for tenure, and the total number of coauthors oni’s

papers. Tenure institution (θ

oni’s papers. Tenure institution (θ

), tenure year (θ

), and field fixed effects (θ

) are also

) are also included

included as tenure standards likely vary over time and by field and department.

as tenure standards likely vary over time and by field and department.

The figure shows that a significant tenure gap exists between men and women even

after controlling for productivity, primary field, tenure institution, and tenure year. While

an additional paper is correlated with a 13-16 percentage point increase in tenure proba-

bility for men and women, women are consistently 10-13 percentage points less likely to

receive tenure than men conditional on having written the same number and quality of

papers. The lower intercept for women could stem from tenure committees starting with a

lower prior about women’s ability. However, if all papers were clear signals of ability and

tenure committees are Bayesian, we would expect the slope of the relationship between

papers and tenure to be steeper for women. Put differently, if men and women received

equal credit for papers, the coefficient onTotPapers

should be significantly larger for

women than for men.

I provide a formal test for the difference in slopes for men and women in Column 1 of

We provide a formal test for the difference in slopes for men and women in Column 1

Table 2, where I present the estimates from

of Table 2, where we present the estimates from

if st

=β

TotPapers

+β

fem

+β

(TotPapers

×fem

) +β

TotPapers

+γ

′

+θ

if st

(2)

This is similar to estimating equation 1 except that I interact total papers with a female

This is similar to estimating equation 1 except that we interact total papers with a female

dummy,fem

rather than splitting the sample. There is no significant difference in the

marginal benefit of an additional paper to men and women.

3.1.2 The Tenure Gap and Paper Composition

To test whether coauthored papers matter differently for men and women, I separate pa-

To test whether coauthored papers matter differently for men and women, we separate

pers into those that are solo-authored and those that are coauthored and estimate

papers into those that are solo-authored and those that are coauthored and estimate

ifst

=β

+β

(fem

×S

) +β

+β

(fem

×CA

) +δ

fem

+γ

′

+θ

ifst

(3)

using OLS. Here,S

andCA

are the number of solo-authored and coauthored papers an

individual has at the time of tenure.

The results are presented in Table 2. An additional solo-authored paper is associated

with a 9.7 percentage point increase in men’s tenure rates and a 15.4 percentage point

increase in women’s tenure rates (who start from a lower base tenure rate). If the lower

initial tenure rate for women is due to employers holding the belief that women are lower

ability, it seems that the signals from solo papers begin to outweigh the employer’s prior.

This is consistent with a model in which employers start with a lower prior about women

and update as they receive clear signals about a woman’s ability, giving women full credit

for this solo work. This is further discussed in the next section.

If coauthored papers are an unclear signal of ability, an employer must make a judg-

If coauthored papers are an

ment call as to how much each coauthor contributed to the paper which could lead to

differential attribution of credit. Indeed, we see that while an additional coauthored pa-

per helps both men and women, men benefit much more than women. Men’s tenure

rates increase by 8.2 percentage points when they produce a coauthored paper whereas

women’s increase by 5.6 percentage points.

However, the fact

Saved diffs

Original text

Open file

Gender Differences in Recognition for Group Work
Heather Sarsons
∗
February 4, 2019
Abstract
Does gender influence how credit for group work is allocated?  Using data from
academic economists’ CVs, I test whether coauthored and solo-authored publications
matter differently for tenure for men and women. Because coauthors are listed alpha-
betically in economics, coauthored papers do not provide specific information about
each contributor’s skills or ability.  Solo-authored papers, on the other hand, provide
a relatively clear signal of ability.  I find that conditional on publication quality and
other observables,  men are tenured at roughly the same rate regardless of whether
they coauthor or solo-author.  Women, however, become less likely to receive tenure
the more they coauthor.  The result is most pronounced for women coauthoring with
men and less pronounced among women who coauthor with other women.  An on-
line experiment finds similar patterns when women perform male-stereotyped tasks.
However, women receive equal credit for joint work with men when the task is per-
ceived to be gender-neutral. Taken together, the results suggest that gender and stereo-
types influence the allocation of credit for group work.
∗
I especially thank Roland Fryer, Claudia Goldin, Larry Katz, David Laibson, and Amanda Pallais for
their guidance and encouragement.  I also thank Mitra Akhtari, Amitabh Chandra, John Coglianese, Oren
Danieli, Ellora Derenoncourt, Florian Ederer, Ben Enke, Raissa Fabregas, Nicole Fortin, Peter Ganong, Ed-
ward Glaeser, Siri Isaksson, Sara Lowes, Rob McMillan, Eduardo Montero, Gautam Rao, Alex Segura, Nihar
Shah, Peter Tu, Justin Wolfers, and participants at SOLE 2016, the Early Behavioural Economics Conference,
the Harvard Business School Gender Initiative, and the Harvard Inequality Seminar for helpful comments
and suggestions. This paper is intentionally solo-authored.
1

1    Introduction
Do employers use gender when allocating credit for group work, particularly when in-
dividual contributions are unobserved?  Organizations increasingly rely on group work
for production (Lazear and Shaw, 2007), yet there is little empirical evidence document-
ing how credit for group work is allocated.  Unless employers can perfectly observe each
worker’s contribution to the team’s output, they must decide how to allocate credit with-
out having full information as to what each worker did. This could leave room for demo-
graphic characteristics, such as gender, to influence the allocation of credit.
In this paper, I test whether uncertainty over an individual’s contribution to a project
leads to differential attribution of credit that contributes to the gender promotion gap. In
many industries,  women are not only hired at lower rates than men are,  they are also
promoted at lower rates.
1
This paper explores whether gender differences in credit for
group work exist and whether they explain part of the promotion gap.
I primarily look at the tenure decisions of academic economists to test whether gender
influences the allocation of credit for coauthored papers.  Economics is a relevant setting
as there is a large tenure gap between men and women, and because the amount of coau-
thoring has risen dramatically in recent years (Ginther and Kahn,  2004;  Hammermesh,
2013).  Using data from economists’ CVs, I track individuals’ career trajectories and com-
pare whether the trajectory is different for individuals who coauthor versus solo-author,
and whether there is a difference by gender.
Within economics,  I find that men and women who solo-author most of their work
have similar tenure rates conditional on a proxy for the quality of papers.  However, an
additional coauthored paper is correlated with an 8.2% increase in tenure probability for
men but only a 5.6% increase for women.  This gap is significantly less pronounced for
women  who  coauthor  with  women,  suggesting  that  the  attribution  of  credit  is  related
to the gender mix of coauthors.  Furthermore, a man who coauthors is no less likely to
receive tenure than a comparable man who solo-authors even though there is presumably
more uncertainty as to how much work he did.  A counterfactual exercise suggests that
this difference in credit allocation explains 60% of the unconditional gender gap in tenure
rates and 84% of the gap that remains after controlling for average paper quality, citations,
tenure and PhD institution ranks, and field.
To ensure that I am not picking up on ability differences between men and women, I
control for the quality of papers using both journal rankings and citations, allowing for
1
Blau and DeVaro (2007), for example, find that across jobs, women are less likely to be promoted than
men conditional on worker’s performance and ability ratings.  In the UK, female managers are nearly 40%
less likely to be promoted than male managers (Elmins et al. 2016).
2

a comparison of men and women with similar research portfolios.  The results are also
robust to including other individual-level controls such as length of time to tenure and
the seniority of one’s coauthors, as well as tenure year, tenure institution, and primary
field fixed effects.
I argue that these results are most consistent with a story of women receiving less credit
for their joint work with men because of bias.  To show this, I first use current CV and
citation data to compare the productivity of men and women who did and did not receive
tenure at the institution where they initially went up for tenure.  While the estimates are
imprecise, I find suggestive evidence that women who coauthor and are denied tenure
produce more solo-authored papers that publish in high-ranking journals than men who
are denied tenure. Data on citations show a similar result.
I then rule out several alternative explanations for the empirical patterns. For example,
several papers have demonstrated that selection into coauthorship in economics is not
random.
2
I test for selection into coauthorship and do not find any evidence that women
coauthor with high ability or more senior men.  I also look at the timing of coauthorship
and find no evidence that women begin coauthoring if they have a slower start to their
careers. The empirical patterns are also inconsistent with taste-basted discrimination.
Because the CV data does not allow me to rule out the possibility that women actually
contribute less to papers that are coauthored with men, I conduct an online experiment
designed to test whether this drives credit allocation.  In the experiment, I first hire indi-
viduals to complete quizzes on topics that are either male or female-stereotyped.  I then
hire participants who act as “predictors” and are randomized into a solo treatment or a
group treatment.   Predictors in the solo treatment are shown two individual’s separate
quiz scores while predictors in the group treatment are shown the combined score of two
individuals. They are then asked to predict the performance of each participant on future
quizzes.
I find that in the group treatment, women are predicted to perform worse than their
male counterparts for male-stereotyped quizzes, suggesting that participants making the
predictions believe that women contributed less to the combined score (that is, they per-
formed worse). However, if pairs performed a female-stereotyped quiz, women and men
are given equal credit.  To understand whether these results are driven by participants’
beliefs about the ability distributions of men and women, I randomly provide some par-
2
Boschini and Sjögren (2007) test whether coauthorship patterns in economics are gender neutral.  They
find that women are more likely to solo-author than men and that gender homophily in coauthoring in-
creases with women’s representation in a field. Garcia and Sherman (2015) argue that the alphabetical pub-
lishing norms in economics influence both the types of people authors work with and the types of projects
they will work on.
3

ticipants with the distribution of scores on the initial quiz by gender.  Women appear to
be given equal credit in the female-stereotyped quiz because participants view it as being
gender-neutral.  That is, they do not realize that women tend to outperform men.  Show-
ing participants the gender distribution of scores corrects this belief and women are then
predicted  to  have  a  better  performance  in  future  female-stereotyped  quizzes.   Surpris-
ingly, this treatment does not affect the predicted performance gap for women and men
performing male-stereotyped tasks.
This paper replicates and builds off of the results in Sarsons (2017), which shows the
basic correlational patterns between paper composition and tenure. In this paper, I repli-
cate the results using more data and then use the C.V. data and an experiment to establish a
channel through which gender influences the allocation of credit. The paper also relates to
a large literature seeking to understand difference in labor market outcomes between men
and women.  Factors such as productivity, personality and behavioural differences (such
as competition aversion), and fertility preferences have been shown to explain some differ-
ences in career choice and progression.
3
In academia in particular, studies have pointed
to both supply-side factors, including differences in subject matter interest (Dynan and
Rouse, 1997) and the availability of role models (Hale and Regev, 2014; Carrell, Page, and
West, 2010); demand-side factors, such as implicit bias (Milkman, Akinola, and Chugh,
2015; Moss-Racusin et al., 2012); and institutional factors (Antecol et al., 2018). This paper
directly tests whether the differential treatment of work output contributes to the gender
gap.
The remainder of the paper is organized as follows.  Section 2 describes the data and
shows that a tenure gap exists between male and female economists. In Section 3, I show
that the tenure gap closes as women produce more solo-authored papers but does not
close as they produce more coauthored papers. Women have a consistently lower proba-
bility of tenure for each additional coauthored paper than men. I show that the results are
robust to accounting for attrition, and to using different journal rankings and definitions
of tenure.  In Section 4, I argue that the results are in line with a story in which women
receive less credit for joint work with men.  I test alternative explanations of the relation-
ship between coauthorship and tenure and argue that none can fully explain the observed
empirical patterns. Section 5 discusses the structure and results of the experiment. Section
6 concludes.
3
There is a large literature documenting gender differences in productivity, attitudes toward different
types of work, and family choices.  See, for example, Niederle and Vesterlund (2007), Antecol et al.  (2016),
Ceci et al. (2014), and Ginther and Kahn (2004).
4

2    Data
To examine the relationship between paper composition and tenure, I construct a dataset
using the CVs of economists who came up for tenure between 1985 and 2014 at one of the
top 35 U.S. PhD-granting universities
4
. The academic progression documented in the CVs
makes it possible to evaluate the relationship between an individual’s research output and
career progression.  I can then compare the degree of collaborative work and reward for
that work, and compare these results for men versus women.
2.1    Sample Selection and Data Overview
I include only PhD-granting institutions in the sample as tenure evaluation at these schools
is primarily based on research output, of which I have a clear measure. Other institutions
like  liberal  arts  colleges  place  greater  weight  on  teaching  ability  for  tenure,  something
that I cannot measure.  I exclude business and public policy schools for similar reasons.
5
It is reasonable to assume that the top 35 economics departments in the U.S. emphasize
research which is measured by the number and quality of papers one produces.
One problem in collecting tenure information is that the CVs of individuals who went
up for tenure, were denied it, and left to industry or government are difficult to find, lead-
ing to a sample selection problem. To deal with this issue, I collected historical faculty lists
from 23 of the 35 schools and locate over 90% of faculty who had ever gone up for tenure
at these 23 institutions.  For the remaining 12 schools that did not have historical faculty
lists available, I looked at the top 75 U.S. institutions, the top 5 Canadian institutions, and
the top 5 European institutions to locate anyone who went up for tenure at a top 35 U.S.
school and then moved to another institution.  I also checked economists’ CVs at the ma-
jor Federal Reserve Boards and other large research institutes, such as Mathematica, in the
U.S. While there might still be a sample selection problem, I show in Section 3.2.1 that the
results are robust to using only the sample for which I have historical faculty lists.
From individuals’ CVs, I code where and when they received their PhDs, their em-
ployment and publication history, and their primary and secondary fields. When looking
at the relationship between publications and tenure in the main analysis, I only include
papers that were published up to and including the year an individual goes up for tenure.
Book chapters are not included in the paper count. In a robustness check, I include papers
that were published one and two years after tenure.
4
The list of institutions are taken from the RePEc/IDEAS Economics Department rankings.  The list of
schools included can be found in Appendix A
5
Business and policy schools might also value teaching differently and put weight on different types of
journals.
5

To control for the quality of a person’s publications, I primarily use the “AER equiv-
alent” ranking measure developed by Kalaitzidakis et al.  (2003).  This measure converts
journal publications into their equivalent number of American Economic Review papers
6
.
Less than 10% of journal articles cannot be converted because the journal does not appear
in the ranking. In these cases I give the publication a ranking of zero.
7
Using the AER-equivalent measure instead of a list journal rank allows for different
distances  between  journal  ranks  and  for  multiple  journals  to  hold  the  same  rank.   For
example, the top field journals can all hold the same rank.  Other journal rankings force
a ranking among these even though the journals might count the same amount toward
tenure  depending  on  one’s  field.   For  robustness,  I  replace  this  paper  quality  measure
with the RePEc/IDEAS ranking of economics journals in Section 3.2.2.
Finally, I include citations, measured in 2015, of pre-tenure papers as a control variable.
These citations were scraped from Google Scholar.
I supplement this dataset with results from a survey designed to measure individuals’
beliefs about the returns to various types of papers. The survey also contains information
on how frequently individuals present their papers. The exact questions and nature of the
survey are discussed in greater detail in Section 4.
2.2    Construction of Tenure
To determine whether someone received tenure, I follow the guidelines on each school’s
website (as of 2015) as to when tenure decisions are made. The majority of schools require
faculty to apply for tenure 7 years after their initial appointment.   I therefore consider
years 6-8 to be the “tenure window” in which someone applies for tenure to account for
people who go up for tenure early or late (because of a leave of absence, for example). I as-
sume that an individual is denied tenure if s/he moves to a university ranked 5 positions
below the initial institution during the tenure window.  Similarly, I assume that an indi-
vidual is denied tenure if he moves from academia to industry during the tenure window.
Defining tenure in this way accounts for the fact that some people switch institutions 2-3
years after their initial appointment, not because they were denied tenure but for personal
preferences, and that some people might choose to move to a comparable school around
the time of tenure even though they were offered tenure at their original institution.  For
6
The  American  Economic  Review  is  regarded  as  one  of  the  top  journals  in  economics.   Most  journal
publications are therefore converted to be some fraction of an AER paper.
7
If someone does not have any solo or coauthored papers, I set the relevant journal ranking to zero and
include a dummy variable indicating that the individual has no solo (or coauthored) papers.  This enables
me to keep using the full sample.
6

example, someone who moves from MIT to Harvard after 7 years was presumably offered
tenure at MIT but chose to move to Harvard for other reasons.
As mentioned, a person who moves 5 or fewer years after his or her initial appointment
is not assumed to have been denied tenure since s/he moved before the tenure window
starts. If someone moves before the tenure window, I use the second institution they were
at to determine tenure.   For example,  if a person’s first job is at University A but s/he
moves to University B after three years, I use University B as the tenure institution but
do not start the tenure clock over.  I do not restart the clock because the data shows that
in over 80% of cases, the individual still appears to go up for tenure within 8 years of his
or her appointment at the first institution.  However, I do extend this tenure clock in a
robustness check.
Individuals who move from an academic institution into industry before the tenure
window are excluded from the sample.
2.3    Summary Statistics
Table 1 presents summary statistics of the data.  Approximately 68% of the full sample
received tenure, but this masks a stark difference between men and women. Only 52% of
women received tenure while 73% of men did.
Total Papers,Solo-authored, andCoauthoredare the number of papers in each group that
an individual had published by the time of tenure.  These publication counts do not in-
clude books or book chapters.   Papers published in non-economics journals (such as a
political science journal) are included but receive a ranking of 0 (the lowest ranking). The
results are robust to excluding publications in non-economics journals.
There is no statistically significant difference in the number of papers that men and
women produce.  Panel B looks at differences in the quality of papers.  Men are no more
likely to publish their papers in "Top 5" journals (American Economic Review, Economet-
rica,  Journal of Political Economy,  Quarterly Journal of Economics,  and The Review of
Economic Studies) than women.  The only statistically significant productivity difference
is that men tend to publish their coauthored papers in slightly higher-ranking journals.
Specifically, men’s coauthored papers have an average ranking of 0.34 AER-equivalents
while women’s coauthored papers have an average ranking of 0.30 AER-equivalents.  I
therefore control for the quality of papers, measured using the AER-equivalent ranking as
well as average citations, throughout the analysis.
Panel C displays differences in coauthoring patterns between men and women.Num-
ber Unique CAsis the number of unique coauthors an individual has had by tenure.  Men
7

and women have roughly the same number of coauthors but there are some differences in
the types of people men and women coauthor with. For example, women are less likely to
coauthor with senior faculty and more likely to coauthor with other assistant professors.
This could in part be driven by the fact that they are also more likely to coauthor with
other women, many of whom are also junior professors.
For  illustrative  purposes,  I  plot  the  number  of  women  and  men  who  have  various
combinations of solo and coauthored papers in Figure 1,  as well as the average proba-
bility of receiving tenure for each paper combination in Figure 2.  Most men and women
have a similar combination of solo and coauthored papers.  Figure 2 illustrates that in-
dividuals with a large number of either solo or coauthored papers are likely to receive
tenure. However, Panel A suggests that women with a higher fraction of their papers that
are solo-authored have a better chance of receiving tenure than women with a mix of solo
and coauthored papers. I examine this claim formally in the next section.
3    Empirical Strategy and Results
3.1    Main Results
I show three main results.   I first establish that a significant tenure gap exists between
men and women.  I then show that the gap becomes more pronounced the more women
coauthor, and that women who solo-author all of their papers have comparable tenure
rates to men. Finally, I show that the gender of a woman’s coauthor matters. Women who
coauthor with other women do not suffer a coauthor penalty.
3.1.1    The Tenure Gap
Figure 3 plots the coefficient
ˆ
β
1
from estimating
T
if st
=β
1
TotPapers
i
+β
2
TotPapers
2
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(1)
separately for men and women using OLS. The dependent variable,T
if st
, is an indicator
that individualiin fieldfat schoolsreceives tenure in yeart.TotPapers
i
is the number of
papers (both coauthored and solo-authored) individualihas at the time he or she went up
for tenure.  A quadratic in the number of papers is included to capture non-linearities in
how publications matter for tenure.  The vector of individual-level controls,Z
i
, includes
average journal rank (measured as average AER-equivalents),  the log of total citations,
the number of years it tookito go up for tenure, and the average number of coauthors
8

oni’s papers.  Tenure institution (θ
s
), tenure year (θ
t
), and field fixed effects (θ
f
) are also
included as tenure standards likely vary over time and by field and department.
The figure shows that a significant tenure gap exists between men and women even
after controlling for productivity, primary field, tenure institution, and tenure year. While
an additional paper is correlated with a 13-16 percentage point increase in tenure proba-
bility for men and women, women are consistently 10-13 percentage points less likely to
receive tenure than men conditional on having written the same number and quality of
papers. The lower intercept for women could stem from tenure committees starting with a
lower prior about women’s ability. However, if all papers were clear signals of ability and
tenure committees are Bayesian, we would expect the slope of the relationship between
papers and tenure to be steeper for women.  Put differently, if men and women received
equal credit for papers,  the coefficient onTotPapers
i
should be significantly larger for
women than for men.
I provide a formal test for the difference in slopes for men and women in Column 1 of
Table 2, where I present the estimates from
T
if st
=β
1
TotPapers
i
+β
2
fem
i
+β
3
(TotPapers
i
×fem
i
) +β
4
TotPapers
2
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(2)
This is similar to estimating equation 1 except that I interact total papers with a female
dummy,fem
i
rather than splitting the sample.  There is no significant difference in the
marginal benefit of an additional paper to men and women.
3.1.2    The Tenure Gap and Paper Composition
To test whether coauthored papers matter differently for men and women, I separate pa-
pers into those that are solo-authored and those that are coauthored and estimate
T
ifst
=β
1
S
i
+β
2
(fem
i
×S
i
) +β
3
CA
i
+β
4
(fem
i
×CA
i
) +δ
1
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
ifst
(3)
using OLS. Here,S
i
andCA
i
are the number of solo-authored and coauthored papers an
individual has at the time of tenure.
The results are presented in Table 2.  An additional solo-authored paper is associated
with a 9.7 percentage point increase in men’s tenure rates and a 15.4 percentage point
increase in women’s tenure rates (who start from a lower base tenure rate).  If the lower
initial tenure rate for women is due to employers holding the belief that women are lower
ability, it seems that the signals from solo papers begin to outweigh the employer’s prior.
This is consistent with a model in which employers start with a lower prior about women
9

and update as they receive clear signals about a woman’s ability, giving women full credit
for this solo work. This is further discussed in the next section.
If coauthored papers are an unclear signal of ability, an employer must make a judg-
ment call as to how much each coauthor contributed to the paper which could lead to
differential attribution of credit.  Indeed, we see that while an additional coauthored pa-
per  helps  both  men  and  women,  men  benefit  much  more  than  women.   Men’s  tenure
rates increase by 8.2 percentage points when they produce a coauthored paper whereas
women’s increase by 5.6 percentage points.
However, the fact that men benefit nearly as much from a coauthored paper as they do
from a solo-authored paper is at odds with the story that employers are dividing credit
for projects among authors. If employers do divide credit, not all men can get 100% of the
credit, particularly for those papers coauthored with other men.
8
This result could point to
an alternative mechanism.  For example, if employers exhibit taste-based discrimination,
they could use joint projects as an excuse to promote men over women. I discuss and test
several such alternative stories in Section 4.
The relationship between paper composition and tenure is summarized in Figure 4.
This figure plots the relationship between the fraction of an individual’s papers that are
solo-authored and tenure,  controlling for the total number of papers,  citations,  journal
quality,  number  of  coauthors,  and  tenure  institution,  year,  and  field  fixed  effects.   For
men,  it  does  not  matter  if  one  coauthors  or  solo-authors:  tenure  rates  are  comparable
conditional on the quality of papers.   Women who write all of their papers alone have
similar tenure rates to men.  However, women who coauthor all of their papers have an
approximately 37% tenure rate, substantially lower than that of men who coauthor all of
their papers ( 72%). The slope for women is
ˆ
β=0.780and is statistically significant at the
1% level (s.e.=0.184).
3.1.3    Does Coauthor Gender Matter?
The probability of receiving tenure is not lower for all women who coauthor. In Table 3, I
categorize coauthored papers into those written with only men, only women, or a mix of
men and women:
8
It could be the case that because tenure committees are evaluating one person, they always assume that
the man they evaluate deserves full credit for the paper (and we do not see the amount of credit they would
have given to the other man). It is impossible to evaluate such theories with these data.
10

T
if st
=β
1
S
i
+β
2
(fem
i
×S
i
) +β
3
CAmale
i
+β
4
(fem×CAmale
i
) +β
5
CAmix
i
+β
6
(fem×CAmix
i
) +β
7
CAfem
i
+β
8
(fem
i
×CAfem
i
) +β
9
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(4)
As before,S
i
is the number of solo-authored papers individualihas at the time of tenure.
CAfem
i
is the number of coauthored papers individualihas in which all of the coauthors
are female. Similarly,CAmale
i
is the number of papersihas in which all of the coauthors
are male andCAmix
i
is the number of papersihas in which the coauthors consist of men
and women.
The estimated coefficients on the interaction terms show that the negative relationship
between coauthoring and tenure for women is driven almost entirely by papers that are
coauthored with men. While a coauthored paper with another man is associated with an
8.7 percentage point increase in tenure probability for a man, it is associated with a 3.1
percentage point increase in tenure probability for a woman.
9
An additional paper with
a woman, however, is associated with an 11.6 percentage point increase in tenure proba-
bility for a woman.  While this estimate is imprecise due to sample size, I can say that an
additional coauthored paper with a woman has a more positive impact on tenure than an
additional coauthored paper with a man. Any explanation as to why women have lower
tenure rates than men when they coauthor must therefore be correlated with coauthor
gender. The estimates are robust to including all of the control variables discussed earlier.
3.1.4    Counterfactual Analysis
I conduct a counterfactual analysis to estimate how much of the gender gap in tenure rates
can be explained by the different treatment of coauthored papers. I first estimate
T
if st
=β
1
S
i
+β
2
CA
i
+δ
1
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(5)
and use the estimates to predict the probability of tenure,
ˆ
T
i
, for everyone in the sample.
I then let the female dummyfem
i
be 0 for everyone and predict tenure rates again (call
this
 ̃
T
i
).   The  difference
ˆ
T
i
−
 ̃
T
i
gives  the  gender  gap  in  tenure  rates  conditional  on  all
observable characteristics but not allowing for differences in the marginal impact of solo
9
These results again show the puzzling pattern that the amount of credit that is divided among male
coauthors adds up to more than one.
11

and coauthored papers for men and women.
10
I then repeat this exercise using the estimates from equation 4, first letting the female
dummy equal one and then predicting tenure rates again letting the female dummy (and
therefore all of the interactions) equal zero.  This second set of predicted tenure probabil-
ities tells us what women’s predicted tenure rate would be if their papers were treated in
the same way that men’s papers are treated.
The unconditional gender gap in tenure rates is 22 percentage points. The conditional
gap in tenure rates from equation 5 is approximately 16 percentage points. Thus, observ-
able characteristics such as differences in time to tenure and paper quality account for
about 27% of the gap.  The results from using equation 4 to predict tenure probabilities
suggest that the gap would close by a further 13.5 percentage points if men and women’s
papers were treated similarly.   The different assignment of credit thus accounts for ap-
proximately 60% of the unconditional tenure gap and 84% of the conditional gap.
3.2    Robustness Checks
One may be concerned that the results are a product of the types of productivity measures
used or are affected by missing data.  In this section, I show that the results are robust to
using only the sample for which I have historical faculty lists, to using different journal
rankings, to accounting for papers published shortly after tenure, and to using different
measures of paper counts.
11
3.2.1    Attrition
The results will be biased if the sample excludes individuals who are denied tenure and
go into industry,  government,  or other institutions where I do not observe them.   This
would be particularly problematic if men who go to industry after being denied tenure
disproportionately coauthored their papers.  If this is true, I would be overestimating the
benefit  of  coauthoring  for  men.   I  would  have  a  similar  problem  if  women  who  go  to
industry after being denied tenure typically wrote solo-authored papers.
As discussed in Section 2.1, I attempted to find such individuals by searching institu-
tions outside of the top 35 U.S. schools, federal reserves, and other research institutes. To
10
Interacting all variables except for the number of solo/coauthored papers with the female dummy does
not substantially change the results.
11
In Appendix Table A1, I also test whether the results vary by school rank and over time. The estimates
suggest that the coauthoring penalty is driven largely by schools outside of the top 10, although the esti-
mates are imprecise.  The coauthorship penalty is also stronger in later years but again the estimates are
imprecise.
12

further allay concerns about sample selection, I run the analysis on the sample for which I
received historical faculty lists. These lists allow me to track who went up for tenure and
find them even if they left academia. The results, presented in Column 1 of Table 4, do not
change when run on the sample for which there should be very few missing observations.
The coefficient on theFemale×Coauthoredinteraction is significant only at the 10% level
due to the smaller sample, but the direction and magnitude do not change.
3.2.2    Journal Rankings
In the main analysis, I use a flexible journal ranking that allows multiple journals to hold
the same rank. However, while the economics profession largely agrees on what the “top”
journals are, rankings of field journals or lower-tier journals have changed over time.  In
Columns 2-4 of Table 4, I show that the results are robust to using three alternative journal
ranking metrics as controls.
In  Column  2,  I  use  the  current  RePEc-IDEAS  journal  ranking.   This  ranking  forces
a linear relationship between journals and tenure but also contains a larger number of
journals. The main results do not change when using this ranking.
In Column 3, I allow journal rankings to change over time.  I use historical rankings
of economics journals (drawn from Laband and Piette, 1994, and combined with current
rankings) and match each paper with its journal ranking at the time it was published. Us-
ing these rankings accounts for journals moving in rank over time as well as new journals
being added.  The coefficient on theFemale×Coauthoredinteraction is slightly smaller
but the same pattern persists.  An additional coauthored paper is associated with an 8.1
percentage point increase in tenure probability for men and a 5.6 percentage point increase
for women. In section 4, I also separate papers into "Top 5s" and "non-Top 5s".
Finally, in Column 4, I divide the AER-equivalent measure into deciles and control for
the number of solo and coauthored papers an individual has in each decile. For example,
if an individual publishes one solo-authored paper in the AER and another in the lowest-
rank journal, she will have one paper in the tenth bin, one in the first bin, and zero in the
others.  Thus, instead of having a single coauthored or solo-authored paper rank control,
I include ten variables controlling for the quality of an individual’s solo-authored papers
(the number of solo papers in each AER-equivalent bin) and ten variables controlling for
the quality of an individual’s coauthored paper (the number of coauthored papers in each
AER-equivalent bin). Again, the results hold.
13

3.2.3    Tenure Definition
In the main analysis, I only consider papers that were published up to and including the
year that an individual goes up for tenure.  If an individual goes up for tenure in 1995,
for example, papers published in 1996 are not included in the paper count even though
they may have been “revise and resubmits” at the time of tenure.  This could affect the
results if men who coauthor have several promising unpublished papers at the time of
tenure but women who coauthor do not, in which case I am not actually comparing people
with similar publication records. In Columns 5 and 6 of Table 4, I include papers that are
published one and two years after a person’s tenure year in the paper count variables. The
magnitude of the coefficients are smaller but the results do not change: women continue
to benefit less from coauthored papers than men do.
3.2.4    Paper Count Variable
While I control for journal quality, the main independent variables (number of solo and
coauthored papers) may not accurately reflect how tenure committees decide on tenure
cases.   For  example,  institutions  might  trade  off  the  quantity  and  quality  of  papers  in
different ways.  In Column 7 of Table 4, I use an alternative measure for the number of
papers.   Specifically,  after  converting  each  publication  to  its  AER-equivalent,  I  add  up
the AER-equivalent measure to give the total number of "AERs" an individual has at the
time of tenure.  For example,  if an individual published two solo-authored papers and
one is worth 0.25 AERs and the other worth 0.8 AERs, the individual will have 1.05 solo-
authored AERs at the time of tenure.
Again, the patterns are the same. An additional coauthored “AER” paper is correlated
with an 8.9 percentage point increase in a man’s tenure probability but a 5.3 percentage
point increase in a woman’s tenure probability.
3.3    Testing Against Other Disciplines and Coauthoring Conventions
Many disciplines use different coauthoring conventions, such as listing authors in order
of contribution. However, these disciplines differ on several other dimensions, such as the
fraction of women in the disciplines and what is most important for tenure (publications,
grants, conference proceedings, etc.).  In Appendix A, I conduct the same analysis for a
sample of sociologists, a discipline that order authors by contribution.  The sample and
results are discussed in more detail in the Appendix, but I do not find evidence of women
being penalized for coauthoring. What matters is being first author on a paper: being first
author is correlated with a 5% increase in tenure probability for both men and women.
14

Because sociology differs from economics in many ways, though, it is difficult to interpret
whether these results suggest that ordering authors by contribution helps eliminate bias
or whether the larger presence of women helps to eliminate it.
4    Channels
The previous section established three facts:
1.  For very few papers, women have a lower tenure probability than men;
2.  As women produce more solo-authored papers, their tenure probability converges
to that of comparable men;
3.  Women benefit less than men from work coauthored with men.
There are several explanations for these patterns. In this section, I argue that the results are
most consistent with a story of women receiving less credit for their joint work with men
rather than a story of women contributing less when they work with men.  I assume that
tenure committees begin with the prior that women are on average lower ability than men,
and that solo-authored papers provide a clear signal of one’s ability whereas coauthored
papers provide an unclear signal. Employers then misattribute credit for work produced
by a man and a woman as the man is assumed to be higher ability.
I first test the claim by comparing the productivity of men and women who were de-
nied tenure. I then explore and rule out several threats to this story. Specifically, I test for
ability and preference-based sorting, women receiving less exposure by presenting less,
and taste-based discrimination.  Finally, I present evidence from an experiment designed
to shut down the possibility that women put in less effort when working with men, and
find that even in this context, women receive less credit than men when they perform a
stereotypically male task.  However, women receive at least as much credit as men when
they perform a stereotypically female task.
4.1    Do Men Get the Credit or Do Women Contribute Less?
If tenure committees hold the prior that women are lower ability than men and if solo-
authored papers provide clear signals of ability, we will see differences in tenure rates for
men and women with few publications. However, additional solo-authored publications
of the same quality will have a larger marginal impact on a woman’s tenure probability
15

than a man’s. As these clear signals begin to dominate the committee’s prior, tenure rates
between men and women will converge.
If committees are biased toward giving men more credit for work coauthored with
women, we would expect to see the following. Assuming that there is some fixed amount
of credit that can be given for a paper, a man will benefit more than a woman from joint
work between them. In addition, both men and women will benefit more from their coau-
thored work with women than their coauthored work with men, as two men who coau-
thor will be assumed to have contributed similarly while a woman will be assumed to
have contributed less.
These two claims largely play out in the data.  Table 2 shows that the marginal solo-
authored paper helps women more than it helps men as they start from a lower baseline
tenure rate.  Table 3 shows that men benefit the most from coauthoring with women (an
increase in tenure probability of 9.7% when coauthoring with a woman vs.  8.7% when
coauthoring with a man) although this difference is insignificant.  Similarly, women ben-
efit more from coauthoring with other women than with men.   One result that it is in-
consistent with a story of credit allocation is the fact that the total amount of credit that
can be allocated, at least when all coauthors are men, seems to add up to more than one.
Men benefit as much from a coauthored paper as they do from a solo-authored paper,
suggesting that tenure committees are either making a mistake when dividing credit (for
example,  each committee assumes that the male author under consideration for tenure
at its school did most of the work), or that there is an alternative mechanism behind the
results. In Section 4.2, I test several potential mechanisms.
We would see these same empirical patterns if women contribute less to projects that
are joint with men.  Comparing the productivity of men and women who were denied
tenure helps to disentangle these two stories. If women who coauthor are given less credit,
then women who coauthor and are denied tenure should on average be more productive
than  men  who  are  denied  tenure.   If  women  who  coauthor  simply  contribute  less,  we
would not expect to see productivity differences between men and women who are denied
tenure.
I use two productivity measures to test whether women who coauthor and are denied
tenure are more productive than men:  the number of solo-authored AER-equivalents an
individual publishes after the tenure decision and the log number of citations an individ-
ual has as of 2015.
1213
Individuals who leave academia and do not publish after tenure
are excluded from the AER-equivalent outcome sample, but including them and setting
12
Citations were scraped from Google scholar in 2015.
13
For the top 5 papers outcome, I do not compare coauthored papers as these can reflect the ability of
one’s coauthors. Citation data includes both solo and coauthored papers as the data came in this structure.
16

their number of post-tenure papers to zero does not change the results.
Table 5 shows the results from estimating
Y
if st
=β
1
fem
i
+β
2
FracCA
it
+β
3
T
i
+β
4
(fem
i
×FracCA
it
) +β
5
(fem
i
×T
i
)
+β
6
(FracCA
it
×T
i
) +β
7
(FracCA
it
×T
i
×Fem
i
) +X
′
i
γ+θ
f
+θ
t
+θ
p
+
if st
(6)
where the outcome variableY
if st
is one of the two productivity measures described above
andT
i
is a tenure dummy. I include a post-tenure institution fixed effect,θ
p
, to account for
the fact that individuals will have access to different resources depending on where they
go after the initial tenure decision.
Column  1  shows  the  results  from  estimating  equation  6  with  the  number  of  solo-
authored AER-equivalents as the outcome.  Women who are denied tenure and coauthor
have 0.4 more solo-authored AER-equivalents than men who are denied tenure and coau-
thor.  Column 2, which has log citations as the outcome variable, shows a similar pattern
although the results are much noisier.   Together,  these results provide some suggestive
evidence that these women receive less credit for joint projects.
4.2    Alternative Stories
There are other possible explanations for the above findings, not all of which can be tested
with  these  particular  data.   Here  I  shed  light  on  four  standard  and  testable  channels:
ability-based sorting, preference-based sorting, women not claiming credit for their work,
and  taste-based  discrimination.   The  empirical  patterns  are  inconsistent  with  all  of  the
proposed explanations.
4.2.1    Ability-Based Sorting
Employers might rationally deny women who coauthor tenure if individuals sort such
that only lower ability women coauthor with men.  This could arise for several reasons.
For example, if coauthoring lowers the cost of producing a paper, but women know that
they receive less credit for papers, high ability women might forego the cost savings and
choose to work alone.  They know they can produce high quality papers by themselves
and send the employer a clearer signal of their ability.  However,  if low ability women
can  only  produce  high  quality  papers  with  the  help  of  a  high  ability  man,  they  might
coauthor even if they receive less credit.   High ability men will agree to coauthor with
them if it reduces the cost of the paper without reducing the quality.  Employers would
17

then know that any woman coauthoring with a man is lower ability.
In what follows, I test whether women anticipate receiving less credit, whether high
ability women sort out of coauthoring with men, and whether men coauthor with women
whose careers begin more slowly.   To do so,  I first present survey evidence suggesting
that women do not know that the returns to coauthoring are lower than solo-authoring.
I then show that women do receive some credit for papers that publish well, suggesting
that employers might believe that there is some assortative matching.  I also provide evi-
dence that even when women tend to work with men who are slightly higher ability than
themselves this unequal match does not explain the gender gap in tenure.
Survey Evidence on Knowledge of Returns to CoauthoringIf women know that their
returns to coauthoring with men are low, it is plausible that high ability women would
choose to solo-author or only work with other women. Here I test whether women antici-
pate receiving less credit for collaborative work using a survey conducted with economists
currently working at the top 35 U.S. economics departments.  The survey was sent to all
professors,  regardless of rank,  at these institutions and received an 32% response rate.
The gender composition of the sample is representative of the profession today, with 89
respondents being female and 300 being male.  In the survey, economists were asked the
following question:
Suppose a solo-authored AER increases your chance of receiving tenure by 15%.  For each of
the following, please give an estimate of how much you think the described paper would increase
your chance of receiving tenure.
Respondents then go through five types of papers (coauthored AER, coauthored AER
with senior faculty, coauthored AER with junior faculty, solo-authored top field, and coau-
thored top field) and record their beliefs about the returns to these papers
14
.
In  Table  6,  I  test  the  difference  in  the  mean  beliefs  of  men  and  women.
15
There  is
no  statistically  significant  difference  in  the  beliefs  of  men  and  women  for  any  type  of
paper. Men believe that a coauthored AER will increase their chance of receiving tenure by
12.1%, and women by 12.2%. Women believe that there are slightly lower returns to AER
papers coauthored with senior faculty (8.8% versus 9.1% for men), but the difference is not
statistically significant. These results suggest that, in this context, women are unaware of
14
I did not ask respondents about paper coauthored with men/women so that they would not be primed
to think about gender
15
Because the survey was anonymous, the answers can not be linked to the CV data. I can therefore only
test for differences in means without controls.
18

the true returns to coauthoring.
C.V. Evidence on Sorting by AbilityA second test of whether women know that they
will receive less credit for papers and sort accordingly is to look at the correlation between
propensity to coauthor and ability.  I first test whether high ability women are less likely
to coauthor than low ability women and then test for assortative matching among coau-
thors. I proxy for ability using the quality of journal that an individual’s job market paper
was published in.  I assume that the job market paper is the first solo-authored paper an
individual publishes after he or she graduates.
If women anticipate discrimination,  ability and the fraction of one’s papers that are
coauthored  will  be  negatively  correlated.   High  ability  women  should  be  less  likely  to
coauthor. In Figure 5.A I plot the coefficients
ˆ
β
1
and
ˆ
β
2
from estimating
FracCA
if st
=β
1
a
i
+β
2
(fem
i
×a
i
) +β
3
fem
i
+β
4
TotPapers
i
+θ
f
+θ
s
+θ
t
+
if st
(7)
whereFracCA
if st
is the fraction of personi’s papers that are coauthored anda
i
is person
i’s ability (job market paper rank).  If high ability women anticipate receiving less credit,
we expect
ˆ
β
2
<0.  In Figure 5.A, however, we see that ability is uncorrelated with the
fraction of papers that are coauthored for both men and women: both estimates are precise
zeros.  There is no evidence that women along the ability distribution act strategically in
their choice to coauthor versus solo author.
I also find no evidence that high ability women strategically coauthor with other women
rather than men. Figure 5.B plots the results from equation 7 using the fraction of papers
that are coauthored with women as the dependent variable.  Women are more likely to
coauthor with other women than men are but there is no sorting by ability.
While women do not seem to be sorting according to ability, it is possible that women
tend  to  work  with  higher-ability  or  more  prominent  coauthors  who  then  receive  more
credit for a paper.  I test for this by correlating a person’s ability with that of his or her
coauthors.  While I do not have the job market paper information for all coauthors in the
dataset, I can see where the coauthors were working at the time the individual went up for
tenure. As a measure of average coauthor ability, I take the average school rank of all of an
individual’s pre-tenure coauthors. For example, ificoauthors withjandkandjworks at
the 5th-ranked institution andkworks at the 15th-ranked institution, the average ability
ofi’s coauthors is 10.
I correlatei’s ability with the average ability of her coauthors in Figure 6.  The line of
best fit is plotted controlling for number of coauthored and solo-authored publications,
19

time until tenure, and field, institution, and tenure year fixed effects.
Men and women both sort positively on ability but women are more likely to collab-
orate with individuals at more highly-ranked institutions than men are.  To see whether
this explains the main results, I estimate
T
if st
=β
1
S
i
+β
2
(fem
i
×S
i
) +β
3
CA
i
+β
4
(fem
i
×CA
i
) +β
5
rank
iJ
+β
6
(CA
i
×rank
iJ
) +β
7
(fem
i
×CA
i
×rank
iJ
) +β
8
(fem
i
×rank
iJ
)
+β
9
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(8)
whererank
iJ
is the average institution rank ofi’s coauthors and all other variables are
defined as before. The results are reported in Table 7. If men receive more credit because
they are coauthoring with lower ability women,
ˆ
β
7
should be negative.  However,
ˆ
β
7
is
close to zero, indicating that the ability or prominence of one’s coauthor is not driving the
tenure gap for coauthoring women.
Returns to Top PapersFor high ability women to receive no credit for their coauthored
papers,  employers would have to believe that there is no assortative matching by abil-
ity.  Otherwise, employers would receive a signal that women who coauthor with high
ability men are also high ability,  and be more likely to promote them.   Figure 6 shows
that assortative matching does occur, but it is possible that employers do not recognize
this.  I test for this by looking at how credit for top 5 publications is allocated.  If employ-
ers know that there is assortative matching, they should believe that women coauthoring
with high-ability men are also likely to be high ability.
Table 8 shows the results from estimating
T
ifst
=β
1
TopS
i
+β
2
(fem
i
×TopS
i
) +β
3
TopCA
i
+β
4
(fem
i
×TopCA
i
) +β
5
NonTopS
i
+β
6
NonTopCA
i
+β
7
(fem
i
×NonTopS
i
) +β
8
(fem
i
×NonTopCA
i
) +β
9
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
ifst
(9)
whereTopS
i
andTopCA
i
are the number of solo and coauthored papers that individuali
has published in a top 5 journal. Similarly,NonTopS
i
andNonTopCA
i
are the number of
solo and coauthored papers the individual has published in non-top 5 journals.  In Table
8, the “nop-top 5” interaction terms are presented in the second column.
Power becomes an issue as (1) there are relatively few people publishing in the top
5  journals,  and  (2)  cutting  by  gender  means  that  there  are  even  fewer  women  in  each
category.
Table 8 shows that coauthored papers published in a top 5 journal help women much
20

more than those published in non-top 5 journals.  Non-top 5 coauthored papers do not
have any positive influence on women’s tenure probability.  It seems that employers re-
ceive some signal when a woman publishes her coauthored papers in top journals which
is at odds with the hypothesis that only low ability women coauthor with men.
Overall,  there is little evidence that ability-based sorting is driving the results.
16
If
anything,  employers  seem  to  recognize  that  high  ability  men  and  women  might  work
together and are therefore more likely to grant these women tenure. However, their tenure
rate is still lower than that of high ability men.
4.2.2    Preference-Based Sorting
If women prefer to coauthor with senior faculty, we could reasonably expect that women
would have lower tenure rates.  Assuming senior faculty are more likely to be credited
for a paper, the fact that most senior faculty are men would drive the correlation between
coauthoring with a man and tenure. That is, women receive less credit because they enjoy
coauthoring with senior faculty and these senior faculty are predominantly male.
The basic summary statistics showed that women were not more likely to coauthor
with senior faculty than men.  However, I conduct an additional test as to whether coau-
thorship with senior faculty could be driving the results. I reestimate equation 3 but con-
trol for the fraction of a person’s coauthors who are senior.  The results are presented in
Column 3 of Table 7.  The seniority of women’s coauthors does not explain the results.
Controlling for seniority, an additional coauthored paper increases a man’s probability of
tenure by 8 percentage points but a woman’s by 5 percentage points.
4.2.3    Timing of Coauthorship
It is possible that men offer to work with women who are struggling to publish.  If this
is the case, we should see women who have few publications in the early years of their
appointment being more likely to coauthor with men. I test for this possibility by looking
at differences in early publications and by testing whether women with a longer time lag
between their initial appointment and first publication are more likely to coauthor with
men.
16
Garcia and Serman (2015) show that there could be selection into coauthorship driven by a desire to
be first author on a paper (that is, depending on where you are in the alphabet relative to your coauthors).
This would be an issue in this setting if, for example, men are more likely to be strategic than woman and
are therefore more likely to be first author on a paper (which is correlated with having more citations).  I
test whether men are more likely to be first author on their papers than women and whether men have a
“higher” author position overall. I find that men in my sample are first author 57% of the time while women
are first author 55% of the time (p-value = 0.907).
21

Figure  7  descriptively  shows  the  timing  of  publications  for  men  and  women,  split
by whether they received tenure at their initial tenure institution.  More formally, I test
whether women have fewer publications early in their careers by estimating
Y
if st
=β
1
Fem
i
+β
2
T
is
+β
3
(Fem
i
×T
is
) +β
4
Papers
i
+β
5
 ̄q
i
+θ
f
+θ
s
+θ
t
+
if st
(10)
whereY
if st
is the number of years between individuali’s initial appointment andi’s first
post-appointment publication.
17
I test whether women who did not receive tenure had
a longer publishing lag by interacting the female dummy term with an indicator for re-
ceiving tenure at schools,T
is
.  I control for the number of papers published pre-tenure
(Papers
i
) and the average quality of those papers (  ̄q
i
) All other variables are defined as
before.
The results are presented in Table 9.   Women who do not receive tenure do have a
longer lag (approximately 0.5 years) between their first appointment and their first publi-
cation although the result is noisily estimated. I test whether women with a longer lag are
more likely to coauthor with men by estimating
FracM
if st
=β
1
Fem
i
+β
2
T
is
+β
3
(Fem
i
×T
is
) +β
4
Y
i
+β
5
(Fem
i
×Y
i
)
+β
6
(Fem
i
×T
is
×Y
i
) +β
4
Papers
i
+β
5
 ̄q
i
+θ
f
+θ
s
+θ
t
+
if st
(11)
where the outcome variable,Y
i
in equation 10 is used as a regressor. If men bring women
with a slow start to publishing onto their projects, we would expect to see
ˆ
β
5
>0.
The  results,  presented  in  Column  2  of  Table  9,  do  not  support  the  hypothesis  that
women who struggle to publish initially are more likely to begin publishing with men.
The coefficient onβ
5
is negative, suggesting that women with a longer publishing lag are
less likely to coauthor with men although this result is again insignificant.
4.2.4    Women Not Claiming Credit for Papers
Women might be given less credit for their work if they are less likely to claim it as their
own.  For example, if women present less frequently than men, people might associate a
paper with the male coauthor who presents it more. The survey discussed in Section 4.2.1
also asked individuals how many times per year they present their work and whether
they are more or less likely to present their coauthored papers than their coauthor. Panel
B of Table 6 shows that men and women report the same likelihood of presenting their
joint papers relative to their coauthors.  Interestingly, though, women present their solo-
17
I exclude papers that were published before the person’s first appointment.
22

authored  papers  fewer  times  per  year  than  men  do.   It  is  possible  that  women  do  not
"advertise" their work as much as men do and this leads to women receiving less recogni-
tion for their work in general.  If this were true, though, women who solo author should
also be less likely to receive tenure.
4.2.5    Taste-Based Discrimination
If some employers have a distaste for tenuring women, as in Becker (1971), we should see
women who write solo-authored papers being denied tenure as well. If employers cannot
plausibly deny a woman who solo-authored several well-published papers, however, they
might be constrained to deny tenure only to those for whom they can make a reasonable
case.  If it can be argued that a woman who coauthors did little of the work, taste-based
discrimination could help to explain the results as employers have an excuse for denying
tenure to coauthoring women. However, as shown in Table 3, only women who coauthor
with men have lower tenure rates.  This would imply that employers have a particular
distaste for tenuring women who coauthor with men, which seems unlikely.
5    Experimental Evidence
In the previous section, I provided suggestive evidence that factors like sorting and taste-
based discrimination do not explain why women who coauthor with men are less likely
to receive tenure. I instead argue that the results are most consistent with women receiv-
ing less credit for joint work with men.  Specifically, because coauthored signals are an
unclear signal of ability, women receive less credit for their joint work with men if they
are believed to be lower ability.  I cannot rule out, though, that real of perceived differ-
ences in effort explain the results.  For example, tenure committees might hold the belief
that women contribute less or lower effort when they work with men, regardless of their
beliefs about a woman’s ability.  In addition, tenure committees might believe that low
ability women choose to work with high ability men even if the empirical evidence sug-
gests otherwise.
To  shed  light  on  whether  effort  or  perceptions  of  effort  and  sorting are  driving  the
results, I run an experiment on mTurk that is designed to shut down these channels. I also
test whether differences in the allocation of credit depend on whether the associated task is
male or female-stereotyped, which speaks to whether beliefs about ability are driving the
results.  In a female-stereotyped or gender-neutral task, we would not expect there to be
differences in beliefs about men and women’s abilities.  Although this setting is different
23

from academia, it provides additional evidence that gender plays a role in the allocation
of credit and that differences in credit are not driven by sorting or effort.
5.1    Experiment Design
The experiment consists of two incentivized parts and follows a two-by-two randomiza-
tion design.  In the first step, mTurk workers are recruited to complete two math quizzes
(male-stereotyped)  or  two  grammar  quizzes  (female-stereotyped),  each  containing  five
questions.
18
I refer to the two quizzes as Quiz 1 and Quiz 2.  These workers, hereafter
called “quiz-takers”, received $0.30 for participating in the study as well as $0.05 for each
question they answer correctly.
In the second step, designed to test whether people misallocate credit for joint work,
a different set of mTurk workers, referred to as “predictors”, are recruited to predict how
well  quiz-takers  will  do  on  Quiz  2.   Predictors  are  randomized  into  a  solo  treatment
or  a  group  treatment,  described  in  greater  detail  below.   In  each  treatment,  predictors
are shown the Quiz 1 questions and correct answers.   They are then either shown two
randomly-drawn  quiz-takers’  individual  scores  on  Quiz  1,  or  a  single  score  that  is  the
sum of the two quiz-takers’ Quiz 1 scores. Upon seeing these Quiz 1 scores, as well as the
questions that were asked on Quizzes 1 and 2, the predictors are asked to estimate how
well the quiz-takers will do on Quiz 2.  Predictors were paid a participation fee of $0.50
and receive $0.10 for each score they correctly predict
19
.
I also run a cross-treatment in which predictors in both the solo and group treatments
are shown the average score of all men and women on each quiz to understand whether
predictors’ guesses are driven by incorrect beliefs.
5.1.1    Treatments
Solo TreatmentThe solo treatment tests whether predictors correctly predict scores when
they see a clear signal of each quiz-taker’s ability.  This parallels the solo-author paper
analysis:  if predictors correctly assign credit when they see a clear signal of ability (i.e.
they assume that the quiz-taker was responsible for his or her quiz score), there should be
no difference in how men and women are evaluated conditional on Quiz 1 scores.
In the solo treatment, predictors are shown a randomly chosen man and woman’s score
on Quiz 1.  Both the man and the woman took the same quiz (either math or grammar).
The predictor is first shown the questions asked on Quiz 1.  Next, the predictor sees each
18
The math and grammar quizzes are shown in Appendix B.
19
Differences in payments reflect differences in the time it took to complete the tasks
24

quiz-taker’s score on Quiz 1, as well as the distribution of quiz scores across all partici-
pants. Note that the distribution of scores is not broken out by gender. Predictors are then
shown the Quiz 2 questions and are asked to predict each quiz-taker’s score on Quiz 2.
The full set of instructions that predictors receive are shown in Appendix D.
Group TreatmentThe group treatment is designed to understand how the predictors
assign credit for performance when they cannot observe individual contributions.  Quiz-
takers  are  randomly  paired  with  a  member  of  the  opposite  sex.   However,  there  is  no
interaction between the two:  each quiz-taker completes the same five-question quiz and
is paid based on his or her individual score. The other person’s score does not affect them
in any way.
In this treatment, though, the predictor is shown the sum of the two scores rather than
the individual scores.  For example, if Person A scored 3/5 and Person B scored 4/5, the
predictor would see the score 7/10 for that pair.  Importantly, the predictor is told that
each quiz-taker was paid based on the number of questions that she answered correctly
on herownquiz.   Thus,  the predictors know that quiz-takers are randomly paired,  but
worked independently and were individually incentivized.   This removes any worry a
predictor might have about selection into pairs (such as high-ability men working with
low-ability women) and free-riding.  The predictor’s guess should therefore only reflect
his or her beliefs about each quiz-taker’s score and ability. To draw a parallel between this
treatment and the main analysis, the individual scores that make up the joint score can be
thought of as each person’s “contribution” to the group, but in this case cannot be driven
by selection or effort.
Cross-Randomization: Gender DistributionTo understand whether predictors’ guesses
are driven by (possibly incorrect) beliefs about ability or taste-based animus, I run a cross-
randomization in which I provide participants with information about the average per-
formance of men and women on the quizzes.  All predictors are shown the overall score
distribution, but in this treatment, they are additionally shown the average male and fe-
male  scores,  shown  in  Figure  B1.   All  other  predictors  are  shown  the  same  figure  but
without the lines indicating the mean performance of men and women.
If predictors exhibit taste-based animus, providing them with information about men
and  women’s  average  performance  will  not  change  their  prediction.   In  addition,  this
treatment also helps to understand whether differences in attribution are driven by incor-
rect beliefs about gender differences in performance. If participants hold incorrect beliefs
about men and women’s average performances, the gender distribution treatment should
25

correct those beliefs and the predictors should adjust their guesses accordingly.
To summarize, the specific steps of the experiment are as follows:
1.  Quiz-takers are randomly assigned to take two math quizzes or two grammar quizzes
2.  Predictors are shown the questions and correct answers from Quiz 1
3.  Predictors are shown the quiz scores of two randomly drawn quiz-takers. If predic-
tors are in the solo treatment, they see each quiz-taker’s score.  If predictors are in
the group treatment, they see the sum of the two scores.
4.  Predictors are shown the distribution of quiz scores. If predictors are in the “gender
distribution” cross treatment, they also see the average scores of all men and women
who took the quiz.
5.  Predictors are shown the Quiz 2 questions.
6.  Predictors guess what each quiz-taker’s score will be on Quiz 2.
5.2    Results
Quiz Results
Women had a lower average score than men on the math quizzes (2.51/5 vs. 2.72/5) and
a higher average score on the grammar quizzes (2.41 vs.  2.17).  The distribution of scores
on  Quiz  1  are  shown  in  Figure  B1.   This  is  the  same  figure  that  predictors  are  shown.
If predictors are in the gender distribution cross-treatment,  they also see the two lines
indicating the mean male and female scores.
Analysis
The main experimental results are presented in Table 10.  Columns (1) and (3) show how
predictors’ guesses vary based on the quiz-taker’s gender and by whether they were in
the gender distribution treatment. Specifically, I estimate
Q2
ij
=β
1
fem
i
+β
2
D
j
+β
3
(fem
i
×D
j
) +β
4
Q1
i
+
ij
(12)
separately for the sample of individuals who took math quizzes (Column 1) and grammar
quizzes (Column 3). The outcome variable,Q2
ij
, is predictorj’s estimate of quiz-takeri’s
Quiz 2 score.  An indicator for the quiz-taker being female,fem
i
, is interacted with an
26

indicator  for  the  predictor  being  in  the  gender  distribution  cross-treatment,D
j
.   I  also
control fori’s Quiz 1 score (Q1
i
).
Women are predicted to have lower average math scores and higher average gram-
mar scores than men, although the gender difference in math scores is insignificant.  The
difference in grammar score predictions is driven by the gender distribution treatment,
suggesting that predictors did not hold the prior that women outperform men on gram-
mar quizzes.
In Columns (2) and (4), I interact all of the independent variables in equation 12 with
an indicator for predictorjbeing in the group treatment. Note that in the group treatment,
predictors do not see the individual Quiz 1 scores as they only see a man and a woman’s
joint score.  I control for this variable regardless and interact it with the group treatment
dummy, as it should only have a significant effect for predictors in the solo treatment.
The predicted score difference for math quizzes is entirely driven by the group treat-
ment. When predictors see only the sum of a man and a woman’s Quiz 1 score, they pre-
dict that women will score 0.35 points lower on the second math quiz than men. However,
there is no significant difference in men and women’s predicted performance in the solo
treatment (if anything, women are predicted to perform slightly better than men). This is
consistent with the finding that women suffer a coauthor penalty when their contribution
to a paper is unobserved but are not discriminated against when their contributions are
observed, as in solo-authored papers.  Predicting that the woman will do worse than the
man in the group treatment suggests that predictors believe that the woman’s first score
was lower; that is, she is worse at the task and therefore contributed less to the joint score.
Puzzlingly, showing predictors the mean scores of male and female quiz-takers does not
change the predictions for the second math quiz. Predictors guess that women will score
0.35 points below men regardless of whether they saw the mean scores.
The grammar results in Column (4),  though,  suggest that the results are not driven
by taste-based animus in which women are always penalized in collaborative situations.
Women in the group treatment are not predicted to perform worse than men on the second
quiz, indicated by a positive, though insignificant, coefficient on Female×Group Treatment.
Predictors in the group treatment who see the actual score distribution also predict that
women will outperform men on the second quiz.
While the experimental context is different from the academic context, the results pro-
vide evidence that individuals make different inferences about men and women’s contri-
butions to a joint project which are rooted in stereotypes and bias.  Even after shutting
down effort and selection channels, and providing the distribution of scores, participants
still used gender to assign credit.  High scores on joint math quizzes were assumed to be
27

driven by men’s individual scores rather than women’s while high scores on join gram-
mar quizzes were assumed to be driven by women’s individual scores.  While there are
gender differences in scores, the actual differences are smaller than the predicted differ-
ences even when predictors saw the actual mean differences. In the individual treatment,
as in the case of solo-authored papers, men and women are evaluated similarly.
6    Conclusion
Women receive tenure at significantly lower rates than men in many academic fields.  As
discussed  in  the  introduction,  this  phenomenon  is  not  exclusive  to  academia.   Several
explanations have been put forward for the gap, but it persists even after accounting for
observable characteristics such as fertility preferences and productivity.
This paper proposes an alternative explanation. I argue that women receive less credit
for group work when employers can not perfectly observe their contribution.  When sig-
nals are noisy, employers have to infer each worker’s ability or productivity. Coauthored
papers provide employers with a noisy signal. The fact that women who work specifically
with men receive tenure at lower rates than comparable women who work alone or with
other women suggests that gender enters into the employer’s inference process. However,
when employers receive clear signals, men and women are treated similarly.  For exam-
ple, men and women receive the same amount of credit for solo-authored papers, which
provide a clear signal of ability.  Evidence from an online experiment suggest that these
results are not explained by sorting or differences in effort. The online experiment further
suggests that this phenomenon is not specific to women. Men also suffer a penalty when
working with women on a female-stereotyped task.
Being aware of this phenomenon is important in a world that is increasingly relying
on group work for production. The tech industry, for example, prides itself on collabora-
tion.  In such male-dominated fields, however, group work could result in fewer women
moving up the career ladder if credit is not properly attributed.  The same could be true
for men in female-dominated industries.  The unequal attribution of credit would then
contribute to and help maintain gender segregation in occupations.
28

References
[1]  Antecol, Heather, Kelly Bedard, and Jenna Stearns. 2016. "Equal but Inequitable: Who
Benefits from Gender-Neutral Tenure Clock Stopping Policies?" IZA Discussion Paper
No. 9904.
[2]  Becker,  Gary. 1971.The Economics of Discrimination.Chicago:  University of Chicago
Press.
[3]  Blau, Francine D. and Jed DeVaro. 2007. "New Evidence on Gender Differences in Pro-
motion Rates: An Empirical Analysis of a Sample of New Hires." Industrial Relations,
46(3): 511-550.
[4]  Boschini, Anne and Ana Sjögren. 2007. "Is Team Formation Gender Neutral? Evidence
from Coauthorship Patterns." Journal of Labor Economics, 25(2): 325-365.
[5]  Carrell, Scott E., Marianne E. Page, and James E. West. 2010. "Sex and Science:  How
Professor  Gender  Perpetuates  the  Gender  Gap."  Quarterly  Journal  of  Economics,
125(3): 1101-1144.
[6]  Ceci,  Stephen J.,  Donna K. Ginther,  Shulamit Kahn,  and Wendy M. Williams. 2014.
"Women in Academic Science:  A Changing Landscape." Psychological Science in the
Public Interest, 15(3): 1-67.
[7]  Dynan, Karen and Cecilia Rouse. 1997. "The Underrepresentation of Women in Eco-
nomics:  A Study of Undergraduate Economics Students." The Journal of Economic
Education, 28(4): 350-368.
[8]  Elmins, William, Robert Joyce, and Monica Costa Bias. 2016. "The Gender Wage Gap."
Institute for Fiscal Studies Briefing Note 186.
[9]  Garcia,  Daniel  and  Joshua  Sherman.  2015.  "Norms  and  Team  Formation:  Evidence
from Research Partnerships." working paper.
[10]  Ginther, Donna K. and Shulamit Kahn. 2004. "Women in Economics:  Moving Up or
Falling Off the Academic Career Ladder?" Journal of Economic Perspectives,  18(3):
193-214.
[11]  Hale,  Galina  and  Tali  Regev.  2014.  "Gender  Ratios  at  Top  PhD  Programs  in  Eco-
nomics." Economics of Education Review, 41: 55-70.
29

[12]  Hammermesh, Daniel. 2013. "Six Decades of Top Economics Publishing:  Who and
How?" Journal of Economic Literature, 51(1): 162-172.
[13]  Kalaitzidakis, Pantelis, Theofanis Mamuneas, and Thanasis Stengos. 2003. "Rankings
of Academic Journals and Institutions in Economics." Journal of the European Eco-
nomic Assocation, 1(6): 1346-1366.
[14]  Lazear,   Edward   P.   and   Kathryn   L.   Shaw.   2007.   "Personnel   Economics:The
Economist’s View of Human Resources." Journal of Economic Perspectives, 21(4): 91-
114.
[15]  MacLeod, W. Bentley. 2003. "Optimal Contracting with Subjective Evaluation." Amer-
ican Economic Review, 93(1): 216-240.
[16]  Milkman, K.L., M. Akinola, and D. Chugh. 2015. "What Happens Before?  A Field
Experiment Exploring How Pay and Representation Differentially Shape Bias on the
Pathway into Organizations." Journal of Applied Psychology, 100(6): 1678-1712.
[17]  Moss-Racusin, Corinne, John Dovidio, Victoria Brescoll, Mark Graham, and Jo Han-
delsman. 2012. "Science Faculty’s Subtle Gender Biases Favor Male Students." PNAS,
109(41): 16474-16479.
[18]  Niederle, Muriel and Lise Vesterlund. 2007. "Do Women Shy Away From Competi-
tion?   Do Men Compete Too Much?" Quarterly Journal of Economics,  122(3):  1067-
1101.
[19]  Prendergast,  Canice.  1999.  "The  Provision  of  Incentives  in  Firms."  Journal  of  Eco-
nomic Literature, 104(5): 7-63.
[20]  Sarsons,   Heather.  2017.  "Recognition  for  Group  Work:Gender  Differences  in
Academia." AER Papers & Proceedings, 107(5): 141-145.
30

Figures
FIGURE1: DISTRIBUTION  OFPAPERCOMBINATIONS
Notes: This figure shows the number of women (Panel A) and men (Panel B) who had various combinations of solo and coauthored
papers at the time of tenure. Each dot represents a specific combination of papers with the number of coauthored papers measured on
the x-axis and the number of solo-authored papers measured on the y-axis. The shading of the dots represents how many individuals
had that combination of papers at the time they went up for tenure, with darker shades indicating a larger number of individuals with
that combination. In the legend, “n” is the minimum and maximum number of individuals who have a specific paper combination.
Panel A is constructed using the full sample of women (N=143) and Panel B is constructed using the full sample of men (N=501).
31

FIGURE2: TENUREPROBABILITIES BYPAPERCOMBINATIONS
Notes: This figure plots the unconditional tenure probability for women (Panel A) and men (Panel B) who have various combinations
of papers at the time they go up for tenure. Coauthored papers are counted along the x-axis and solo-authored papers are countred
along the y-axis. A darker shade indicates a higher probability of receiving tenure. For example, if a dot is the darkest shade, it
indicates that individuals with that combination of solo and coauthored papers receives tenure with probability one. Panel A is
constructed using the full sample of women (N=143) and Panel B is constructed using the full sample of men (N=501).
32

FIGURE3: TOTALPAPERS ANDTENURE
Notes: This binned scatterplot shows the correlation between the total number of publications an individual has at the time they go up
for tenure and the probability of receiving tenure. The y-variable, tenure, is a binary variable that equals one if an individual received
tenure at their initial institution of employment. For more details on how the tenure variable is constructed, see Section 2. To
construct the plot, tenure is first residualized with respect to the following controls: number of years it took to go up for tenure,
average journal rank of pre-tenure publications, log citations, total coauthors, and tenure school, tenure year, and field fixed effects.
The x-variable, number of publications, is then divided into twenty equal-sized groups. Within each of these groups, I plot the mean
of the y-variable (tenure) residuals against the mean of the x-variable (also within each bin). I then add back the unconditional mean
of Tenure to help with the interpretation of the line of best fit. The lines of best fit are estimated using the full sample (N=621) and
have slopes ofβ=0.132(s.e. = 0.016) for men andβ=0.165(s.e. = 0.043) for women.
33

FIGURE4: RELATIONSHIPBETWEENPAPERCOMPOSITION ANDTENURE
Notes: This figure is a binned scatterplot of the correlation between tenure and the fraction of an individual’s papers that are
solo-authored, split by gender. The y-variable is a binary variable indicating whether an individual received tenure. To construct the
plot, tenure is first residualized with respect to the following controls: total number of papers an individual published by the time of
tenure, number of years it took to go up for tenure, average journal rank of pre-tenure publications, log citations, total coauthors, and
tenure school, tenure year, and field fixed effects. The x-variable, fraction of papers that are solo-authored, is then divided into twenty
equal-sized groups. Within each of these groups, I plot the mean of the y-variable (tenure) residuals against the mean of the x-variable
(also within each bin). I then add back the unconditional mean of Tenure to help with the interpretation of the line of best fit. The line
of best fit using OLS is shown separately for men and women. The lines of best fit are estimated using the full sample (N=621) and
have slopes ofβ=0.521(s.e. = 0.158) for women andβ=0.023(s.e. = 0.748) for men.
34

FIGURE5: ABILITY  ANDSORTING
Notes: This binned scatterplot shows the correlation between an individual’s ability and the propensity to coauthor (Panel A) or the
propensity to coauthor with women (Panel B). The outcome variable in Panel A is the fraction of an individual’s papers that were
published by tenure that are coauthored. The outcome variable in Panel B is the fraction of an individual’s pre-tenure papers that are
coauthored with only women. I proxy for an individual’s ability with the rank of the journal in which the individual’s job market
paper was published. The plot is constructed as described in Figure 3 with the y-variable residualized on the following controls
before plotting: total solo and coauthored papers, the number of years it took to go up for tenure, log citations, and tenure school,
tenure year, and field fixed effects. The lines of best fit using OLS are shown separately for men and women. The estimates for Fig. 5A
areβ=−0.0001(s.e. = 0.0003) for women andβ=0.0002(s.e. = 0.0002) for men. The estimates for Fig. 5B areβ=−0.00004(s.e. =
0.0008) for women andβ=0.0002(s.e. = 0.0003) for men.
35

FIGURE6: ASSORTATIVEMATCHING
Notes: This binned scatterplot shows the correlation between an individual’s ability, proxied by the journal in which their job market
paper is published, and their coauthor’s ability, proxied by the average school rank of their coauthors. The school rank of coauthors
are measured at the time that individualiwent up for tenure. School rankings are taken from IDEAS/RePEc. The plot is constructed
as described in Figure 3 with the y-variable residualized on the following controls before plotting: total solo and coauthored papers,
the number of years it took to go up for tenure, log citations, and tenure school, tenure year, and field fixed effects. The line of best fit
using OLS is shown separately for men and women. The lines of best fit are estimated on the full sample and have slopes of
β=0.062(s.e. = 0.091) for women andβ=0.109(s.e. = 0.056) for men.
36

FIGURE7: TIMING  OFPUBLICATIONS
Notes: This figure shows the average number of publications an individual has in the years surrounding his or her initial appointment
as an assistant professor. Year 0 is the year that the individual begins working at his/her tenure institution (tenure institutions are
defined in Section 2). The blue bars represent publications that are coauthored with men. The red bars represent all other publications
(either solo-authored or coauthored with women). Panels A and B show the timing of publications for women and men who were
denied tenure. Panels C and D show the timing of publications for women and men who received tenure.
37

Tables
TABLE1: SUMMARYSTATISTICS
FullMaleFemalep-value
Panel A:
Tenure0.680.730.520.001
(0.47)(0.44)(0.50)
Years to tenure6.86.67.30.001
(1.6)(1.6)(1.8)
Total papers8.38.48.00.262
(3.9)(4.1)(3.3)
Solo-authored3.03.03.00.879
(2.4)(2.4)(2.3)
Coauthored5.35.45.00.189
(3.6)(3.7)(3.1)
Panel B:
Top 5 Solo0.670.660.680.900
(0.98)(0.99)(0.92)
Top 5 Coauthored1.31.31.20.570
(1.4)(1.4)(1.4)
AER Equivalent:
Solo Pubs.0.340.340.330.500
(0.24)(0.23)(0.25)
Coauthored Pubs.0.330.340.300.039
(0.20)(0.21)(0.18)
Panel C
Number Unique CAs4.524.554.470.767
(2.79)(2.78)(2.83)
Frac. coauthors who are:
Full Professor0.460.470.410.052
(0.35)(0.33)(0.38)
Associate Professor0.160.150.160.810
(0.24)(0.23)(0.28)
Assistant Professor0.250.230.280.060
(0.24)(0.22)(0.30)
Graduate Student0.0170.0150.0210.239
(0.067)(0.056)(0.095)
Female0.130.0940.2700.001
(0.23)(0.179)(0.309)
Observations644501143
This table displays the average tenure rate, pre-tenure productivity, and pre-tenure authorship patterns
of men and women who went up for tenure at one of top 35 U.S. economics departments between 1985
and 2014.  The top 35 institutions are taken determined according to the RePEc/IDEAS economics de-
partment rankings.  In Panel A, Tenure is an indicator that equals one if an individual was promoted to
associate or full professor 6-8 years after his or her initial appointment.  Years to tenure is the number
of years between an individual’s PhD graduation year and the year s/he went up for tenure.  All paper
counts are measured as the number of papers an individual had published at the time of tenure.  Top 5
Solo/Coauthored are the number of publications an individual had published in one of the top 5 eco-
nomics journals:  AER, QJE, Econometrica, JPE, and ReStud.AER Equivalentis a measure that converts
an individual’s publications into the number of AER-equivalent publications they correspond to.  For
more details on this variable, see Section 2.  Number Unique CAs is the number of different coauthors
an individual had published with by the time s/he went up for tenure.  Coauthor positions (full, asso-
ciate, assistant, and graduate student) are the positions an individual’s coauthors had at the time that
individual went up for tenure.
38

TABLE2: RELATIONSHIPBETWEENPAPERS& TENURE
Outcome Variable: Tenure
(1)(2)(3)(4)(5)
Sample:FullFullFullFemaleMale
Total papers0.142
∗∗∗
(0.016)
Fem x Papers-0.005
(0.012)
Solo-authored0.094
∗∗∗
0.097
∗∗∗
0.196
∗∗∗
0.095
∗∗∗
(0.013)(0.019)(0.055)(0.024)
Fem x Solo0.048
∗∗∗
0.057
∗∗∗
(0.018)(0.015)
Coauthored0.085
∗∗∗
0.082
∗∗∗
-0.0310.090
∗∗∗
(0.016)(0.014)(0.054)(0.016)
Fem x Coauthored-0.030
∗
-0.026
∗
(0.016)(0.015)
Total coauthors-0.0050.0010.0030.025-0.001
(0.004)(0.005)(0.005)(0.016)(0.006)
Total Papers Sq-0.004
∗∗∗
(0.001)
Solo Papers Sq-0.005
∗∗∗
-0.005
∗∗∗
-0.007-0.005
∗∗
(0.001)(0.002)(0.005)(0.002)
Coauthored Sq-0.003
∗∗∗
-0.003
∗∗∗
0.000-0.003
∗∗∗
(0.001)(0.001)(0.003)(0.001)
Log Citations0.059
∗∗∗
0.031
∗∗
0.065
∗∗∗
0.098
∗
0.058
∗∗∗
(0.012)(0.012)(0.013)(0.054)(0.016)
AER Equiv. Ranking0.533
∗∗∗
(0.116)
AER Equiv. Solo0.139
∗
0.331
∗∗∗
0.3210.416
∗∗∗
(0.071)(0.069)(0.206)(0.091)
AER Equiv. CA0.201
∗∗
0.325
∗∗∗
0.486
∗
0.280
∗∗∗
(0.099)(0.073)(0.248)(0.089)
Female-0.135-0.166-0.205
∗
(0.105)(0.121)(0.109)
Tenure Inst. FEYNYYY
Tenure Year FEYNYYY
Field FEYNYYY
Observations625629621139482
R-squared0.4170.2870.4250.5210.421
This table shows the relationship between publications and tenure.  The dependent variable, Tenure, is binary and
indicates whether an individual received tenure 6-8 years after being hired at the initial tenure institution. Total papers
is the number of papers an individual published by the time s/he went up for tenure. Solo-authored and Coauthored
are the number of solo or coauthored papers s/he had published at the time of tenure.  AER Equiv.  Ranking, AER
Equiv. Solo, and AER Equiv. CA are journal quality measures described in Section 2. Total coauthors is the number of
coauthors an individual had on the papers s/he had published by the time of tenure. Tenure length is the number of
years it took the individual to go up for tenure. Citations are from Google Scholar and measured in 2017. The equations
are estimated using a linear probability model.  Bootstrapped standard errors are clustered by tenure institution and
reported in parentheses. (*=p<0.10, **=p<0.05 ,***=p<0.01)
39

TABLE3: COAUTHORGENDER
(1)
×Female
Solo-authored0.093
∗∗∗
0.049
∗∗
(0.019)(0.015)
CA with only fem CAs0.097
∗∗∗
0.019
(0.024)(0.020)
CA with only male CAs0.087
∗∗∗
-0.056
∗∗∗
(0.015)(0.015)
Pubs. with M and F CAs0.087
∗∗
0.033
(0.026)(0.042)
Female-0.156
(0.101)
Total coauthors-0.001
(0.004)
Log Citations0.064
∗∗∗
(0.014)
AER Equiv. CA0.332
∗∗∗
(0.073)
AER Equiv. Solo0.328
∗∗∗
(0.065)
Tenure Inst. FEYes
Tenure Year FEYes
Field FEYes
Observations621
This table presents the results of one regression where the variables that are in-
teracted with Female (a dummy indicating that the researcher is a woman) are
displayed  in  the  right-hand  column.Papers  with  only  fem  CAsis  the  number
of publications an individual has in which all coauthors are female.   Similarly,
Papers with only male CAsandPapers with male and fem  CAsare the number of
publications with only male coauthors and with a mix of male and female coau-
thors respectively. Controls for tenure length; quadratics in the number of papers;
and tenure institution, year, and field fixed effects are also included.  The equa-
tions is estimated using a linear probability model. Bootstrapped standard errors
are reported in parentheses and are clustered by tenure institution.   (*=p<0.10,
**=p<0.05 ,***=p<0.01)
40

T
ABLE
4: R
OBUSTNESS
C
HECKS
Faculty
Journal Rankings
Publication Count
Total AERs
List Sample
RePEc
Over Time    AER Bins    Tenure +1    Tenure +2
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Solo-authored
0.115
∗∗∗
0.083
∗∗∗
0.078
∗∗∗
0.078
∗∗∗
0.058
∗∗∗
0.038
∗∗∗
0.060
∗∗∗
(0.024)
(0.018)
(0.017)
(0.019)
(0.017)
(0.013)
(0.019)
Fem x Solo
0.050
∗∗
0.055
∗∗∗
0.052
∗∗∗
0.045
∗∗∗
0.044
∗∗∗
0.055
∗∗∗
0.091
∗
(0.019)
(0.015)
(0.015)
(0.015)
(0.014)
(0.013)
(0.024)
Coauthored
0.092
∗∗∗
0.079
∗∗∗
0.081
∗∗∗
0.069
∗∗∗
0.038
∗∗∗
0.031
∗∗∗
0.089
∗∗∗
(0.023)
(0.015)
(0.016)
(0.020)
(0.011)
(0.008)
(0.011)
Fem x Coauthored
-0.032
∗
-0.026
∗
-0.025
∗
-0.029
∗
-0.022
∗
-0.013
-0.036
∗
(0.019)
(0.014)
(0.014)
(0.015)
(0.013)
(0.013)
(0.019)
Years to Tenure
-0.046
∗∗∗
-0.054
∗∗∗
-0.055
∗∗∗
-0.051
∗∗∗
-0.046
∗∗∗
-0.045
∗∗∗
-0.050
∗∗∗
(0.009)
(0.008)
(0.008)
(0.009)
(0.008)
(0.008)
(0.011)
Total Coauthors
-0.001
0.004
0.004
-0.004
0.008
∗
0.010
∗∗
0.003
(0.007)
(0.005)
(0.005)
(0.006)
(0.004)
(0.004)
(0.005)
Log Citations
0.056
∗∗∗
0.070
∗∗∗
0.074
∗∗∗
0.079
∗∗∗
0.072
∗∗∗
0.074
∗∗∗
0.094
∗∗∗
(0.017)
(0.013)
(0.013)
(0.013)
(0.014)
(0.013)
(0.016)
CA Paper Rank
0.310
∗∗∗
0.003
∗∗
0.003
∗∗∗
0.345
∗∗∗
0.345
∗∗∗
(0.084)
(0.001)
(0.001)
(0.081)
(0.081)
Solo Paper Rank
0.457
∗∗∗
0.002
∗
0.004
∗∗∗
0.291
∗∗∗
0.299
∗∗∗
(0.077)
(0.001)
(0.001)
(0.069)
(0.065)
Female
-0.158
-0.193
∗
-0.197
∗∗
-0.175
∗
-0.197
∗
-0.280
∗∗∗
-0.199
∗∗
(0.128)
(0.102)
(0.099)
(0.100)
(0.105)
(0.110)
(0.048)
Observations
369
621
621
621
621
621
621
The dependent variable in all columns is an indicator for receiving tenure.  Column 1 restricts the sample to those schools I received a historical faculty list from.
Column 2 uses RePEc journal rankings as the paper quality measure. The ranking used can be found at https://ideas.repec.org/top/top.journals.all.html. Column3 uses historical journal rankings from RePEc to allow for rankings to change over time and to account for new journals entering. Column 4 controls for the numberof papers within each of 10 AER “bins”.  For this analysis, the AER-equivalent measure is divided into deciles.  For each individual, I then add up the number ofsolo and coauthored papers within each decile and include the number of papers in each bin as controls. In Columns 5 and 6, I include papers that were publishedone and two years after an individual went up for tenure in the paper counts.  In Column 7, I use the AER Equivalent measure of journal ranking to calculate thetotal number of AER equivalents (solo and coauthored) an individual had at the time of tenure. I use this measure in place of the solo and coauthored paper counts(the main independent variables).  All regressions control for a quadratic in the number of papers as well as tenure institution, tenure year, and field fixed effects.Bootstrapped standard errors are reported in parentheses and are clustered by tenure institution. (*=p<0.10, **=p<0.05 ,***=p<0.01)
41

TABLE5: FUTUREPRODUCTIVITY
Outcome Var:Post TenureLog Citations
Solo AER Equivalents
(1)(2)
PoissonOLS
Fraction Coauthored-1.45
∗∗∗
0.533
(0.500)(0.390)
Female-0.232-0.151
(0.380)(0.414)
Female×Frac. Coauthored1.057
∗
0.742
(0.576)(0.660)
Tenured0.1940.496
∗
(0.352)(0.289)
Tenured×Frac. Coauthored0.0020.408
(0.006)(0.486)
Female×Tenured0.1850.210
(0.528)(0.509)
Fem×Tenured×Frac. Coauthored-0.991-0.740
(1.131)(0.769)
Top 5 Coauthored0.013
∗∗
(0.007)
Total papers0.071
∗∗∗
(0.015)
Tenure Inst. FENY
Post-Tenure Inst. FEYN
Tenure Year FEYY
Field FEYY
Observations621621
Column 1 shows the results from estimating equation 6 using a zero-inflated Poisson model, where the outcome vari-
able is the number of solo-authored AER equivalents an individual published after the tenure decision (measured as
of 2017).  “Top 5 Coauthored” is the number of coauthored AER equivalents the individual published after tenure.
Post-tenure institution is the institution the individual went to following the tenure decision.   For people who re-
ceived tenure, this the same as the tenure institution. Column 2 shows the results from estimating the same equation
using OLS where log citations is the outcome variable.  Citations are measured in 2015.  Robust standard errors are
reported in parentheses and are clustered at the tenure institution or post-tenure institution level. (*=p<0.10, **=p<0.05
,***=p<0.01)
42

TABLE6: SURVEYRESULTS
(1)(2)(3)
Men    Women    p-value
Panel A: Beliefs about Returns to Papers
Coauthored AER12.112.20.939
Coauthored AER, Sr. Faculty9.18.80.528
Coauthored AER, Jr. Faculty13.313.40.796
Solo Top Field8.08.20.669
Coauthored Top Field6.36.80.223
Panel B: Frequency of Presenting Papers
Times Presented3.12.20.07
Present More Freq. than CA0.370.440.20
Observations30089
This table presents the mean responses for men and women to the following survey questions:
Panel A: "Suppose a solo authored AER increases your chance of receiving tenure by 15 percent. By
how much do you think each of the following increases your change of receiving tenure?" Panel B:
"How many times per year do you typically present your solo-authored papers?  Are you more or
less likely than your coauthors to present a joint paper?"Present More Freq.  than CAis the fraction
of respondents who reported that they are more likely than their coauthors to present a joint paper.
The survey was conducted with a sample of academic economists currently working at a top 35 U.S.
economics department. Respondents were anonymous.
43

TABLE7: ACCOUNTING  FORSORTING
Dep. Variable: Tenure
(1)(2)(3)
Solo-authored0.086
∗∗∗
0.084
∗∗∗
0.090
∗∗∗
(0.017)(0.017)(0.017)
Fem x Solo0.064
∗∗∗
0.067
∗∗∗
0.062
∗∗∗
(0.017)(0.018)(0.017)
Coauthored0.087
∗∗∗
0.089
∗∗∗
0.082
∗∗∗
(0.014)(0.014)(0.015)
Fem x Coauthored-0.032
∗
-0.032
∗
-0.033
∗∗
(0.016)(0.016)(0.015)
Female-0.220-0.348
∗
-0.262
∗
(0.121)(0.133)(0.136)
Rank Difference0.001
(0.002)
Fem×Rank Difference-0.001
(0.002)
Avg. Coauthor Rank-0.002
(0.001)
Fem×Avg. Coauthor Rank0.003
(0.002)
Frac. Full Prof.-0.057
(0.071)
Fem×Frac. Full Prof.0.194
(0.067)
Observations595595595
The dependent variable in all columns is an indicator for receiving tenure.   Column (1)
shows the relationship between solo and coauthored papers and tenure when controlling
for the difference between individuali’s institution rank and the average institution rank of
his or her coauthors. Column (2) controls for the average institution rank of an individual’s
coauthors, and column (3) controls for the fraction of an individual’s coauthors who are full
professors. Only coauthors that an individual coauthored with up until tenure are included.
All regressions control for tenure length, journal rank (AER equivalent measure), and log
citations.   They  also  include  tenure  institution,  tenure  year,  and  field  fixed  effects.   The
sample size is smaller in this analysis because individuals with no coauthors are excluded.
(*=p<0.10, **=p<0.05 ,***=p<0.01)
44

TABLE8: PAPERSPLIT  BYTOP5
Dep Var: Tenure
(1)
Top 5Non-Top 5
Solo0.067
∗∗∗
0.033
∗∗∗
(0.019)(0.007)
Coauthored0.086
∗∗
0.031
∗∗∗
(0.016)(0.007)
Fem x Solo0.0200.055
∗∗
(0.037)(0.019)
Fem x Coauthored-0.007-0.035
∗∗
(0.031)(0.017)
Female-0.171
(0.108)
Total coauthors-0.002
(0.005)
Years to tenure-0.049
∗∗∗
(0.008)
Log Citations0.079
∗∗∗
(0.012)
Tenure Inst. FEY
Tenure Year FEY
Field FEY
Observations621
R-squared0.415
This  table  presents  the  results  from  estimating  equation  9.   The
results  in  the  able  are  from  this  single  regression,  but  solo  and
coauthored papers are split into those published in the top 5 jour-
nals  (Column  1)  and  journals  below  the  top  5  (Column  2).   Top
5  papers  are  those  published  in  the  American  Economic  Review,
Econometrica, the Journal of Political Economy, Quarterly Journal
of Economics, or the Review of Economic Studies.  The dependent
variable  is  an  indicator  for  receiving  tenure.   The  regression  in-
cludes tenure institution, tenure year, and field fixed effects. Robust
standard errors are clustered by tenure institution and reported in
parentheses. (*=p<0.10, **=p<0.05 ,***=p<0.01)
45

TABLE9: TIMING OFCOAUTHORSHIP WITHMEN
Years to First    Fraction of Papers
Publicationwith Men
(1)(2)
Female0.0540.062
(0.186)(0.080)
Tenure-0.0020.060
(0.130)(0.047)
Female×Tenure-0.151-0.258
∗∗
(0.237)(0.089)
Years to 1st Pub.0.016
(0.015)
Fem×Years to 1st Pub.-0.013
(0.032)
Tenure×Years to 1st Pub.-0.030
(0.018)
Fem×Tenure×Years to 1st Pub.0.017
(0.041)
Total papers-0.119
∗∗∗
0.009
∗
(0.015)(0.003)
AER Equiv.-0.3800.251
∗∗
(0.344)(0.081)
School FEYY
Tenure Year FEYY
Primary Field FEYY
Observations603594
This table tests whether there are gender differences in the timing of an individual’s first publication (Col-
umn 1) and whether women who take a longer time to publish their first paper are more likely to coauthor
with men (Column 2).  The outcome variable in Column 1 is the number of years it takes an individual to
publish his or her first paper after graduating, and is measured as the year of the individual’s first pub-
lication minus the year of the individual’s initial faculty appointment.  Articles published before the first
appointment (i.e.  during graduate school) are not counted.  The outcome variable in Column 2 is the frac-
tion of an individual’s papers published by tenure that are coauthored with men.  The independent vari-
able,Y earsto1stPubis  the  outcome  variable  in  Column  1.   Both  regressions  include  tenure  institution,
tenure year, and field fixed effects. Robust standard errors are reported in parentheses. (*=p<0.10, **=p<0.05
,***=p<0.01)
46

TABLE10: EXPERIMENT: PREDICTEDSCORE BYQUIZTYPE
Outcome: Predicted Quiz 2 Score
MathGrammar
(1)(2)(3)(4)
Female-0.0960.1110.055-0.021
(0.097)(0.096)(0.103)(0.121)
Gender Distribution0.0620.145-0.389
∗∗∗
-0.241
∗∗
(0.101)(0.096)(0.096)(0.112)
Female×Gender Distribution-0.085-0.1080.424
∗∗∗
0.194
(0.133)(0.144)(0.144)(0.176)
Female×Group Treatment-0.354
∗∗
0.124
(0.166)(0.179)
Female×Group Treatment×Gender Distr.0.0010.545
∗∗
(0.230)(0.255)
Group Treatment×Gender Distr.-0.069-0.215
(0.165)(0.173)
Group Treatment3.294
∗∗∗
2.957
∗∗∗
(0.418)(0.474)
Quiz 1 Score0.416
∗∗∗
0.735
∗∗∗
0.385
∗∗∗
0.725
∗∗∗
(0.068)(0.072)(0.072)(0.089)
Group Treatment×Quiz 1 Score-0.696
∗∗∗
-0.693
∗∗∗
(0.116)(0.131)
Constant1.713
∗∗∗
0.1371.784
∗∗∗
0.289
(0.259)(0.261)(0.286)(0.327)
Observations516516493493
R-squared0.0810.2990.0920.296
This table presents the results from an mTurk experiment in which participants predict how well an individual will do on a math
or grammar quiz based on that individual’s performance on an earlier quiz.  Columns 1 and 2 show the results for the sample of
participants who predicted math quiz scores and Columns 3-4 show the results for the sample of participants who predicted grammar
quiz scores. In the experiment, participants were randomized into one of two treatments: (1) a group that saw each individual’s score
on a previous quiz, or (2) a group that saw the sum of two individuals’ scores. Group Treatment is a dummy variable indicating that
participants saw the sum of two scores (Treatment 2) rather than individual scores (Treatment 1).  Gender Distribution is a dummy
indicating that participants were told the average quiz scores of all men and women. (*=p<0.10, **=p<0.05 ,***=p<0.01)
47

Appendix A    Additional Tables
TABLEA1: RESULTS BYINSTITUTION  ANDYEAR
Panel A: Tenure Institution
Institution Rank:Top 10Top 20Top 35
(1)(2)(3)
Solo-authored0.031
∗∗∗
0.053
∗∗∗
0.039
∗∗
(0.006)(0.018)(0.018)
Coauthored0.035
∗∗
0.052
∗∗∗
0.023
∗∗∗
(0.013)(0.016)(0.007)
Fem x Coauthored0.002-0.048
∗∗
-0.048
∗
(0.027)(0.020)(0.026)
Fem x Solo0.074
∗
0.071
∗
0.104
∗∗∗
(0.035)(0.037)(0.035)
Female-0.471
∗
-0.048-0.245
(0.247)(0.173)(0.243)
Observations211157155
Panel B: Tenure Year
Tenure Year:1985-19951996-20052006-2014
(1)(2)(3)
Solo-authored0.034
∗∗∗
0.043
∗∗
0.033
∗
(0.010)(0.021)(0.019)
Coauthored0.018
∗
0.049
∗∗∗
0.047
∗∗∗
(0.010)(0.011)(0.015)
Fem x Coauthored0.011-0.047
∗∗
-0.053
∗
(0.041)(0.022)(0.027)
Fem x Solo0.145
∗∗∗
0.079
∗∗∗
0.054
(0.037)(0.029)(0.042)
Female-0.787
∗∗∗
-0.219-0.003
(0.275)(0.160)(0.202)
Observations141157215
Panel A shows the relationship between coauthoring and tenure by tenure institution rank.
Schools are divided into the top 10, top 20, and top 35 departments, according to the RePEc
rankings.  All regressions include the following controls:  time until tenure, number of coau-
thors,  log  citations,  solo  and  coauthored  journal  rankings,  and  tenure  year  and  field  fixed
effects.  Panel B shows the relationship splitting the sample by time period.  The year groups
are the years that an individual went up for tenure.   All regressions include the following
controls:  time until tenure, number of coauthors, log citations, solo and coauthored journal
rankings, and tenure rank and field fixed effects. (*=p<0.10, **=p<0.05 ,***=p<0.01)
48

Sociology Results
The sociology sample consists of randomly sampled faculty at the top 20 sociology PhD-
granting departments in the U.S
20
. There are 250 sociologists in the sample, 40% of whom
are female.  Summary statistics are presented in Table A2.  There is no statistically signifi-
cant difference between men and women’s tenure rates (with the mean tenure rate being
76%) although men seem to publish more solo-authored articles than women.
TABLEA2: SOCIOLOGYSUMMARYSTATISTICS
MenWomenp-value
Tenure0.7520.7760.547
(0.433)(0.419)
Total papers12.1510.180.033
(7.808)(5.726)
Total coauthored6.4095.9590.567
(6.641)(4.999)
Solo papers5.7454.2240.003
(4.451)(2.892)
Time to tenure7.5847.5200.686
(1.607)(1.724)
Books0.7790.5710.139
(1.185)(0.799)
Observations150100
This table presents summary statistics for the full sample of sociologists and sepa-
rately for men and women.  All paper and book count variables (Total Papers,Solo-
authored,Coauthored, andTop 5s) are the number of papers or books an individual
had published at the time of tenure.
To test whether men and women are treated differently,  I reestimate equation 3 but
include measures of the number of papers that researcheriis first author on.  The results
are presented in Table A3. I include the number and fraction of papers a researcher is first
author on in Columns 1 and 2 respectively, along with female dummy interaction terms.
20
Ranking from U.S. News Education
49

TABLEA3: SOCIOLOGY: PAPERS ANDTENURE
Dep Var: TenureProbitProbit
(1)(2)
Total first author0.050
∗∗
(0.017)
Fem x First Author0.026
(0.040)
Fraction first author0.403
∗∗∗
(0.043)
Fem x Frac. First Author-0.042
(0.172)
Solo papers0.0080.000
(0.006)(0.006)
Fem x Total Solo0.0020.007
(0.011)(0.011)
Total Coauthored-0.010
∗
0.009
(0.004)(0.007)
Fem x Total CA-0.0200.001
(0.017)(0.015)
Books0.063
∗
0.058
(0.032)(0.035)
Book chapters0.0070.005
(0.013)(0.012)
Female0.0260.010
(0.114)(0.163)
School FEYesYes
Tenure Year FEYesYes
Observations237209
This table shows the relationship between the number and types of papers
an individual publishes and tenure for a sample of sociologists.   The de-
pendent variable is a binary variable indicating whether the individual re-
ceived tenure 6-7 years after being hired at the initial tenure institution.To-
tal first authoris the number of papers an individual is first author on while
Fraction first authoris the fraction of an individual’s papers that s/he was
first author on.  The equations are estimated using a probit model and the
marginal probabilities calculated at the mean are displayed.  Standard er-
rors, reported in parentheses, are clustered by tenure institution. (*=p<0.10,
**=p<0.05 ,***=p<0.01)
50

Appendix B    Additional Figures
FIGUREB1: EXPERIMENT: QUIZ1 RESULTS
Notes: These bar graphs show the distribution of scores on first math and grammar quizzes. The lines mark the means score of men
(dashed line) and women (solid lines). The experiment participants who predicted scores saw these distributions with or without the
lines, depending on whether they were in the gender distribution treatment.
51

Appendix C    Institutions List
Received faculty list:Brown, Columbia, Cornell, Duke, Harvard, Michigan State Uni-
versity, New York University, Northwestern, Ohio State University, Penn State, Rutgers,
Stanford, UC Berkeley, UC Davis, UC San Diego, UCLA, University of Virginia, University
of Maryland, University of Michigan, University of Minnesota, University of Pennsylva-
nia, University of Wisconsin-Madison
No faculty list:Boston College,  Boston University,  California Institute of Technology,
Georgetown,  MIT,  Princeton,University  of  Southern  California,  University  of  Chicago,
University of Texas - Austin, University of Rochester, Vanderbilt, Yale
52

Appendix D    Experiment Information
D.1    Quizzes
Grammar Quiz 1
1.  The storm prevented ....... on a picnic.
(a)  us to going
(b)  us going
(c)  us to go
(d)  us from going
2.  A man’s concept of liberty is different from ........ .
(a)  a woman’s
(b)  womens
(c)  a woman
(d)  woman’s
3.  ........ hour went by before we received ........ invitation
(a)  an/an
(b)  a/a
(c)  an/a
(d)  a/an
4.  When a subordinate clause is followed by the main clause, what is required?
(a)  a dash
(b)  a semi-colon
(c)  a period
(d)  a comma
5.  ........ are used around a relative clause that defines the noun it follows.
(a)  Only commas
(b)  No commas
(c)  Semi-colons
(d)  Quotation marks
53

Math Quiz 1
1.  Which of the following is a subset of {b,c,d}?
(a)  { }
(b)  {a}
(c)  {1,2,3}
(d)  {a,b,c}
2.  A man’s regular pay is $3 per hour up to 40 hours.  Overtime is twice the payment
for regular time. If we was paid $168, how many hours overtime did he work?
(a)  8
(b)  16
(c)  28
(d)  48
3.  3 4/5 expressed as a decimal is
(a)  3.40
(b)  3.45
(c)  3.50
(d)  3.80
4.  Which of the following is the highest common factor of 18, 24, and 36?
(a)  6
(b)  18
(c)  36
(d)  72
5.  Given thataandbare integers, which of the following is not necessarily an integer?
(a)2a−5b
(b)a
7
(c)b
a
(d)ab
55

Math Quiz 2
1.  Items bought by a trader for80areoldfor100.  The project expressed as a percentage
of cost price is
(a)  2.5%
(b)  20%
(c)  25%
(d)  50%
2.  A man bought a shirt at a sale.  He saves30onthenormalpricewhenhepaid120 for the
shirt. What was the percentage discount on the shirt?
(a)  20
(b)  25
(c)  33.33
(d)  80
3.  How many subsets does {a,b,c,d,e} have?
(a)  2
(b)  4
(c)  10
(d)  32
4.  What is the median of the given data: 13, 16, 12, 14, 19, 14, 13, 14
(a)  14
(b)  19
(c)  12
(d)  14.5
5.  In coordinate geometry, what is the equation of the x-axis?
(a)y=0
(b)x=y
(c)x=0
(d)y=1
56

D.2    Experiment Instructions
Screen 1
INSTRUCTIONS: Please read all the way through
This project seeks to understand how well individuals can predict a person’s future
performance on a task based on his/her past performance.
We recruited a group of people to complete two math [grammar] quizzes.  Each quiz
had five questions.  Participants had one minute to complete each quiz.  In what follows,
we will show two participants’ scores from the first quiz. We then ask you to predict each
participant’s score on the second quiz. We will provide you with some basic information
on each individual.
You will be paid $0.50 for your participation but will also be paid a bonus of $0.10 if
you correctly guess a participant’s score on the second quiz.
Screen 2
We will first show you the distribution of scores on the first quiz. Each bar represents the
fraction of people who obtained that score. For example, 30% of people scored 4/5 on the
first quiz.
Gender Distribution TreatmentParticipants also see: The average score of female par-
ticipants (2.5/5) is shown by the solid line. The average score of male participants (2.8/5)
is shown by the dashed line.
Screen 3
Solo TreatmentBelow we are showing you two individual’s test scores on Quiz 1, as
well  as  some  basic  demographic  information  about  each  of  them.   Please  predict  each
individual’s score on Quiz 2. You can view each quiz by clicking on the link below.
Group TreatmentBelow we are showing you a team’s score on Quiz 1. Recall that each
team member worked on the questions independently.  We then take the sum of the two
scores and assign it to the team. For example, if Person A scored 3/5 and Person B scored
4/5, their team score would be 7/10.  We provide you with some basic demographic in-
formation about each team member.
Based on the team’s performance,  please predict each individual’s score on Quiz 2.
You can view each quiz by clicking on the link below.
57

Changed text

Open file

Gender Differences in Recognition for Group Work
Heather Sarsons, Klarita Gërxhani, Ernesto Reuben, and Arthur Schram
∗
September 15, 2019
Abstract
Does gender influence how credit for group work is allocated?  Using data from
academic economists’ CVs,  we test whether coauthored and solo-authored publica-
tions matter differently for tenure for men and women.  Because coauthors are listed
alphabetically in economics, coauthored papers do not provide specific information
about each contributor’s skills or ability.   Solo-authored papers,  on the other hand,
provide  a  relatively  clear  signal  of  ability.   We  find  that  conditional  on  publication
quality and other observables, men are tenured at roughly the same rate regardless of
whether they coauthor or solo-author. Women, however, become less likely to receive
tenure the more they coauthor. The result is most pronounced for women coauthoring
with men and less pronounced among women who coauthor with other women. Two
experiments suggest that both stereotypes surrounding a task as well as the evalua-
tors’ gender affect who receives credit.  Taken together, our results are best explained
by gender and stereotypes having a noticeable influence on the allocation of credit for
group work.
∗
Sarsons,  University of Chicago Booth (heather.sarsons@chicagobooth.edu);  Gërxhani,  European Uni-
versity Institute; Reuben, New York University Abu Dhabi and LISER; Schram, Amsterdam School of Eco-
nomics and European University Institute.  Sarsons especially thanks Roland Fryer, Claudia Goldin, Larry
Katz,  David Laibson,  and Amanda Pallais for their guidance and encouragement.   We also thank Mitra
Akhtari, Amitabh Chandra, John Coglianese, Oren Danieli, Ellora Derenoncourt, Florian Ederer, Ben Enke,
Raissa Fabregas, Nicole Fortin, Nickolas Gagnon, Peter Ganong, Edward Glaeser, Siri Isaksson, Emir Ka-
menica, Sara Lowes, Rob McMillan, Eduardo Montero, Gautam Rao, Alex Segura, Nihar Shah, Peter Tu,
Jeroen van de Ven, Justin Wolfers, and various conference and seminar participants for their helpful com-
ments and suggestions.
1

1    Introduction
Do employers use gender when allocating credit for group work, particularly when in-
dividual contributions are unobserved?  Organizations increasingly rely on group work
for production (Lazear and Shaw, 2007), yet there is little empirical evidence document-
ing how credit for group work is allocated.  Unless employers can perfectly observe each
worker’s contribution to the team’s output, they must decide how to allocate credit with-
out having full information as to what each worker did. This could leave room for demo-
graphic characteristics, such as gender, to influence the allocation of credit.
In this paper, we test whether uncertainty over an individual’s contribution to a project
leads to differential attribution of credit that contributes to the gender promotion gap. In
many industries,  women are not only hired at lower rates than men are,  they are also
promoted at lower rates.
1
This paper explores whether gender differences in credit for
group work exist and whether they explain part of the promotion gap.
We primarily look at the tenure decisions of academic economists to test whether gen-
der  influences  the  allocation  of  credit  for  coauthored  papers.   Economics  is  a  relevant
setting as there is a large tenure gap between men and women, and because the amount
of coauthoring has risen dramatically in recent years (Ginther and Kahn, 2004; Hammer-
mesh, 2013).  Using data from economists’ CVs, we track individuals’ career trajectories
and compare whether the trajectory is different for individuals who coauthor versus solo-
author, and whether there is a difference by gender.
Within economics, we find that men and women who solo-author most of their work
have similar tenure rates conditional on a proxy for the quality of papers.  However, an
additional coauthored paper is correlated with an 8.2% increase in tenure probability for
men but only a 5.6% increase for women.  This gap is significantly less pronounced for
women  who  coauthor  with  women,  suggesting  that  the  attribution  of  credit  is  related
to the gender mix of coauthors.  Furthermore, a man who coauthors is no less likely to
receive tenure than a comparable man who solo-authors even though there is presumably
more uncertainty as to how much work he did.  A counterfactual exercise suggests that
this difference in credit allocation explains 60% of the unconditional gender gap in tenure
rates and 84% of the gap that remains after controlling for average paper quality, citations,
tenure and PhD institution ranks, and field.
To ensure that we are not picking up on ability differences between men and women,
we control for the quality of papers using both journal rankings and citations, allowing
1
Blau and DeVaro (2007), for example, find that across jobs, women are less likely to be promoted than
men conditional on worker’s performance and ability ratings.  In the UK, female managers are nearly 40%
less likely to be promoted than male managers (Elmins et al., 2016).
2

for a comparison of men and women with similar research portfolios. The results are also
robust to including other individual-level controls such as length of time to tenure and the
seniority of one’s coauthors, as well as tenure year, tenure institution, and primary field
fixed effects.
We argue that these results are most consistent with a story of women receiving less
credit for their joint work with men because of bias.  To show this, we first use current
CV and citation data to compare the productivity of men and women who did and did
not receive tenure at the institution where they initially went up for tenure.  While the
estimates are imprecise, we find suggestive evidence that women who coauthor and are
denied tenure produce more solo-authored papers that publish in high-ranking journals
than men who are denied tenure. Data on citations show a similar result.
We then rule out several alternative explanations for the empirical patterns.  For ex-
ample, several papers have demonstrated that selection into coauthorship in economics
is not random.
2
We test for selection into coauthorship and do not find any evidence that
women coauthor with high ability or more senior men. We also look at the timing of coau-
thorship and find no evidence that women begin coauthoring if they have a slower start
to their careers. The empirical patterns are also inconsistent with taste-basted discrimina-
tion.
Because the CV data do not allow us to rule out the possibility that women actually
contribute less to papers that are coauthored with men, we conduct two experiments de-
signed to test whether real or perceived differences in contributions drive credit allocation.
In the first experiment, we first hire individuals to complete quizzes on topics that are ei-
ther male or female-stereotyped.  We then hire participants who act as “predictors” and
are randomized into an individual treatment or a joint treatment.  Predictors in the indi-
vidual treatment are shown two individual’s separate quiz scores while predictors in the
joint treatment are shown the combined score of two individuals. They are then asked to
predict the performance of each participant on future quizzes.
In the joint treatment, women are predicted to perform worse than their male counter-
parts for male-stereotyped quizzes, suggesting that predictors believe that women con-
tributed less to the combined score (that is,  they performed worse).   However,  if pairs
performed a female-stereotyped quiz, women and men are given equal credit. To under-
stand whether these results are driven by participants’ beliefs about the ability distribu-
tions of men and women, we randomly provide some participants with the distribution of
scores on the initial quiz by gender. Women appear to be given equal credit in the female-
stereotyped quiz because participants view it as being gender-neutral. That is, they do not
2
See, for example, Boschini and Sjögren (2007), Garcia and Sherman (2015), and Bikard et al (2015).
3

realize that women tend to outperform men. Showing participants the gender distribution
of scores corrects this belief and women are then predicted to have a better performance
in future female-stereotyped quizzes but it does not affect the predicted performance gap
for women and men performing male-stereotyped tasks.
The second experiment is conducted in a more natural setting with human resources
personnel.  Following a similar design, we again test whether women are less likely than
men to receive credit for good group performance. We additionally elicit the HR person-
nels’ beliefs about male and female performance and find that differences in the allocation
of credit are largely driven by differences in beliefs. We also find that male HR personnel
are more likely to hire in favor of men, and women in favor of women.
This paper replicates and builds off of the results in Sarsons (2017), which shows the
basic  correlational  patterns  between  paper  composition  and  tenure.   In  this  paper,  we
replicate the results using more data and then use the C.V. data and two experiments to es-
tablish a channel through which gender influences the allocation of credit. The paper also
relates to a large literature seeking to understand difference in labor market outcomes be-
tween men and women. Factors such as productivity, personality and behavioural differ-
ences (such as competition aversion), and fertility preferences have been shown to explain
some differences in career choice and progression.
3
In academia in particular, studies have
pointed to both supply-side factors, including differences in subject matter interest (Dy-
nan and Rouse, 1997) and the availability of role models (Hale and Regev, 2014; Carrell et
al., 2010); demand-side factors, such as implicit bias (Milkman et al., 2015; Moss-Racusin et
al., 2012); and institutional factors (Antecol et al., 2018). This paper directly tests whether
the differential treatment of work output contributes to the gender gap.
The remainder of the paper is organized as follows.  Section 2 describes the data and
shows that a tenure gap exists between male and female economists.   In Section 3,  we
show that the tenure gap closes as women produce more solo-authored papers but does
not close as they produce more coauthored papers.   Women have a consistently lower
probability of tenure for each additional coauthored paper than men.  We show that the
results are robust to accounting for attrition, and to using different journal rankings and
definitions of tenure. In Section 4, we argue that the results are in line with a story in which
women receive less credit for joint work with men. We test alternative explanations of the
relationship between coauthorship and tenure and argue that none can fully explain the
observed empirical patterns. Section 5 discusses the design and results of the experiments.
Section 6 concludes.
3
There is a large literature documenting gender differences in productivity, attitudes toward different
types of work, and family choices.  See, for example, Niederle and Vesterlund (2007), Buser et al.  (2014),
Antecol et al. (2018), Ceci et al. (2014), Reuben et al. (2017), and Ginther and Kahn (2004).
4

2    Data
To examine the relationship between paper composition and tenure, we construct a dataset
using the CVs of economists who came up for tenure between 1985 and 2014 at one of the
top 35 U.S. PhD-granting universities
4
. The academic progression documented in the CVs
makes it possible to evaluate the relationship between an individual’s research output and
career progression. We can then compare the degree of collaborative work and reward for
that work, and compare these results for men versus women.
2.1    Sample Selection and Data Overview
We  include  only  PhD-granting  institutions  in  the  sample  as  tenure  evaluation  at  these
schools is primarily based on research output, of which we have a clear measure.  Other
institutions like liberal arts colleges place greater weight on teaching ability for tenure,
something that we cannot measure.  We exclude business and public policy schools for
similar  reasons.
5
It  is  reasonable  to  assume  that  the  top  35  economics  departments  in
the U.S. emphasize research which is measured by the number and quality of papers one
produces.
One problem in collecting tenure information is that the CVs of individuals who went
up for tenure, were denied it, and left to industry or government are difficult to find, lead-
ing to a sample selection problem.  To deal with this issue, we collected historical faculty
lists from 23 of the 35 schools and locate over 90% of faculty who had ever gone up for
tenure at these 23 institutions.  For the remaining 12 schools that did not have historical
faculty lists available, we looked at the top 75 U.S. institutions, the top 5 Canadian institu-
tions, and the top 5 European institutions to locate anyone who went up for tenure at a top
35 U.S. school and then moved to another institution. We also checked economists’ CVs at
the major Federal Reserve Boards and other large research institutes, such as Mathemat-
ica, in the U.S. While there might still be a sample selection problem, we show in Section
3.2.1  that  the  results  are  robust  to  using  only  the  sample  for  which  we  have  historical
faculty lists.
From individuals’ CVs, we code where and when they received their PhDs, their em-
ployment and publication history, and their primary and secondary fields. When looking
at the relationship between publications and tenure in the main analysis, we only include
4
The list of institutions are taken from the RePEc/IDEAS Economics Department rankings.  The list of
schools included can be found in Appendix C.
5
Business and policy schools might also value teaching differently and put weight on different types of
journals.
5

papers that were published up to and including the year an individual goes up for tenure.
Book chapters are not included in the paper count.   In a robustness check,  we include
papers that were published one and two years after tenure.
To control for the quality of a person’s publications, we primarily use the “AER equiv-
alent” ranking measure developed by Kalaitzidakis et al.  (2003).  This measure converts
journal publications into their equivalent number of American Economic Review papers.
6
Less than 10% of journal articles cannot be converted because the journal does not appear
in the ranking. In these cases we give the publication a ranking of zero.
7
Using the AER-equivalent measure instead of a list journal rank allows for different
distances  between  journal  ranks  and  for  multiple  journals  to  hold  the  same  rank.   For
example, the top field journals can all hold the same rank.  Other journal rankings force
a ranking among these even though the journals might count the same amount toward
tenure depending on one’s field.  For robustness, we replace this paper quality measure
with the RePEc/IDEAS ranking of economics journals in Section 3.2.2.
Finally, we include citations, measured in 2015, of pre-tenure papers as a control vari-
able. These citations were scraped from Google Scholar.
We supplement this dataset with results from a survey designed to measure individu-
als’ beliefs about the returns to various types of papers. The survey also contains informa-
tion on how frequently individuals present their papers.  The exact questions and nature
of the survey are discussed in greater detail in Section 4.
2.2    Construction of Tenure
To determine whether someone received tenure, we follow the guidelines on each school’s
website (as of 2015) as to when tenure decisions are made. The majority of schools require
faculty to apply for tenure 7 years after their initial appointment.  We therefore consider
years 6-8 to be the “tenure window” in which someone applies for tenure to account for
people who go up for tenure early or late (because of a leave of absence, for example).
We assume that an individual is denied tenure if s/he moves to a university ranked 5
positions below the initial institution during the tenure window.  Similarly,  we assume
that an individual is denied tenure if s/he moves from academia to industry during the
tenure window. Defining tenure in this way accounts for the fact that some people switch
institutions 2-3 years after their initial appointment, not because they were denied tenure
6
The  American  Economic  Review  is  regarded  as  one  of  the  top  journals  in  economics.   Most  journal
publications are therefore converted to be some fraction of an AER paper.
7
If someone does not have any solo or coauthored papers, we set the relevant journal ranking to zero and
include a dummy variable indicating that the individual has no solo (or coauthored) papers.  This enables
us to keep using the full sample.
6

but for personal preferences, and that some people might choose to move to a comparable
school around the time of tenure even though they were offered tenure at their original
institution.  For example,  someone who moves from MIT to Harvard after 7 years was
presumably offered tenure at MIT but chose to move to Harvard for other reasons.
As mentioned, a person who moves 5 or fewer years after his or her initial appointment
is not assumed to have been denied tenure since s/he moved before the tenure window
starts.  If someone moves before the tenure window, we use the second institution they
were at to determine tenure. For example, if a person’s first job is at University A but s/he
moves to University B after three years, we use University B as the tenure institution but
do not start the tenure clock over. We do not restart the clock because the data shows that
in over 80% of cases, the individual still appears to go up for tenure within 8 years of his
or her appointment at the first institution.  However, we do extend this tenure clock in a
robustness check.
Individuals who move from an academic institution into industry before the tenure
window are excluded from the sample.
2.3    Summary Statistics
Table 1 presents summary statistics of the data.  Approximately 68% of the full sample
received tenure, but this masks a stark difference between men and women. Only 52% of
women received tenure while 73% of men did.
Total Papers,Solo-authored, andCoauthoredare the number of papers in each group that
an individual had published by the time of tenure.  These publication counts do not in-
clude books or book chapters.   Papers published in non-economics journals (such as a
political science journal) are included but receive a ranking of 0 (the lowest ranking). The
results are robust to excluding publications in non-economics journals.
There is no statistically significant difference in the number of papers that men and
women produce.  Panel B looks at differences in the quality of papers.  Men are no more
likely to publish their papers in “Top 5” journals (American Economic Review, Economet-
rica,  Journal of Political Economy,  Quarterly Journal of Economics,  and The Review of
Economic Studies) than women.  The only statistically significant productivity difference
is that men tend to publish their coauthored papers in slightly higher-ranking journals.
Specifically, men’s coauthored papers have an average ranking of 0.34 AER-equivalents
while women’s coauthored papers have an average ranking of 0.30 AER-equivalents. We
therefore control for the quality of papers, measured using the AER-equivalent ranking as
well as average citations, throughout the analysis.
7

Panel C displays differences in coauthoring patterns between men and women.Num-
ber Unique CAsis the number of unique coauthors an individual has had by tenure.  Men
and women have roughly the same number of coauthors but there are some differences in
the types of people men and women coauthor with. For example, women are less likely to
coauthor with senior faculty and more likely to coauthor with other assistant professors.
This could in part be driven by the fact that they are also more likely to coauthor with
other women, many of whom are also junior professors.
For illustrative purposes, we plot the number of women and men who have various
combinations of solo and coauthored papers in Figure 1, as well as the average probability
of receiving tenure for each paper combination in Figure 2.  Most men and women have
a similar combination of solo and coauthored papers.  Figure 2 illustrates that individu-
als with a large number of either solo or coauthored papers are likely to receive tenure.
However,  Panel A suggests that women with a higher fraction of their papers that are
solo-authored have a better chance of receiving tenure than women with a mix of solo
and coauthored papers. We examine this claim formally in the next section.
3    Empirical Strategy and Results
3.1    Main Results
We show three main results. We first establish that a significant tenure gap exists between
men and women. We then show that the gap becomes more pronounced the more women
coauthor, and that women who solo-author all of their papers have comparable tenure
rates to men.  Finally, we show that the gender of a woman’s coauthor matters.  Women
who coauthor with other women do not suffer a coauthor penalty.
3.1.1    The Tenure Gap
Figure 3 plots the coefficient
ˆ
β
1
from estimating
T
if st
=β
1
TotPapers
i
+β
2
TotPapers
2
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(1)
separately for men and women using OLS. The dependent variable,T
if st
, is an indicator
that individualiin fieldfat schoolsreceives tenure in yeart.TotPapers
i
is the number of
papers (both coauthored and solo-authored) individualihas at the time he or she went up
for tenure.  A quadratic in the number of papers is included to capture non-linearities in
how publications matter for tenure.  The vector of individual-level controls,Z
i
, includes
8

average journal rank (measured as average AER-equivalents),  the log of total citations,
the number of years it tookito go up for tenure, and the total number of coauthors oni’s
papers. Tenure institution (θ
s
), tenure year (θ
t
), and field fixed effects (θ
f
) are also included
as tenure standards likely vary over time and by field and department.
The figure shows that a significant tenure gap exists between men and women even
after controlling for productivity, primary field, tenure institution, and tenure year. While
an additional paper is correlated with a 13-16 percentage point increase in tenure proba-
bility for men and women, women are consistently 10-13 percentage points less likely to
receive tenure than men conditional on having written the same number and quality of
papers. The lower intercept for women could stem from tenure committees starting with a
lower prior about women’s ability. However, if all papers were clear signals of ability and
tenure committees are Bayesian, we would expect the slope of the relationship between
papers and tenure to be steeper for women.  Put differently, if men and women received
equal credit for papers,  the coefficient onTotPapers
i
should be significantly larger for
women than for men.
We provide a formal test for the difference in slopes for men and women in Column 1
of Table 2, where we present the estimates from
T
if st
=β
1
TotPapers
i
+β
2
fem
i
+β
3
(TotPapers
i
×fem
i
) +β
4
TotPapers
2
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(2)
This is similar to estimating equation 1 except that we interact total papers with a female
dummy,fem
i
rather than splitting the sample.  There is no significant difference in the
marginal benefit of an additional paper to men and women.
3.1.2    The Tenure Gap and Paper Composition
To test whether coauthored papers matter differently for men and women, we separate
papers into those that are solo-authored and those that are coauthored and estimate
T
ifst
=β
1
S
i
+β
2
(fem
i
×S
i
) +β
3
CA
i
+β
4
(fem
i
×CA
i
) +δ
1
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
ifst
(3)
using OLS. Here,S
i
andCA
i
are the number of solo-authored and coauthored papers an
individual has at the time of tenure.
The results are presented in Table 2.  An additional solo-authored paper is associated
with a 9.7 percentage point increase in men’s tenure rates and a 15.4 percentage point
increase in women’s tenure rates (who start from a lower base tenure rate).  If the lower
initial tenure rate for women is due to employers holding the belief that women are lower
9

ability, it seems that the signals from solo papers begin to outweigh the employer’s prior.
This is consistent with a model in which employers start with a lower prior about women
and update as they receive clear signals about a woman’s ability, giving women full credit
for this solo work. This is further discussed in the next section.
If coauthored papers are an unclear signal of ability, an employer must make a judg-
ment call as to how much each coauthor contributed to the paper which could lead to
differential attribution of credit.  Indeed, we see that while an additional coauthored pa-
per  helps  both  men  and  women,  men  benefit  much  more  than  women.   Men’s  tenure
rates increase by 8.2 percentage points when they produce a coauthored paper whereas
women’s increase by 5.6 percentage points.
However, the fact that men benefit nearly as much from a coauthored paper as they do
from a solo-authored paper is at odds with the story that employers are dividing credit
for projects among authors. If employers do divide credit, not all men can get 100% of the
credit, particularly for those papers coauthored with other men.
8
This result could point to
an alternative mechanism.  For example, if employers exhibit taste-based discrimination,
they could use joint projects as an excuse to promote men over women.  We discuss and
test several such alternative stories in Section 4.
The relationship between paper composition and tenure is summarized in Figure 4.
This figure plots the relationship between the fraction of an individual’s papers that are
solo-authored and tenure,  controlling for the total number of papers,  citations,  journal
quality,  number  of  coauthors,  and  tenure  institution,  year,  and  field  fixed  effects.   For
men,  it  does  not  matter  if  one  coauthors  or  solo-authors:  tenure  rates  are  comparable
conditional on the quality of papers.   Women who write all of their papers alone have
similar tenure rates to men.  However, women who coauthor all of their papers have an
approximately 37% tenure rate, substantially lower than that of men who coauthor all of
their papers ( 72%). The slope for women is
ˆ
β=0.521and is statistically significant at the
1% level (s.e.=0.158).
3.1.3    Does Coauthor Gender Matter?
The probability of receiving tenure is not lower for all women who coauthor.  In Table 3,
we categorize coauthored papers into those written with only men, only women, or a mix
of men and women:
8
It could be the case that because tenure committees are evaluating one person, they always assume that
the man they evaluate deserves full credit for the paper (and we do not see the amount of credit they would
have given to the other man). It is impossible to evaluate such theories with these data.
10

T
if st
=β
1
S
i
+β
2
(fem
i
×S
i
) +β
3
CAmale
i
+β
4
(fem×CAmale
i
) +β
5
CAmix
i
+β
6
(fem×CAmix
i
) +β
7
CAfem
i
+β
8
(fem
i
×CAfem
i
) +β
9
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(4)
As before,S
i
is the number of solo-authored papers individualihas at the time of tenure.
CAfem
i
is the number of coauthored papers individualihas in which all of the coauthors
are female. Similarly,CAmale
i
is the number of papersihas in which all of the coauthors
are male andCAmix
i
is the number of papersihas in which the coauthors consist of men
and women.
The estimated coefficients on the interaction terms show that the negative relationship
between coauthoring and tenure for women is driven almost entirely by papers that are
coauthored with men. While a coauthored paper with another man is associated with an
8.7 percentage point increase in tenure probability for a man, it is associated with a 3.1
percentage point increase in tenure probability for a woman.
9
An additional paper with a
woman, however, is associated with an 11.6 percentage point increase in tenure probabil-
ity for a woman.  While this estimate is imprecise due to sample size, we can say that an
additional coauthored paper with a woman has a more positive impact on tenure than an
additional coauthored paper with a man. Any explanation as to why women have lower
tenure rates than men when they coauthor must therefore be correlated with coauthor
gender. The estimates are robust to including all of the control variables discussed earlier.
3.1.4    Counterfactual Analysis
We conduct a counterfactual analysis to estimate how much of the gender gap in tenure
rates can be explained by the different treatment of coauthored papers. We first estimate
T
if st
=β
1
S
i
+β
2
CA
i
+δ
1
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(5)
and use the estimates to predict the probability of tenure,
ˆ
T
i
, for everyone in the sample.
We  then  let  the  female  dummyfem
i
be  0  for  everyone  and  predict  tenure  rates  again
(call this
 ̃
T
i
). The difference
ˆ
T
i
−
 ̃
T
i
gives the gender gap in tenure rates conditional on all
observable characteristics but not allowing for differences in the marginal impact of solo
9
These results again show the puzzling pattern that the amount of credit that is divided among male
coauthors adds up to more than one.
11

and coauthored papers for men and women.
10
We then repeat this exercise using the estimates from equation 4, first letting the female
dummy equal one and then predicting tenure rates again letting the female dummy (and
therefore all of the interactions) equal zero.  This second set of predicted tenure probabil-
ities tells us what women’s predicted tenure rate would be if their papers were treated in
the same way that men’s papers are treated.
The unconditional gender gap in tenure rates is 22 percentage points. The conditional
gap in tenure rates from equation 5 is approximately 16 percentage points. Thus, observ-
able characteristics such as differences in time to tenure and paper quality account for
about 27% of the gap.  The results from using equation 4 to predict tenure probabilities
suggest that the gap would close by a further 13.5 percentage points if men and women’s
papers were treated similarly.   The different assignment of credit thus accounts for ap-
proximately 60% of the unconditional tenure gap and 84% of the conditional gap.
3.2    Robustness Checks
One may be concerned that the results are a product of the types of productivity measures
used or are affected by missing data. In this section, we show that the results are robust to
using only the sample for which we have historical faculty lists, to using different journal
rankings, to accounting for papers published shortly after tenure, and to using different
measures of paper counts.
11
3.2.1    Attrition
The results will be biased if the sample excludes individuals who are denied tenure and
go into industry, government, or other institutions where we do not observe them.  This
would be particularly problematic if men who go to industry after being denied tenure
disproportionately coauthored their papers.  If this is true, we would be overestimating
the benefit of coauthoring for men.  We would have a similar problem if women who go
to industry after being denied tenure typically wrote solo-authored papers.
As discussed in Section 2.1, we attempted to find such individuals by searching insti-
tutions outside of the top 35 U.S. schools, federal reserves, and other research institutes.
10
Interacting all variables except for the number of solo/coauthored papers with the female dummy does
not substantially change the results.
11
In Appendix Table A1, we also test whether the results vary by school rank and over time.  The esti-
mates suggest that the coauthoring penalty is driven largely by schools outside of the top 10, although the
estimates are imprecise. The coauthorship penalty is also stronger in later years but again the estimates are
imprecise.
12

To further allay concerns about sample selection, we run the analysis on the sample for
which we received historical faculty lists.  These lists allow me to track who went up for
tenure and find them even if they left academia. The results, presented in Column 1 of Ta-
ble 4, do not change when run on the sample for which there should be very few missing
observations. The coefficient on theFemale×Coauthoredinteraction is significant only at
the 10% level due to the smaller sample, but the direction and magnitude do not change.
3.2.2    Journal Rankings
In the main analysis, we use a flexible journal ranking that allows multiple journals to hold
the same rank. However, while the economics profession largely agrees on what the “top”
journals are, rankings of field journals or lower-tier journals have changed over time.  In
Columns  2-4  of  Table  4,  we  show  that  the  results  are  robust  to  using  three  alternative
journal ranking metrics as controls.
In Column 2, we use the current RePEc-IDEAS journal ranking.  This ranking forces
a linear relationship between journals and tenure but also contains a larger number of
journals. The main results do not change when using this ranking.
In Column 3, we allow journal rankings to change over time.  We use historical rank-
ings  of  economics  journals  (drawn  from  Laband  and  Piette,  1994,  and  combined  with
current rankings) and match each paper with its journal ranking at the time it was pub-
lished. Using these rankings accounts for journals moving in rank over time as well as new
journals being added.  The coefficient on theFemale×Coauthoredinteraction is slightly
smaller but the same pattern persists. An additional coauthored paper is associated with
an 8.1 percentage point increase in tenure probability for men and a 5.6 percentage point
increase for women. In section 4, we also separate papers into “Top 5s” and “non-Top 5s”.
Finally, in Column 4, we divide the AER-equivalent measure into deciles and control
for the number of solo and coauthored papers an individual has in each decile.  For ex-
ample, if an individual publishes one solo-authored paper in the AER and another in the
lowest-rank journal,  she will have one paper in the tenth bin,  one in the first bin,  and
zero in the others.  Thus, instead of having a single coauthored or solo-authored paper
rank control, we include ten variables controlling for the quality of an individual’s solo-
authored papers (the number of solo papers in each AER-equivalent bin) and ten variables
controlling for the quality of an individual’s coauthored paper (the number of coauthored
papers in each AER-equivalent bin). Again, the results hold.
13

3.2.3    Tenure Definition
In the main analysis, we only consider papers that were published up to and including the
year that an individual goes up for tenure. If an individual goes up for tenure in 1995, for
example, papers published in 1996 are not included in the paper count even though they
may have been “revise and resubmits” at the time of tenure. This could affect the results if
men who coauthor have several promising unpublished papers at the time of tenure but
women who coauthor do not, in which case we are not actually comparing people with
similar publication records.  In Columns 5 and 6 of Table 4, we include papers that are
published one and two years after a person’s tenure year in the paper count variables. The
magnitude of the coefficients are smaller but the results do not change: women continue
to benefit less from coauthored papers than men do.
3.2.4    Paper Count Variable
While we control for journal quality, the main independent variables (number of solo and
coauthored papers) may not accurately reflect how tenure committees decide on tenure
cases.   For  example,  institutions  might  trade  off  the  quantity  and  quality  of  papers  in
different ways.  In Column 7 of Table 4, we use an alternative measure for the number
of papers.  Specifically, after converting each publication to its AER-equivalent, we add
up the AER-equivalent measure to give the total number of “AERs” an individual has
at the time of tenure.  For example, if an individual published two solo-authored papers
and one is worth 0.25 AERs and the other worth 0.8 AERs, the individual will have 1.05
solo-authored AERs at the time of tenure.
Again, the patterns are the same. An additional coauthored “AER” paper is correlated
with an 8.9 percentage point increase in a man’s tenure probability but a 5.3 percentage
point increase in a woman’s tenure probability.
3.3    Testing Against Other Disciplines and Coauthoring Conventions
Many disciplines use different coauthoring conventions, such as listing authors in order
of contribution. However, these disciplines differ on several other dimensions, such as the
fraction of women in the disciplines and what is most important for tenure (publications,
grants, conference proceedings, etc.). In Appendix A, we conduct the same analysis for a
sample of sociologists, a discipline that order authors by contribution. The sample and re-
sults are discussed in more detail in the Appendix, but we do not find evidence of women
being penalized for coauthoring. What matters is being first author on a paper: being first
author is correlated with a 5% increase in tenure probability for both men and women.
14

Because sociology differs from economics in many ways, though, it is difficult to interpret
whether these results suggest that ordering authors by contribution helps eliminate bias
or whether the larger presence of women helps to eliminate it.
4    Channels
The previous section established three facts:
1.  For very few papers, women have a lower tenure probability than men;
2.  As women produce more solo-authored papers, their tenure probability converges
to that of comparable men;
3.  Women benefit less than men from work coauthored with men.
There are several explanations for these patterns. In this section, we argue that the results
are most consistent with a story of women receiving less credit for their joint work with
men rather than a story of women contributing less when they work with men. We assume
that tenure committees begin with the prior that women are on average of lower ability
than men, and that solo-authored papers provide a clear signal of one’s ability whereas
coauthored papers provide an unclear signal. Employers then misattribute credit for work
produced by a man and a woman as the man is assumed to be of higher ability.
We first test the claim by comparing the productivity of men and women who were
denied tenure.  We then explore and rule out several threats to this story.  Specifically, we
test for ability and preference-based sorting, women receiving less exposure by presenting
less, and taste-based discrimination.  Finally, we present evidence from two experiments
designed to shut down the possibility that women put in less effort when working with
men, and find additional evidence that women receive less credit than men when they
perform a stereotypically male task. Moreover, the gender of the person attributing credit
matters in this context, and we also find that women receive at least as much credit as men
when they perform a stereotypically female task.
4.1    Do Men Get the Credit or Do Women Contribute Less?
If tenure committees hold the prior that women are lower ability than men and if solo-
authored papers provide clear signals of ability, we will see differences in tenure rates for
men and women with few publications. However, additional solo-authored publications
of the same quality will have a larger marginal impact on a woman’s tenure probability
than a man’s. As these clear signals begin to dominate the committee’s prior, tenure rates
between men and women will converge.
15

If committees are biased toward giving men more credit for work coauthored with
women, we would expect to see the following. Assuming that there is some fixed amount
of credit that can be given for a paper, a man will benefit more than a woman from joint
work between them. In addition, both men and women will benefit more from their coau-
thored work with women than their coauthored work with men, as two men who coau-
thor will be assumed to have contributed similarly while a woman will be assumed to
have contributed less.
These two claims largely play out in the data.  Table 2 shows that the marginal solo-
authored paper helps women more than it helps men as they start from a lower baseline
tenure rate.  Table 3 shows that men benefit the most from coauthoring with women (an
increase in tenure probability of 9.7% when coauthoring with a woman vs.  8.7% when
coauthoring with a man) although this difference is insignificant.  Similarly, women ben-
efit more from coauthoring with other women than with men.   One result that it is in-
consistent with a story of credit allocation is the fact that the total amount of credit that
can be allocated, at least when all coauthors are men, seems to add up to more than one.
Men benefit as much from a coauthored paper as they do from a solo-authored paper,
suggesting that tenure committees are either making a mistake when dividing credit (for
example,  each committee assumes that the male author under consideration for tenure
at its school did most of the work), or that there is an alternative mechanism behind the
results. In Section 4.2, we test several potential mechanisms.
We would see these same empirical patterns if women contribute less to projects that
are joint with men.  Comparing the productivity of men and women who were denied
tenure helps to disentangle these two stories. If women who coauthor are given less credit,
then women who coauthor and are denied tenure should on average be more productive
than  men  who  are  denied  tenure.   If  women  who  coauthor  simply  contribute  less,  we
would not expect to see productivity differences between men and women who are denied
tenure.
We use two productivity measures to test whether women who coauthor and are de-
nied tenure are more productive than men: the number of solo-authored AER-equivalents
an individual publishes after the tenure decision and the log number of citations an indi-
vidual has as of 2015.
12
Individuals who leave academia and do not publish after tenure
are excluded from the AER-equivalent outcome sample, but including them and setting
their number of post-tenure papers to zero does not change the results.
12
Citations were scraped from Google scholar in 2015. For the top 5 papers outcome, we do not compare
coauthored papers as these can reflect the ability of one’s coauthors.  Citation data includes both solo and
coauthored papers as the data came in this structure.
16

Table 5 shows the results from estimating
Y
if st
=β
1
fem
i
+β
2
FracCA
it
+β
3
T
i
+β
4
(fem
i
×FracCA
it
) +β
5
(fem
i
×T
i
)
+β
6
(FracCA
it
×T
i
) +β
7
(FracCA
it
×T
i
×Fem
i
) +γ
′
Z
′
i
+θ
f
+θ
t
+θ
p
+
if st
(6)
where the outcome variableY
if st
is one of the two productivity measures described above
andT
i
is a tenure dummy. We include a post-tenure institution fixed effect,θ
p
, to account
for the fact that individuals will have access to different resources depending on where
they go after the initial tenure decision.
Column  1  shows  the  results  from  estimating  equation  6  with  the  number  of  solo-
authored AER-equivalents as the outcome.  Women who are denied tenure and coauthor
have 0.4 more solo-authored AER-equivalents than men who are denied tenure and coau-
thor (
ˆ
β
2
+
ˆ
β
4
). Column 2, which has log citations as the outcome variable, shows a similar
pattern although the results are much noisier.  Together, these results provide some sug-
gestive evidence that these women receive less credit for joint projects.
4.2    Alternative Stories
There are other possible explanations for the above findings, not all of which can be tested
with these particular data.  Here we shed light on four standard and testable channels:
ability-based sorting, preference-based sorting, women not claiming credit for their work,
and  taste-based  discrimination.   The  empirical  patterns  are  inconsistent  with  all  of  the
proposed explanations.
4.2.1    Ability-Based Sorting
Employers might rationally deny women who coauthor tenure if individuals sort such
that only lower ability women coauthor with men.  This could arise for several reasons.
For example, if coauthoring lowers the cost of producing a paper, but women know that
they receive less credit for papers, high ability women might forego the cost savings and
choose to work alone.  They know they can produce high quality papers by themselves
and send the employer a clearer signal of their ability.  However,  if low ability women
can  only  produce  high  quality  papers  with  the  help  of  a  high  ability  man,  they  might
coauthor even if they receive less credit.   High ability men will agree to coauthor with
them if it reduces the cost of the paper without reducing the quality.  Employers would
then know that any woman coauthoring with a man is lower ability.
17

In what follows, we test whether women anticipate receiving less credit, whether high
ability women sort out of coauthoring with men, and whether men coauthor with women
whose careers begin more slowly.  To do so, we first present survey evidence suggesting
that women do not know that the returns to coauthoring are lower than solo-authoring.
We then show that women do receive some credit for papers that publish well, suggesting
that employers might believe that there is some assortative matching.  We also provide
evidence that even when women tend to work with men who are slightly higher ability
than themselves this unequal match does not explain the gender gap in tenure.
Survey Evidence on Knowledge of Returns to CoauthoringIf women know that their
returns to coauthoring with men are low, it is plausible that high ability women would
choose to solo-author or only work with other women.   Here we test whether women
anticipate  receiving  less  credit  for  collaborative  work  using  a  survey  conducted  with
economists  currently  working  at  the  top  35  U.S.  economics  departments.   The  survey
was sent to all professors, regardless of rank, at these institutions and received an 32%
response rate.  The gender composition of the sample is representative of the profession
today, with 89 respondents being female and 300 being male.  In the survey, economists
were asked the following question:
Suppose a solo-authored AER increases your chance of receiving tenure by 15%.  For each of
the following, please give an estimate of how much you think the described paper would increase
your chance of receiving tenure.
Respondents then go through five types of papers (coauthored AER, coauthored AER
with senior faculty, coauthored AER with junior faculty, solo-authored top field, and coau-
thored top field) and record their beliefs about the returns to these papers.
13
In  Table  6,  we  test  the  difference  in  the  mean  beliefs  of  men  and  women.
14
There
is no statistically significant difference in the beliefs of men and women for any type of
paper. Men believe that a coauthored AER will increase their chance of receiving tenure by
12.1%, and women by 12.2%. Women believe that there are slightly lower returns to AER
papers coauthored with senior faculty (8.8% versus 9.1% for men), but the difference is not
statistically significant. These results suggest that, in this context, women are unaware of
the true returns to coauthoring.
13
We  did  not  ask  respondents  about  paper  coauthored  with  men/women  so  that  they  would  not  be
primed to think about gender
14
Because the survey was anonymous, the answers can not be linked to the CV data.  We can therefore
only test for differences in means without controls.
18

Evidence on Sorting by Ability from CVsA second test of whether women know that
they will receive less credit for papers and sort accordingly is to look at the correlation
between  propensity  to  coauthor  and  ability.   We  first  test  whether  high  ability  women
are less likely to coauthor than low ability women and then test for assortative matching
among coauthors.  We proxy for ability using the quality of journal that an individual’s
job market paper was published in. We assume that the job market paper is the first solo-
authored paper an individual publishes after he or she graduates.
If women anticipate discrimination,  ability and the fraction of one’s papers that are
coauthored  will  be  negatively  correlated.   High  ability  women  should  be  less  likely  to
coauthor. In Figure 5.A we plot the coefficients
ˆ
β
1
and
ˆ
β
2
from estimating
FracCA
if st
=β
1
a
i
+β
2
(fem
i
×a
i
) +β
3
fem
i
+β
4
TotPapers
i
+θ
f
+θ
s
+θ
t
+
if st
(7)
whereFracCA
if st
is the fraction of personi’s papers that are coauthored anda
i
is person
i’s ability (job market paper rank).  If high ability women anticipate receiving less credit,
we expect
ˆ
β
2
<0.  In Figure 5.A, however, we see that ability is uncorrelated with the
fraction of papers that are coauthored for both men and women: both estimates are precise
zeros.  There is no evidence that women along the ability distribution act strategically in
their choice to coauthor versus solo author.
We also find no evidence that high ability women strategically coauthor with other
women rather than men.  Figure 5.B plots the results from equation 7 using the fraction
of papers that are coauthored with women as the dependent variable.  Women are more
likely to coauthor with other women than men are but there is no sorting by ability.
While women do not seem to be sorting according to ability, it is possible that women
tend  to  work  with  higher-ability  or  more  prominent  coauthors  who  then  receive  more
credit for a paper.  We test for this by correlating a person’s ability with that of his or her
coauthors.  While we do not have the job market paper information for all coauthors in
the dataset, we can see where the coauthors were working at the time the individual went
up for tenure. As a measure of average coauthor ability, we take the average school rank
of all of an individual’s pre-tenure coauthors.  For example, ificoauthors withjandk
andjworks at the 5th-ranked institution andkworks at the 15th-ranked institution, the
average ability ofi’s coauthors is 10.
We correlatei’s ability with the average ability of her coauthors in Figure 6.  The line
of best fit is plotted controlling for number of coauthored and solo-authored publications,
time until tenure, and field, institution, and tenure year fixed effects.
Men and women both sort positively on ability but women are more likely to collab-
19

orate with individuals at more highly-ranked institutions than men are.  To see whether
this explains the main results, we estimate
T
if st
=β
1
S
i
+β
2
(fem
i
×S
i
) +β
3
CA
i
+β
4
(fem
i
×CA
i
) +β
5
rank
iJ
+β
6
(CA
i
×rank
iJ
) +β
7
(fem
i
×CA
i
×rank
iJ
) +β
8
(fem
i
×rank
iJ
)
+β
9
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
if st
(8)
whererank
iJ
is the average institution rank ofi’s coauthors and all other variables are
defined as before. The results are reported in Table 7. If men receive more credit because
they are coauthoring with lower ability women,
ˆ
β
7
should be negative.  However,
ˆ
β
7
is
close to zero, indicating that the ability or prominence of one’s coauthor is not driving the
tenure gap for coauthoring women.
Returns to Top PapersFor high ability women to receive no credit for their coauthored
papers, employers would have to believe that there is no assortative matching by ability.
Otherwise, employers would receive a signal that women who coauthor with high abil-
ity men are also high ability, and be more likely to promote them.  Figure 6 shows that
assortative matching does occur, but it is possible that employers do not recognize this.
We test for this by looking at how credit for top 5 publications is allocated.  If employ-
ers know that there is assortative matching, they should believe that women coauthoring
with high-ability men are also likely to be high ability.
Table 8 shows the results from estimating
T
ifst
=β
1
TopS
i
+β
2
(fem
i
×TopS
i
) +β
3
TopCA
i
+β
4
(fem
i
×TopCA
i
) +β
5
NonTopS
i
+β
6
NonTopCA
i
+β
7
(fem
i
×NonTopS
i
) +β
8
(fem
i
×NonTopCA
i
) +β
9
fem
i
+γ
′
Z
i
+θ
f
+θ
s
+θ
t
+
ifst
(9)
whereTopS
i
andTopCA
i
are the number of solo and coauthored papers that individuali
has published in a top 5 journal. Similarly,NonTopS
i
andNonTopCA
i
are the number of
solo and coauthored papers the individual has published in non-top 5 journals.  In Table
8, the “nop-top 5” interaction terms are presented in the second column.
Power becomes an issue as (1) there are relatively few people publishing in the top
5  journals,  and  (2)  cutting  by  gender  means  that  there  are  even  fewer  women  in  each
category.
Table 8 shows that coauthored papers published in a top 5 journal help women much
more than those published in non-top 5 journals.  Non-top 5 coauthored papers do not
have any positive influence on women’s tenure probability.  It seems that employers re-
20

ceive some signal when a woman publishes her coauthored papers in top journals which
is at odds with the hypothesis that only low ability women coauthor with men.
Overall,  there is little evidence that ability-based sorting is driving the results.
15
If
anything,  employers  seem  to  recognize  that  high  ability  men  and  women  might  work
together and are therefore more likely to grant these women tenure. However, their tenure
rate is still lower than that of high ability men.
4.2.2    Preference-Based Sorting
If women prefer to coauthor with senior faculty, we could reasonably expect that women
would have lower tenure rates.  Assuming senior faculty are more likely to be credited
for a paper, the fact that most senior faculty are men would drive the correlation between
coauthoring with a man and tenure. That is, women receive less credit because they enjoy
coauthoring with senior faculty and these senior faculty are predominantly male.
The basic summary statistics showed that women were not more likely to coauthor
with  senior  faculty  than  men.   However,  we  conduct  an  additional  test  as  to  whether
coauthorship with senior faculty could be driving the results.  We reestimate equation 3
but control for the fraction of a person’s coauthors who are senior.  The results are pre-
sented in Column 3 of Table 7.  The seniority of women’s coauthors does not explain the
results. Controlling for seniority, an additional coauthored paper increases a man’s prob-
ability of tenure by 8 percentage points but a woman’s by 5 percentage points.
4.2.3    Timing of Coauthorship
It is possible that men offer to work with women who are struggling to publish. If this is
the case, we should see women who have few publications in the early years of their ap-
pointment being more likely to coauthor with men. We test for this possibility by looking
at differences in early publications and by testing whether women with a longer time lag
between their initial appointment and first publication are more likely to coauthor with
men.
Appendix Figure B1 descriptively shows the timing of publications for men and women,
split by whether they received tenure at their initial tenure institution. More formally, we
15
Garcia and Serman (2015) show that there could be selection into coauthorship driven by a desire to
be first author on a paper (that is, depending on where you are in the alphabet relative to your coauthors).
This would be an issue in this setting if, for example, men are more likely to be strategic than woman and
are therefore more likely to be first author on a paper (which is correlated with having more citations). We
test whether men are more likely to be first author on their papers than women and whether men have a
“higher” author position overall.  We find that men in our sample are first author 57% of the time while
women are first author 55% of the time (p=0.907).
21

test whether women have fewer publications early in their careers by estimating
Y
if st
=β
1
Fem
i
+β
2
T
is
+β
3
(Fem
i
×T
is
) +β
4
Papers
i
+β
5
 ̄q
i
+θ
f
+θ
s
+θ
t
+
if st
(10)
whereY
if st
is the number of years between individuali’s initial appointment andi’s first
post-appointment publication.
16
We test whether women who did not receive tenure had
a longer publishing lag by interacting the female dummy term with an indicator for re-
ceiving tenure at schools,T
is
. We control for the number of papers published pre-tenure
(Papers
i
) and the average quality of those papers (  ̄q
i
).  All other variables are defined as
before.
The results are presented in Table 9.   Women who do not receive tenure do have a
longer lag (approximately 0.5 years) between their first appointment and their first publi-
cation although the result is noisily estimated. We test whether women with a longer lag
are more likely to coauthor with men by estimating
FracM
if st
=β
1
Fem
i
+β
2
T
is
+β
3
(Fem
i
×T
is
) +β
4
Y
i
+β
5
(Fem
i
×Y
i
)
+β
6
(Fem
i
×T
is
×Y
i
) +β
4
Papers
i
+β
5
 ̄q
i
+θ
f
+θ
s
+θ
t
+
if st
(11)
where the outcome variable,Y
i
in equation 10 is used as a regressor. If men bring women
with a slow start to publishing onto their projects, we would expect to see
ˆ
β
5
>0.
The  results,  presented  in  Column  2  of  Table  9,  do  not  support  the  hypothesis  that
women who struggle to publish initially are more likely to begin publishing with men.
The coefficient onβ
5
is negative, suggesting that women with a longer publishing lag are
less likely to coauthor with men although this result is again insignificant.
4.2.4    Women Not Claiming Credit for Papers
Women might be given less credit for their work if they are less likely to claim it as their
own.
17
For example, if women present less frequently than men, people might associate a
paper with the male coauthor who presents it more. The survey discussed in Section 4.2.1
also asked individuals how many times per year they present their work and whether
they are more or less likely to present their coauthored papers than their coauthor. Panel
B of Table 6 shows that men and women report the same likelihood of presenting their
joint papers relative to their coauthors.  Interestingly, though, women present their solo-
authored  papers  fewer  times  per  year  than  men  do.   It  is  possible  that  women  do  not
16
We exclude papers that were published before the person’s first appointment.
17
Isaksson (2019) finds experimental evidence that women often claim less credit than men for their con-
tributions to solving puzzles.
22

“advertise” their work as much as men do and this leads to women receiving less recog-
nition for their work in general. If this were true, though, women who solo author should
also be less likely to receive tenure.
4.2.5    Taste-Based Discrimination
If some employers have a distaste for tenuring women, as in Becker (1971), we should see
women who write solo-authored papers being denied tenure as well. If employers cannot
plausibly deny a woman who solo-authored several well-published papers, however, they
might be constrained to deny tenure only to those for whom they can make a reasonable
case.  If it can be argued that a woman who coauthors did little of the work, taste-based
discrimination could help to explain the results as employers have an excuse for denying
tenure to coauthoring women. However, as shown in Table 3, only women who coauthor
with men have lower tenure rates.  This would imply that employers have a particular
distaste for tenuring women who coauthor with men, which seems unlikely.
5    Experimental Evidence
In  the  previous  section,  we  provided  suggestive  evidence  that  factors  like  sorting  and
taste-based discrimination do not explain why women who coauthor with men are less
likely to receive tenure. We instead argue that the results are most consistent with women
receiving  less  credit  for  joint  work  with  men.   Specifically,  because  coauthored  papers
are an unclear signal of ability, women receive less credit for their joint work with men
if they are believed to be of lower ability (Correll and Ridgeway, 2003).  We cannot rule
out, though, that real or perceived differences in effort explain the results.  For example,
tenure  committees  might  hold  the  belief  that  women  contribute  less  or  provide  lower
effort when they work with men, regardless of their beliefs about a woman’s ability.  In
addition, tenure committees might believe that low ability women choose to work with
high ability men even if the empirical evidence suggests otherwise.
To shed light on whether different contributions to group work (or perceptions of dif-
ferential contributions) and sorting are driving the results, we run two experiments de-
signed to shut down these channels.  The experiments also allow us to assess the role of
beliefs about ability more directly. The first experiment is an artefactual experiment run on
mTurk. The second is a framed field experiment for which we recruited individuals who
work in human resources and whose job is to recruit personnel.  Although these settings
are different from academia, they provide additional evidence that gender plays a role in
23

the allocation of credit due to differences in beliefs about the ability of men and women.
The first does so in a relatively abstract setting with high control, while the second adds
more context from the process of hiring candidates (see Harrison and List, 2004).
Both experiments consist of two incentivized parts.  In the first step, workers are re-
cruited to complete tasks individually. In the second step, designed to test whether people
misallocate credit for joint work, another set of individuals are recruited to either predict
how well the workers will do on a second set of related tasks (Experiment I) or to choose
a worker to hire to perform a second set of tasks (Experiment II). In both experiments, we
vary whether the predictors/hirers see workers’ individual scores in the first task, or the
sum of two or more individuals’ scores.
5.1    Experiment I
The first experiment consists of two incentivized parts.  In the first step, mTurk workers,
henceforth referred to as “workers” are recruited to complete two related quizzes (Quiz 1
and Quiz 2).
18
We then recruit 506 mTurk participants, referred to as “predictors”, to pre-
dict the Quiz 2 scores of a randomly-chosen man and a randomly-chosen woman on Quiz
2. Before making their predictions, the predictors are told that the workers completed the
two quizzes on their own, and are shown information on Quiz 1. Specifically, they see the
questions asked, the overall score distribution (not broken out by gender), and informa-
tion about the Quiz 1 scores of the two workers they will be making predictions about.
Predictors are then shown the Quiz 2 questions and are asked to estimate the score of
both workers in Quiz 2.  Predictors are paid a participation fee of $0.50 and receive $0.10
for each score they correctly predict.  The instructions given to predictors are available in
Appendix D. This experiment uses a 2x2x2 treatment design, described in detail below.
5.1.1    Treatments
Individual vs. Joint ScoresPredictors are randomized into an Individual treatment or
a Joint treatment.  In the Individual treatment, predictors are shown the individual score
of each of the two workers in Quiz 1.  This treatment tests whether predictors correctly
predict scores when they see a clear signal of each worker’s ability. This parallels the solo-
author paper analysis: if predictors correctly assign credit when they see a clear signal of
ability, there should be no difference in how men and women are evaluated conditional
on Quiz 1 scores.
18
Workers receive a participation fee of $0.30 plus $0.05 for each question they answer correctly.   The
quizzes contain five questions each and are available in Appendix D.
24

In the Joint treatment, predictors are shown thesumof the scores of the two workers.
For example, if worker A scored 3 out of 5 and worker B scored 4 out of 5, the predic-
tor  would  see  the  score  7  out  of  10  for  that  pair.   Importantly,  predictors  are  told  that
there was no interaction between workers:  each worker completed the same quiz and
was paid according to his or her individual score.  Thus, predictors know that workers
are randomly paired with a member of the opposite sex but worked independently and
were individually incentivized. This treatment is designed to understand how predictors
assign credit for performance when they cannot observe individual contributions, but in
a setting where there is no selection into pairs (such as high-ability men working with
low-ability women) or free-riding. Therefore, the predictors’ estimates should reflect only
their  beliefs  about  each  worker’s  score  and  ability  since  they  know  that  workers  com-
pleted the quizzes individually and were individually incentivized.  To draw a parallel
between this treatment and the main analysis, the individual scores that make up the joint
score can be thought of as each person’s “contribution” to a group project that, in this case,
is unaffected by selection or effort.
No-Information vs. Gender-InformationTo understand whether predictors’ estimates
are driven by (possibly incorrect) beliefs about ability or taste-based animus, we provided
some  predictors  with  information  about  the  performance  of  men  and  women.   In  the
No-Information treatment, the only aggregate information predictors receive is the over-
all score distribution.  In the Gender-Information treatment,  predictors are additionally
shown the average score of male and female workers.
19
If predictors exhibit taste-based animus, providing them with information about men
and women’s average performance will not change their predictions. In addition, compar-
ing these treatments helps to understand whether differences in attribution are driven by
incorrect beliefs about gender differences in performance.  If participants hold mistaken
beliefs about men and women’s average performances, the gender information treatment
should correct those beliefs, and the predictors should adjust their estimates accordingly.
Male-Stereotyped vs. Female-Stereotyped quizzesTo evaluate whether differences in
credit  for  joint  work  depends  on  the  type  of  task  that  is  being  performed,  workers  in
the Male-Stereotyped treatment completed math quizzes while workers in the Female-
Stereotyped treatment completed grammar quizzes.
19
The overall score distribution was presented as a histogram.  In the Gender-Information treatment, the
histogram contained lines indicating the mean performance of men and women. See Figure D1 in Appendix
D.
25

5.1.2    Results
The main experimental results are presented in Table 10, which shows how predictors’
guesses vary based on the quiz-taker’s gender and the treatment. Specifically, we estimate
Q2
ij
=β
1
fem
i
+β
2
D
j
+β
3
(fem
i
×D
j
) +β
4
Q1
i
+
ij
(12)
separately for the sample of individuals who took math quizzes (Columns 1 and 2) and
grammar quizzes (Columns 3 and 4),  and by joint/individual treatment.  The outcome
variable,Q2
ij
, is predictorj’s estimate of quiz-takeri’s Quiz 2 score. An indicator for the
quiz-taker being female,fem
i
, is interacted with an indicator for the predictor being in
the Gender-Information treatment,D
j
. We also control fori’s Quiz 1 score (Q1
i
).
In the Individual treatment, there is no significant difference in men’s and women’s es-
timated performance in the math quiz. In this treatment, predictors base their estimations
on the observed individual scores.
20
By contrast, in the Group treatment, where predictors
see only the sum of a man’s and a woman’s Quiz 1 score, they predict that women scored
less than men on the second math quiz. This mirrors the finding that women suffer a coau-
thor penalty when their contribution to a paper is unobserved but are not discriminated
against when their contributions are observed, as in solo-authored papers. Predicting that
the woman will do worse than the man in the Joint treatment suggests that predictors be-
lieve that the woman’s first score was lower; that is, she is worse at the task and therefore
contributed less to the joint score. Puzzlingly, showing predictors the mean scores of male
and female workers does not change the predictions for the second math quiz.  Together,
these results suggest that predictors hold a prior that men are better than women at the
math quiz, and the evidence that men are only slightly better does not affect this belief.
The results for the grammar quiz in Columns 3 and 4 suggest that the results are not
driven by taste-based animus in which women are always penalized in collaborative situ-
ations. Here, women are not predicted to perform differently than men on the second quiz
in both the Individual and Joint treatments.  In addition, seeing that the mean grammar
score of female workers is higher than that of male workers creates a gender difference in
predicted scores in favor of women.
20
Women  had  a  lower  average  score  than  men  on  the  math  quizzes  (2.51/5  vs.  2.72/5)  and  a  higher
average score on the grammar quizzes (2.41 vs. 2.17).  The distribution of scores on Quiz 1 are shown in
Appendix Figure D1.  This is the same figure that predictors are shown.  If predictors are in the Gender-
Information treatment, they also see the two lines indicating the mean male and female scores.
26

5.2    Experiment II
The  second  experiment  was  designed  to  study  attribution  of  credit  for  joint  work  in  a
setting  that  more  closely  approximates  a  hiring  scenario.   In  addition,  this  experiment
allows us to test for gender differences based on the recruiter’s gender, and to test more
directly whether beliefs affect credit attribution.
Before conducting the experiment, we collected individual characteristics from univer-
sity students (henceforth referred to as job candidates), along with their performance in
two incentivized real-effort tasks.  The experiment itself was conducted with 479 actual
human resource workers whose job is to recruit personnel.  The HR workers, henceforth
“recruiters”, were asked to choose job candidates for a task based on short resumes.
The “recruiters” complete an incentivized online survey. Each recruiter is sequentially
shown three sets of four candidates’ resumes. Recruiters pick one candidate from each set
and are paid according to the chosen candidate’s score in the real-effort task.
21
After their choices, the recruiters’ belief about relative gender differences in ability is
elicited.   Specifically,  they are asked to indicate the degree to which they think men or
women are better at the real-effort task.
22
Therefore, Experiment II does not try to induce
beliefs by providing information about scores, as in Experiment I. Instead, it elicits beliefs
to observe whether beliefs are biased and to evaluate the extent to which individual beliefs
affect the recruiters’ choices. The experiment uses a 2x2 design.
5.2.1    Treatments
Individual vs. Joint scoresAs in Experiment I, recruiters are randomized into an In-
dividual or a Joint treatment.  In the Individual treatment, recruiters see the individual
scores of all four candidates in a set.  In the Joint treatment, recruiters see two summed
scores (the sum of candidate 1’s and 2’s score, and the sum of candidates 3’s and 4’s score).
The sets are chosen such that one of the summed scores is obviously superior to the other
to give recruiters a strong incentive to choose one of these two candidates. These superior
candidate pairs always include a male and a female whose resumes are otherwise alike.
23
The pair of inferior candidates may vary on all characteristics but had much lower joint
21
Excluding the participation fee, recruiters earned an average of $6 to complete the ten-minute experi-
ment.
22
Answers ranged from a difference in means of 4 or more points in favor of men to 4 or more points in
favor of women.  Choosing the correct answer is rewarded with $1.50.  The correct answer was calculated
based on the actual scores of candidates in the tasks. For more details, see Appendix D.
23
In addition to their scores, the resume of each candidate shows the candidate’s field of study, degree
length (from three to five years), age, gender, and geographic region of origin.  See Appendix D for more
details and an example of a set of candidates.
27

scores.
Search vs. Vocabulary tasksFor the real-effort tasks, recruiters were randomized to pick
candidates who performed a vocabulary task (finding words using a set of provided let-
ters)  or  a  numeric-search  task  (finding  the  highest  numbers  in  each  of  two  10x10  ma-
trices and adding them up, as in Weber and Schram, 2017).  The tasks are described in
more detail in Appendix D.
24
Compared to Experiment I, these tasks are arguably less
stereotypical and have been shown to exhibit little to no gender difference in performance
(Schram et al.,  2019;  Shurchkov,  2012).   Because men and women perform similarly on
these tasks, using them provides us with a stronger test of whether incorrect beliefs about
performance drive credit allocation.
5.2.2    Results
When recruiters are informed about the individual scores,  the candidate’s gender does
not affect the recruiters’ choices, which are primarily determined by individual scores (for
details, see Appendix D). This mirrors the result in the Individual treatment in Experiment
I and the observation that women and men receive equal credit for solo-authored papers.
To investigate credit attribution in the Joint treatment, we use McFadden’s random-
utility model to explain the binary choice of whether or not to select a candidate under
the  restriction  that  only  one  out  of four  candidates  can  be  chosen  in  a  set  (McFadden,
1974). More specifically, we assume that the utility of recruiterjfrom choosing candidate
iin setkis given by
u
jik
=β
1
fem
ik
+β
2
(fem
ik
×Belief
j
) +β
3
Score
ik
+γ
′
Z
ik
+θ
jk
+
jik
,(13)
wherefem
ik
is an indicator that candidateiin setkis female,Score
ik
is candidatei’s joint
score in the task,  andBelief
j
is recruiterj’s belief about the difference in mean scores
between men and women (constructed such that zero implies a belief of no gender differ-
ences in mean scores, and positive (negative) values imply a belief that men (women) are
better). The vector of controls,Z
ik
, include all the other elements of candidatei’s resume,
whileθ
jk
correspond to fixed effects for each recruiter-set combination.  Recruiteripicks
the candidatejthat gives the highest utility in setk. The random variable
jik
is assumed
to have an extreme value distribution, which allows us to estimate the model using a con-
ditional logistic regression. The estimation results are presented in Table 11 as odds ratios.
24
Recruiters received $0.06 for each point in the vocabulary task by the chosen candidate or $0.15 for each
correct addition in the numerical search task.
28

Column 1 contains the results for the search task and Column 2 for the vocabulary task.
The results for all recruiters show that they are much more likely to pick candidates
in pairs with a high joint score (i.e., from the superior pair in the set).  On average, the
gender  of  the  candidate  does  not  have  a  significant  impact  on  the  likelihood  of  being
chosen  in  either  task.   However,  this  is  no  longer  the  case  once  recruiters  are  divided
according to their gender. Columns 3 to 6 show the estimation results for male recruiters
and Columns 7 to 10 for female recruiters. Columns 3 and 7 reveal that male recruiters are
less likely to pick female candidates, though the odds ratio is not significantly different
from one in the numeric-search task. This aligns with the results of the male-stereotyped
quiz in Experiment I and the data on co-authorship in Economics, where men are given
more credit for joint work.  By contrast, Columns 7 and 9 show that female recruiters are
significantlymorelikely to pick a female than a male candidate.   We will return to this
surprising result below.  For both male and female recruiters, the task does not appear to
matter much as the odds ratios for the female indicator are quite similar across the search
and vocabulary tasks.  Columns 4, 6, 8, and 10 show the estimation results once beliefs
about gender differences in scores are introduced in the regressions. In all cases, recruiters
who believe men are better than women at the task are significantly less likely to pick a
female candidate (and vice-versa).  Correcting for beliefs brings all the odds ratios of the
female indicator closer to 1 (a significant effect for gender remains only for male recruiters
in the vocabulary task), which suggests that the observed biases in credit attribution for
joint work are largely mediated by the recruiters’ beliefs about which gender is better.
25
5.3    Discussion
While the experimental context is different from the academic context, the results provide
evidence that, even after shutting down effort and selection channels, individuals make
different inferences about men and women’s contributions to a joint project that are rooted
in beliefs.
Our two experiments differ along two dimensions: the stereotypical nature of the tasks
used (higher in Experiment I) and the specificity concerning the hiring context (higher in
Experiment II). In addition, we distinguish between the gender of the recruiter in Exper-
iment II. Despite these differences, both experiments provide evidence of a bias against
women when attributing credit to joint work.  Experiment I shows that this can depend
on the task under consideration, with women receiving less credit for stereotypically male
25
Our data show (see Appendix D) that the beliefs of male recruiters are not significantly biased towards
either gender in either task. By contrast, on average, female recruiters expect female candidates to do better
than male candidates in both tasks.
29

tasks and men receiving less credit for stereotypically female tasks. Experiment II, on the
other hand,  shows that the credit-attribution bias can depend on the gender of the re-
cruiter. Male recruiters exhibit this bias against women. Female recruiters, however, show
the opposite bias by attributing more credit to women than to men for joint work. In both
experiments, the differences can be explained by beliefs about which gender is better at a
task.
26
The patterns observed in our two experiments can be used to re-evaluate the re-
sults on tenure decisions in economics. In fact, if we assume that such decisions are made
primarily by men and that economics is seen as a stereotypically-male discipline, then our
experimental results would predict the bias observed in attributing credit to co-authored
papers.
27
Our experimental results further suggest that this bias is caused by (incorrect)
beliefs about the male and female co-authors contributions to joint work.
6    Conclusion
Women receive tenure at significantly lower rates than men in many academic fields.  As
discussed  in  the  introduction,  this  phenomenon  is  not  exclusive  to  academia.   Several
explanations have been put forward for the gap, but it persists even after accounting for
observable characteristics such as fertility preferences and productivity.
This paper proposes an alternative explanation.   We argue that women receive less
credit for group work when employers can not perfectly observe their contribution. When
signals are noisy,  employers have to infer each worker’s ability or productivity.  Coau-
thored papers provide employers with a noisy signal.  The fact that women who work
specifically with men receive tenure at lower rates than comparable women who work
alone or with other women suggests that gender enters into the employer’s inference pro-
cess.  However, when employers receive clear signals, men and women are treated simi-
larly. For example, men and women receive the same amount of credit for solo-authored
papers, which provide a clear signal of ability.  Evidence from two experiments suggests
that these results are not explained by sorting or differences in effort to group work. The
experiments further suggest that this phenomenon is not specific to women as men also
suffer a penalty when working with women on a female-stereotyped task.   Finally,  the
26
However, even after we account for beliefs, by providing information about Quiz 1 scores in Experiment
I or controlling for measured beliefs in Experiment II, we observe too little credit is attributed to women by
predictors in the math quiz of Experiment I and male recruiters in the vocabulary task of Experiment II. In
other words, we find suggestive evidence that both beliefs about ability and taste-based animus can play a
role.
27
These results are also consistent with the lack of evidence of women being penalized for coauthoring in
sociology (see Appendix A), a discipline with relatively more women and thus less likely stereotypically-
male than economics.
30

gender of the person assigning credit also influences credit attribution.
Being aware of this phenomenon is important in a world that is increasingly relying
on group work for production. The tech industry, for example, prides itself on collabora-
tion.  In such male-dominated fields, however, group work could result in fewer women
moving up the career ladder if credit is not properly attributed.  The same could be true
for men in female-dominated industries.  The unequal attribution of credit would then
contribute to and help maintain gender segregation in occupations.
31

References
[1]  Antecol, Heather, Kelly Bedard, and Jenna Stearns. 2018. "Equal but Inequitable: Who
Benefits from Gender-Neutral Tenure Clock Stopping Policies?" American Economic
Review, 108 (9): 2420-41.
[2]  Becker,  Gary. 1971.The Economics of Discrimination.Chicago:  University of Chicago
Press.
[3]  Bikard, Michael, Fiona Murray, and Joshua Gans. 2015. "Exploring Trade-offs in the
Organization of Scientific Work:  Collaboration and Scientific Reward." Management
Science, 61(7): 1473-1740.
[4]  Blau, Francine D. and Jed DeVaro. 2007. "New Evidence on Gender Differences in Pro-
motion Rates: An Empirical Analysis of a Sample of New Hires." Industrial Relations,
46(3): 511-550.
[5]  Boschini, Anne and Ana Sjögren. 2007. "Is Team Formation Gender Neutral? Evidence
from Coauthorship Patterns." Journal of Labor Economics, 25(2): 325-365.
[6]  Buser, Thomas, Muriel Niederle, and H Oosterbeek. 2014. "Gender, Competitiveness,
and Career Choices." Quarterly Journal of Economics, 129 (3): 1409-47.
[7]  Carrell, Scott E., Marianne E. Page, and James E. West. 2010. "Sex and Science:  How
Professor  Gender  Perpetuates  the  Gender  Gap."  Quarterly  Journal  of  Economics,
125(3): 1101-1144.
[8]  Ceci,  Stephen J.,  Donna K. Ginther,  Shulamit Kahn,  and Wendy M. Williams. 2014.
"Women in Academic Science:  A Changing Landscape." Psychological Science in the
Public Interest, 15(3): 1-67.
[9]  Correll, Shelley J. and Cecilia L. Ridgeway. 2003. "Expectation States Theory." Pages
29-51  inThe  Handbook  of  Social  Psychology,  ed.  John  Delamater.  New  York:  Kluwer
Academic Press.
[10]  Dynan, Karen and Cecilia Rouse. 1997. "The Underrepresentation of Women in Eco-
nomics:  A Study of Undergraduate Economics Students." The Journal of Economic
Education, 28(4): 350-368.
[11]  Elmins,  William,  Robert  Joyce,  and  Monica  Costa  Dias.  2016.  "The  Gender  Wage
Gap." Institute for Fiscal Studies Briefing Note 186.
[12]  Garcia, Daniel and Joshua Sherman. 2015. "Norms and Team Formation:  Evidence
from Research Partnerships." Working paper.
[13]  Ginther, Donna K. and Shulamit Kahn. 2004. "Women in Economics:  Moving Up or
Falling Off the Academic Career Ladder?" Journal of Economic Perspectives,  18(3):
193-214.
32

[14]  Hale,  Galina  and  Tali  Regev.  2014.  "Gender  Ratios  at  Top  PhD  Programs  in  Eco-
nomics." Economics of Education Review, 41: 55-70.
[15]  Hammermesh, Daniel. 2013. "Six Decades of Top Economics Publishing:  Who and
How?" Journal of Economic Literature, 51(1): 162-172.
[16]  Harrison,  Glenn  W.  and  John  List.  2004.  "Field  Experiments"  Journal  of  Economic
Literature, 42(4): 1009-1055.
[17]  Isaksson,  Siri.  2019.  "It  Takes  Two:  Gender  Differences  in  Group  Work."  Working
paper.
[18]  Kalaitzidakis, Pantelis, Theofanis Mamuneas, and Thanasis Stengos. 2003. "Rankings
of Academic Journals and Institutions in Economics." Journal of the European Eco-
nomic Assocation, 1(6): 1346-1366.
[19]  Lazear,  Edward and Kathryn Shaw. 2007. "Personnel Economics:  The Economist’s
View of Human Resources." Journal of Economic Perspectives, 21(4): 91-114.
[20]  McFadden, Daniel. 1973. "Conditional logit analysis of qualitative choice behavior."
In P. Zarembka (ed.),Frontiers in Econometrics. New York: Academic Press.
[21]  Milkman, K.L., M. Akinola, and D. Chugh. 2015. "What Happens Before?  A Field
Experiment Exploring How Pay and Representation Differentially Shape Bias on the
Pathway into Organizations." Journal of Applied Psychology, 100(6): 1678-1712.
[22]  Moss-Racusin, Corinne, John Dovidio, Victoria Brescoll, Mark Graham, and Jo Han-
delsman. 2012. "Science Faculty’s Subtle Gender Biases Favor Male Students." PNAS,
109(41): 16474-16479.
[23]  Niederle, Muriel and Lise Vesterlund. 2007. "Do Women Shy Away From Competi-
tion?   Do Men Compete Too Much?" Quarterly Journal of Economics,  122(3):  1067-
1101.
[24]  Reuben,  Ernesto,  Matthew Wiswall,  and Basit Zafar. 2017. "Preferences and Biases
in Educational Choices and Labour Market Expectations:  Shrinking the Black Box of
Gender." The Economic Journal, 127 (604): 2153-86.
[25]  Sarsons,   Heather.  2017.  "Recognition  for  Group  Work:Gender  Differences  in
Academia." AER Papers & Proceedings, 107(5): 141-145.
[26]  Schram, Arthur, Jordi Brandts, and Klarita Gërxhani. 2019. "Social-status ranking:  a
hidden  channel  to  gender  inequality  under  competition."  Experimental  Economics,
22(2): 396-418
[27]  Shurchkov, Olga. 2012. "Under Pressure:  Gender Differences in Output Quality and
Quantity  Under  Competition  and  Time  Constraints."  Journal  of  the  European  Eco-
nomic Association, 10: 1189-1213.
[28]  Weber, Matthias and Arthur Schram. 2017. "The Non-Equivalence of Labour Market
Taxes: A Real-Effort Experiment." The Economic Journal, 127(604): 2187-2215.
33

FIGURE2: TENUREPROBABILITIES BYPAPERCOMBINATIONS
Notes: This figure plots the unconditional tenure probability for women (Panel A) and men (Panel B) who have various combinations
of papers at the time they go up for tenure. Coauthored papers are counted along the x-axis and solo-authored papers are counted
along the y-axis. A darker shade indicates a higher probability of receiving tenure. For example, if a dot is the darkest shade, it
indicates that individuals with that combination of solo and coauthored papers receives tenure with probability one. Panel A is
constructed using the full sample of women (N=143) and Panel B is constructed using the full sample of men (N=501).
35

FIGURE3: TOTALPAPERS ANDTENURE
Notes: This binned scatterplot shows the correlation between the total number of publications an individual has at the time they go up
for tenure and the probability of receiving tenure. The y-variable, tenure, is a binary variable that equals one if an individual received
tenure at their initial institution of employment. For more details on how the tenure variable is constructed, see Section 2. To
construct the plot, tenure is first residualized with respect to the following controls: number of years it took to go up for tenure,
average journal rank of pre-tenure publications, log citations, total coauthors, and tenure school, tenure year, and field fixed effects.
The x-variable, number of publications, is then divided into twenty equal-sized groups. Within each of these groups, we plot the
mean of the y-variable (tenure) residuals against the mean of the x-variable (also within each bin). We then add back the
unconditional mean of Tenure to help with the interpretation of the line of best fit. The lines of best fit are estimated using the full
sample (N=621) and have slopes ofβ=0.132(s.e. = 0.016) for men andβ=0.165(s.e. = 0.043) for women.
36

FIGURE4: RELATIONSHIPBETWEENPAPERCOMPOSITION ANDTENURE
Notes: This figure is a binned scatterplot of the correlation between tenure and the fraction of an individual’s papers that are
solo-authored, split by gender. The y-variable is a binary variable indicating whether an individual received tenure. To construct the
plot, tenure is first residualized with respect to the following controls: total number of papers an individual published by the time of
tenure, number of years it took to go up for tenure, average journal rank of pre-tenure publications, log citations, total coauthors, and
tenure school, tenure year, and field fixed effects. The x-variable, fraction of papers that are solo-authored, is then divided into twenty
equal-sized groups. Within each of these groups, we plot the mean of the y-variable (tenure) residuals against the mean of the
x-variable (also within each bin). We then add back the unconditional mean of Tenure to help with the interpretation of the line of
best fit. The line of best fit using OLS is shown separately for men and women. The lines of best fit are estimated using the full sample
(N=621) and have slopes ofβ=0.521(s.e. = 0.158) for women andβ=0.023(s.e. = 0.748) for men.
37

FIGURE5: ABILITY  ANDSORTING
Notes: This binned scatterplot shows the correlation between an individual’s ability and the propensity to coauthor (Panel A) or the
propensity to coauthor with women (Panel B). The outcome variable in Panel A is the fraction of an individual’s papers that were
published by tenure that are coauthored. The outcome variable in Panel B is the fraction of an individual’s pre-tenure papers that are
coauthored with only women. We proxy for an individual’s ability with the rank of the journal in which the individual’s job market
paper was published. The plot is constructed as described in Figure 3 with the y-variable residualized on the following controls
before plotting: total solo and coauthored papers, the number of years it took to go up for tenure, log citations, and tenure school,
tenure year, and field fixed effects. The lines of best fit using OLS are shown separately for men and women. The estimates for Fig. 5A
areβ=−0.0001(s.e. = 0.0003) for women andβ=0.0002(s.e. = 0.0002) for men. The estimates for Fig. 5B areβ=−0.00004(s.e. =
0.0008) for women andβ=0.0002(s.e. = 0.0003) for men.
38

TABLE2: RELATIONSHIPBETWEENPAPERS& TENURE
Outcome Variable: Tenure
(1)(2)(3)(4)(5)
Sample:FullFullFullFemaleMale
Total papers0.142
∗∗∗
(0.016)
Fem x Papers-0.005
(0.012)
Solo-authored0.094
∗∗∗
0.097
∗∗∗
0.196
∗∗∗
0.095
∗∗∗
(0.013)(0.019)(0.055)(0.024)
Fem x Solo0.048
∗∗∗
0.057
∗∗∗
(0.018)(0.015)
Coauthored0.085
∗∗∗
0.082
∗∗∗
-0.0310.090
∗∗∗
(0.016)(0.014)(0.054)(0.016)
Fem x Coauthored-0.030
∗
-0.026
∗
(0.016)(0.015)
Total coauthors-0.0050.0010.0030.025-0.001
(0.004)(0.005)(0.005)(0.016)(0.006)
Total Papers Sq-0.004
∗∗∗
(0.001)
Solo Papers Sq-0.005
∗∗∗
-0.005
∗∗∗
-0.007-0.005
∗∗
(0.001)(0.002)(0.005)(0.002)
Coauthored Sq-0.003
∗∗∗
-0.003
∗∗∗
0.000-0.003
∗∗∗
(0.001)(0.001)(0.003)(0.001)
Log Citations0.059
∗∗∗
0.031
∗∗
0.065
∗∗∗
0.098
∗
0.058
∗∗∗
(0.012)(0.012)(0.013)(0.054)(0.016)
AER Equiv. Ranking0.533
∗∗∗
(0.116)
AER Equiv. Solo0.139
∗
0.331
∗∗∗
0.3210.416
∗∗∗
(0.071)(0.069)(0.206)(0.091)
AER Equiv. CA0.201
∗∗
0.325
∗∗∗
0.486
∗
0.280
∗∗∗
(0.099)(0.073)(0.248)(0.089)
Female-0.135-0.166-0.205
∗
(0.105)(0.121)(0.109)
Tenure Inst. FEYNYYY
Tenure Year FEYNYYY
Field FEYNYYY
Observations625629621139482
R-squared0.4170.2870.4250.5210.421
This table shows the relationship between publications and tenure.  The dependent variable, Tenure, is binary and
indicates whether an individual received tenure 6-8 years after being hired at the initial tenure institution. Total papers
is the number of papers an individual published by the time s/he went up for tenure. Solo-authored and Coauthored
are the number of solo or coauthored papers s/he had published at the time of tenure.  AER Equiv. Ranking, AER
Equiv. Solo, and AER Equiv. CA are journal quality measures described in Section 2. Total coauthors is the number of
coauthors an individual had on the papers s/he had published by the time of tenure. Tenure length is the number of
years it took the individual to go up for tenure. Citations are from Google Scholar and measured in 2017. The equations
are estimated using a linear probability model.  Bootstrapped standard errors are clustered by tenure institution and
reported in parentheses. (*=p<0.10, **=p<0.05 ,***=p<0.01)
41

T
ABLE
4: R
OBUSTNESS
C
HECKS
Faculty
Journal Rankings
Publication Count
Total AERs
List Sample
RePEc
Over Time    AER Bins    Tenure +1    Tenure +2
(1)
(2)
(3)
(4)
(5)
(6)
(7)
Solo-authored
0.115
∗∗∗
0.083
∗∗∗
0.078
∗∗∗
0.078
∗∗∗
0.058
∗∗∗
0.038
∗∗∗
0.060
∗∗∗
(0.024)
(0.018)
(0.017)
(0.019)
(0.017)
(0.013)
(0.019)
Fem x Solo
0.050
∗∗
0.055
∗∗∗
0.052
∗∗∗
0.045
∗∗∗
0.044
∗∗∗
0.055
∗∗∗
0.091
∗
(0.019)
(0.015)
(0.015)
(0.015)
(0.014)
(0.013)
(0.024)
Coauthored
0.092
∗∗∗
0.079
∗∗∗
0.081
∗∗∗
0.069
∗∗∗
0.038
∗∗∗
0.031
∗∗∗
0.089
∗∗∗
(0.023)
(0.015)
(0.016)
(0.020)
(0.011)
(0.008)
(0.011)
Fem x Coauthored
-0.032
∗
-0.026
∗
-0.025
∗
-0.029
∗
-0.022
∗
-0.013
-0.036
∗
(0.019)
(0.014)
(0.014)
(0.015)
(0.013)
(0.013)
(0.019)
Years to Tenure
-0.046
∗∗∗
-0.054
∗∗∗
-0.055
∗∗∗
-0.051
∗∗∗
-0.046
∗∗∗
-0.045
∗∗∗
-0.050
∗∗∗
(0.009)
(0.008)
(0.008)
(0.009)
(0.008)
(0.008)
(0.011)
Total Coauthors
-0.001
0.004
0.004
-0.004
0.008
∗
0.010
∗∗
0.003
(0.007)
(0.005)
(0.005)
(0.006)
(0.004)
(0.004)
(0.005)
Log Citations
0.056
∗∗∗
0.070
∗∗∗
0.074
∗∗∗
0.079
∗∗∗
0.072
∗∗∗
0.074
∗∗∗
0.094
∗∗∗
(0.017)
(0.013)
(0.013)
(0.013)
(0.014)
(0.013)
(0.016)
CA Paper Rank
0.310
∗∗∗
0.003
∗∗
0.003
∗∗∗
0.345
∗∗∗
0.345
∗∗∗
(0.084)
(0.001)
(0.001)
(0.081)
(0.081)
Solo Paper Rank
0.457
∗∗∗
0.002
∗
0.004
∗∗∗
0.291
∗∗∗
0.299
∗∗∗
(0.077)
(0.001)
(0.001)
(0.069)
(0.065)
Female
-0.158
-0.193
∗
-0.197
∗∗
-0.175
∗
-0.197
∗
-0.280
∗∗∗
-0.199
∗∗
(0.128)
(0.102)
(0.099)
(0.100)
(0.105)
(0.110)
(0.048)
Observations
369
621
621
621
621
621
621
The dependent variable in all columns is an indicator for receiving tenure.  Column 1 restricts the sample to those schools we received a historical faculty list
from.  Column 2 uses RePEc journal rankings as the paper quality measure.  The ranking used can be found at https://ideas.repec.org/top/top.journals.all.html.Column 3 uses historical journal rankings from RePEc to allow for rankings to change over time and to account for new journals entering.  Column 4 controls forthe number of papers within each of 10 AER “bins”.  For this analysis, the AER-equivalent measure is divided into deciles.  For each individual, we then add upthe number of solo and coauthored papers within each decile and include the number of papers in each bin as controls.  In Columns 5 and 6, we include papersthat were published one and two years after an individual went up for tenure in the paper counts.  In Column 7, we use the AER Equivalent measure of journalranking to calculate the total number of AER equivalents (solo and coauthored) an individual had at the time of tenure.  We use this measure in place of the soloand coauthored paper counts (the main independent variables). All regressions control for a quadratic in the number of papers as well as tenure institution, tenureyear, and field fixed effects. Bootstrapped standard errors are reported in parentheses and are clustered by tenure institution. (*=p<0.10, **=p<0.05 ,***=p<0.01)
43

TABLE6: SURVEYRESULTS
(1)(2)(3)
Men    Women    p-value
Panel A: Beliefs about Returns to Papers
Coauthored AER12.112.20.939
Coauthored AER, Sr. Faculty9.18.80.528
Coauthored AER, Jr. Faculty13.313.40.796
Solo Top Field8.08.20.669
Coauthored Top Field6.36.80.223
Panel B: Frequency of Presenting Papers
Times Presented3.12.20.071
Present More Freq. than CA0.370.440.203
Observations30089
This table presents the mean responses for men and women to the following survey questions:
Panel A: “Suppose a solo authored AER increases your chance of receiving tenure by 15 percent. By
how much do you think each of the following increases your change of receiving tenure?” Panel B:
“How many times per year do you typically present your solo-authored papers?  Are you more or
less likely than your coauthors to present a joint paper?”Present More Freq.  than CAis the fraction
of respondents who reported that they are more likely than their coauthors to present a joint paper.
The survey was conducted with a sample of academic economists currently working at a top 35 U.S.
economics department. Respondents were anonymous.
45

TABLE10: EXPERIMENTI PREDICTEDSCORE BYQUIZTYPE
Dep. Var.: Predicted Quiz 2 ScoreMathGrammar
Ind.JointInd.Joint
(1)(2)(3)(4)
Female0.111-0.243
∗∗
-0.0210.103
(0.081)(0.118)(0.071)(0.114)
Gender-Information0.1450.076-0.241
∗∗
-0.456
∗∗∗
(0.097)(0.134)(0.112)(0.130)
Female×Gender-Information-0.108-0.1060.1940.738
∗∗∗
(0.115)(0.154)(0.131)(0.153)
Quiz 1 Score0.735
∗∗∗
0.0200.725
∗∗∗
0.016
(0.057)(0.051)(0.066)(0.053)
Constant0.1373.432
∗∗∗
0.2893.246
∗∗∗
(0.212)(0.361)(0.248)(0.386)
Observations250266231262
Predictors125133116131
R-squared0.2980.0410.2390.139
This table presents the results from Experiment I in which participants predict how well an individual did on a
math or grammar quiz based on that individual’s performance on an earlier quiz. Columns 1 and 2 show the results
for the math quiz and Columns 3 and 4 show the results from the grammar quiz.  In the experiment, participants
were randomized into the Individual treatment, where participants saw each individual’s score on a previous quiz
(Columns 1 and 3), or the Joint treatment, where participants saw the sum of two individuals’ scores (Columns 2
and 4).  Gender-Information is a dummy indicating that participants were told the average quiz scores of all men
and women. (*=p<0.10, **=p<0.05 ,***=p<0.01)
49

T
ABLE
11: E
XPERIMENT
II O
DDS
R
ATIOS OF
B
EING
P
ICKED BY
T
ASK  AND
R
ECRUITER
G
ENDER
Dep. Var.:
All recruiters
Male recruiters
Female recruiters
Picked by recruiter
Search
Vocab.
Search
Vocabulary
Search
Vocabulary
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
Female
1.120
1.041
0.890
0.922
0.750
∗∗
0.723
∗∗
1.341
∗∗
1.157
1.390
∗∗
1.237
(0.108)
(0.099)
(0.136)
(0.142)
(0.104)
(0.098)
(0.167)
(0.152)
(0.181)
(0.178)
Female
×
Belief
0.694
∗∗∗
0.724
∗∗∗
0.760
∗∗∗
0.844*
(0.072)
(0.072)
(0.081)
(0.086)
Joint Score
1.173
∗∗∗
1.032
∗∗∗
1.151
∗∗∗
1.149
∗∗∗
1.036
∗∗∗
1.036
∗∗∗
1.190
∗∗∗
1.189
∗∗∗
1.032
∗∗∗
1.032
∗∗∗
(0.026)
(0.003)
(0.036)
(0.037)
(0.005)
(0.005)
(0.036)
(0.037)
(0.005)
(0.005)
Candidate resumé controls
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Set
×
recruiter fixed effects
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Yes
Observations
3144
2172
1368
1368
996
996
1776
1776
1176
1176
Recruiters
262
181
114
114
83
83
148
148
98
98
This table presents the results from Experiment II in which human resource recruiters pick one candidate out of four for a search or vocabulary task based on short resumes. Columns 1, 3-4, and
7-8 show the results for the search task and Columns 2, 5-6, and 9-10 show the results for the vocabulary task.  Results are shown separately depending on the recruiter’s gender:  all recruiters inColumns 1-2, male recruiters in Columns 3-6, and female recruiters in Columns 7-10.  All regressions include fixed effects for each set-recruiter combination and controls for other variables in thecandidates’ resumes. Results are presented as odds ratios. Standard errors are clustered at the recruiter level. (*=p<0.10, **=p<0.05 ,***=p<0.01)
50

Sociology Results
The sociology sample consists of randomly sampled faculty at the top 20 sociology PhD-
granting departments in the U.S.
28
There are 250 sociologists in the sample, 40% of whom
are female.  Summary statistics are presented in Table A2.  There is no statistically signifi-
cant difference between men and women’s tenure rates (with the mean tenure rate being
76%) although men seem to publish more solo-authored articles than women.
TABLEA2: SOCIOLOGYSUMMARYSTATISTICS
MenWomenp-value
Tenure0.7520.7760.547
(0.433)(0.419)
Total papers12.1510.180.033
(7.808)(5.726)
Total coauthored6.4095.9590.567
(6.641)(4.999)
Solo papers5.7454.2240.003
(4.451)(2.892)
Time to tenure7.5847.5200.686
(1.607)(1.724)
Books0.7790.5710.139
(1.185)(0.799)
Observations150100
This table presents summary statistics for the full sample of sociologists and
separately for men and women. All paper and book count variables (Total Pa-
pers,Solo-authored,Coauthored, andTop 5s) are the number of papers or books
an individual had published at the time of tenure.
To test whether men and women are treated differently, we reestimate equation 3 using
a  probit  model  but  include  measures  of  the  number  of  papers  that  researcheriis  first
author on.  The results are presented in Table A3.  We include the number and fraction of
papers a researcher is first author on in Columns 1 and 2 respectively, along with female
dummy interaction terms.
28
Ranking from U.S. News Education
52

TABLEA3:SOCIOLOGY:PAPERSAND
TENURE
Dep Var: TenureProbitProbit
(1)(2)
Total first author0.050
∗∗
(0.017)
Fem x First Author0.026
(0.040)
Fraction first author0.403
∗∗∗
(0.043)
Fem x Frac. First Author-0.042
(0.172)
Solo papers0.0080.000
(0.006)(0.006)
Fem x Total Solo0.0020.007
(0.011)(0.011)
Total Coauthored-0.010
∗
0.009
(0.004)(0.007)
Fem x Total CA-0.0200.001
(0.017)(0.015)
Books0.063
∗
0.058
(0.032)(0.035)
Book chapters0.0070.005
(0.013)(0.012)
Female0.0260.010
(0.114)(0.163)
School FEYesYes
Tenure Year FEYesYes
Observations237209
This table shows the relationship between the number and types of
papers  an  individual  publishes  and  tenure  for  a  sample  of  sociolo-
gists. The dependent variable is a binary variable indicating whether
the individual received tenure 6-7 years after being hired at the initial
tenure institution.Total first authoris the number of papers an individ-
ual is first author on whileFraction first authoris the fraction of an indi-
vidual’s papers that s/he was first author on.  The equations are esti-
mated using a probit model and the marginal probabilities calculated
at the mean are displayed.  Standard errors, reported in parentheses,
are clustered by tenure institution. (*=p<0.10, **=p<0.05 ,***=p<0.01)
53

Appendix B    Additional Figures
FIGUREB1: TIMING OFPUBLICATIONS
Notes: This figure shows the average number of publications an individual has in the years surrounding his or her initial appointment
as an assistant professor. Year 0 is the year that the individual begins working at his/her tenure institution (tenure institutions are
defined in Section 2). The blue bars represent publications that are coauthored with men. The red bars represent all other publications
(either solo-authored or coauthored with women). Panels A and B show the timing of publications for women and men who were
denied tenure. Panels C and D show the timing of publications for women and men who received tenure.
54

Appendix D    Experiment I
This section provides additional information for Experiment I. As mentioned in the main
body of the paper, the first experiment was conducted with participants from the mTurk
online platform.  First, 80 participants were recruited to play the role of “workers” and
perform two five-question quizzes (21 men and 19 women completed the math quizzes
while 23 men and 17 women completed the grammar quizzes).  Workers received a par-
ticipation  fee  of  $0.30  plus  $0.05  for  each  question  they  answer  correctly.   On  average,
workers earned $0.55. The quizzes used are provided below.
For  the  main  part  of  the  experiment,  505  participants  were  recruited  to  predict  the
scores of one randomly-chosen male worker and a randomly-chosen female worker in a
task.  Predictors were paid a participation fee of $0.50 and received $0.10 for each score
they correctly predicted. The number of predictors in each treatment was as follows: 242
recruiters were assigned to the Individual treatment, of which 120 were assigned to the
No-Information treatment (62 for math quizzes and 58 for grammar quizzes) and 122 to
the Gender-Information treatment (63 for math quizzes and 59 for grammar quizzes), and
264 recruiters were assigned to the Joint treatment, of which 138 were assigned to the No-
Information treatment (70 for math quizzes and 68 for grammar quizzes) and 126 to the
Gender-Information treatment (63 for math quizzes and 63 for grammar quizzes).
D.1    Quizzes used
Grammar Quiz 1
1.  The storm prevented ....... on a picnic.
(a) us to going (b) us going (c) us to go (d) us from going
2.  A man’s concept of liberty is different from ........ .
(a) a woman’s (b) womens (c) a woman (d) woman’s
3.  ........ hour went by before we received ........ invitation
(a) an/an (b) a/a (c) an/a (d) a/an
4.  When a subordinate clause is followed by the main clause, what is required?
(a) a dash (b) a semi-colon (c) a period (d) a comma
5.  ........ are used around a relative clause that defines the noun it follows.
(a) Only commas (b) No commas (c) Semi-colons (d) Quotation marks
56

Grammar Quiz 2
1.  I am dizzy and need to ........ down
(a) lie (b) lay (c) lye (d) go lay
2.  Which of these is not an article?
(a) The (b) A (c) It (d) An
3.  His idea is ........ mine
(a) different to (b) different from (c) different than (d) different then
4.  Adverbs can modify which of the following?
(a) nouns (b) adjectives (c) pronouns (d) none of the above
5.  ........ did you bump into?
(a) Who (b) Whose (c) Who’s (d) Whom
Math Quiz 1
1.  Which of the following is a subset of {b,c,d}?
(a) { } (b) {a} (c) {1,2,3} (d) {a,b,c}
2.  A man’s regular pay is $3 per hour up to 40 hours.  Overtime is twice the payment
for regular time. If we was paid $168, how many hours overtime did he work?
(a) 8 (b) 16 (c) 28 (d) 48
3.  3 4/5 expressed as a decimal is
(a) 3.40 (b) 3.45 (c) 3.50 (d) 3.80
4.  Which of the following is the highest common factor of 18, 24, and 36?
(a) 6 (b) 18 (c) 36 (d) 72
5.  Given thataandbare integers, which of the following is not necessarily an integer?
(a)2a−5b(b)a
7
(c)b
a
(d)ab
57

Math Quiz 2
1.  Items bought by a trader for $80 are sold for $100.  The project expressed as a per-
centage of cost price is
(a) 2.5% (b) 20% (c) 25% (d) 50%
2.  A man bought a shirt at a sale. He saves $30 on the normal price when he paid $120
for the shirt. What was the percentage discount on the shirt?
(a) 20 (b) 25 (c) 33.33 (d) 80
3.  How many subsets does {a,b,c,d,e} have?
(a) 2 (b) 4 (c) 10 (d) 32
4.  What is the median of the given data: 13, 16, 12, 14, 19, 14, 13, 14
(a) 14 (b) 19 (c) 12 (d) 14.5
5.  In coordinate geometry, what is the equation of the x-axis?
(a)y=0(b)x=y(c)x=0(d)y=1
D.2    Instructions
Below are the instructions for the Joint and Gender-Information treatments.  Instructions
for the Individual and No-Information treatments are almost identical and are available
upon request.
Instructions screen 1
INSTRUCTIONS: Please read all the way through.
This project seeks to understand how well individuals can predict a person’s future
performance on a task based on his/her past performance.
We recruited a group of people to complete two math [grammar] quizzes.  Each quiz
had five questions.  Participants had one minute to complete each quiz.  In what follows,
we will show two participants’ scores from the first quiz. We then ask you to predict each
participant’s score on the second quiz. We will provide you with some basic information
on each individual.
You will be paid $0.50 for your participation but will also be paid a bonus of $0.10 if
you correctly guess a participant’s score on the second quiz.
58

Instructions screen 2
We will first show you the distribution of scores on the first quiz. Each bar represents the
fraction of people who obtained that score. For example, 30% of people scored 4/5 on the
first quiz. The average score of female participants (2.5/5) is shown by the solid line. The
average score of male participants (2.8/5) is shown by the dashed line.
Instructions screen 3
Below  we  are  showing  you  a  team’s  score  on  Quiz  1.   Recall  that  each  team  member
worked  on  the  questions  independently.   We  then  take  the  sum  of  the  two  scores  and
assign it to the team. For example, if Person A scored 3/5 and Person B scored 4/5, their
team score would be 7/10.  We provide you with some basic demographic information
about each team member.
Based on the team’s performance,  please predict each individual’s score on Quiz 2.
You can view each quiz by clicking on the link below.
Histograms
The histograms seen by recruiters containing the distribution of scores are seen below in
Figure D1.
59

FIGURED1: DISTRIBUTION  OF SCORES INQUIZ1
Notes: These bar graphs show the distribution of scores on first math and grammar quizzes. The lines mark the means score of men
(dashed line) and women (solid lines). The experiment participants who predicted scores saw these distributions with or without the
lines, depending on whether they were in the Gender-Information treatment.
60

Appendix E    Experiment II
This section provides additional information and analysis for Experiment II.
E.1    Candidates
Before  running  Experiment  II,  the  sets  of  candidates  are  constructing  using  data  from
students who took part in laboratory experiments run in Bologna and Abu Dhabi.
29
In
Bologna, 68 students completed one of the two tasks (16 men and 20 women completed
the search task while 12 men and 20 women completed the vocabulary task),  while in
Abu Dhabi,  90 students completed both tasks (42 men and 48 women).  Students were
paid according to their performance in the tasks.
Vocabulary Task
Students are asked to solve Word-in-a-Word puzzles.  They are given ‘large’ words, one
at a time. The task is to find smaller Italian (Bologna) or English (Abu Dhabi) words that
can be formed out of the letters of the large word.  The task lasts 15 minutes.  There is
a maximum of 24 large words and participants can freely move to the next word at any
time,  but cannot return to previous words.   The following rules apply:  (i) words must
consist of four letters or more, (ii) each letter of the large word can only be used once, (iii)
proper nouns (names, etc.)  are not allowed, and (iv) plurals and verb conjugations are
allowed.  Points are awarded to submitted words ofnletters according to the following
rules:  (i) each word found in a dictionary adds(n−3)points to the score, (ii) words not
found in a dictionary subtract(n−3)points from the score, (iii) words that are too short
subtract 1 point from the score; and (iv) words submitted more than once have no impact
on the score. Points were converted to cash at an exchange rate of 0.10 euros per point in
Bologna (around $0.11 per point) and 1 Emirati dirham per point in Abu Dhabi (around
$0.27 per point).
Search Task
Students are shown two 10x10 matrices.  Each cell is filled with a two-digit number.  The
task is to find the highest number in each matrix, add these up, and enter the sum.  Each
correct answer increases the score by one point.  After entering a number, a new pair of
29
We thank BLESS for allowing us to use their facilities in Bologna.  The experimental software used in
Bologna was developed in PHP-MySQL with the help of Ailko van Veen and Joep Sonnemans, and was
later adapted for the use in Qualtrics by Manu Muñoz. In Abu Dhabi, the experiment was run using zTree.
61

matrices appear,  irrespective of whether the sum is correct.   The task lasts 15 minutes.
Points were converted to cash at an exchange rate of 0.50 euros for every point in Bologna
(around $0.55 per point) and 4 Emirati dirhams per point in Abu Dhabi (around $1.09 per
point).
Resumes
In addition to performing the tasks, students in Bologna and Abu Dhabi answered a few
questions about their demographics and studies.  This information and their scores are
used to construct eight sets of “candidates” for each task. Each set consists of the resumes
of four candidates.  The resume of each candidate includes information about their score
in the real effort task as well as their field of study, degree length (from three to five years),
age, gender, and geographic region of origin. The score is shown for each candidate in the
Individual treatment or as sums of two pairs of candidates in the Joint treatment.   An
example from the Vocabulary task treatment is provided in Figure E1.  The other sets of
this treatment and those of the Search task treatment are available upon request.
FIGUREE1: EXAMPLE OF ONE SET OF CANDIDATE RESUMES
Individual treatment
Joint treatment
62

Note that the sets are constructed such that the summed score of one pair of candidates
in the Joint Treatment is obviously better to that of the other pair (e.g., candidates 1 and
2 in Figure E1).   The candidate pairs with the high score always include a male and a
female whose resumes are otherwise alike.  Specifically, the field of study, degree length,
and geographic region of origin is always identical while age is allowed to vary but within
a narrow range.  The characteristics of the pair of candidates with lower joint scores are
permitted to vary.  This design is used to mask the purpose of the study to recruiters by
giving them multiple characteristics to base their decision on, while at the same time keep
these characteristics constant within the relevant pair of candidates.
E.2    Procedures
For Experiment II, human resource workers from the United States and India were re-
cruited from Qualtrics’ panel of participants to complete an incentivized online experi-
ment.
30
Only respondents who are involved in their firm’s hiring decisions and those that
passed a set of attention checks are considered.  Respondents who complete the experi-
ment receive a participation fee set by Qualtrics plus additional incentives based on their
choices. In total, 479 human resource workers (212 in the U.S. and 267 in India) took part
in the experiment as “predictors”.
Predictors are randomly assigned to the Vocabulary task treatment or the Search task
treatment, and subsequently, to the Individual treatment (top example in Figure E1) or
the  Joint  treatment  (bottom  example  in  Figure  E1).   The  number  of  predictors  in  each
treatment was as follows:  281 predictors were assigned to the Search task treatment, of
which 19 were assigned to the Individual treatment (10 men and 9 women) and 262 to
the Joint treatment (114 men and 148 women), and 198 predictors were assigned to the
Vocabulary task treatment, of which 17 were assigned to the Individual treatment (6 men
and 11 women) and 181 to the Joint treatment (83 men and 98 women).  More predictors
were assigned to the Joint treatment because that is the treatment of interest.
Predictors first complete a simplified version of the task they are assigned to and earn
$0.06 per point in the vocabulary task or $0.15 per point in the Search task.  Thereafter,
in the main part of the experiment, each predictor sees three sets of four candidates and
is required to pick one student from each set.  The sets are shown sequentially and are
picked at random from the eight constructed sets.  The picked students’ scores are paid
30
Throughout the paper, we pool the data from the U.S. and India.  However, our results are unaffected
by further controlling for the recruiter’s country. Running regressions like the ones in Table 11 including an
interaction between the gender dummyfem
ik
and a country indicator results in insignificant coefficients
for the interaction term.
63

out to the predictor at a rate of $0.06 per point in the Vocabulary task or $0.15 per point in
the Search task.  Finally, predictors are asked whether they think that men or women are
better at the task they have participated in. Responses are in five categories and choosing
the correct answer (based on the students’ actual scores in the task) is rewarded with $1.50.
Instructions for the experiment are provided below.
E.3    Instructions
Below are the instructions for the Joint treatment with the Search task.  Instructions for
the Individual treatment and the Vocabulary task are very similar and are available upon
request.
Instructions welcome screen
Thank you for taking part in this survey! The survey will take around 20 minutes to com-
plete.  We would like to see how people make choices when they have to select someone
based on task performance. We will explain this in much more detail later.
You will be compensated for participating in this survey in the usual way. In addition,
you may makeextra earnings, depending on the answers you give and choices you make.
How you can make extra earnings will be made clear in subsequent instructions. All extra
earnings you make will be calculated in US dollars. Your total earnings in dollars will be
paid to you as panel points in the usual manner.  Once again, these extra earnings come
on top of your compensation for participating.
Your decisions in the study are private and anonymous.  They will not be linked to
your name in any way.  We are interested in your own decisions.  We kindly request that
you do not communicate with other people while taking part in the study.
The study consists of three parts. Part 2 will be explained after you have finished Part
1 and Part 3 will be explained after you have finished Part 2. Next, we will explain Part 1.
Instructions part 1 screen
In this first part, we ask you to do a simple addition task with which you can earn money.
When you start, you will see two matrices on the screen.  Each matrix has 6 rows and
6 columns and is filled with randomly generated numbers. Your task is to find the largest
number in each of the two matrices and then to add them up. We will give you an example
below.
For each correct addition, you will receive $0.15. You will have five minutes to do this
task.  Irrespective of whether your answer is correct or incorrect, a new pair of matrices
64

will appear after you enter your answer.  This means that, for each pair, you have only
one attempt to provide the correct answer. At the top of the screen you can see how many
correct answers you have so far.
As mentioned, you will have five minutes in total.  You will see the time that remains
in the upper right corner of the screen. You will be allowed at most 40 addition attempts.
This is much more than anyone can actually add up.
After you have finished reading these instructions, you will see a link.  Click on this
link to complete the addition task.  Note that the addition task will open a new window
in your browser.  Once you have completed the task, you will be given a code.  You will
need this code to complete the study and receive your payment.  Please write it down.  If
you accidentally close the window, you can click on the link again and it will show you
the code.
Perform the addition task:Below is a10-digit number.  Please write it down and then
click on the link to perform the addition task.  When you click on the link, anew window
will appear where you will have to enter your 10-digit code.  Note that if you enter the
wrong code, we won’t be able to pay you for your performance in Part 1.
Once you are done with the task, you will receive a password. You will have to come
back to this page and enter the password below.  This will confirm that you have com-
pleted the addition task.
Instructions part 2 screen
Before we instruct you about Part 2, we would like to inform you of the following:
Between 2016 and 2017, a large number of university students from all over the world per-
formed an addition task like the one you have just performed.
There are two differences between your addition task and the one performed by the
university students: students faced larger matrices (10x10 instead of 6x6) and were given
more time to perform the task (15 min instead of 5 min). These changes were made for you
to be able to experience the same task without taking too much of your time.  However,
despite  these  changes,  the  nature  of  the  task  remains  the  same.   This  means  that  your
experience with the task should give you a sense of what is needed to do well.
Your choices in Part 2:We will present to you three different sets of profiles describ-
ing some of the characteristics of students who did this previous task.  Each set contains
profiles of four different students.  For each set, we would like you tochoose one student.
65

Your choice gives you money. More precisely, you will receive$0.15 for each correct addition
performed by the student you choosewhen he or she did the task.
Because we will give you three sets of profiles to choose a student from, you need to
make a choice three times.  This means you will earn money three times.  Note that once
you have made a choice you won’t be able to go back and change it.
The  profiles  we  give  you  will  contain  background  information  about  the  students.
Specifically,  their age,  field of study,  gender,  type of university degree they purse,  and
the region of the world they come from.  We will also give you an indication of the score
obtained by the students when they did the task. However, you will not be told each stu-
dent’s own score. Instead, we have grouped the students in pairs. Below is an example of
how a set of four students will be presented.  [Here the instructions included an example
similar to the ones in Figure E1]
Please continue to make your three choices.
Instructions decision screen
Examine the profiles closely and choose one student.  Remember, you will receive $0.15
for each correct addition performed by the student you choose when he or she did the
task. The profiles of four different students are below.
Instructions part 3 screen
In Part 3, we ask you to estimate whether female students or male students were better
in the previously-described task.  More precisely, we calculated the average score of all
female students and the average score of all male students who participated in the task
across  all  the  regions  of  the  world.   We  ask  you  to  estimate  whether  females  or  males
scored better on average by answering the question below.  If you estimate correctly, we
you will earn an additional $1.50. I estimate that:
•Female students are much better (the average score of female students is 4 more than
that of male students)
•Female students are slightly better (the average score of female students is between
1 and 3.99 more than that of male students)
•Male and female students are about the same (the average score of male and female
students differs by less than 1)
•Male students are slightly better (the average score of male students is between 1
and 3.99 more than that of female students)
66

•Male students are much better (the average score of male students is 4 more than
that of female students)
E.4    Additional analysis
This  subsection  contains  the  additional  analysis  of  Experiment  II  that  could  not  be  in-
cluded in the main body of the paper due to space constraints.
Individual Treatment
Table E1 shows results from analyzing recruiters’ choices in the Individual treatment. Like
in the main body of the paper, we use McFadden’s random-utility model to explain the
choice of whether or not to select one candidate out of four in each set.   Columns 1 to
4 contain the results for the Individual treatment (Columns 1 and 2 for male recruiters
and Columns 3 and 4 for female recruiters) and Columns 5 to 8 for the Joint treatment
for comparison (Columns 5 and 6 for male recruiters and Columns 7 and 8 for female
recruiters).  The regressions include data form the search and vocabulary tasks to have
enough independent observations. The only difference between this specification and that
in the paper is that instead of the candidates’ joint score, we use an indicator for having
the highest individual or joint score in a set. The estimation results are presented as odds
ratios.
In contrast to the Joint treatment,  the results show that in the Individual treatment,
the gender of the candidate does not have a significant impact on the likelihood of being
chosen irrespective of the gender of the recruiter. Moreover, the higher odds ratio for the
indicator of the highest score shows that, compared to the Joint treatment, recruiters focus
relatively more on scores when making a decision in the Individual treatment.  Finally,
one can also see that including the recruiters’ beliefs concerning the mean scores of men
and women has a smaller effect in the Individual treatment vis-à-vis the Joint treatment,
implying that Joint evaluation makes these beliefs a more important part of the decision.
Beliefs
The recruiters were asked to report their belief about the difference in mean scores of men
and women in either the search or the vocabulary task.   Answers were given in a five
categories, which we code as: (-2) women much better (women’s mean score is more than
4 points larger), (-1) women slightly better (women’s mean score is between 1 and 3.99
points larger), (0) about the same (mean scores differ by less than 1 point), (1) men slightly
67

TABLEE1: EXPERIMENTII ODDSRATIOS OFBEINGPICKED
Dep. Var.:Individual treatmentJoint treatment
Picked byMale recruitersFemale recruitersMale recruitersFemale recruiters
recruiter(1)(2)(3)(4)(5)(6)(7)(8)
Female1.4451.5971.0451.0360.8560.8571.276***1.105
(0.465)(0.537)(0.474)(0.457)(0.082)(0.081)(0.109)(0.104)
Female×Belief0.504*0.9650.717***0.793***
(0.190)(0.238)(0.053)(0.060)
Highest score0.016***0.014***0.080***0.080***0.256***0.254***0.139***0.138***
(0.014)(0.014)(0.030)(0.030)(0.034)(0.034)(0.020)(0.020)
Observations1921922402402364236429522952
Recruiters16162020197197246246
This  table  presents  results  from  Experiment  II.  Columns  1-4  show  the  results  for  the  Individual  treatment  and  Columns  5-8  for  the
Joint treatment.  Results are shown separately depending on the recruiter’s gender:  male recruiters in Columns 1-2 and 5-6, and female
recruiters in Columns 3-4 and 7-8. All regressions include fixed effects for each set-recruiter combination and controls for other variables
in the candidates’ resumes. Results are presented as odds ratios. Standard errors clustered on recruiters. (*=p<0.10, **=p<0.05 ,***=p<0.01)
better (men’s mean score is between 1 and 3.99 points larger), and (2) men much better
(men’s mean score is more than 4 points larger).
Figure E2 shows the distribution of the recruiters’ beliefs depending on the task and the
gender of the recruiter. The modal belief of male recruiters is that the performance of men
and women is about the same in both tasks.  Moreover, the remaining answers are more
or less evenly distributed among the remaining options, implying that the beliefs of male
recruiters are not systematically biased in favor of either male or female candidates. This
is confirmed by sign tests evaluating whether the median of the distribution is zero (p=
0.348for the search task andp=0.597for the vocabulary task).  By contrast, the modal
answer for female recruiters is that the performance of women is slightly better than that
of men,  reflecting a slight bias by female recruiters in favor of female candidates (sign
testsp <0.001in both tasks).  ).  Finally, there are no significant differences in the beliefs
distributions depending on the task (Fisher’s exact tests:p=0.283for male recruiters and
p=0.726for female recruiters),  which confirms that neither task is perceived as more
stereotypically male (female) than the other.
68

FIGUREE2: DISTRIBUTION  OF  RECRUITERS’BELIEF  OF  GENDER  DIFFERENCES  IN  PER-
FORMANCE INEXPERIMENTII
69