Citation Gender Gaps in Top Economics Journals

J. IGNACIO CONDE-RUIZ

MIGUEL DÍAZ SALAZAR

JUAN JOSÉ GANUZA

MANU GARCÍA

Documento de Trabajo 2025/07

Junio de 2025

fedea

Las opiniones recogidas en este documento son las de sus autores y no coinciden necesariamente con las de Fedea.

J.Ignacio Conde-Ruiz,^a,c Miguel Díaz Salazar,^a Juan-José Ganuza,^b and Manu García^d†

^aFedea

^bUniversitat Pompeu Fabra and Barcelona GSE ^cUniversidad Complutense de Madrid and ICAE ^dFederal Reserve Bank in St. Louis

May 2025

Abstract

This paper investigates the existence and drivers of gender citation gaps in the five leading journals in economics. Using a comprehensive dataset of 7,244 articles published between 1999 and 2023, we examine whether female-authored papers are cited more frequently than male-authored ones, and whether this pattern persists after controlling for diferences in research topics. We apply Structural Topic Modeling (STM) to abstracts to estimate latent research themes and complement this approach with field classifications based on JEL codes. Our results show that female-authored papers initially display a citation premium—receiving up to 16 log points more citations—but this advantage becomes statistically insignificant once we control for research field composition using either STM topics or JEL codes. These findings suggest that horizontal gender diferences in thematic specialization, rather than bias in citation behavior, account for most of the observed citation gap. Our analysis highlights the importance of accounting for field heterogeneity when assessing academic recognition and contributes to ongoing discussions about fairness and diversity in economics publishing.

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

JEL Classification: I20, J16, Z13.

^∗Thanks to Christian Zimmermann for helpful comments. José Ignacio Conde-Ruiz acknowledges the support of the Research Project of the Ministry of Science and Innovation, PID2023-148090NB-I00. Juan José Ganuza acknowledges the support of the Barcelona School of Economics and the Research Project of the Ministry of Science and Innovation PID2023-153318NB-I00 and from the Spanish Agencia Estatal de Investigación (AEI), through the Severo Ochoa Programme for Centres of Excellence in R&D (Barcelona School of Economics CEX2024-001476-S). The opinions and analyses are the responsibility of the authors and do not necessarily reflect those of the Federal Reserve Bank of St. Louis or the Federal Reserve System.

^†Corresponding Author: Juan-Jose Ganuza, Universitat Pompeu Fabra, Ramon Trias Fargas 27, 08005, Spain; E-mail: juanjo.ganuza@gmail.com

1 Introduction

During the past few decades, significant eforts have been made to improve gender representation in various fields, particularly in academia. However, despite these eforts, women remain underrepresented at the highest levels of the profession, including in prestigious economic journals and faculty positions. Gender disparities in academic publishing and career progression continue to be a pressing concern, especially in fields like economics, where such gaps can have long-lasting impacts on professional advancement and recognition.

In academia, promotions and career progression are heavily influenced by publications in the most prestigious journals, often referred to as the “Top 5" (American Economic Review, Quarterly Journal of Economics, Journal of Political Economy, Econometrica, and Review of Economic Studies). These journals set the benchmark for academic excellence and play a pivotal role in shaping research careers. A strong publication record in these journals is frequently regarded as a prerequisite for tenure, promotions, and broader academic recognition, making it crucial to examine gender disparities in this domain. Women are also underrepresented in Top 5 publications. In our sample, they are around 15% of the authors.

This contrasts with the findings of several academic studies that have documented persistent gender gaps in citations in economics. For instance, Card et al. (2020) found that women-authored papers in top economics journals receive more citations than men-authored papers with comparable referee scores, suggesting that women face higher publication standards. Similarly, Hengel and Moon (2023) extends this analysis, demonstrating that femaleauthored papers are cited 12 log points more than male-authored ones, with the citation premium rising to 20 log points when adjusting for the “ Matthew efect "¹. Other studies, such as Kofi (2021) and Ductor and Prummer (2024), highlight the importance of collaboration networks and research focus, showing that while women-authored papers receive a citation premium overall, this dynamic varies across subfields and journal prestige.

Taking citations as a proxy for quality, the gender citation gap raises the important question of whether women are held to higher standards in academic publishing. If femaleauthored papers face stricter thresholds for acceptance, those published in the Top 5 journals should, on average, exhibit higher quality (more citations). However, citations are also influenced by the research agenda, as some fields attract more citations than others. Therefore, the key question is not simply whether female-authored papers are cited more, but whether they receive more citations after controlling for diferences in research areas.

¹The “Matthew efect"—a term coined by sociologist Robert K. Merton—refers to the phenomenon whereby well-known researchers or highly visible papers tend to accumulate more citations simply due to their existing prominence. In academic publishing, this creates a cumulative advantage, where recognition reinforces itself regardless of intrinsic quality (see (Merton, 1968)).

In this paper, we investigate whether gender disparities exist in citation patterns within the Top 5 economics journals. Using an extended dataset covering articles published from 1999 to 2023, we analyze whether female-authored papers are cited more frequently than male-authored ones, after controlling for research topics. We apply the same methodology introduced by Conde-Ruiz et al. (2022a), employing a Structural Topic Model (STM) to identify and control for latent research topics, allowing us to account for thematic diferences across papers. Consistent with Conde-Ruiz et al. (2022a), we show that men and women exhibit diferent patterns when choosing research topics in economics. To measure citations, we complement the dataset with citation information from RePEc (Research Papers in Economics), a comprehensive database that tracks bibliographic data and citations for economics research. This addition enables us to systematically evaluate the impact of gender on citation patterns. Our results show that when controlling for research topics, the citation gaps between male- and female-authored papers narrow significantly, often becoming statistically insignificant. This indicates that the perceived citation disparity is closely tied to the thematic content and distribution of topics among authors, rather than systemic bias in how citations are allocated.

Similarly to us, Kofi (2021), Card et al. (2020), and Ductor and Prummer (2024), have also shown that the gender citation premium narrows when controlling for the JEL codes of the published articles. The Journal of Economic Literature (JEL) classification system has been widely used in the literature to capture field-level heterogeneity in the published papers to analyze the trends of the economic research², gender heterogeneity regarding the research fields,³ and also, as we have seen, the gender citation premium. We complement our analysis by replicating our machine learning exercise using JEL codes, and we will discuss the relationship between the two methodologies. For doing so, we enriched our dataset by incorporating the Journal of Economic Literature (JEL) codes assigned to each published article. These codes are taken from the metadata available in the RePEc database. When JEL codes were not listed in the published version, we retrieved them from the corresponding working paper version, if available. This step allowed us to systematically map each paper to one or more JEL categories and to compare these author-assigned classifications with the latent topics extracted through text analysis. We document persistent horizontal gender diferences across primary JEL categories, confirming that men and women tend to specialize in diferent research fields. This horizontal diferences seem to be very aligned with ones obtained with our STM approach ⁴. In fact, we show that there is some relationship between estimated research topics and JEL codes, as well as in the allocation of papers across estimated research topics and JEL codes. By linking the two methodologies, we highlight their complementarities, and provide a more comprehensive characterization of research specialization across gender. After undertaking this complementarity analysis, we find that both methodologies have their advantages and limitations, but perform similarly well in addressing our research questions.

²For example, Angrist et al. (2017) explore the classification of economics research by combining JEL codes with machine-learning techniques to examine long-term trends across fields and research styles, demonstrating how empirical work has gained prominence over time. Meanwhile, Kosnik (2014)leverages textual analysis on a large corpus of economics publications to document the stability and shifts in research focus across JEL categories over the past five decades, revealing a decline in macroeconomic research and an increasing emphasis on empirical methodologies.

The structure of the paper is as follows. Section 2 describes the dataset and presents descriptive statistics on publication patterns, author gender composition, and citation trends across the Top 5 economics journals. Section 3 analyzes horizontal gender diferences in research focus using two complementary approaches: a Structural Topic Model (STM) to estimate latent research topics from abstracts, and the Journal of Economic Literature (JEL) classification system. The section also discusses the limitations and advantages of JEL codes and explores the correspondence between both taxonomies. Section 4 examines gender disparities in citation outcomes, controlling for field specialization using both STM topics and JEL codes, and assesses the extent to which topic composition and JEL codes explain the observed citation gap. Finally, Section 5 concludes and ofers policy implications related to diversity and field representation in academic publishing.

³JEL codes have been used to highlight persistent gender diferences in research fields. Lundberg and Stearns (2019) analyzes PhD dissertations in Economics from 1991 to 2017, using JEL codes to identify the research area. They find that women are more likely to focus on research fields as Labor and Public Economics than in Macro and Finance. Their results also show that this pattern has remained stable over time.

⁴Relatedly, Hospido and Sanz (2021) show that gender gaps in conference acceptance rates are larger in male-dominated fields like finance, yet field fixed efects do not fully account for these disparities.

2 Data: Articles, Journals, Authors and Citations

In this section, we analyze the publication patterns in the top five economics journals, focusing on the number of articles published, the gender composition of authorship teams, and the citations received per paper. We use a database similar to that employed by Conde-Ruiz et al. (2022a), which includes articles published in the Top 5 economics journals (American Economic Review, Econometrica, Journal of Political Economy, Quarterly Journal of Economics, and Review of Economic Studies) in the period 2002-2019. However, for this article we extend the period of analysis to include all articles published from 1999 to 2023, which increases the number of observations from 5,311 articles in the original dataset to 7,244 articles in our updated version. Additionally, we complement this database with information on the citations received by each article, sourced from the RePEc (Research Papers in Economics) database. RePEc is a comprehensive database that aggregates bibliographic information on economics research, including article metadata, working papers, and citation counts. This extensive database covers virtually all relevant journals and a significant number of working papers (pre-prints) from various institutions, with more than 60,000 economists registered⁵. It allows us to track the impact of each article across a wide range of economics publications, providing a detailed picture of citation dynamics over time.

Figure 1 shows the number of articles published annually in the top five economics journals over time. The left axis displays the number of articles published per journal, while the right axis represents the total number of articles published across all five journals combined. The data reveals an upward trend in the total number of articles published until around 2015. Notably, the American Economic Review consistently publishes the highest number of articles, around 35-40% of total.

⁵For more details, see http://repec.org/. RePEc, dedicated to enhancing the dissemination of economic research, compiles metadata from over 2,000 publishers, encompassing academic and commercia publishing houses, research organizations, policy institutions, and think tanks. Additional applications of this data in economics are explored in Zimmermann (2013) and Cabrales et al. (2024)

Table 1 : Descriptive Statistics by Journal Figure 1 : Number of Articles Published per year in Top 5 Journals.

American Economic Review Econometrica JPE QJE RES

Note: Publications exclude notes (without abstract), comments, announcements, and Papers and Proceedings (P&P). Red line corresponds to the total number of articles.

Figure 2 reveals that articles with two authors consistently dominate the publications, maintaining the highest share throughout the time period. Articles authored by a single author exhibit a declining trend, indicating a shift away from solo-authored research over time. Articles with three and four authors have steadily increased, suggesting a growing trend toward collaborative work. Papers with five authors or six or more authors remain relatively rare but show a modest increase in recent years, reflecting a gradual rise in larger

collaborative teams.

Figure 2 : Number of Articles Published per year in Top Journals by Number of Authors. Note: Publications exclude notes (without abstract), comments, announcements, and Papers and Proceedings (P&P).

Figure 3 illustrates the number of articles published annually in Top 5 economics journals, categorized by the gender composition of authorship teams: All Female, All Male, Majority Female, Majority Male, and Equally Distributed. The figure shows that articles authored by all-male teams dominate throughout the period, maintaining the highest share of publications by a significant margin, though their numbers peak around 2021 and decline slightly thereafter. Teams with a majority of male authors consistently contribute the second-largest share of articles, showing a slight upward trend over time. Equally distributed teams exhibit a modest but steady increase in publications, reflecting a gradual shift toward more gender-balanced collaborations. In contrast, all-female teams and teams with a majority of female authors represent a much smaller share of publications, with limited growth over the years.

When analyzing annual citations per article in the top economics journals (Figure 4), clear diferences emerge. The Quarterly Journal of Economics (QJE) consistently registers the highest average citation counts, with notable peaks in the early 2000s and around 2015, underscoring its influence. The American Economic Review (AER) maintains relatively high and stable citation levels over time, likely reflecting its broad readership. In contrast, Econometrica, the Journal of Political Economy (JPE), and the Review of Economic Studies (RES) show comparatively lower averages, with RES persistently garnering the fewest citations per article. These patterns highlight the varying scope, audience, and impact of each journal’s publications.

Figure 3 : Number of Articles per year in Top Journals by Gender Composition. Note: Publications exclude notes (without abstract), comments, announcements, and Papers and Proceedings (P&P).

Figure 4 : Trends in Annual Citations per Article Across Leading Economics Journals. Note: Publications exclude notes (without abstract), comments, announcements, and Papers and Proceedings (P&P).

Figure 5 : Annual Citations per article in Top Journals by Gender Composition of Authors. Note: Publications exclude notes (without abstract), comments, announcements, and Papers and Proceedings (P&P).

Finally, Figure 5 illustrates the annual citations per article and per year in the Top economics journals, categorized by the gender composition of authorship teams: All Female, All Male, Majority Female, Majority Male, and Equally Distributed. The figure highlights that diferences in citation patterns across gender compositions are relatively small. Most categories, including All Male, Equally Distributed, and Majority Male teams, follow similar trends with only slight variations over time.

Our analysis highlights persistent disparities in gender representation, with male-dominated teams consistently accounting for the largest share of publications, while female-dominated teams remain underrepresented. Additionally, diferences in citation patterns emerge, with some variation across journals and gender composition, although these diferences tend to be relatively small on a per-paper basis. These results provide a foundation for the subsequent sections, where we investigate the drivers of these disparities and explore potential mechanisms underlying the observed pattern.

3 Gender Horizontal Diferences in Research

3.1 Gender Horizontal Diferences in Latent Estimated Resarch Topics

As in Conde-Ruiz et al. (2022a) we use the Structural Topic Model (STM), developed by Roberts et al. (2019), to identify the research topics in our extended data based on the abstracts of articles published in top economics journals from 1999 to 2023. This method identifies latent topics in the text, ofering a probabilistic, low-dimensional representation (topics) of high-dimensional data (abstracts) while preserving as much informational content as possible.⁶

Our dataset comprises 7,244 abstracts. After extracting the full set of words, we apply text cleaning procedures designed to reduce the vocabulary and emphasize terms with greater informational value. These steps include removing stop words, performing stemming, and filtering out infrequent terms. As a result, the initial vocabulary of 13,835 words shrinks to a more focused corpus of 4,241 unique terms⁷. These words were then organized into a document-term matrix, which served as the input for the STM algorithm. STM identifies k topics that best fit the document-term matrix where each topic is a probability distribution over words. Intuitively, certain words tend to appear more frequently in texts discussing specific topics than in others. An abstract (document) is treated as a collection of words, each with diferent probabilities of belonging to one or more latent topics. Using this probabilistic relationship between words and topics, the STM allocates each document d to the various topics by estimating a distribution

⁶Compared to the baseline Latent Dirichlet Allocation (LDA), STM improves the estimation by incorporating covariates such as publication year and journal name. LDA is the foundational algorithm for topic modeling and one of the most widely used machine learning methods for reducing the dimensionality of textual data. For a technical overview of LDA, see Blei et al. (2003). Furthermore, Hansen et al. (2017), Bansak et al. (2024) and Beneito et al. (2021) provide examples of its application in economic research.

⁷Our cleaning process of the text is as follows: We have converted all text to lowercase. We have removed common stop words (e.g., “for,” “in”) based on the SMART list developed at Cornell University, which is widely used in text analysis to exclude non-informative words. We have also reduced words to their linguistic roots (e.g., “educ” instead of “education”. Finally, we have eliminated terms that appear only once or twice in the entire dataset. We have followed the same preprocessing steps than in Conde-Ruiz et al. (2022a). See this article for more technical details.

3.1.1 Estimation of the Structural Topic Model (STM)

The first step of the estimation is to determine the number of topics k that best fits our text data. The optimal number of topics, k, in topic modeling using the Structural Topic Model (STM), represents the number of distinct latent topics that best balance statistical fit and interpretability. Each topic corresponds to a cluster of frequently co-occurring words that capture thematic patterns in the dataset, such as research fields or methodologies in a set of academic abstracts. Determining the optimal k ensures that the model identifies meaningful and manageable topics without being overly broad or fragmented.

In line with Conde-Ruiz et al. (2022a), we aim to identify the optimal number of topics for our analysis using Structural Topic Models (STM). Determining the appropriate number of topics is a critical step in STM applications, as it directly influences the interpretability and robustness of the resulting topics. Choosing too few topics may oversimplify the underlying thematic structure, while too many topics can lead to over-fragmentation and reduced clarity, ultimately hindering the usefulness of the model in deriving meaningful insights.

To identify the optimal number of topics (k), we estimate STM models with k ranging from 15 to 65 and assess their performance using held-out likelihood. The held-out likelihood provides a measure of the model’s ability to generalize to unseen data by partitioning the dataset into training and test sets. Specifically, the model is trained on a subset of the data, and the likelihood of the held-out portion is evaluated based on the inferred topicword distributions and document-topic proportions. A higher held-out likelihood indicates that the model captures the data’s underlying structure efectively without overfitting. Our analysis shows that the held-out likelihood achieves its maximum value between k = 49 and k = 60, indicating that this range provides the best statistical fit.

In addition to held-out likelihood, we evaluate the quality of the topics using two complementary metrics: exclusivity and semantic coherence. Exclusivity measures the uniqueness of the words associated with each topic, ensuring that the top words in a topic are not frequently shared with other topics. This enhances the interpretability of the model by ensuring the distinctiveness of topics. Semantic coherence assesses the co-occurrence of the top words in each topic within the original documents, with higher coherence indicating that the topics are more meaningful and contextually grounded.

Balancing these metrics, we select as the optimal number of topics. This choice reflects a trade-of between statistical fit, as indicated by the held-out likelihood, and interpretability, guided by the exclusivity and coherence metrics. By selecting , we achieve a parsimonious model that captures the thematic diversity of the data while maintaining the clarity and interpretability of individual topics. This approach aligns with best practices in the application of STM, ensuring that the model provides both robust quantitative results and actionable qualitative insights.

Figure 6 : Optimal K Topics Ranked by Prevalence in the Corpus (k = 50).

Figure 6 displays the keywords associated with each of the 50 latent topics identified using the Structural Topic Model. The words within each row are arranged from left to right based on their probability of appearing in each topic. To facilitate the interpretation of the latent topics, Table 2 assigns each topic a corresponding JEL code. For this exercise, we have employed ChatGPT (OpenAI, 2024) that has analyzed the top-ranked keywords associated with each topic and proposed the most semantically appropriate JEL code.

However, assigning labels to the estimated latent topics based on commonly recognized fields in Economics is not the primary goal of the analysis. Instead, latent topics may capture a broader range of dimensions, including not only research fields but also methodologies or writing styles, ofering a more comprehensive understanding of the structure and diversity within the dataset.

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 distribution.

Figure 7 represents the network of latent topics, illustrating the connections across the identified topics. Each circle represents a latent topic, and its size reflects the proportion of documents associated with that topic. The connections between topics indicate their semantic similarity, meaning that topics that are connected share overlapping keywords and are more likely to co-occur within the same documents. Topics that are positioned closer together in the figure are more semantically related, often representing similar research fields, methodologies, or writing styles (see, for example, that topics 16, 43 and 23 are connected).

Figure 8 represents the network of latent topics for female-authored papers, which shows the gender-specific distribution across the identified topics. Similar to Figure 7, each circle represents a latent topic, and its size reflects the proportion of all female-authored documents associated with that topic. Significant diferences are observed between the two network figures. This is primarily because women account for less than 20% of the authors in the sample, and also because distinct gender-specific patterns exist regarding research topics. However, this latter aspect is better visualized through the conditional distributions of research topics by gender.

Figure 9 shows the empirical density distributions of topics across male and female authors, conditional on having published a paper in a Top 5 economics journal. These distributions reflect the probability that a paper written by a male or female author belongs to one of the 50 latent topics identified through the Structural Topic Model (STM). Figure 9 highlights significant “horizontal” diferences in research focus between male and female authors, as evidenced by the empirical density distributions of topics. The figure shows that male and female authors are not evenly distributed across the 50 latent topics, with certain areas displaying pronounced gender disparities or important “horizontal” diferences.

The horizontal diferences in research between men and women are best observed by subtracting these two conditional distributions, as illustrated in Figure 13.

Women show a relatively higher propensity than men to make research on topics with positive mass. This analysis leads us to identify topics 23 and 16 as those where, in relative terms, women are more likely to engage, while topic 8 exhibits a similar pattern for men.⁸ child welfare, with words like “children," “family," “parent," “marriage," “education," and “birth," pointing to research areas such as family dynamics, the efects of marriage, child development, and access to education, likely within development economics, social policy, and education economics. These topics highlight applied areas of research where female authors are more represented.

⁸Alternatively, instead of analyzing the diference in topic mass between the conditional distributions,

, we could have examined the ratio , which, in this particular case, would have led to the same conclusions regarding the topics that are relatively more prevalent. This approach of analyzing ratios to identify the most distinctive characteristic that separates one group from another was proposed by Bordalo et al. (2016) as a formalization of the concept of stereotype originally introduced by Kahneman and Tversky.

Figure 9 : Empirical distributions across topics between males and females (conditional on having published an article in Top 5). Note: This figure shows the empirical distribution of estimated research topics by gender, conditional on having published in a Top 5 economics journal. Each bar represents the percentage of articles by male (blue) and female (pink) authors assigned to each of the 50 STM-estimated topics.

Similarly, Topic 8 plays the same role for men. Figure 17) shows the word cloud for Topic 8, which has the highest diference of conditional probability between the males and females, . The most prominent terms, such as “equilibrium," “game," “strategy," “player," and “payof," indicate that this topic is centered on game theory and related areas in economic theory. The presence of terms like “Nash," “equilibria," “stochastic," and “dynamic" suggests a focus on strategic behavior, repeated games, and mathematical modeling.

Figure 10 : Diference between the Empirical distributions across topics between females and males (conditional of having published an article in Top 5).

3.2 Correspondence between Estimated Latent Research Topics and JEL codes Having established the presence of horizontal diferences in research focus through the estimation of latent topics, this section examines how these topics relate to the Journal of Economic Literature (JEL) classification system. The JEL system is a standardized taxonomy developed by the American Economic Association (AEA) to categorize economic research based on its thematic focus. It consists of 20 primary categories, each representing a broad field of economics. Within these, there are 146 secondary categories, which further refine the classification into specific research areas. Additionally, the system includes 856 tertiary categories, providing a granular breakdown of subfields within each secondary classification. This hierarchical structure allows for a systematic organization of economic research, fa cilitating literature searches, enhancing comparability across studies, and enabling a more detailed analysis of research trends and academic contributions. Given its widespread use in top-tier economics journals, the JEL classification system provides a valuable framework for examining potential horizontal diferences in research focus across gender lines.⁹

⁹The classification of economic research through Journal of Economic Literature (JEL) codes has been extensively analyzed in the literature, highlighting its historical evolution, methodological implications,

To begin our analysis, we focus on horizontal diferences using the primary JEL categories, which represent broad research fields in economics. First, we examine the overall distribution of papers across these primary fields, identifying the number of publications within each category. Next, we assess the gender composition of each field by calculating the proportion of female-authored papers, providing a preliminary view of gender disparities in research focus. Finally, we analyze the conditional distributions by gender, determining how male and female authors allocate their research eforts across diferent fields.

and role in shaping the discipline. Cherrier (2017) and Cherrier (2015) provide a comprehensive historical account of the JEL classification system, emphasizing how its revisions reflect deeper debates about the boundaries between theoretical and applied economics, as well as the structuring of economic knowledge.

Table 3 presents the distribution of papers and authors across primary JEL codes, along with the percentage of female authors in each category. The data reveal significant heterogeneity in research output and gender representation across fields. Notably, Macroeconomics and Monetary Economics (E) and Mathematical and Quantitative Methods (C) account for a substantial share of publications, yet exhibit relatively low female representation. In contrast, fields such as Health, Education, and Welfare (I) and Labor and Demographic Economics (J) show a higher proportion of female authors, consistent with prior findings on gender disparities in economics research.

Table 3 : Percentage of Authors and Papers per Primary JEL Code

Figure 13 (similar to Figure 9 done with latent topics) illustrates the empirical distribution of male and female authors across primary JEL codes, conditional on having published in a Top 5 economics journal. The figure reveals notable horizontal diferences in research specialization by gender. Male authors (blue bars) are disproportionately concentrated in Microeconomics (D), which exhibits the highest gender gap, as well as in Macroeconomics and Monetary Economics (E) and in Mathematical and Quantitative Methods (C). In contrast, female authors (pink bars) are more prevalent in Health, Education, and Welfare (I) and Labor and Demographic Economics (J). These patterns align with prior evidence on the gendered division of research fields, where women tend to be more represented in applied microeconomics areas, while men dominate in more theoretical and quantitatively intensive fields.

Figure 13 : Empirical distributions across Primary JEL codes between males and females (conditional on having published an article in Top 5).

Similarly to Figure 13, Figures A 1 and A 2 in the Appendix display the diference in empirical distributions across primary JEL codes (one-digit level), comparing male and female authors conditional on having published in a Top 5 economics journal. Positive values indicate research fields where female are relatively overrepresented (in the conditional distribution), while negative values correspond to fields with a higher proportion of male authors.

The results reveal pronounced horizontal segregation across research specializations. Men are more prevalent in aproximately 50% of the secondary JEL codes and women in the other 50%. J1 and D8 are the JEL codes that, in relative terms, are most frequently used by female and male, respectively. The JEL code J1 falls under Demographic Economics. More specifically, it covers research fields like fertility, family planning, child care, children, and youth. On the other hand, JEL code D8 refers to topics related to Information, Knowledge, and Uncertainty. This broad category includes sub-categories like decision-making under risk and uncertainty, asymmetric information, search and learning. Notice that the JEL codes most representative of men and women exhibit a pattern similar to the most representative latent topics discussed earlier, where women appear relatively more inclined toward applied research, while men tend to focus more on theoretical work. This seems to suggest that there is a relationship between the latent topics estimated from text analysis and the JEL codes assigned to papers.

To visualize the relationships between topics and JEL codes, we plot a correlation heatmap with rows as topics and columns as JEL codes, thereby identifying which topics appear most prominently under certain JEL codes.¹⁰ In particular, Figure 14 shows how STM-derived topics (rows) map onto 0-digit JEL codes (columns). Each cell in the heatmap represents the weight of a given topic for a particular JEL code, with darker shades indicating stronger associations. The distribution reveals a clear many-to-many relationship: most JEL codes are composed of multiple topics, and conversely, many topics contribute to more than one JEL category. Some topics exhibit sharp concentration around specific JEL codes—such as V35 with code D (Microeconomics), V14 with code B (History of Economic Thought), and V06 with code H (Public Economics)—suggesting a high degree of topical coherence in those cases. In contrast, other topics, such as V02 or V39, are more difusely distributed across multiple JEL codes, indicating broader thematic overlap. This visualization highlights both the granularity and complementarity between the latent topics inferred from textual content and the standardized classification provided by the JEL taxonomy.¹¹

¹⁰In the the Appendix, we theoretically explain how this correlation heatmap is built to capture the relationships between STM topics and JEL codes.

¹¹Appendix Figure 3 replicates this analysis using the 1-digit JEL classification. At this higher level of disaggregation, the mapping between topics and JEL codes becomes notably noisier, with few strong, isolated correspondences. The figure suggests that many topics span across multiple subfields within broader categories (e.g., C2, D8, E3), and that the 1-digit JEL taxonomy may obscure rather than reveal the topica

Figure 14 : Correspondences between JEL Codes (0-Digit) and Topics.

3.2.1 Advantages and Limitations of Estimated Latent Research Topics and JEL Codes

We find that both latent topics estimated through text analysis and JEL codes assigned to papers are useful for capturing horizontal gender diferences in economic research. Each methodological approach has its own advantages and limitations, and their suitability depends on the specific application. For example, the mapping between texts and latent topics is fully automated, whereas the JEL classification system relies on self-assignment, with authors independently selecting the JEL codes for their papers. This introduces subjectivity and potential misclassification, either due to a lack of incentives for careful coding or to strategic behavior aimed at maximizing visibility and citations¹². Similarly, while each document is assigned to multiple latent topics to maximize the statistical fit of the model, the number of JEL codes assigned to each paper is also left to the discretion of the authors. This means that, beyond the subjectivity in selecting JEL codes, there is also inconsistency in how broadly or narrowly a paper is classified.¹³

Moreover, the Structural Topic Model provides a distribution of topic weights for each paper, while the JEL classification system does not establish a clear hierarchy among the assigned codes, making it dificult to determine their relative importance within a given article. When a paper is classified under multiple JEL codes, there is no indication of which one best represents its core contribution. This poses a challenge for empirical analysis, as researchers must typically assume equal weights across all assigned codes. In practice, this leads to a symmetry assumption—for example, a paper with three JEL codes is assumed to devote one-third of its focus to each—an assumption that may not hold if some codes reflect central contributions while others are secondary. That said, the JEL system remains a widely used and standardized classification scheme, well understood by editors and researchers alike. Its structure enables consistent tracking of research trends over time and facilitates comparability across studies.

¹²Kosnik (2017) provides empirical evidence of this issue by analyzing the discrepancy between authorassigned and editor-assigned JEL codes in the American Economic Review (AER) between 1990 and 2008. In this dataset, the AER editorial team reassigned JEL codes to published papers, altering 43% of them, revealing substantial inconsistencies. These changes suggest that authors often misclassify their research, either inadvertently or as a strategic response to perceived trends in citation behavior. The study also finds that some JEL categories, such as C (Mathematical and Quantitative Methods) and D (Microeconomics), tend to be overused by authors, while others are underrepresented, potentially distorting the perceived composition of economic research. These findings underscore a major shortcoming of the JEL system: while it provides a structured taxonomy, its reliance on self-assignment introduces biases that may weaken its reliability as a tool for tracking and analyzing research trends

¹³In this sense, Kosnik (2017) compares author-assigned and editor-assigned JEL codes in the American Economic Review (AER). The study finds that authors assigned, on average, more JEL codes to their papers than the editors ultimately retained, suggesting a tendency among researchers to overclassify their work. This discrepancy indicates that the incentives guiding authors’ choices may difer from the editoria standards applied in formal classification. The overuse of JEL codes by authors can dilute the precision of the system, making it harder to track research specialization and trends accurately.

The text analysis approach ofers a valuable alternative for addressing research questions traditionally explored using JEL codes. In this paper, our main contribution is to analyze gender citation gaps using latent research topics derived from text analysis, and to compare their performance with that of JEL-based classifications. We show that, in this application, the latent topic approach yields results that are closely aligned with those obtained using JEL codes, supporting its validity as a complementary tool for studying horizontal gender diferences in academic publishing.

4 Citation Gender Gaps

In this section, we examine whether female-authored papers receive systematically diferent citation counts in Top 5 economics journals after accounting for diferences in research topics. Building on the methodology introduced in earlier sections, we estimate citation regressions that control for the publication year, the journal, the number of authors and either the JEL classification or the latent research topics identified through the Structural Topic Model (STM). This setup allows us to assess whether the observed citation gap by gender reflects diferences in thematic specialization or whether it persists after conditioning on field composition.

We first explain our empirical strategy. To formally assess whether female economists experience systematic diferences in citation rates, we rely on a standard linear regression framework commonly used in the literature (see for example Kofi (2021)). In line with previous studies, our empirical model relates the number of citations received by a paper to the gender composition of its authors, controlling for key confounding factors. Specifically, we estimate the following specification:

\[C _ {p, t} = \beta_ {0} + \beta_ {F} F _ {p} + \beta_ {3} X _ {p} + \gamma_ {j t} + \lambda_ {f} + \epsilon_ {p, p ^ {\prime}, t},\tag{1}\]

where is the cumulative citations received by paper Following standard practice, we apply the inverse hyperbolic sine transformation, asinh , to the citation variable to ensure that the estimated coeficients can be interpreted as semi-elasticities, even when denotes the proportion of female co-authors on paper while captures the total number of authors. represents fixed efects for research fields, operationalized either through latent topics or JEL codes. The error term captures unobserved heterogeneity. We estimate the model under two diferent specifications: one including calendar-year fixed efects to absorb common temporal trends in citation practices, and another including journal-by-year fixed efects to account for time-varying shocks specific to each outlet. The latter specification enables more precise comparisons by restricting variation to articles published in the same journal and year, and—when combined with topic or JEL fixed efects—within the same research field. This approach helps isolate the role of author gender from confounding diferences across journals, time periods, or subject areas.

Under specification, the coeficient gives us a clean estimate of the gender citation gap: if it’s positive, it suggests that papers with more female authors tend to receive more citations; if it’s negative, it suggests a penalty. To ensure our statistical inference is reliable, we cluster standard errors by year (or by journal-year) to allow for the possibility that papers published around the same time might be exposed to similar citation patterns or shocks. Our empirical approach is closely related to that of Kofi (2021), who also examines the relationship between citation outcomes and the gender composition of author teams, controlling for research fields using JEL classification codes. We extend this framework by incorporating not only JEL codes but also latent research topics estimated via structural topic modeling. As discussed in earlier sections, the latent topics show strong alignment with JEL codes while also uncovering additional dimensions of research focus that conventional classifications may overlook—an advantage that proves especially valuable for analyzing citation dynamics.

Table 4 presents our main results. In line with previous studies such as Kofi (2021), Card et al. (2020), and Hengel and Moon (2023), we find that papers written by women tend to receive more citations. Specifically, we estimate a citation advantage of about 16.3 log points when controlling for year, and about 12.1 log points when controlling for both year and journal. This suggests that, on average, female-authored papers are cited more often than comparable male-authored ones, even after adjusting for outlet and time efects.

¹⁴The year subscript t is retained for consistency with the fixed-efects structure, yet each article appears only once—at its publication year—so represents the cumulative stock of citations tallied at the end of the sample.

Table 4 : Regression Results ( Letter + 2 digits)

. Standard errors are clustered by year in Models ( 1 ) ( 3 ) and ( 5 ) and by j ournal-yea in Mo dels ( 2 ) (4) and ( 6 ) . The inverse- hyp erb olic- sine transform is asin a log-like function that is well-defined at zero .

However, once we control for research fields—either through estimated topics or JEL codes—the gender citation gap largely disappears. The coeficient on the proportion of female authors is positive and statistically significant when the model includes only year fixed efects (Model 1), or both year and journal fixed efects (Model 2). Yet, when we further account for field diferences—by including 49 latent topic dummies (Models 3–4) or detailed JEL code fixed efects (Models 5–6)—the estimated gender efect becomes statistically insignificant.¹⁵ In our most comprehensive specification, increasing the share of female co-authors by one standard deviation is associated with a change in expected citations of less than 0.025 log points—a negligible magnitude well within the margin of statistical uncertainty. These results suggest that the initial citation premium for female-authored papers is largely driven by diferences in field specialization between men and women, rather than a direct gender efect.¹⁶

Goodness-of-fit statistics support this interpretation. The overall R²—which measures how much of the variation in citation counts our model explains—increases substantially, from 0.36 in the basic specification to 0.56 when we control for both research topics and JEL codes. This suggests that a large share of citation diferences across papers is linked to the specific field of research. Even more telling is the change in the within-group , which rises from almost zero to 0.29. This means that once we account for when and where a paper was published, most of the remaining variation in citations is still driven by the topic or field the paper belongs to. In other words, the field of research matters—a lot. These results confirm the importance of using the richest possible set of controls when estimating the gender citation gap. Simpler models that ignore field diferences tend to overstate the role of gender because they inadvertently capture persistent diferences in the kinds of topics men and women work on.

An additional and novel feature of our analysis is the inclusion of team size as a control variable. While previous studies have focused on author gender and field of research, few have systematically accounted for the number of co-authors when analyzing citation dynamics. Across all specifications, we find a strong and robust association between team size and citation outcomes: the estimated elasticity ranges between 0.14 and 0.22. In practical terms, adding one co-author increases expected citations by approximately 15–20 percent, holding other factors constant. This efect remains stable across all models, suggesting that collaboration enhances scholarly visibility or impact independently of research field or journal. One plausible interpretation is that larger teams benefit from greater specialization, broader dissemination networks, or complementary skills.

¹⁵In Appendix Tables 10 and 12, we report the estimated coeficients for the field fixed efects—both STM-estimated topics and JEL codes—included in the citation regressions. To maintain clarity and focus, we display only those coeficients that are statistically significant at the 5% level in Table 10 and statistically significant at the 1% level in Table 12 . These results provide additional insight into field-specific citation patterns and complement the main analysis presented in the paper.

¹⁶For completeness, Appendix C replicates the analysis using JEL codes aggregated at two alternative levels: (i) the letter plus one-digit level (Table 9) and (ii) the letter-only level (Table 11). The findings remain broadly consistent, reinforcing the robustness of our conclusions across diferent specifications of field controls. Table 9 shows the JEL code significant coeficients at the 1% level in Table 11, Model 6.

For robustness, we also consider an alternative gender categorization. Instead of using the proportion of female authors in each coauthor team, we classify author gender composition into three groups: papers with no female authors, papers with a minority share of female authors (0–50%), and papers with a majority of female authors (above 50%). Using this more flexible classification, we replicate our main regression with these categorical indicators. The results remain highly consistent with those obtained using a continuous measure of female author share, providing additional robustness to our findings. We report this alternative specification in Table 5.

¹⁶For large values of the dependent variable, the inverse hyperbolic sine transformation asinh(x) approximates ln(2x). Therefore, the estimated coeficients can be interpreted as approximate semi-elasticities, especially for papers with moderate to high citation counts.

Table 5 : Regression Results (Letter + 2 digits) and 3 Female Groups

Notes : + p < 0 . 1 * p < 0 . 05 * * p < 0 . 0 1 * * * p < 0 . 00 1 . Standard errors are clustered by year in Models ( 1 ) (3) and ( 5 ) and by j ournal-year in Models ( 2 ) , (4) , and ( 6 ) . The inverse-hyperbolic-sine transform is asin a log-like function that is well-defined at zero

Our analysis shows that both latent research topics and JEL-based classifications are useful methodologies for analyzing gender citation gaps, as they yield similar results. However, it is also relevant to ask whether one of these approaches ofers superior explanatory power when modeling citation outcomes. This question cannot be addressed directly using Table 4, since the sample of papers with JEL codes is smaller. To ensure a fair comparison, we address this question by holding the sample constant.

Specifically, we compare model fit using either JEL codes or STM topics as field controls, restricting the dataset to the subset of papers for which JEL codes are available (as in regressions (5) and (6)). We then re-estimate regression (4) from Table 4 on this same subsample and compute three standard model fit statistics: the adjusted , the Bayesian Information Criterion (BIC), and the Akaike Information Criterion (AIC). The results are presented in Table 6.

Table 6 : Model Fit Comparison Using Diferent Sets of Controls

While the adjusted captures the proportion of variance explained, the AIC and BIC incorporate penalties for model complexity, with lower values indicating a better model. These criteria allow us to evaluate whether any improvement in fit justifies the additional parameters. In our case, the model using topic controls outperforms the one based on JEL codes, with substantially lower AIC and BIC values. This suggests that the STM-based specification ofers a better overall balance between explanatory power and parsimony.

5 Conclusions

This paper provides new evidence on gender disparities in citation patterns within the Top 5 economics journals by combining two complementary approaches to classify research content: a data-driven method based on STM-estimated topics, and a standardized taxonomy based on JEL codes. Our main contribution is to show that female-authored papers exhibit a citation premium on average, but that this premium largely disappears once we control for field specialization using either classification. The consistency of results across both systems reinforces the conclusion that horizontal gender diferences in research focus—rather than diferential treatment in citation behavior—explain most of the observed citation gap.

We do not find direct evidence of gender-based citation bias. However, our results highlight persistent horizontal gender diferences in field specialization, and there are plausible mechanisms through which these diferences may lead to indirect structural barriers in academic careers. Theoretical models of statistical discrimination in evaluation processes show that if women are underrepresented in editorial boards or evaluation committees, diferences in research focus across genders could negatively afect promotion, tenure, and long-term academic recognition. Conde-Ruiz et al. (2022b) introduce the concept of homo-accuracy bias, whereby evaluators assess candidates more accurately when they share similar research interests. Similarly, Siniscalchi and Veronesi (2020) describe a form of self-image bias, in which evaluators tend to favor candidates who resemble their younger selves. Both models suggest that thematic underrepresentation may reinforce academic inequality, even in the absence of explicit bias.

Data Availability

This study is based on publicly available data, except for the citation data, which was provided by RePEc (Research Papers in Economics). Access to these citation data is subject to RePEc’s terms of use and data-sharing policies.

References

1. Angrist, Joshua, Pierre Azoulay, Glenn Ellison, Ryan Hill, and Susan Feng Lu, “Economic Research Evolves: Fields and Styles,” American Economic Review: Papers & Proceedings, 2017, 107 (5), 293–297.
2. Bansak, Cynthia, Wendy Dunn, Ellen Meade, and Martha Starr, “Impact versus Inclusion in the Economics Profession: Insights from the Papers and Proceedings,” AEA Papers and Proceedings, 2024, 114, 292–299.
3. Beneito, Pilar, José E. Boscá, Javier Ferri, and Manu García, “Gender Imbalance across Subfields in Economics: When Does It Start?,” Journal of Human Capital, 2021, 15 (3), 469–511.
4. Blei, David M., Andrew Y. Ng, and Michael I. Jordan, “Latent Dirichlet Allocation,” J. Mach. Learn. Res., March 2003, 3 (null), 993–1022.
5. Bordalo, Pedro, Katherine B Cofman, Nicola Gennaioli, and Andrei Shleifer, “Stereotypes,” The Quarterly Journal of Economics, 2016, 131 (4), 1753–1794.
6. Cabrales, Antonio, Manu García, David Ramos Muñoz, and Angel Sánchez, “The Interactions of Social Norms about Climate Change: Science, Institutions and Economics,” Working Papers 2024-036, Federal Reserve Bank of St. Louis November 2024.
7. Card, David, Stefano DellaVigna, Patricia Funk, and Nagore Iriberri, “Are referees and editors in economics gender neutral?,” The Quarterly Journal of Economics, 2020, 135 (1), 269–327.
8. Cherrier, Beatrice, “JEL Codes and the Structuring of Economic Knowledge,” SSRN Working Paper, 2015.
9. , “Classifying Economics: A History of the JEL Codes,” Journal of Economic Literature, 2017.

Conde-Ruiz, J. Ignacio, Juan J. Ganuza, Manu García, and Luis A. Puch, “Gender distribution across topics in the top five economics journals: a machine learning approach,” SERIEs: Journal of the Spanish Economic Association, May 2022, 13 (1), 269–308.

, Juan José Ganuza, and Paola Profeta, “Statistical discrimination and committees,” European Economic Review, 2022, 141 (C).

Ductor, Lorenzo and Anja Prummer, “Gender homophily, collaboration, and output,” Journal of Economic Behavior & Organization, 2024, 221, 477–492.

Hansen, Stephen, Michael McMahon, and Andrea Prat, “Transparency and Deliberation Within the FOMC: A Computational Linguistics Approach,” The Quarterly Journal of Economics, 10 2017, 133 (2), 801–870.

Hengel, Erin and Eunyoung Moon, “Gender and quality at top economics journals,” mimeo 202001, University of Liverpool, Department of Economics February 2023.

1. Hospido, Laura and Carlos Sanz, “Gender Gaps in the Evaluation of Research: Evidence from Submissions to Economics Conferences,” Oxford Bulletin of Economics and Statistics, 2021, 83, 590–618.
2. Kofi, Marlène, “Gendered Citations at Top Economic Journals,” AEA Papers and Proceedings, 2021, 111, 60–64.

Kosnik, Lea, “What Have Economists Been Doing for the Last 50 Years? A Text Analysis of Published Academic Research from 1960-2010,” SSRN Working Paper, 2014.

1. Kosnik, Lea-Rachel, “A Survey of JEL Codes: What Do They Mean and Are They Used Consistently?,” Journal of Economic Surveys, 2017, 00, 1–24.

Lundberg, Shelly and Jenna Stearns, “Women in Economics: Stalled Progress,” Journal of Economic Perspectives, February 2019, 33 (1), 3–22.

Merton, Robert K., “The Matthew Efect in Science,” Science, 1968, 159 (3810), 56–63.

1. Roberts, Margaret E., Brandon M. Stewart, and Dustin Tingley, “stm: An R Package for Structural Topic Models,” Journal of Statistical Software, Articles, 2019, 91 (2), 1–40.
2. Siniscalchi, Marciano and Pietro Veronesi, “Self-image bias and lost talent,” Technical Report DP15621, CEPR Discussion Paper Series 2020.
3. Zimmermann, Christian, “Academic Rankings with RePEc,” Econometrics, 2013, 1 (3), 249–280.

A Conditional Empirical Distribution across Topics between Male and Female.

Table 7 : Conditional Distribution of Topics by Gender

Note:

Figure A. 1 : Diference of the Empirical distributions across Primary JEL codes + 1 Digit between males and females (conditional on having published an article in Top 5). Part I.

Figure A. 2 : Diference of the Empirical distributions across Primary JEL codes + 1 Digit between males and females (conditional on having published an article in Top 5). Part II.

B Linking Latent Estimated Research Topics to JEL codes

In this appendix, we aim to disentangle the existing relationship between the topics uncovered by a structural topic model (STM) and the JEL codes assigned to articles in our sample. Suppose we have a set of n documents. The STM estimation yields a topic-distribution vector for each document,

\[\theta_ {d} = \left(\theta_ {d, 1}, \theta_ {d, 2}, \dots , \theta_ {d, k}\right), \quad \sum_ {t = 1} ^ {k} \theta_ {d, t} = 1,\]

where k is the total number of extracted topics, and indicates the proportion of document d’s content assigned to topic t. We can rewrite the topic-distribution vector as a matrix . Similarly, each document d is allocated to a set of JEL codes.

\[\gamma_ {d} = \left(\gamma_ {d, 1}, \gamma_ {d, 2}, \dots , \gamma_ {d, \mathcal {J}}\right)\]

Where is the number of JEL codes under consideration (i.e. primary, secondary or tertiary JEL classification) and is equal to 1 if the JEL code has been assigned to document d and it is 0 otherwise. We can rewrite this document-JEL codes vector as a matrix

To connect topics to the JEL codes, for each document we multiply the topicdistribution matrix by JEL codes matrix this generates a matrix matrix , where is for document d the topic-t weight if the JEL code j has been also assigned to d (and 0 otherwise). We define an aggregate matrix M by summing over all documents, the matrices

We normalize M to by homogenizing the total weight of the columns to 1, in order to facilitate interpretation. Under this column normalization, each column j becomes a probability distribution over topics:

\[\widetilde {M} _ {t, j} = \frac {M _ {t , j}}{\sum_ {t = 1} ^ {k} M _ {t , j}},\]

so that represents the topic t weight for JEL code . In efect, this transformation

Figure A. 3 : Correspondences between JEL Codes (1-Digit) and Topics. reveals how each JEL code is composed of the various STM-derived topics. Figure 14 in the main text and Figure 3 in the appendix, display the normalized correspondence matrix between STM-derived topics (rows) and JEL codes (columns) as heatmaps.

C Additional regressions

Table 8 : Regression Results ( Just letter)

N_o t_{es :} Standard errors in arentheses Models ( 5 ) – ( 6 ) include JEL code fixed efects (no rep orted) _.

Table 9 : Regression Results ( Letter + 1 digit )

N_o t_{es :} Standard errors in arentheses Models ( 5 ) – ( 6 ) include JEL code fixed efects (no rep orted) _.

Table 10 : Topic significant coeficients at the 5% level in Model 4, Table 4.

Note: ^∗∗p < 0.05. Model 4 includes Journal and Year as combined fixed efects, and also includes Topic as a control. Only topics coeficients statistically significant at 5% are reported in the table.

Table 11 : JEL code significant coeficients at the 1% level in Model 6, Table 8.

Table 12 : JEL code significant coeficients at the 1% level in Model 6, Table 4

Note: ^∗∗∗p < 0.01. Model 6 includes Journal and Year as combined fixed efects, and also includes JEL codes as a control. Only statistically significant at 1% level coeficients are reported in the table.