ISSN: 2341-2356 
WEB DE LA COLECCIÓN:  
https://www.ucm.es/icae/working-papers  
Copyright © 2021 by ICAE. 
Working papers are in draft form and are distributed for discussion. It may not be reproduced without permission of the 
author/s. 

 

 

 

 

 

 

 

 

  

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Gender Distribution across Topics in the Top 5 
Economics Journals: A Machine Learning Approach 

ICAE Working Paper nº 2109 

 
Keywords: Machine Learning; Gender Gaps; Structural Topic Model; 
Gendered Language; Research Fields. 

 

Abstract 

June, 2021 

JEL Classification I20, J16, Z13. 

 

J.Ignacio Conde-Ruiz  
Fedea 

Universidad Complutense de Madrid and ICAE 
 

Juan-José Ganuza  
Universitat Pompeu Fabra and Barcelona GSE 

 
Manu García 

Washington University in St. Louis and ICAE 
 

Luis A. Puch 
Universidad Complutense de Madrid and ICAE 

We analyze all the articles published in the top five (T5) Economics journals be- 
tween 2002 and 2019 in order to find gender differences in their research 
approach. We implement an unsupervised machine learning algorithm: the 
Structural Topic Model (STM), so as to incorporate gender document-level meta-
data into a probabilistic text model. This algorithm characterizes jointly the set of 
latent topics that best fits our data (the set of abstracts) and how the 
documents/abstracts are allocated to each latent topic. Latent topics are mixtures 
over words where each word has a probability of belonging to a topic after 
controlling by journal name and publication year (the meta-data). Thus, the topics 
may capture research fields but also other more subtle characteristics related to the 
way in which the articles are written. We find that fe- males are unevenly 
distributed along the estimated latent topics, by using only data driven methods. 
This finding relies on “automatically” generated built-in data given the contents in 
the abstracts of the articles in the T5 journals, without any arbitrary allocation of 
texts to particular categories (as JEL codes, or research areas). 
 



Gender Distribution across Topics in the Top 5

Economics Journals: A Machine Learning Approach∗

J.Ignacio Conde-Ruiz,a,c Juan-José Ganuza,b Manu Garćıad and Luis A. Puchc†

aFedea

bUniversitat Pompeu Fabra and Barcelona GSE

cUniversidad Complutense de Madrid and ICAE

dWashington University in St. Louis and ICAE

June 2021

Abstract

We analyze all the articles published in the top five (T5) Economics journals be-
tween 2002 and 2019 in order to find gender differences in their research approach. We
implement an unsupervised machine learning algorithm: the Structural Topic Model
(STM), so as to incorporate gender document-level meta-data into a probabilistic text
model. This algorithm characterizes jointly the set of latent topics that best fits our
data (the set of abstracts) and how the documents/abstracts are allocated to each
latent topic. Latent topics are mixtures over words where each word has a probability
of belonging to a topic after controlling by journal name and publication year (the
meta-data). Thus, the topics may capture research fields but also other more subtle
characteristics related to the way in which the articles are written. We find that fe-
males are unevenly distributed along the estimated latent topics, by using only data
driven methods. This finding relies on “automatically” generated built-in data given
the contents in the abstracts of the articles in the T5 journals, without any arbitrary
allocation of texts to particular categories (as JEL codes, or research areas).

Keywords: Machine Learning; Gender Gaps; Structural Topic Model; Gendered Language;

Research Fields.

JEL Classification: I20, J16, Z13.

∗We thank Antonio Cabrales, Pedro Delicado and Nagore Iriberri for helpful comments, and Elvira Alonso
for excellent research assistance. We also thank the Editor and two anonymous referees for their suggestions,
as well as session participants at Computing in Economics & Finance Conference, Tokyo (virtual) 2021.
José Ignacio Conde-Ruiz and, Manu Garćıa and Luis Puch, respectively, acknowledge the Spanish Ministry
of Science and Innovation for financial support through projects PID2019-105499GB-I00 and PID2019-
107161GB-C32. Juan-José Ganuza gratefully acknowledges the financial support from the Spanish Agencia
Estatal de Investigación, through the Severo Ochoa Programme for Centres of Excellence in R&D (CEX2019-
000915-S) and the Spanish Ministry of Education and Science Through Project ECO2017-89240-P.
†Corresponding Author: Juan-Jose Ganuza, Universitat Pompeu Fabra, Ramon Trias Fargas 27, 08005,

Spain; E-mail: juanjo.ganuza@gmail.com



1 Introduction

Despite the efforts undertaken for the whole economic profession to fight against discrimi-

nation, women are underrepresented in academia. Lundberg and Stearns (2019) make an

assessment of the presence of female economists in the profession and they report a very

slow improvement in the last two decades. The picture is as follows. In the beginning of

this century, 35% percent of PhD students and 30% of Assistant Professors were female.

Since then, these numbers have not increased.1 Additionally, Siniscalchi and Veronesi (2020)

summarizing Chevalier (2019) (Report of the Committee on the Status of Women in the

Economics Profession) point out that the proportion of women assistant professors in the

“top 10” schools has declined to less than 20% by 2019. They document also that female

have been less successful in promoting to tenured associate or full professors.

In Economics, the tenure path often requires to publish in the top five (Top 5, or just

T5) journals, namely: American Economic Review (AER), Econometrica (ECA), Journal of

Political Economy (JPE ), Quarterly Journal of Economics (QJE ) and Review of Economic

Studies (REStud). Heckman and Moktan (2020) analyze the tenure decisions of the top

35 Economics departments in the U.S. and they conclude that T5 publications are a very

powerful explanatory variable of the promotion to Tenure. Publishing in a T5 is becoming

the main goal of young professors in Economics because their professional career may de-

pend on succeeding on this target. In addition, the content published in these journals is

also determining the path of research in Economics. As a consequence of these facts the

competition to publish in any of these journals have increased in recent years. Card and

DellaVigna (2013) analyze the publication records in the Top 5 from 1970 to 2012 showing

that the acceptance rate has fallen from 15% (1970) to 6% (2012). They explain this fact as

a combination of the increasing number of submissions and a declining number of published

papers. Card et al. (2019) further analyze the publication records from two of the T5 jour-

nals (the QJE and REStud), together with the Journal of European Economic Association

and the Review of Economics and Statistics. They report that the current proportion of

1Boustan and Langan (2019) analyze the performance of women across PhD programs in Economics.
They report that in 2017, women were a 32% of entering PhD students in economics, This proportion of
women in economics is below many other fields including science, technology, engineering, and mathematics
(see also Bayer and Rouse (2016)).

1



accepted papers is 3%. Is the T5 entry barrier harder for women? The answer provided by

Card et al. (2019) to this question is ambiguous. On the one hand, these authors do not

find any gender biases in the refereeing process, and editors decisions are gender-neutral

conditional on the referee advises. On the other hand they find that conditional on ref-

eree process, female authored papers end up accumulating more citations in later years.2

A potential explanation for this second result is that journals hold female-authored papers

to higher standards. Hengel (2020) uses readability scores and finds that female-authored

papers are better written and improve during peer review and as they publish more papers.

These results could be related to some “horizontal” features or characteristics of female-

authored papers that lead to more citations or better writing standards, but not to higher

acceptance rates in the editorial process. As Card et al. (2019) control by research fields

(JEL codes), their results may be linked to more subtle “horizontal” differences, for exam-

ple, that in the same research field, males choose a more theoretical approach and females

a more applied perspective (tend to be more cited or subject to less complicated wording).

We use a methodology that allows us to identify these subtle gender “horizontal” research

differences in the form of word-level research topics.

Several papers have pointed out persistent gender differences in the choice of research

fields in Economics. Dolado et al. (2012) analyze the gender distribution of research fields in

the Top-50 Economics departments in 2005, and show that women are unevenly distributed

across fields. Similarly, Chari and Goldsmith-Pinkham (2017) use data from submissions to

the National Bureau of Economic Research Summer Institute (2001-2016), and show that

the distribution of female researchers is not uniform across fields. From these, we learnt that

women are particularly underrepresented in macro, finance and economic theory, and more

prevalent in labor or applied microeconomics fields. Beneito et al. (2018) find similar results

using data from the annual AEA meetings from 2010-2016, while Lundberg and Stearns

(2019) focus on PhD dissertations in Economics from 1991-2017, in almost all major PhD-

granting departments in the United States. Using the JEL code for identifying the research

area, they find that women are more prone to study topics in Labor and Public Economics

than in Macro and Finance. They also show that this pattern has not changed over time.

2Hengel and Moon (2020) analyze publications in T5 and they also find that female authors published
articles are more cited.

2



In this paper, we want to contribute to this literature in two directions. First, we focus on

exploring gender “horizontal” distribution across research topics in the leading Economics

journals. More importantly, we do so, using a new methodological approach based on

Machine Learning techniques. This classifies our abstracts’ database into latent topics.

We collect all the articles published in T5 journals for the period 2002-2019. We obtain

5,311 articles, and we keep track for each article of the authors’ names, year of publication,

journal and the abstract. With this information, we can provide a very accurate picture of

the performance of men and women while publishing in these leading journals. Our goal is

to describe what these latent topics are and the distribution by gender across these topics.

Second, from the universe of algorithms for topic modelling we implement and develop

the Structural Topic Model (STM) developed by Roberts et al. (2019). This choice is

because the algorithm allows to incorporate document-level meta-data into a probabilistic

text model. Precisely, we keep track of journal names and publication years as covariates

to improve the estimation of the prevalence of topics in our data. Our abstracts come

from different sources and different periods of time, so it is natural to allow this meta-

data to affect the frequency with which a topic appears. The output of the algorithm is

a stochastic model that generates latent topics and allocate the documents to them in a

probabilistic way. The main advantage of this unsupervised machine learning approach is

that “latent topics” are mixtures over words where each word has a probability to belong

to the different topics. Therefore, these topics can capture, conditional on covariates and

without human intervention, research fields, information regarding the style of writing,

methodology, conversational patterns or even different ways of thinking.

We start by identifying the number of latent topics for which the stochastic model fits

best our data. Our main result is that female are unevenly distributed across latent topics.

One key aspect is that female prevalence dispersion is higher across these topics. Moreover,

we show that although the proportion of females is slightly increasing among the popula-

tion of T5 authors over the years, the identified “horizontal” differences persist. We have

computed the empirical distribution of latent topics by gender and we show some striking

differences between male and female expected proportions. We want to emphasize the im-

portance of these results, not only because latent topics may capture subtle “horizontal”

3



differences, but also because the differences about gender we estimate are “automatically”

generated given the documents, without any arbitrary allocations to particular categories

(as JEL codes, or declared research areas), and thus, they are possibly more robust.

Notwithstanding, the choice of the number of latent topics, even if optimal as we discuss,

is subject to clustering issues. Thus, we also choose to reduce the number of topics the

algorithm has to generate in order to try to capture the mixtures of words that more

closely relate to research areas. There is a trade off when choosing ex-ante the number of

latent topics. On the one hand, a relatively high number of topics usually fits better the

data. On the other hand, a lower number of latent topics facilitates the broad semantic

interpretation of them. In our setting, a lower number of topics turns out to make them

closer indeed to traditional research fields. Consistently with our main finding above, we

also characterize an uneven distribution of topic/research fields by gender, and very much

in line with the existing literature cited above. However, here we can also discuss the

link between the existing findings and our class of probabilistic results. In a nutshell,

our approach provides complementary evidence from previous literature over “horizontal”

research differences between males or females. The estimated larger set of research topics

may allow to identify more precisely the gender gaps, and what is more important, may

help to understand the driving forces behind these gaps.

There are several channels for which the gender differences in the choice of research topic

that we identify in this paper can have an impact on the probability of publishing in top

journals, earning tenure and in general on career success. Conde-Ruiz et al. (2017, 2021)

and Siniscalchi and Veronesi (2020) provide two dynamic mechanisms that may explain how

“horizontal” gender differences, together with an initially uneven distribution of gender re-

searchers, may generate an unintentional discrimination trap linked with the functioning

of academic organizations (journals, departments, etc.). Conde-Ruiz et al. (2017, 2021)

analyzes a promotion setting in which workers’ skills are assessed by committees whose

members have different abilities to evaluate workers’ signals (they are better at evaluating

workers from the same group). This “homo-accuracy” assumption naturally translates to

the present academic setting, where promotions and editorial processes are done by “com-

mittees” and where evaluators making research in the same research field are able to assess

4



better the underlying quality of the candidate. Under this “homo-accuracy bias”, the group

that is most represented in the evaluation committee generates more accurate signals, and,

consequently, has a greater incentive to invest in human capital. This gives rise to a dis-

crimination trap. If, for some exogenous reason, one group is initially poorly evaluated (less

represented into evaluation committees), this translates into lower investment in human

capital of individuals of such group, which leads to lower representation in the evaluation

committee in the future, generating a persistent discrimination process. Siniscalchi and

Veronesi (2020) focus specifically on the academic labor market and point out a similar un-

intentional discrimination trap linked to the so-called “self image bias”. Research evaluation

is biased towards young researchers with similar characteristics to them. The authors build

up an overlapping-generations model with two groups of researchers with equally desirable

(but a little bit different) research characteristics and identical ex-ante productivity distri-

butions. If one group is slightly over-represented into the evaluation group, this group (and

its specific research characteristics) may dominate forever. These theoretical results go in

line with the empirical findings of Dolado et al. (2012) that show that the probability for a

female researcher to work on a given field is positively related to the share of women already

working on that field (path-dependence). The proportions these authors find based on JEL

codes are very similar to what we find automatically at the same level of aggregation, but

we can set forth a lot more field idiosyncrasy. At the end of the paper we discuss various

issues for further research in related applications.

The paper is organized as follows: the next section presents the raw data and the descrip-

tive analysis of the patterns of publication in T5 journals. Section 3 presents the Structural

Topic Model. Section 4 studies the gender differences in the latent estimated topics. Sec-

tion 5 extends the model to analyze topics as research fields. Last section concludes and

in the Appendix we explore several extensions and provide details about the functioning of

the Structural Topic Model (STM) algorithm.

2 Raw Data and Descriptive Analysis

We collect the publicly available information from all articles published between 2002 and

2019 in the T5 leading journals in economics, as already indicated: The American Economic

5



Figure 1: Number of Articles Published per Year in T5.

Note: Publications exclude notes (without abstract), comments, announcements, and Papers and
Proceedings (P&P).

Review, Econometrica, The Journal of Political Economy, The Quarterly Journal of Eco-

nomics, and The Review of Economic Studies. For each article we collect the information

about the journal, year of publication, authors and the abstract of the paper.

We have 5,311 articles in total over the period 2002-2019, the average number of papers

published in Top-5 journals per year is 295, with a maximum of 351 (on year 2017), and a

minimum of 234 (on year 2002). Figure 1 shows that the distribution of published papers by

journal is uneven. AER accounts for 34.3% while JPE only represent 13.4% of the sample.

AER publishes regular articles as well as shorter papers.3 We include in our sample the

shorter papers (as long as they have abstract) since their editorial processes is similar to

regular articles. We exclude the articles published in AER as Papers and Proceedings since

their requirements and editorial processes are different.4 We want to compare this descriptive

information with Card and DellaVigna (2013) who analyze all the articles published in the

T5 from 1970 to 2012. They obtain several interesting facts, among them, that the total

number of articles published in these journals declined from 400 per year in the late 1970s

to 300 per year in 2012. They also show that one journal, the American Economic Review,

3AER stopped publishing shorter papers in 2018.
4In Appendix E we add P&P articles to our data and we replicate the analysis for these extended data.

6



Figure 2: Number of Authors of Published Papers in T5.

accounted in 2012 for 40% of T5 publications, up from 25% in the 1970s. In our updated

sample, as it is shown in the figure, we find that this trend has stabilized after 2012.

Card and DellaVigna (2013) also find that the number of authors per paper has increased

from 1.3 in 1970 to 2.3 in 2012. We observe the same trend in the recent years, in particular

in 2019 the average number of authors was above 2.5. Figure 2 reports the share of articles

by number of authors, one to five or more. Clearly the steepest trend downward is for

solo authorship, whereas the three authors case (or even the four authors case) exhibits

the opposite pattern. The two authors case share has remained fairly stable over the entire

sample at around 40% of articles (base, not augmented). Five or more authors in Economics’

articles at leading journals are still a rare event.

Next we move to analyze gender issues. We do not observe directly gender in our data.

For solving that problem, we classify authors by gender according to their first name. We

rely on three different databases: the first-names’ database published by the U.S. Social

Security Administration, created using data from Social Security card applications; the

database constructed by Tang et al. (2011), who use Facebook to collect data on first names

and self-reported gender; and finally, the names’ database developed by Bagues and Campa

(2017). We check manually any candidate who (a) falls within the [0.05 0.95] probability

interval of being male/female or (b) cannot be found in any of the databases.

We convert the original sample of articles into an articles-authors sample. We transform

7



Figure 3: Number of article-author observations by gender, and the share of female articles.

the original 5,311 articles to a total sample of 11,721 (with implied 9,840 articles-men

authors, and 1,881 articles-women authors). Except otherwise indicated all measures below

are computed over this augmented articles-authors sample.

Figure 3 depicts the share of female authors (right axis), which has been steadily increas-

ing (with fluctuations) at a rate of 6.2% per year, (compared to men’s share average rate

at 3.7%), reaching 20% share during a couple of years in the recent past. Despite female

authors are increasing at a higher rate, and that there have been an important improvement

in the last decades, women are clearly under-represented in T5 publications. This data is

consistent with the data from the report of the Committee on the Status of Women in the

Economics Profession, Chevalier (2020). Figure 4 compares the evolution of the share of

women in the different professor categories of the top 20 Schools of Economics in the United

States in 2020 with the proportion of female authors in Top 5. Notice that the share of

female authors is very similar to the 20,4% share of women in the faculty of the top 20

Schools in the United States on 2020. In line with Heckman and Moktan (2020), the rate

of increase of female coauthors in T5 seems to be very similar to the rate of increase of

female full Professors in these Departments. The average proportion of females that are full

professor in Spain and the EU average are very similar5

5See Auriol, Friebel and Wilhelm (2019).

8



Source: CSWEP Report, 2020 and own elaboration.

Figure 4: The Pipeline for Top 20 Economcs Departments: Percent and Numbers of
Faculty and Students who are Women.

We have split the description of the data into two figures, one for single gender groups

and another for mixed teams. Figure 5(a) shows the corresponding co-authorships pattern

when the set of co-authors are single gender groups. The more salient feature of these data

are that, while the share of sole maleauthors has been declining from 30% of total, to slightly

above 10%, the share of sole female articles has been stable over the entire sample, at a

share close to 5%. We want also to point out that despite the slow decline, two males is the

most common co-authors team.

The equal share of male-female authors has been fairly stable at about 12% (92.7% of

these articles are, in particular, one male-one female). Alternatively, the share of articles

with at least one woman and at least two men has been increasing from nearly 5% over total

to around 14%. Thus, the strongest trend in data seems to be associated to the participation

of female authors in articles with more male authors.

9



(a) Percentage of T5 articles coauthored by single gender teams.

(b) Percentage of T5 articles coauthored mixed gender teams.

Figure 5: Co-authorships patterns in T5 journals.

10



Figure 6: Distribution of number of T5 papers published by gender.

Figure 6 shows the distribution of the number of published papers by gender. Condi-

tioning on having published in T5 journals, females are more likely than males to publish

only one or two papers, while the proportion of authors that have published more than three

papers is greater for males than for females. Clearly though, more than 80% of either female

(15% of the distribution) or male authors have published less than two T5 over the last 20

years. This is an important fact for understanding the role of superstars in the profession

as well as the formation of networks of coauthors.

3 The Empirical Model: Structural Topic Model (STM)

Our empirical strategy is to use unsupervised machine learning techniques to uncover the

hidden structure of our text documents.6 By unsupervised we denote the absence of human

intervention in order to identify the latent topics behind the abstracts of articles published in

the T5 journals during the period 2002-2019. For us, an abstract is a set of words and these

words have different probabilities to belong to one or several latent topics. Informally, when

6For an excellent non technical introduction to machine learning, see Hansen et al. (2017)

11



we are writing on a particular topic there are words that are used more often than others.

Our objective is to provide a low-dimensional representation (topics) of a high dimensional

object (abstracts) while retaining as much as possible its informational content.

The baseline for topic modelling is the LDA algorithm (Latent Dirichlet Allocation)

developed by Blei et al. (2003) and also the most popular machine learning algorithm in

reducing the dimensionality of text documents.7 In this paper, we use an algorithm called

STM (Structural Topic Model) developed by Roberts et al. (2019), which can be understood

as a refinement for this LDA algorithm. This topic model is said to be structural because it

allows the use of “covariates” to inform about the structure (partial pooling of parameters).

These covariates in our case are going to be the different journal names and the different

years in the sample. The idea is to better capture along these dimensions the changing

relationship between words in abstracts and the latent topics. Next we want to explain

the algorithm and the outcome variables, and in Appendix A we provide a more technical

discussion over STM and LDA.

We start by describing the inputs. We have our 5,311 abstracts (or documents) to extract

all the words. First, we have to “clean” this set of words in order to reduce the vocabulary

and select terms with more informational content. This helps us for a better estimation of

more semantically meaningful topics. The corpora is the set of unique words that we obtain,

after converting to lower case and remove from the original raw text common stop-words,8

as “for” or “in”. Also, we prune the words until we get their original linguistic root (”educ”

instead of ”education”), and eliminate the words that appears one or two times only.9 In

our case, we start with a set of 13,835 different terms and end up in a corpora of 4,241 of

unique words.

The second step is to represent our text data in a document-term matrix of D rows (5,311

abstracts) and V columns (4,182 unique words in our corpus) where the element (d, v) of

the matrix is the number of times the vth unique word appears in the dth abstract. This

document-term matrix that reduces the dimensionality of our original text variables is the

input of the algorithm. Our objective is to find a probabilistic topic model that is able to

7For technical description of the LDA algorithm, see the original article of Blei et al. (2003) and also
Hansen et al. (2017) that is the first paper that uses the LDA algorithm in the economic literature.

8In particular, we remove the stop-words from the SMART list, developed at Cornell University in 1960.
9See Appendix B for the details of this pre-processing.

12



explain the document-term-matrix in two additional steps. First by identifying K topics in

our corpora and then by representing documents as a combination of those topics. What is

a topic? The topic k is a probability distribution βk over all the unique words of our corpus,

where βv
k is the probability that topic k generates word v. Each document d has its own

distribution over the set of topics θd. This captures that each document/abstract can refer to

several topics. Then, θkd would mean the weight of topic k in document d. The probabilistic

topic model is described by these topic βk and document θd distributions. Given that, we

can compute the probability that an arbitrary word in the document d coincides with the

vth term is pdv =
∑

k β
v
kθ

k
d . Using these probabilities, we can obtain the total likelihood

of our data,
∏
d

∏
v

p
nd,v

d,v , where the nd,v corresponds to the elements in the document-term

matrix (the number of times the vth unique word appears in the dth abstract).

This total likelihood is our “objective” function. In a nutshell, The LDA and the STM

algorithms are designed for finding numerically the stochastic model of latent topics (the

distributions βk and θd) that better suit our document-term matrix, that is that maximizes

this total likelihood. We are going to skip here further details on the algorithms we use,

and we refer the interested reader to the appendix A (and also to Roberts et al. (2014)).

However we want to make two important observations.

First, as indicated above, we are implementing STM instead of LDA. The main ad-

vantage of STM for our data is that we can use very relevant covariate information about

our documents in order to improve parameter estimation.10 In particular, for each docu-

ment/abstract we interact the year of publication as well as the journal name. We take

advantage of the variability of the abstract along the time and across journals for improving

the estimation of our stochastic model in particular of the distribution θd).

The second important observation refers to the determination of the number of topics.

We can follow two strategies. One, it is to find the number of topics that better fits the data,

which usually leads to a large (optimal) K. The alternative is to force the algorithm to use a

given number of topics for facilitating the interpretation of those. For our baseline analysis

we use the first approach and we work with 54 topics, but we also pursue the estimation of

9See Hansen et al (2018) for a precise description of the computation of the total likelihood.
10In Cabrales et al. (2018) there is an attempt to impute also gender as an additional covariate for the

articles published in the British press by looking for female names in the body text of this articles

13



our stochastic model using a fixed number of topics to facilitate comparison with the results

in existing literature.

Previous literature, using JEL codes (for example, in Card et al. (2019)) or research

areas in top departments (for example, in Dolado et al. (2012)) have concentrated in a

broad definition of topics as fields of research, say, Labor or Econometrics. However, the

unsupervised learning methodology we use allow us to go beyond pre-labelled research areas

so as to capture more subtle differences, such as writing style, particular methodologies, or

the variation in research questions. For example, our methodology allow us, when identifying

latent topics, to separate two papers of labor economics, but one more applied and other

with a theoretical contribution. We consider our approach a promising tool to analyze if

there are horizontal gender differences in economics research, that is, whether or not male

and female write different articles even within the same research field. For this reason, in

the next section we will analyze our stochastic model with K = 54 topics, while in Section

5, we will be focusing on estimating our stochastic model with K = 15 topics. In addition

to these two exercises, in the appendix we extend our original sample for including the

abstracts of 1,117 articles published as Papers and Proceeding in AER, between 2011 and

2018 (before 2011 these types of papers do not have abstracts and after 2018 are published

in a different journal). We will show that for this extended sample the optimal number of

topics increases to K = 70. While we have preferred to exclude these papers of the main

baseline analysis because these are very short papers with very different editorial processes

than regular submissions, this extended sample generates interesting new insights.

4 Gender Differences in Latent Estimated Topics

As we said above the number of topics that best fits the text data is 54.11 We estimate

probabilities for each document to belong to this set of built-in latent topics using the

Structural Topic Model. The STM output is summarized by the latent topics displayed in

Figure 7 that shows the key words associated to each of the 54 topics. The words within each

row are ordered left to right by the probability they appear in each latent topic. Eventually,

we could assign some labels to latent topics, based on well known fields names in Economics.

11In Appendix C we provide a formal discussion about the optimal number of topics.

14



For instance, we can associate the more prevalent topic in the sample in expectation, topic

28, to international trade. Likewise, the second more prevalent topic in the distribution,

topic 9, may be associated to Econometric Theory. However, this is not the goal of the

analysis as we have indicated above. The important thing is that latent topics may be

related to something beyond research fields, as methodology or style of writing. These

latent characteristics hide gender differences too.

4.1 Topic Prevalence

Once we have identified the estimated latent topics, we can analyze how our documents/

abstracts are distributed among them. In allocating an abstract to a particular topic we

consider our underlying θd distribution. Then we assign document d to different topics

with different probability weights. Following this approach, Figure 8 shows latent estimated

topics in a way that also illustrates the number of documents in each topic, notice that in

Figure 8 the size of the circle is proportional to the expected number of documents in the

topic (we have also reproduced numerically this information in a column in Figure 7). As

we cannot make a mapping of our 54 topics to particular fields of research, it is difficult to

interpret the information of Figure 8 regarding the size of the topics. For example, topics

11, 9 and 21, in Figure 8 are related to “Econometric Theory”, and are relatively large

compared with other topics. However, if the algorithm would have introduced more topics

within “Econometric Theory”, each topic would have had a smaller mass, the weight of the

research field being the same. In other words, our perception of the successful topics is

affected by how the research field is split into topics.

Figure 8 also contains information over the connectedness between topics. For example,

if the latent topic k is closer to k′ than k′′, it means that the distribution βk is more alike

to the distribution βk′ than to distribution βk′′ . Looking at Figure 7 and the description

of the latent topics in Figure 8, some interesting patterns arise. For example, the previous

discussed topics 11, 9 and 21 (“Econometric Theory”) are in someway isolated from the rest

of topics. In Figure 8 we can also identify some other clusters of topics, for example (East in

Figure 8) 51,34, 23, 2, etc are topics related to Macro-Finance, closer to those in Econometric

theory, but not that much; (West in the Figure 8) 50 is a central node of a set of topics

15



import
qualiti
project

rule
conflict
search

vote
save
rate

inform
group
firm

econom
debt
tax

work
public
bank
social
state

auction
mechan
inform

technolog
market
capit
incom
optim
cycl

household
region
percent
consum

risk
financi
polici

contract
return
welfar
market
women
price
game
belief
shock

equilibrium
wage
school

test
prefer
experi
condit
estim
trade

use
delay
effort

demand
increas

unemploy
news
citi

exchang
vote

ethnic
contract

studi
default
reform

program
regul
credit

network
unit
bid

implement
coordin
innov
match
human
earn
alloc
busi
hous

econom
health
firm

avers
invest
polit

agent
firm
cost

inform
children

cost
player
agent
polici

dynam
worker
student
statist
choic

subject
variabl
method
countri

addit
probabl
team
set

violenc
worker
voter
retir

interest
signal

member
ownership

name
borrow
incom
labor
enforc
polici

individu
right

bidder
incent
action

product
stabl
invest
inequ
effici

product
consumpt

area
insur

product
consumpt
constraint

govern
princip
stock

benefit
trade

parent
adjust

strategi
expect

monetari
general
employ
effect

asymptot
decis

experiment
function
sampl

product

data
accept
perform
ration
crime

job
media

account
currenc

voter
trust
vertic

correct
govern

rate
suppli
good
fund

incent
issu

buyer
transfer

communic
new

friction
skill
data

distort
industri
spend
local

increas
demand
ambigu
recess
parti

commit
manag
insur
price
femal
chang
payoff
prior
inflat

equilibria
firm
educ

distribut
util

behavior
identif
data

export

sever
fee

redistribut
problem

war
distribut
candid
popul

countri
aggreg

evid
integr
bias

credit
increas

hour
privat
crisi

interact
econom

seller
type

strateg
firm

competit
growth
differ

economi
fluctuat
incom
growth
hospit
market

util
shock
elect
optim
asset
gain
asset
men
data

equilibrium
ration

aggreg
exist

product
colleg

method
individu

treatment
identifi

asymptot
intern

relat
order

outcom
yield
polic
durat
elect
life
real
bias

segreg
adopt
black
bond
taxat

increas
law
lend

opportun
protect

valu
design
payoff
patent
labor

accumul
measur
privat
chang

expenditur
land
estim
good

discount
asset
voter

hazard
equiti
polici
valu
famili
firm
play

probabl
respons
economi

job
score

paramet
make

predict
restrict

paramet
import

support
card
win

constitut
outsid
wage
estim

increas
patient
privat

countri
industri
measur
fiscal

margin
transfer
provis
liquid

depend
problem

price
compat
game

research
agent
differ

survey
condit

demand
increas

agricultur
care
price
prefer
firm

power
incent

investor
estim
trader
educ

demand
bargain
signal
money
condit
increas

test
confid
altern
learn
estim

consist
firm

analys
offer

competit
optim
option
rate

committe
german
donor
strateg
increas

cost
data

sovereign
chang
time

punish
loan

connect
institut
revenu
post

outcom
adopt
labour
labor

distribut
ineffici
volatil
effect
locat

patient
profit

expect
aggreg

politician
moral

portfolio
loss

privat
marriag

good
repeat

util
real

stochast
labor

teacher
propos

behavior
evid

distribut
use

sector

find
paper
one

function
effect

employ
newspap
individu
regim
elect
cultur

supplier
signific
market
optim

particip
legal

financi
link

properti
privat
agent
sender

knowledg
side

account
use

resourc
aggreg
respons
develop

drug
advertis

asset
credit
elector
inform
predict
reduc
advers
child

markup
cooper

set
nomin
solut
plant

program
forecast

set
theori

instrument
error
factor

limit
higher
prize
util

attack
benefit

bias
rate

transplant
larg

chang
exclus
racial
matur
effect
home
cost

market
secur

resourc
inform

problem
signal
spillov
type
life

mobil
polici
entri
data
data
use

competit
intertempor

financ
public

problem
size
use

select
birth
relat

equilibria
learn
volatil
uniqu
skill

assign
bootstrap

maker
differ

bound
bias

develop

18.7%

17%

14.5%

15.3%
14.8%

10.4%

17.5%
14.7%

13.4%

10.9%

15.5%

19.9%

18%

10.5%

18.7%

10.1%

14.5%

14.4%

17.8%

16.9%

15%
14.1%

16.8%

14.8%

19.5%
15.6%

17.4%

17.8%

17.5%

21.8%

19.4%

11.6%
18.4%

15.6%

17.1%

15%

10.8%

19.8%

13.6%

14.1%

22%

18%

15.1%

14.9%

14.4%

19.4%

15.1%

15.4%

32.8%

13.5%

13.2%

14.5%

15.1%

15.7%

1.3%

2.3%

2%

1.5%
1.6%

2.3%

1.2%
1.2%

3.5%

1.6%

3.3%

1.5%

1.4%

1%

2.2%

2.3%

1.5%

1.9%

2.6%

1.4%

2.7%
2.7%

1.4%

0.8%

1.7%
1.7%

1.7%

3.8%

2.8%

1.3%

1.2%

2.1%
2.2%

2%

1.1%

1.3%

2.5%

1.3%

1.2%

1.8%

1.9%

1.5%

1.8%

1.6%

1.8%

0.9%

1.9%

2.6%

2.2%

2%

2.4%

1.7%

2.2%

0.3%Topic 54
Topic 24
Topic 46
Topic 14
Topic 35
Topic 8
Topic 7

Topic 31
Topic 39
Topic 36
Topic 38
Topic 30
Topic 1

Topic 23
Topic 20
Topic 13
Topic 42
Topic 17
Topic 12
Topic 4
Topic 5

Topic 10
Topic 44
Topic 25
Topic 26
Topic 52
Topic 27
Topic 40
Topic 45
Topic 43
Topic 18
Topic 41
Topic 47
Topic 3

Topic 34
Topic 50
Topic 32
Topic 33
Topic 15
Topic 53
Topic 49
Topic 2
Topic 6

Topic 16
Topic 51
Topic 37
Topic 48
Topic 19
Topic 21
Topic 22
Topic 29
Topic 11
Topic 9

Topic 28

Topic
Prop.

Female
Prop.

5 10 15 20
Word Prevalence (%)

10 20 30

Topic Proportions (%)
(White = median Female Prop.)

Figure 7: Optimal K Topics Ranked by Prevalence in the corpus.

related to Political Economy and Institutions), (South-West in Figure 8) 29,32,22, etc., are

topics related to Microeconomics (contract theory, decision theory, etc.). Finally, applied

areas as labor, international-development, or public economics are located around topics 19,

16



Figure 8: Connectedness between topics and the fraction documents/abstracts in each
topic (θd distribution).

49, 28, and 48 (north in Figure 8). In Appendix D we undertake a more formal analysis

of the distance between topics using a Simple Correspondence Analysis of the probability

matrix for documents to belong to the different latent topics. We find the corpus organized

along two dimensions: Dimension 1 can be interpreted as going from Applies to Theory,

whereas Dimension 2 goes from, say, Economics to Econometrics.

17



Figure 9: Connectedness between topics and the female authors documents/abstracts in
each topic.

Using our classification of authors’ names by gender and the allocation of documents to

latent topics, we can build up a similar figure with information about the gender distribution.

Figure 9 shows latent topics where the sizes of circles are proportional to the percentage of

female authors working in such topics (we have also reproduced numerically this information

in the last column in Figure 7).

Figure 9 provides interesting evidence of the main message of this paper, male and

female display different patterns when doing research. Independently of the grade of under-

representation of women in the profession, if there were not significant gender horizontal

18



differences we would expect that sizes of latent topics measure for the proportion of females

were similar. On the contrary, we observe an uneven distribution of sizes.

There is a small subset of topics (North in the figure 9), specially topic 49, with a relative

high proportion of females, that moreover seem to be closely connected (according to the

terminology for applied economics fields). On the contrary, there is other set of topics (for

example South-West in Figure 9) that are also closely connected and where the presence of

females is scarce (around terms common to economic theory research questions).

4.2 Topic analysis and the gender distribution

As we said above, it is difficult to describe the precise semantic meaning of the latent topics

when we are working with K = 54. We are able, however, to look closer to the latent

topics where females are more or less prevalent and its potential implications. In particular,

Figure 10 shows that the latent topic with the highest proportion of female authors is topic

49 (32.8% as indicated in Figure 7). On the contrary topic 16 turns out to be the topic with

the lowest proportion of females (10.1% as indicated in Figure 7). As a simple illustration,

Figure 10 represents these topics as word clouds, where the size of terms in the cloud is

equivalent to its probability in the latent topic distribution βk.

(a) Topic 49 (highest prop. of female authors). (b) Topic 16 (lowest prop. of female authors).

Figure 10: Topic Word Clouds: Topic 49 vs Topic 16

19



The words that seem to be more prominent in the cloud 49 are women, men, parent,

children, health, etc. These words could be easily linked to research fields, as gender or

health economics, traditionally associated to women. Similarly, the word cloud of topic 16

seems to be related to Micro theory that has been often labeled (while not statistically) as

an area where there are less female than average.

Latent topics may differ in other dimensions beside semantic content. For instance,

Hengel (2020) uses readability scores to measure the quality of writing of article abstracts.12

We have implemented E. Hengel’s Python module Textatistic to compute readability

results over the article abstracts across our latent topics. The finding is that scores across

more female topics are better rated than across more male topics. However, it is hard to

disentangle the role of the prevalence of female authors face to face the wording within a

topic. Moreover, scores that are outliers should be properly treated to ease comparisons.

We leave the study of these readibility issues implying fundamental gender differences for

further research.

Rather, Figure 11 shows the mean of the presence of women authors by topic, together

with the standard deviation of this presence over the sample of years. For some latent topics

the proportion of females is larger than the average (which is 15, 9% over the period 2002-

2019), reaching a proportion of 33% for topic 49. On the contrary, females are specially

underrepresented in other topics, as topic 16, with only a 10%. Dispersion over time differs

also across topics, and it seems that is higher for topics with higher proportion of females (the

correlation between dispersion and the proportion of females is 0.35). While it is true that

the proportion of female authors has been increasing in the last two decades from around

13% on 2002 to 19% on 2019, we do not see a trend in the dispersion of the proportion of

females by topic. Consequently we see the prevalence of females across topics as a signal of

gender “horizontal” differences in research..

Nevertheless, for having a more accurate picture of this “horizontal” differences, we need

to add the information regarding the relative prevalence of the topics. It could be possible

that females are unrepresented in a particular topic, and this circumstance having little

impact as far as this topic contains very few published papers.

12As E. Hengel discusses in detail, abstract readability is strongly positively correlated with the readability
of other sections of a paper.

20



Figure 11: On the presence of women, by topic: mean and one standard deviation across
time.

Figure 12 shows the distribution between males and females across topics normalized

for having the same size. This gives us the propensity that, say, a female authored paper

belongs to any of the 54 topics. We rank the topics according to probability of being

chosen by a male author. This figure provides evidence that male and female authors either

have different preferences or follow different strategies when pursuing and publishing their

research. We observe that topics with higher “demand” by males are also highly demanded

by females. However, there is a set of topics, for which the proportion of published papers

for men are high, which are less attractive (o more difficult to publish) for females. In

general, male and female distributions are different, with the salient feature of topic 49 for

females, that it is a clear spike in the female distribution of published papers.

We confirm this evidence with a complementary Figure 13 representing the dispersion

of published female authored papers across topics, but accounting also for the prevalence

of latent topics. In particular, for each topic we have the proportion of published papers

by female authors (taken from Figure 12) minus the proportion of published papers in this

topic overall. Conditioning on having published a paper, male and female would be equally

21



Figure 12: Empirical distributions across topics between males and females (conditional
of having published an article in Top 5).

22



likely to publish a paper in a specific topic, this difference would be zero. Then, we can

interpret this difference as the excess propensity to publish a paper in a particular topic by

females. These differences can be positive or negative, and the sum over all topics is zero.

The figure shows that there are topics for which the propensity of publishing papers by

females is higher than males, and the opposite. Again topic 49 but also topics 41 (health)

and 30 (applied IO) are in one side. While theory topics as 16 or 37 are in the other side.

In order to analyze the pattern of coauthor-ships we have pooled the articles in three

groups, papers written by male authors, by female authors, and gender mixed team of

authors. The main results are summarises in Figure 14 that shows that there is a important

difference between the pattern of latent topics between sole male teams and sole female

teams, while mixed teams generate an intermediate distribution over the latent topics.

Finally, we want to address a related but different question, how male and female diver-

sify across topics. For example, when writing an article, an author may contribute to a single

latent topic or several, authors that have published several papers may have written similar

articles or they could have been more diverse: are these diversification patterns different

for males and females? For addressing this question, the first step is to choose a measure

of latent topic dispersion/concentration. A natural candidate is the Herfindahl-Hirschman

Index (HHI) that is used to measure the concentration in a market.

The HHI index is calculated by squaring the market share of the firm (the topic) that

compete in a single market and then summing up the resulting numbers HHI =
∑N

i=1 s
2
i .

We apply this index to our problem as follows. For each author (the market), we identify

all the latent topics that she has contributed to (the firms). For each article the algorithm

computes a probability distribution over the latent topics. We repeat the process for all

articles of the same author. Then, the cumulative probability divided by the number of

articles is the contribution of the author to this particular latent topic (the market share,

si). For example, if an author publishes very similar papers related to a single or a few

latent topics, her HHI will be high. On the contrary, authors with a more diverse research

agenda will have a lower HHI. Figure 15 shows the corresponding average HHI for males

and females.

We have computed the HHI controlling for the number of papers by author. It is clear,

23



Figure 13: Relative propensity of publishing papers by females over topics.

24



0 1 2 3 4 5 6 7 8 9

54
24
46
31
35
38
7
14
8
30
39
36
13
23
1
20
12
42
49
41
17
25
4
5
27
26
44
52
10
43
40
45
47
18
3
33
34
15
50
2
53
32
19
51
48
29
6
16
21
22
37
11
28
9

Male Female Mixed

Figure 14: Empirical distributions across topics between males, females and mixed au-
thorship (conditional of having published an article in Top 5).

25



Figure 15: Diversify across latent topics by gender (HHI).

that an author that has published more papers is likely to have contributed to a larger set of

latent topics and therefore she must have a lower HHI. Interestingly, the figure shows some

differences between genders in terms of diversification. Females are more diverse (lower

HHI) when publishing one or two papers, but less (higher HHI) when publishing a larger

number of papers in the Top 5.13

5 Topics as Research Fields

In this section we estimate the stochastic model with a lower number of topics, with two

objectives. On one hand, a low K facilitates the semantic interpretation of topics and then

to analyze, for instance, whether or not, the weight of a particular field in the T5 has

increased over time. On the other hand, a low number of topics will allow us to frame our

results with previous literature that has used a small number of categories linked to JEL

codes and research areas in top departments. After estimating the model for a range of

K ∈ 10, ...., 20, we have found that K = 15 is a number of topics for which the estimated

13The HHI is a first approximation as measure of research diversification. In the future, we want to
improve the measure by taking into in account that some latent topics are close to others.

26



model performs better in terms of fitting to the data and the semantic content of the latent

topics at the same time. The model with K = 15 latent topics is summarized in Figure 16.

technolog

effect

polit

wage

polici

social

agent

price

increas

countri

equilibrium

prefer

asset

product

estim

innov

school

vote

worker

tax

experi

contract

market

household

growth

game

choic

financi

firm

test

invest

treatment

voter

labor

rate

individu

optim

consum

percent

econom

inform

decis

invest

trade

distribut

regul

program

polici

employ

welfar

group

effici

inform

insur

incom

player

util

bank

industri

condit

adopt

student

elect

market

monetari

perform

mechan

competit

health

state

equilibria

risk

risk

import

paramet

institut

test

govern

job

optim

network

alloc

cost

estim

data

payoff

expect

return

cost

method

right

outcom

parti

earn

govern

incent

incent

demand

hous

capit

action

subject

market

data

function

resourc

random

candid

unemploy

inflat

inform

problem

auction

women

across

learn

theori

credit

countri

use

increas

assign

crime

increas

respons

effort

match

good

incom

citi

signal

behavior

rate

sector

variabl

enforc

use

power

skill

shock

manag

condit

profit

children

develop

belief

individu

debt

export

asymptot

16.3%

17.3%

15.4%

18.8%

15.3%

11.3%

16.6%

16.1%

14.3%

12.7%

17.6%

18.2%

23.4%

14%

14.7%

7%

8.4%

7.9%

4.5%

6.9%

7.2%

4.6%

5.3%

10.3%

6.8%

4.6%

6.2%

6.9%

7.9%

5.5%

Topic 4

Topic 11

Topic 7

Topic 8

Topic 15

Topic 12

Topic 10

Topic 5

Topic 13

Topic 1

Topic 6

Topic 14

Topic 3

Topic 2

Topic 9

Topic
Prop.

Female
Prop.

2 4 6 8
Word Prevalence (%)

5 10 15 20

Topic Proportions (%)
(White = median Female Prop.)

Figure 16: Latent topics ranked by prevalence in the corpus with k = 15.

27



Figure 17: A topic with “labor”: topic 8 in the set with K = 15

The reader may then wonder what additional information is contained in the unrestricted

version of the Structural Topic Model (STM). One way to illustrate on the importance of

an adequate selection of the number of topics is to explore in detail the composition effects

we already discussed above. We proceed as follows. First, we consider the stem “labor”,

and we look for it among the fifteen more frequent words within the restricted version of the

STM, that is, the version with just 15 latent topics (K = 15). We only find that particular

word under the required frequency within topic 8 in Figure 16. Figure 17 depicts the word

cloud for that topic 8 in the restricted version of the model with K = 15. Clearly, in this

particular case, one may say this cloud describes well the research field corresponding to

JEL code J, which is, Labor and Demographic Economics.

The key idea with the Structural Topic Model is that a field like, ”Labor”, can fit many

research lines in the unrestricted version of the model, in our case the one with 54 latent

topics. When we look for the stem ”labor” within the 54 latent topics, we find it among

the fifteen more frequent words in as many as six topics. Figure 18 illustrates on the most

prevalent among these topics which are: Labor Search, Labor Supply, Human Capital, or

Productivity Analysis. Notice, in particular, that there are important differences on the

prevalence of females across these different subtopics, from 18 per cent in the more policy

oriented topic which is “labor supply” to 14 per cent in the more theoretical “labor search”

(go back to Figure 7 for these shares). Important variability can be washed out when the

28



(a) Labor Search (14% fm) (b) Labor Supply-Policy (18% fm)

(c) Human capital (14% fm) (d) Productivity (16% fm)

Figure 18: Word clouds for topics with the stem “labor” among the fifteen more frequent
words in the set with K = 54

methodology used account for the research field environment rather than for the research

topic environment.

As we have anticipated, the reduction of the number of topics to K = 15 makes easier to

label the latent topics as meaningful research fields, though. Following our previous analysis,

Figure 19(a) plots the latent topics showing the relative semantic distance between topics

29



as well as their weight in terms of the fraction of documents/abstracts that they contain.

If we compare Figure 7 (with K = 54) and Figure 19(a) (with K = 15), they have a

similar “geography” in terms of general areas of knowledge. Therefore, similar patterns in

terms of the distances between topics arise. For example, “Econometric Theory” seems to

be isolated, whereas applied fields as Labor and Public Economics, are closely connected.

Figure 19(b) (as Figure 8 with K = 54) provides evidence of the “horizontal” differ-

ences between males and females in doing research. The results go in line with the previous

literature as in Dolado et al. (2012), Chari and Goldsmith-Pinkham (2017), Beneito et al.

(2018) and Lundberg and Stearns (2019) that point out that females are unevenly dis-

tributed across fields. We concur with previous literature that females are over-represented

in Applied-Micro fields, specially Health-Gender, Experimental and Education and under-

represented in Econometric and Economic Theory fields, Macro-Monetary and Finance.

For example, Dolado et al. (2012) use the classification of women by research areas

(JEL 20 fields) in the top 50 economic departments in 2005. The proportions they find

are very similar to ours: i) I-Health, Education and Welfare, 25%, ii) D-Microeconomics,

14%; iii) J-Labour and Demographic Economics, 15% or iv) C2-Econometrics, 14.3%. In

our analysis we found that the percentage of female authors are, for example: i) Health

and Gender, 23%; ii) Decision Theory (13.6%), Game Theory (11.4%); iii) Macroeconomics

and Monetary, 14.2%; or iv) Econometrics, 14.4%. Having said that, the distribution of

the proportion of females across these restricted topics seems to be slightly less disperse

than those identified in the previous literature with other sources of data. This can be

due to the fact that our methodology is more “continuous” than allocating females to fixed

categories, and as far as the probabilistic model allocates females’ articles to latent topics

with statistical weights.

Figure 20 analyzes together the evolution of the prevalence of the topics and the pro-

portion of females authors. For building this figure, we have computed the growth rate

of topics’ prevalences and topics’ female proportions from the averages in the latest seven

years (2013-2019) and the first seven years (2002-2008) of the sample. First, we can observe

that the proportion of females have increased in all topics but Finance (−6.6%). Regarding

the prevalence, only four topics have decreased their weight in terms of prevalence, Mecha-

30



(a) Connectedness between topics and the fraction docu-
ments/abstracts in each topic (θd distribution).

(b) Connectedness between topics and the female authors doc-
uments/abstracts in each topic.

Figure 19: Connectedness for K = 15

nism Design (−10.3%) , Econometrics (−29%), Game Theory (−22.5%) and Experimental

(−8.4%). On the one hand, the topics where the percentage of women authors have risen

31



Figure 20: Growth rates of prevalence and female proportion by topics.

more are Political Economy (+67.7%), Decision Theory (+42.5%), Macroeconomics and

Monetary (+32.3%), Experimental (+40%) or Labor (+35%). In all of them the women

were clearly underrepresented. On the other hand, the topics where the percentage of women

has grown the least, besides Finance, have been Health and Gender (+11.4%), Econometrics

(+9.4%), and IO (+9.2%)).

Finally, there is no clear relationship between the growth rate of topic prevalence and the

increase in female prevalence. This is surprising. We do not have data about the seniority of

authors, but as the proportion of female is increasing, we can expect that the proportion of

females among the new entrants in the T5 market should be relatively large. New entrants

should be more likely to work in “hot” topics rather than in declining ones. The combination

of both effects should lead to a positive correlation between the increase in the prevalence of

a topic and the increase in female representation, something that we do not observe clearly

in the data. However, another alternative explanation to the increase of the proportion of

women in some topics is that females that already have published in top five in the past,

have extended their network of male coauthors and getting more papers published.

32



6 Conclusions

Using unsupervised machine learning techniques and a new data base composed by the

abstracts of all articles published in T5 journals in Economics for the period (2002-2019), we

have shown that there are persistent and significant horizontal differences in the way males

and females approach research in Economics. Using the Structural Topic Model we have

identified latent topics for which the distribution of female authors is more uneven than with

research fields. These findings are important for several reasons, because: i) T5 publications

are key for research careers and also for determining the path of economic research; ii) The

results are robust in the sense that they are automatically generated with a probabilistic

model without any deterministic allocation of papers to pre-established categories or fields

of research; iii) Finally, recent theoretical results by Conde-Ruiz et al. (2017, 2021) and

Siniscalchi and Veronesi (2020) show that “horizontal” gender differences in the choice of

research topic may lead to a gender discriminatory trap.

Beyond the scope of the present paper, we plan to extend our analysis in several di-

rections. Firstly, we want to recollect more information about the authors, in order to be

able to capture dynamic effects. For instance, we want to differentiate between the research

patterns by senior and junior authors. We want also to investigate how male and female

build the network of coauthors and how this process determines the choice of latent top-

ics. Secondly, we want to show the usefulness of the methodology and the latent topics we

have identified by reviewing research questions analyzed by previous literature in academic

gender gaps. For example, Hengel (2020) analyzes the differences in quality of writing of

papers. She shows that female-authored manuscripts are better written and concludes that

female are subject to higher writing standards. The reason might be an unwelcome gendered

culture through the entire editorial process at the time of deciphering complicated texts.

We are currently implementing Hengel’s readability scores methodology to the latent topics.

Our preliminary findings suggest that those papers belonging to topics with more prevalence

of females are better written. Although, this evidence can be interpreted as supporting the

view that female-authored articles are better written than equivalent articles by men. It

can be also the case that the results are driven by the particular topics. In other words,

we need a deeper econometric analysis to disentangle if the written quality of the papers is

33



driven by gender of the author or by the choice of the latent topics.

Likewise, Card et al. (2019) shows that female authored papers have more citations, sug-

gesting that journals hold female-authored papers to higher standards. They have obtained

this result controlling for research field. We plan to collect data on citations and review this

result but controlling by latent topic. Finally, we want also to use algorithms (for example,

LASSO a widely used regression analysis machine learning method) for testing if the dif-

ferences between gender research patterns are important enough, for building a predictive

model of gender given an observed abstract.

34



References

Bagues, Manuel and Pamela Campa, “Can Gender Quotas in Candidate Lists Empower

Women? Evidence from a Regression Discontinuity Design,” 2017, (12149).

Bayer, Amanda and Cecilia E. Rouse, “Diversity in the Economics Profession: A New

Attack on an Old Problem,” Journal of Economic Perspectives, Nov. 2016, 30 (4), 221–42.

Beneito, P., J. E. Boscá, J. Ferri, and M. Garćıa, “Women across Subfields in

Economics: Relative Performance and Beliefs,” Fedea WP, June 2018, (2018 - 06).

Blei, David M., Andrew Y. Ng, and Michael I. Jordan, “Latent Dirichlet Allocation,”

J. Mach. Learn. Res., March 2003, 3 (null), 993 – 1022.

Boustan, Leah and Andrew Langan, “Variation in Women’s Success across PhD Pro-

grams in Economics,” Journal of Economic Perspectives, February 2019, 33 (1), 23–42.

Buckley, Chris, “Implementation of the SMART Information Retrieval System,” Technical

Report, USA 1985.

Cabrales, A., M. Garćıa, and L. A. Puch, “Gendered Language in the British Press,”

Mimeo COSME Gender, at 2018 Meetings of the Spanish Economic Association, 2018.

Card, David and Stefano DellaVigna, “Nine Facts about Top Journals in Economics,”

Journal of Economic Literature, March 2013, 51 (1), 144–61.

, , Patricia Funk, and Nagore Iriberri, “Are Referees and Editors in Economics

Gender Neutral?*,” The Quarterly Journal of Economics, 11 2019, 135 (1), 269–327.

Chari, Anusha and Paul Goldsmith-Pinkham, “Gender Representation in Economics

Across Topics and Time: Evidence from the NBER Summer Institute,” Working Paper

23953, National Bureau of Economic Research October 2017.

Chevalier, Judy, “The 2020 Report of the Committee on the Status of Women in the

Economics Profession,” 2020.

35



Conde-Ruiz, J. Ignacio, Juan-José Ganuza, and Paola Profeta, “Statistical Dis-

crimination and the Efficiency of Quotas,” Fedea Working Papers, 2017.

, Juan José Ganuza, and Paola Profeta, “Statistical Discrimination and Commit-

tees,” Fedea Working Papers, February 2021, (2021-06).

Dolado, Juan, Florentino Felgueroso, and Miguel Almunia, “Are men and women-

economists evenly distributed across research fields? Some new empirical evidence,” SE-

RIEs: Journal of the Spanish Economic Association, September 2012, 3 (3), 367–393.

Hansen, Stephen, Michael McMahon, and Andrea Prat, “Transparency and De-

liberation Within the FOMC: A Computational Linguistics Approach,” The Quarterly

Journal of Economics, 10 2017, 133 (2), 801–870.

Heckman, James J. and Sidharth Moktan, “Publishing and Promotion in Economics:

The Tyranny of the Top Five,” Journal of Economic Literature, June 2020, 58 (2), 419–70.

Hengel, E., “Publishing while Female. Are women held to higher standards? Evidence

from peer review,” Cambridge Working Papers in Economics 1753, Faculty of Economics,

University of Cambridge December 2020.

Hengel, Erin and Eunyoung Moon, “Gender and quality at top economics journals,”

Working Papers 202001, University of Liverpool, Dept. of Economics February 2020.

Lundberg, Shelly and Jenna Stearns, “Women in Economics: Stalled Progress,” Jour-

nal of Economic Perspectives, February 2019, 33 (1), 3–22.

Mimno, David, Hanna Wallach, Edmund Talley, Miriam Leenders, and Andrew

McCallum, “Optimizing Semantic Coherence in Topic Models,” 2011, pp. 262 – 272.

Roberts, Margaret E., Brandon M. Stewart, and Dustin Tingley, “stm: An R

Package for Structural Topic Models,” Journal of Statistical Software, 2019, 91 (2), 1–40.

, , , Christopher Lucas, Jetson Leder-Luis, Shana Kushner Gadarian,

Bethany Albertson, and David G. Rand, “Structural Topic Models for Open-Ended

Survey Responses,” American Journal of Political Science, 2014, 58 (4), 1064–1082.

36



Siniscalchi, Marciano and Pietro Veronesi, “Self-image Bias and Lost Talent,” De-

cember 2020, (28308).

Tang, Cong, Keith Ross, Nitesh Saxena, and Ruichuan Chen, “What’s in a Name:

A Study of Names, Gender Inference and Gender Behavior in Facebook,” in “Xu J., Yu

G., Zhou S., Unland R. (eds) Database Systems for Advanced Applications Lecture Notes

in Computer Science, vol 6637,” Springer Berlin Heidelberg, 2011, pp. 344 – 356.

37



Appendix A The topic Model

We implement and develop the Structural Topic Model (STM) to incorporate document-

level meta-data into a probabilistic text model. The topic model is said to be structural

because “covariates” inform about structure (partial pooling of parameters). We keep track

of journal names and publication years as covariates to estimate the prevalence of topics.

The starting point to understand the STM probabilistic model is the LDA (Latent

Dirichlet Allocation) generative model. According to LDA, the Data Generating Process

for document d ∈ D assigns terms in vocabulary V to positions Nd in the document-term

matrix, where the element (d, v) of the matrix is the number of times the vth unique word

appears in the dth abstract. The algorithm follows the steps below

1. Draw a K-dim Dirichlet vector θd containing the expected fraction of words in d

attributed to topic k ∈ K.

2. For each word (position) in d, sample the indicator zd,n from MultK(θd, 1) that indi-

cates the position n associated to a topic.

3. Sample the indicator wd,n from MultV (Bzd,n , 1), where matrix B has distributions βk

over vocabulary V; [βk] is frequency with which terms are generated from k.

STM in its turn builds upon identifying covariates to improve the estimation of the

topics. Covariates affect i) the proportion of a d devoted to a k (topic prevalence-TP), and

ii) how much a word is used in k (topical content-TC). To this purpose:

• for TP, Dirichlet θd draws of document-level attention to each topic are replaced with a

logistic-normal with a mean vector parameterized as a function of document covariates.

• for TC, βk distribution is proportional to a Multinomial logistic regression parameter-

ized as indicated below.

A (partially collapsed) variational expectation-maximization algorithm is implemented

to approximate the posterior (inference). Then posterior predictive checks [cf. Gelman et

al., 1996] and tools for model selection as in Roberts et al. (2014) are used. Beyond TP and

TC functions of document metadata, the structural topic model can be summarized as:

38



1. Given parameters: i) a variance-covariance matrix for topics Σ, ii) a matrix of ob-

served document-level covariates X (journals names and years), and iii) a vector γk

(of prevalence of each topic) for each covariate,

γk ∼ N (0, σ2
k Ip),

sample the topic proportion in each document, vector θd, that is,

θd ∼ LogisticNormalK−1(Γ
′ x′d,Σ), Γ = [γ1|...|γK ]

as a substitute for the Dirichlet conjugate prior, to conform the topic prevalence

model.

2. The core language model given the topic proportion per document θd consists of:

• sampling the probability zd,n that a word is in a topic: zd,n ∼ MNK(Θd), with

K outcomes

• conditional on topic, choose a word from βzd,n , that is wd,n ∼MNV (βzd,n), overB =

[β1|...|βK ] matrix of distributions over vocabulary V.

3. The topical content model samples the topic word distribution βd,k,v,. By now we

do not use covariates to explain topical content of documents.

39



Appendix B Details of this Pre-processing Data

Pre-processing of the abstracts that conform our database is essential in order to organize

the words that form the texts in an homogeneous way. The main goal of this process is

to reduce the dimensionality by reducing the set of words, but at the same time trying to

maximize the information contained in the words used by the authors by selecting the terms

with more informational content. This helps us for a better estimation of more semantically

meaningful topics.

First step is tokenization so as to differentiate words by selecting only single words

(monograms), instead of bigrams, trigrams, paragraphs, etc. Then we eliminate punctu-

ation, and capital letters are converted to small letters. This allows as to remove dupli-

cates, for example ”Education” and ”education” are different words in our database if we

don’t convert all the words to lowercase. Once this is done we eliminate numbers and

stopwords. By stopwords we refer to those words without any informational content: ”com-

mon’ words such as ”and”, ”for”, ”in”, etc. We removed the stop words from the list

SMART developed by Buckley (1985), a public list with more than 500 words. Addition-

ally, we remove some custom stopwords that were very common in our database but not in-

formationally relevant. These are: ‘download’,‘slides’,‘slide’,‘jel’,‘abstract’,‘paper’,‘author’,

‘literature’, ‘among’, ‘whether’,‘authors’, ‘model’, ‘show’, ‘showed’, ‘shows’, ‘find’, ‘can’,

‘matter’,‘model’, ‘models’, ‘may’, ‘effect’, ‘find’, ‘can’, ‘show’, ‘paper’, ‘also’, ‘provide’, ‘ap-

proach’, ‘thus’, ‘main’, ‘obtain’,‘obtained’, ‘without’, ‘modelling’, ‘modeling’, ‘modeled’,

‘modelled’, ‘use’, ‘result’, ‘results’, ‘resulting’, ‘resulted’, ‘discuss’, ‘discussed’, ‘discussing’,

‘recent’, ‘recently’,‘give’, ‘gives’, ‘given’, ‘review’, ‘reviewing’, ‘reviews’,‘require’, ‘required’.

We end by stemming the tokens so as to retain only the roots of words in the same

family,so as to unify the information contained in related words. For example “education”,

“educative”, and “educated”, are all related with education, so we just keep the root “educ”

for all of them. The use of these stems relax dimensionality problems, and groups all

probabilities for families of words into one.

In our sample were initially 13,835 different terms. After this process without loss of

generality, we reduce the number of unique terms to 4,241 in the corpora with which we

build the document term matrix.

40



Appendix C The optimal number of topics

To run the model involves a choice of hyperparameters as discussed in Apendix A above,

and one of those parameters is the number of this latent topics existing in our corpus. As

this can be interpreted as an arbitrary prior, we run some automatic tests in order to choose

this optimal K without human intervention, in order to classify texts in the best possible

way. This approach gives us the advantage of automatically selecting the number of topics

that better fits data. Arbitrary choosing too few topics means to cluster several topics into

a single one. Choosing too many topics means would tend to identify patterns in language

rather than topics.

−6.84

−6.80

−6.76

25 50 75 100
Number of Topics

H
el

d−
ou

t l
ik

el
ih

oo
d 

es
tim

at
io

n

Figure A. 1: Held-out likelihood estimation

We learn a lot on the different patterns of the data when choosing various alternatives for

a fixed number of topics, as we will discuss below. However, our primary selection strategy

for automatic selection focuses on the held-out likelihood estimated. Figure A.1 reports the

log-likelihood of the model evaluated at the estimated parameters on the test set for each

K between 15 and 100. The likelihood is maximized between 49 and 54 topics.

Figure A.2, in its turn displays the number of iterations to convergence of the model,

41



20

40

60

80

25 50 75 100
Number of Topics

Ite
ra

tio
ns

 to
 c

on
ve

rg
en

ce

Figure A. 2: Number of iterations to convergence of the model

which sharply drops at 54 topics and remains at that number of iterations (except for a

small spike at 60) beyond 62 topics.

Finally, Figure A.3 reports the semantic coherence which is decreasing and stable after

59 topics. Semantic coherence is maximized when the more frequent words in a given topic

co-occur together Mimno et al. (2011). High semantic coherence is reached when in the end

there is less topics dominated each by few words. On the other hand, average exclusivity

is large when a particular word frequency corresponds to each topic. We follow Roberts et

al. (2014) to use the FREX metric for this criteria. As showed in Figure A.4 there are two

maximums in 51 and 54 topics.

With our data, we found reasonable to assume that the result is in the neighborhood of

52 topics given the held-likelihood procedure, and given the additional tests, we select the

highest number of topics in this neighborhood, corresponding to 54 topics.

42



−120

−115

−110

−105

−100

25 50 75 100
Number of Topics

S
em

an
tic

 C
oh

er
en

ce

Figure A. 3: Semantic Coherence

9.75

9.80

9.85

25 50 75 100
Number of Topics

E
xc

lu
si

vi
ty

Figure A. 4: Exclusivity

43



Appendix D The topics profile

Given that we have chosen automatically the number of latent topics, it can be helpful

to try to disentangle their nature. As an alternative to Figures 7 and 8, we use Simple

Correspondence Analysis to measure the distance between topics. This is a descriptive

technique to explore relationships among categorical variables. In our application we use

the matrix of probabilities (the matrix θd obtained from STM) for each and every document

to belong to any particular built-in topic in order to measure the distance between topics.

The rows in this matrix are probabilities that add up to one. The clustering of rows measures

the distance between topics (the columns of the matrix). This is the so-called chi-square

distance:

θcolij =
r∑

i=1

(pai − paj)2 ,

where r is the total number of rows, and the measure we compute and represent gives the

euclidean distance between columns i, j (col), for each and every row a (abstract).

Figure A.5(a) depicts the two larger coordinates of the distance matrix computed through

Classical Multidimensional Scaling (MDS), so as to obtain the coordinates of the column

category. The coordinates are given by the order of largest-to-smallest variance. We find

the corpus organized along two dimensions: Dimension 1 can be interpreted as going from

Applied to Theory, whereas Dimension 2 goes from, say, Economics to Econometrics. We

think this is apparent from casual inspection of Figure A.5(a),which involves square distances

between [−4,+4].

Clearly though, outliers (understood as the topics far away from the origin) are very

important in this representation. First, we identify outliers 21, 9,11, that we have associated

to Econometric Theory in the fields of estimation (“estim”, “asymptot”,.... are the keywords

in this case) and testing (“test”, “asymptot”,...), together with structural econometrics

(“identifi”, “instrument”,...) respectively. These actually are are among the top 10 more

prevalent topics. Moreover, topics 9 and 11 are 2nd and 3rd most prevalent. These outliers

are located North East in the diagram in terms of the language they use.

The second set of outliers are located South East and are equally far from the center,

while not isolated. These topics can be associated to Economic Theory texts. On top of

44



(a) Whole Sample

(b) Zoom-in Sample

Figure A. 5: Larger coordinates of the distance matrix computed through Classical Mul-
tidimensional Scaling (MDS) 45



those we find topic 5, and then not that further away from the center, topic 6, 16 and

10. These are, respectively, auction theory (auction, bid,...), together with game (game,

player,...) and information theory (belief, signal,...), as well as mechanism design (mechan,

implement,...). These topics are relatively less prevalent in the sample than the Econometric

Theory topics above as we discussed in the main text.

Finally, there are some outliers at the North West corner of the diagram. We find

here topics that seems to be mostly empirically oriented (applied), and according to our

representation, nearly as distant from Econometric than from Economic Theory. These are

particularly topics 19 and 49, that we have associated before with Education and Gender

issues, and for which female authors’ presence is relatively more prevalent.

There is finally a negative correlation between the two coordinates, suggesting that

distance values are larger than under the hypothesis of independence between these two

key dimensions. This finding would require a treatment that goes beyond the scope in this

paper. We leave further analysis of the nature of latent topics in leading economic journals

for future research. The interested reader can check the center of the representations at

square distances between [−1,+1] in Figure A.5(b).

46



Appendix E Analysis with the abstracts of the Papers Proceed-

ing Papers (P&P)

In this section, we extend our original sample with the Papers and Proceedings (P&P)

articles published in AER in the especial issue of May during the period 2011-2018.14 These

P&P articles are very short (for example, they could be just an extension of a full article

submitted to a different journal) and they are selected from the papers presented in the

annual January meeting of the American Economic Association’s (AEA). Part of the papers

are selected directly for the committee’s members of the AEA meetings and others are chosen

from external proposals of special sessions in AEA meetings.15 Interestingly for our analysis,

papers in P&P are linked to the meeting sessions, and then, they come in groups of 3 or

4 papers of a specific topic. Then, the editorial process of this P&P is very different from

regular submissions and the set of topics is likely to be more diverse, since some of the special

sessions in AEA meeting may be relevant for current policy debate but not necessarily for

research. For example, in the issue of May 2020, among others, we can find two sessions

and the corresponding articles over ”The economics of the health epidemics” or ”Is United

States deficit policy playing with fire?”.

With these additional P&P papers, our sample contains 6,428 abstracts/documents, that

generates 253,312 tokens and 12,936 unique terms. The number of topics that best fits the

these extended sample is 70. The larger number of latent topics can be related to the larger

number of unique words and documents, but also to the selection process of P&P described

above, sessions unrelated to standard research with a small number of (”seed”) papers very

related among themselves.

As in the main text, we estimate these 70 latent topics using the STM algorithms. Figure

A.6 presents the latent topic ranked by prevalence in the corpus with k = 70.

Figure A.7 show the STM output (the estimated latent topics) and also how the docu-

ments are allocated among them.

As in the main text, in the Figure A.7 the size of the circle is proportional to the number

14Before 2011 the P&P articles did not have abstract and after 2018 the P&P articles are included in a
different journal.

15For more information about the about the AEA Papers and Proceedings go to:
https://www.aeaweb.org/journals/pandp/about-pandp

47



of documents in the topic. The most salient feature of the Figure A.7 is that in addition

to the larger number of topics, there are some of them with very small size that could be

related to the ”seeds” described above, sessions of the AEA meetings, with very related

papers among themselves but quite different to research papers closer to them.

Figure A.8 reinforce the evidence of the main message of this paper, male and female

display different pattern when doing research. There is a subset of topics (South-East in

the figure A.8) with a relative high proportion of females, that moreover seems to be closely

connected. On the contrary, there is other set of topic (South-West in the Figure A.8) that

is is also closely connected and where the present of females is relatively scarce.

Now, we want to look closer the content of some particular topics. In this larger sample,

it is easier to see that the latent topics go beyond standard research fields. In particular,

Figure A.9 points out that the latent topics with higher proportions of female authors are

topic 41 and topic 19. In the following figure we can see the distributions over terms that

each of this two topic induces are represented as words clouds, where the size of term in the

cloud is approximately proportional to its probability in the latent topic distribution βk.

Clearly, topic 41 in related with family economics and topic 19 with gender discrimination.

48



use
law

select
mobil

conflict
includ

exchang
process
crime
news
distort

program
communic

bargain
right

student
work
save

technolog
vote
innov

women
tax

bank
regul
credit
decis

search
distribut
match
group

percent
auction
educ
state
hous
health
price
capit
choic
polit

asset
behavior
econom
predict
social
school
employ
experi

risk
region
invest
polici

measur
welfar
game
incom
market
dynam

util
inform

equilibria
shock
test

price
trade
agent
firm

variabl
estim

develop
promot
advers
exclus

bad
mani
organ
mean
black
media
capit

particip
receiv
commit
integr
colleg
time
citi

manag
voter

patent
gender
margin
fund
cost

borrow
rule

unemploy
two

network
individu

year
bid

children
unit
area
insur
adjust
skill

prefer
govern
investor
prefer
empir
theori
public
effect

worker
subject
return

econom
financi
debt
differ
cost

player
household
competit

equilibrium
prefer
belief
exist

aggreg
asymptot
consum
countri
contract
product
function
method

effect
minimum

associ
user
self

sever
coordin
converg

white
advertis
optim
target
inform
delay

ownership
univers
suppli
age

adopt
bias

research
gap
rate

financi
emiss
default
make
job
first
stabl

treatment
increas
bidder
parent
resourc

local
care

chang
human
altern
parti

market
concern
research

data
privat

student
wage

experiment
avers
ethnic
growth
govern
countri
benefit

equilibrium
consumpt

effici
economi
expect
learn
condit
cycl

statist
demand
intern
optim

industri
condit
data

compar
polici

distinguish
platform

pool
wide
rate

frequenc
racial

violenc
tax

evalu
disclosur

option
contract

educ
hour
popul

perform
candid
spillov
femal
incom
crisi

electr
loan

individu
wage

second
interact

intervent
point

mechan
famili

centuri
increas
hospit
real

wage
set

polici
stock
differ

question
empir
enforc
score
labor
field
rate

agricultur
capit

monetari
across
insur
payoff
wealth

equilibrium
polici

ambigu
signal

set
busi

paramet
good
sector
incent
produc
identif
sampl

part
behavior

favor
farm

strateg
rang

donor
long

crimin
effect

economi
subsidi
sender
offer
valu

minor
labor
retir

increas
elect

knowledg
men
optim
run

environment
secur

member
friction
order
agent
expect

percentag
valu
child

protect
effect

increas
rate

occup
choos
power
trade
fair
use
test

good
test
job

treatment
discount
migrat
financ
rate

account
gain

strategi
inequ
ineffici
general

uncertainti
agent
class
volatil
confid
sale

import
mechan

cost
identifi

use

substanti
water

surpris
entri
good
varieti

transplant
revers
collus

newspap
redistribut

impact
report

agreement
vertic

graduat
transfer

rate
practic
polici
new
male
elast
lend

energi
debt

maker
labour
size

structur
mortal
rate
price

marriag
war

counti
patient

cost
inequ
ration
institut
liquid

explain
replic

forecast
provis

outcom
earn

report
equiti
histor
flow

interest
per

polici
action
data
trade

determin
function
privat
finit

fluctuat
interv

purchas
export

problem
entri

restrict
consist

place
evid

probabl
contract
success

set
group
time
polic

increas
progress

effect
strateg
outsid

acquisit
cours

increas
growth

new
prefer
growth
marri
taxat
liquid
climat

constraint
collect
worker
third

substitut
rate

immigr
valuat
fertil

capac
impact
spend
inflat

worker
attent

democraci
price
peopl

correct
consist
ident

assign
employe
random
premium
institut

economi
fiscal
correl
reform
play

expenditur
alloc

stochast
probabl
observ

equilibrium
labor

distribut
profit

foreign
princip
qualiti

instrument
bias

empir
principl

new
incumb

will
hold
effici
jump
rate

content
polici
train

messag
power

properti
enrol
home

increas
chang
one
data
like

effect
system
pollut
rate
bias

market
three
sort

control
estim

revenu
birth

system
neighborhood

drug
nomin
increas
maxim

politician
portfolio

other
import
time

contribut
teacher
increas
design
time

conflict
constraint

spend
develop
reduc
cooper
earn
seller
condit

set
common
continu

macroeconom
procedur
increas

tariff
implement

market
bound

paramet

allow
chang
entir
data

control
fact

kidney
observ
race

internet
incom
elig
truth

negoti
supplier
major
forc
plan

improv
major
effect
less

increas
reserv
estim

mortgag
made
durat
joint

heterogen
health
averag
buyer
invest
world

exposur
medic
data

chang
reveal
legisl

money
individu

data
theoret
norm
high
data
studi
high
data

recess
lower
relat
loss

strateg
use

buyer
monetari
character

ration
general
respons
method

retail
domest

alloc
input

structur
error

new
effect
sourc

increas
leader
analysi
donat

depend
arrest
evid
taxat
applic

set
compens
industri
admiss
week

individu
farmer

campaign
technolog

differ
chang
deposit
carbon
interest

case
match
simpl

complementar
life

annual
privat
health
earli
estim

medicar
index

accumul
random
elector
bond
theori

analysi
implic
role

achiev
industri
effect

intertempor
popul
crise
effect
data

increas
nash
share

outcom
endogen
represent

prior
comput
output
propos
quantiti
factor
design
higher

unobserv
number

studi
reduc
kind
deal
fear

describ
increas
memori

judg
estim

govern
disabl

lie
period

sharehold
perform
suggest

larg
advanc
elector
impact

discrimin
estat
feder
polici
card

status
rate
tail

form
effect

decreas
seller

mother
econom

locat
qualiti
relat

return
problem
support

hold
relat
name

inconsist
cultur
impact

hire
evid
riski

effect
declin
inflat
indic

improv
repeat
increas
qualiti

framework
subject
expect
time

suppli
bootstrap

durabl
global
effort
plant

nonparametr
panel

base
prohibit

evid
use
peac

particular
central

rate
defend
suggest
labour
receiv

manipul
fee

profit
rank
leav

urban
manageri

win
product

evid
use

market
gas

market
propos

equilibrium
domin
type

random
use

winner
intergener
american
segreg
coverag
respons
growth
plan
elect
valu
refer

statist
develop
institut
peer

sector
laboratori

stock
variat
friction

sovereign
signific
provid

outcom
survey
reput
chang

individu
type
valu

idiosyncrat
infer
effect

develop
project

heterogen
treatment
approxim

incorpor
data

potenti
profit
group
import
decentr
constitut

data
polit
studi
aid

disclos
probabl
structur

onlin
respons
german
diffus

strateg
inventor
suggest
respons

effect
chang
increas
influenc
vacanc
upper
link

increas
data
post
age

europ
pollut
estim

markup
abil

inattent
public

fundament
suggest

economist
use

motiv
attend
transit
differ
term

growth
develop

zero
variat

analysi
type
top

incent
regim

properti
structur
game
persist
base

margin
good
studi

export
use

base

document
new
isol
time
lead

perspect
number
propens
sentenc

time
pareto
random

cost
will

acquir
like

cash
germani
frontier
extrem

use
posit

formula
balanc

fuel
home
differ
offer

establish
observ
incent
period
design
young

problem
live

premium
evid
differ

consider
corrupt
volatil
like

studi
futur
level
use

manufactur
conduct
higher
religi
net

central
poor

potenti
deviat

individu
studi

fundament
avers

asymmetr
perfect

uncertainti
sampl

surplus
relat
moral
size

assumpt
asymptot

consist
increas

advantag
store
even

multipl
larg
base
applic
elect

generat
well

condit
accept
land
abil

declin
account
technic
expert
activ
rate
evas
failur
plant

lender
major
effect

present
central
peopl
reduc

equilibrium
attain
chang
data
plan
larg
earn
best
alloc
trader
endow
theoret

new
monitor
random

loss
natur

expect
exploit
boom

commit
gdp
use

equilibria
chang
trader
futur
lotteri
precis

theorem
elast
valid
lower

special
hazard
increas

set
factor

28.6%

15.9%

17.6%

17.8%

16.8%

17%

13.8%

12.5%

16.7%

13.4%

10%

18%

12.7%

16.7%

12.7%

17.4%

14%

16.2%

42.2%

15.2%

15.4%

15.7%

18.7%

14.7%

18.1%

12.6%

17%

15%

21.4%

15.8%

29.4%

21.5%

21.3%

15.3%

17.8%

19.3%

19.2%

15.1%

18.8%

22.5%

31.7%

17.5%

19.9%

21.5%

14%

11%

24.2%

21.4%

15.4%

16.4%

14.7%

16.5%

14.5%

16.8%

14.9%

20%

18.8%

16%

25.7%

25.9%

18.8%

15.8%

21%

21.2%

16.4%

17.6%

18.2%

17.3%

16.1%

17.9%

0.8%

1.3%

0.6%

1.8%

1.4%

1.2%

0.8%

0.6%

1.7%

0.9%

2.1%

1.2%

2.1%

1.3%

2.6%

1.3%

1.3%

1.6%

1.1%

1.5%

2.2%

2.8%

0.8%

2.1%

2.7%

1.8%

0.7%

1.1%

1.7%

1.1%

1.3%

1.1%

1%

1%

1.1%

1.6%

1.7%

1.2%

1.8%

1.3%

1.3%

2.3%

1%

1%

0.8%

2.1%

1%

1.2%

2.2%

1.6%

0.8%

1.8%

1.1%

2.7%

1.5%

1.7%

0.6%

2%

0.9%

1.4%

1.9%

1.5%

0.5%

1.3%

1.8%

1.4%

1.7%

1.6%

2.1%

0.3%Topic 70
Topic 63
Topic 3

Topic 57
Topic 8

Topic 27
Topic 51
Topic 7
Topic 1

Topic 23
Topic 45
Topic 59
Topic 10
Topic 34
Topic 44
Topic 47
Topic 33
Topic 43
Topic 32
Topic 53
Topic 35
Topic 19
Topic 30
Topic 28
Topic 48
Topic 12
Topic 6

Topic 38
Topic 2

Topic 16
Topic 31
Topic 40
Topic 17
Topic 41
Topic 14
Topic 64
Topic 60
Topic 5

Topic 66
Topic 20
Topic 62
Topic 55
Topic 68
Topic 18
Topic 50
Topic 36
Topic 29
Topic 67
Topic 37
Topic 9

Topic 56
Topic 52
Topic 65
Topic 4

Topic 39
Topic 26
Topic 61
Topic 58
Topic 13
Topic 46
Topic 24
Topic 11
Topic 69
Topic 21
Topic 49
Topic 42
Topic 15
Topic 25
Topic 54
Topic 22

Topic
Prop.

Female
Prop.

10 20 30
Word Prevalence (%)

10 20 30 40

Topic Proportions (%)
(White = median Female Prop.)

Figure A. 6: Latent topics ranked by prevalence in the corpus with k = 70. Extended
sample with P&P articles.

49



Figure A. 7: Connectedness between topics and the fraction documents/abstracts in each
topic (θd distribution). Extended sample with P&P articles.

50



Figure A. 8: Connectedness between topics and the female authors documents/abstracts
in each topic. Extended sample with P&P articles.

51



(a) Topic 41

(b) Topic 19

Figure A. 9: Topic Word Clouds in the extended sample with P&P articles

52


	ICAE_WP_CGGP.pdf
	Introduction
	Raw Data and Descriptive Analysis
	The Empirical Model: Structural Topic Model (STM)
	Gender Differences in Latent Estimated Topics
	Topic Prevalence
	Topic analysis and the gender distribution

	Topics as Research Fields
	Conclusions
	References
	The topic Model
	Details of this Pre-processing Data
	The optimal number of topics
	The topics profile
	Analysis with the abstracts of the Papers Proceeding Papers (P&P)