Is science still an old boys’ network? A structural approach for assessing patterns of gender heterophily in a large research network

Abstract

Introduction:

Modern social and health sciences are inherently collaborative through team science, scientific collaboration between at least two people with an interdependent relationship. Despite evidence supporting the case for team science, some groups experience limited collaborative opportunities. In 2006 the National Institutes of Health’s Clinical and Translational Science Awards (CTSA) established programs with a goal of promoting translational science by facilitating collaborative research among diverse, multidisciplinary teams.

Methods:

We take advantage of 15 years of collaboration data from members of the Washington University in St. Louis CTSA, the Institute of Clinical and Translational Sciences (ICTS). We describe (1) characteristics of ICTS researchers and their collaboration networks, (2) differences in network positions by gender (male/female), (3) patterns of scientific collaboration by gender across 15 years, and (4) develop a multivariate statistical network model that identifies predictors of collaboration, including gender. We combined descriptive analyses with social network analysis (SNA) to describe collaboration patterns by gender of co-authorship networks across years. We also built statistical network models of collaboration ties (using exponential random graph models) to explore how gender homophily influences co-authorship collaboration.

Results:

We found gender differences in productivity, number of collaborators, and researcher characteristics. The number of publications and collaborators were higher among males than females. Females were more represented in allied health and social sciences, while males were concentrated in clinical and basic sciences, with very few males in allied health. Females were more represented in lower academic ranks and less in higher ranks, whereas males were more represented in higher ranks and less in lower ranks. The ERGM results indicated collaboration was more likely in clinical and social science disciplines and at higher academic ranks. These differences are reflected in the ERGM results for the homophily terms, which showed collaboration was less likely when involving a female researcher compared to male–male collaborations.

Discussion:

By pairing SNA with descriptive methodologies and statistical modeling, we gained a nuanced understanding complex interactions within a scientific collaboration network, providing insights into how relationships and collaboration patterns evolve, especially with respect to gender. This methodological approach offers a valuable framework for exploring similar questions in other contexts.

1 Introduction

Modern social and health sciences are inherently collaborative; the proportion of social science publications with more than one author increased from approximately 10% in 1959 to over 60% in 2013 and the proportion of all publications with at least two authors increased to 90% (Hilton and Cooke, 2015). Publications with multiple authors are also cited more than single-author publications (Wuchty et al., 2007). Team science is defined as scientific collaboration between at least two people with an interdependent relationship (Hilton and Cooke, 2015). Team science collaborators may vary in gender, race, ethnicity, discipline, goals, organization, or geography composition (Bennett and Gadlin, 2012; Love et al., 2022). Diverse teams are successful when diverse members are in roles to contribute to innovation rather than participating at the periphery (Bear and Woolley, 2011). Inter-gender teams are also more likely to develop innovative ideas and publications than same-gender teams (Hall et al., 2018; Yang et al., 2022; Zhang et al., 2024).

Despite evidence supporting the case for team science, some groups experience limited collaborative opportunities. Empirical research has evaluated gaps by gender or sex. In the pursuit of equity in team science, it is important to differentiate between evaluations of gender and sex and recognize that proxy measures can result in misrepresentations (Heidari et al., 2016). Empirical research found the probability of a project lead selecting a female principal investigator when forming a team in a multi-disciplinary funding opportunity was only 17.5%, which is far from 50% (Pezzoni and Visentin, 2024). An evaluation of UK studies published in The BMJ and the Lancet over 20 years found estimated gender identity based on first names finding that women were underrepresented with women being the first or last author in, at most, 43% of The BMJ articles and only 35% in the Lancet (Abdalla et al., 2023). The lower proportion of women or female researchers may also limit collaborative opportunities for women based on based on gender. The homophily effect suggests collaborators are more likely to work with their same gender or sex (Holman and Morandin, 2019). Further study of gender (name, photograph, or biographical information) in co-authorship networks found that differences in network characteristics were associated with academic rank rather than gender (Walker et al., 2024). This pattern may widen the gap, as women are more likely to leave academia before promotion to higher positions while men receive greater recognition for scientific contributions (Athanasiou et al., 2016; Cain and Leahey, 2014; Goulden et al., 2011; Heijstra et al., 2015; Rossiter, 1993). Additionally, women’s opportunities to lead teams may also be limited by systemic biases in funding opportunities, where women receive smaller amounts of funding or are less likely to receive funding than men (Sato et al., 2020; Schmaling and Gallo, 2023).

Efforts are underway to promote gender and other kinds of equity in team science by both funders and the research community. Funding agencies such as the National Institutes of Health (NIH) can also support team science approaches through funding mechanisms such as grants using a multiple PI model, which most NIH funding opportunities support. The NIH has also launched funding initiatives to address gender diversity and equity (Ten Hagen et al., 2022). Additionally, facets of the research community are supporting this effort by promoting best practices for team science, such as those developed by the Patient-Centered Outcomes Research Institute (PCORI) and The Science of Team Science (SciTS) (Lotrecchiano et al., 2023).

Team science takes a structural approach for understanding how scientific collaboration works, and what kinds of social processes may drive inequities in scientific collaboration (Borner et al., 2010). However, we lack empirical knowledge on how to develop and implement practice and policy solutions that will actively balance the scientific playing field.

Despite growing recognition of these issues, important gaps remain in understanding how gender shapes collaboration as it occurs in a structural manner (Gray, 1994; Santos et al., 2024). Existing literature relies on descriptive statistics or individual-level analyses that cannot account for the interdependent nature of collaboration networks or assess whether gendered patterns persist after controlling for rank, discipline, and institutional context. There is a need for network analyses that examine how collaboration ties form and evolve over time within policy-relevant settings designed to promote team science. Addressing this gap is essential for identifying the mechanisms through which gender inequities are reproduced in collaborative research and for informing institutional strategies aimed at fostering more equitable scientific collaboration.

In this paper, we take advantage of 15 years of collaboration data from members of the Institute of Clinical and Translational Sciences (ICTS) at Washington University in St. Louis (WashU). ICTS is funded by the National Institutes of Health’s Clinical and Translational Science Awards (CTSA) program. A major goal of the CTSA program is to promote translational science by facilitating collaborative research among diverse, multidisciplinary teams. Earlier evaluation of these collaboration data at WashU documented the growth of interdisciplinary collaboration during the first 5 years of CTSA funding (Dhand et al., 2016; Luke et al., 2015). We have also used network models to explore how engagement with ICTS research services and supports was associated with subsequent research productivity (i.e., grants and publications) (Combs et al., 2021). For this paper, we focus on the role of gender in scientific collaboration, using 15 years of ICTS data and co-authorships among ICTS members on publications during this time.

More specifically, the research goals of this paper are to (1) describe characteristics of ICTS researchers and their collaboration networks, (2) describe differences in network positions by gender (male/female), (3) describe patterns of scientific collaboration by gender during this time, and to (4) develop a multivariate statistical network model that identifies predictors of collaboration, including gender.

2 Materials and methods

2.1 Participants

This study uses 15 years of evaluation data from the WashU Institute of Clinical and Translational Sciences (ICTS). Participants include all scientists meeting four criteria in at least one year between 2008 and 2022: (1) ICTS members, (2) faculty or staff at WashU, (3) self-identified as male or female, and (4) published at least one paper. Individuals become ICTS members by registering online to gain access to core facilities, training, internal funding, and collaboration opportunities. IRB approval was not required for secondary analysis of WashU administrative data and publicly available publication information.

2.2 Network construction

Medical librarians at Becker Medical Library (WashU) used Elsevier Scopus to identify ICTS member publications including articles, conference papers, reviews, and short surveys. They also disambiguated author name conflicts. The research team created 2-mode affiliation networks, connecting authors to publications (Rawlings et al., 2023). We then produced annual single-mode co-authorship networks, e.g., one network for 2008 publications, one for 2009 publications, etc. Thus, an annual co-authorship network contained all ICTS members who had published at least one paper that year, and members would be connected if they had been co-authors on specific publications that year. Network data were managed with R igraph (v 2.0.3).

2.3 Other measures

Data for gender and rank was obtained from WashU Human Resources. At the time of hiring, employees were asked their gender with response options of male, female, or not declared. Gender was coded as female or male. This analysis is limited to the response options that reflect self-reported sex, which is a different construct than gender identity (Becker, 2022). Evaluation of gender identity was not available for this analysis. Rank is coded annually in one of four categories: Non-faculty (staff, fellows, post-docs, and instructors), Assistant Professor, Associate Professor, or Professor. Investigators self-selected their discipline from a list of 205 NIH fields (U.S. Department of Health and Human Services, n.d.) when signing up for ICTS membership. These fields were collapsed into four broader categories: Clinical, Social Science/Statistics, Allied Health, and Basic Science. Allied health included other health sciences professionals, including public health, social work, and nursing. Productivity was measured by the total number of publications for each member within a given year. Engagement was defined as the formal usage of ICTS research services and support (Combs et al., 2021). The two types of engagement used here were (1) the use of core services (e.g., biostatistical consulting, use of tissue banks) and (2) the receipt of internal ICTS funding through various types of pilot grants.

2.4 Analyses plan

2.4.1 ICTS researcher and network characteristics

We calculated frequencies and percentages for participant demographics as well as their academic, engagement, and productivity measures for each year of the data set. We explored trends by gender both within and across years. To assess representation, we analyzed the average percent of members within each rank and discipline who were female or male. This provided gender composition within each category, revealing potential imbalances across academic rank as well as discipline.

Network descriptive statistics were calculated for each year. Structural statistics included network size (number of people in the network), average degree (average number of connections per person), and modularity. Modularity quantifies the extent to which groups form around a shared characteristic (Rawlings et al., 2023). The modularity scale ranges from −0.5 to 1, where a value of 0 indicates no grouping based on the characteristic, and a value of 1 signifies complete separation by that characteristic. In this study, we calculated modularity using gender and discipline to assess group formation patterns.

2.4.2 Differences in network positions by gender (male/female)

We used centrality measures to examine how gender differences manifest in network positions within the overall collaboration network. Centrality measures were used to identify nodes located in strategic places within the network. Degree is a measure of the number of collaborators a researcher had that year. It is a local measure that does provide insight into node position within the full network. Betweenness, however, provides the network-wide context. Betweenness is the number of times a node occurs on the shortest path connecting other nodes (Wasserman and Faust, 1994). Betweenness is often conceptualized as bridging. An individual with a high betweenness connects many people and parts of the network that would have otherwise been separated. Degree and betweenness centrality were chosen because they are commonly used, relatively easy to understand, and are relevant to our core question of how gender may drive collaboration patterns. In particular, degree centrality is a local measure of centrality, and this kind of centrality measure makes sense for co-authorship networks where we are interested in how a particular researcher is directly connected to her co-authors.

2.4.3 Patterns of collaboration by gender over time

To understand the role of gender in co-authorship collaboration, the frequencies and percentages of the three possible dyad combinations (Male–Male, Female–Female, and Male–Female) were calculated for each year.

2.4.4 Network modeling of collaboration ties

Statistical modeling of networks is a powerful tool for identifying factors related to network ties and structural characteristics. These modeling approaches, particularly exponential random graph models (ERGMs), allow statistical inference and deductive hypothesis testing while appropriately handling the non-independent nature of network data (Cranmer and Desmarais, 2011; Goodreau, 2007). This approach is particularly important because ERGMs allow us to examine multiple interdependent social processes that general global network characteristics (Kim et al., 2016).

In this study, we operationalized collaboration ties as co-authorship within individual publications. We used exponential random graph modeling (ERGM) to identify predictors of co-authorship collaboration and investigate the role gender plays in collaborations (Harris, 2013). We used a bottom-up model-building approach, adding explanatory covariates in blocks (Kutner et al., 2005). A separate set of models was built for each calendar year (i.e., 15 model sets from 2008 to 2022). The initial null model included only the network edges to serve as a baseline for subsequent models. We then added academic and engagement predictors. Academic rank was entered as a scale variable. Academic discipline was entered as a categorical variable with Clinical as the reference group. Core utilization and funding status were entered as binary variables with “No” as the reference group. Our final model entered gender as a homophily term, allowing us to examine the likelihood of collaborations between scientists of the same or different genders, using the largest category, “Male–Male” as the reference group. By introducing gender in the last block, we isolate the impact of gender on collaboration likelihood while statistically controlling for all other predictors. Nested models were compared using the Akaike Information Criterion (AIC) to evaluate relative fit. Final model coefficient significance and direction were extracted and compared to highlight potential differences over time. Odds ratios for the gender homophily term were calculated and presented for each annual model.

All analyses were conducted in R (Version 4.4.1), and the network analyses used the statnet (Version 2019.6) and igraph (Version 2.1.1) network packages.

3 Results

3.1 ICTS researcher and network characteristics

ICTS membership grew rapidly after receiving initial CTSA funding in 2007. The total number of ICTS members in 2008 and 2022 were 275 and 1855, respectively, of those, 242 (88%) and 1,523 (82%) published at least one paper in those years. Those members constituted the publication networks for this analysis. Over the 15-year period, publication networks experienced growth in the number members across discipline, rank, and gender. The number of members who published increased each year, as did the annual publication count, shown in Figure 1. Network structure changed dramatically over the 15 year period. The 2008 and 2022 networks are visualized in Figure 2 (visualizations for all years are available in Appendix A). These network diagrams show the co-authorship ties between ICTS members for that specific year. Specifically, the nodes are the members, and pairs of members (dyads) are connected to each other if they co-authored one or more papers in that specific year.

Figure 1

Figure 2

Both the 2008 and 2022 networks contain isolates and multiple components, including a large central component in each year. In the 2008 network, the largest component exhibits clustered collaborations and branching structures but remains relatively sparse. By contrast, the 2022 network’s largest component is more densely interconnected, with numerous overlapping collaborations and tighter clustering.

The distribution of discipline and rank over time are shown in Figure 3. For the most part, the distribution of ICTS-member discipline remained stable over time, with the majority of members being clinical scientists. There is a gradual increase of the proportion of basic scientists, until plateauing at about 30% around 2018.

Figure 3

While there was consistent growth in the number of members in each academic rank, non-faculty grew at a faster rate, moving from 11% of members in 2008 and growing to 22% in 2022. The distribution of associate professors was relatively stable from 2008 to 2022, with an average of 20%. The distribution of assistant professors began to decrease in 2015 from 30 to 24% in 2022. The distribution of full professors initially declined from 38% in 2008 until stabilizing at 33% in 2012. This pattern is consistent with a growing focus over time in ICTS with training (and attracting) junior clinical researchers.

3.1.1 Researcher characteristics stratified by gender

As ICTS membership increased since 2008, the gender distribution was also changing. Female members accounted for 30% of the overall membership but increased dramatically to approximately 40% in 2022 (see Figure 4).

Figure 4

Researcher characteristics stratified by gender for 2008 and 2022 are presented in Table 1 (data for all years is provided in Appendix B). The table includes the total number of male and female researchers, along with the counts for each attribute. Percentages are calculated relative to the total for each gender. For publications, the median number is reported separately for male and female researchers.

3.1.1.1 Academic characteristics

The distribution of rank and discipline within each gender is shown in Figure 5. Female researchers are more evenly distributed across ranks compared to their male counterparts. On average, 23.3% of female members were non-faculty, 33.2% held the rank of assistant professor, 21.4% were associate professors, and 22.3% were full professors. In contrast, male members showed a stronger concentration at the full professor level, with an average of 40.2% occupying this rank, while 14.1% were non-faculty, 25.1% were assistant professors, and 20.6% were associate professors. Within the full professor rank, females consistently make up a smaller proportion over time compared to males. While lower academic ranks, such as non-faculty, show a more balanced gender composition.

Figure 5

The distribution of discipline also revealed notable differences. Among male members, the largest proportion is found in clinical fields (56.6%), followed by basic sciences (30.1%), with smaller percentages in allied health (9.7%) and social sciences/statistics (3.5%). While there was a large clinical presence for female members, they were more evenly represented across the other disciplines, with 46.5% in clinical fields, 23% in basic sciences, 22.5% in allied health, and 8% in social sciences/statistics. This distribution suggests that male members are particularly concentrated in the clinical and basic sciences, while female members show a broader representation across multiple fields, including higher proportions in allied health and social sciences.

3.1.1.2 Engagement characteristics

Core users are ICTS members who report using one or more of the ICTS research support services (e.g., biostatistical consulting, tissue bank procurement, etc.). The use of the core services by female members of ICTS increases gradually over time (Figure 6). This increase mirrors quite closely the general rise in the proportion of ICTS members who are female. That is, core usage by females rises from around 33% at the beginning of ICTS funding to 40% in 2022. If males and females were using cores services in notably different ways, then we would not see the close correspondence of the two lines if Figure 6.

Figure 6

Another metric for ICTS engagement is receipt of internal ICTS research funding. The number of researchers receiving ICTS funding showed a considerable increase for both genders. ICTS funding was first available in 2009. In 2009, eight male researchers and six female researchers were funded compared to 2022 where 52 male researchers and 38 female researchers were funded. The percentage of funded researchers within each gender is similar, even though the absolute number of funded male researchers consistently exceeds that of female researchers.

3.1.1.3 Scientific productivity

We found consistent, if small, differences in scientific productivity for male and female researchers. The median number of publications for male researchers ranged from 4 to 5 annually with interquartile ranges (IQR) of 2 to 9 publications. For female researchers, the medians are slightly lower, ranging from 3 to 4 publications, with IQRs of 1–6 publications. These differences are shown by the distribution of publication counts stratified by gender in Figure 7.

Figure 7

3.1.2 Collaboration network characteristics

Basic co-authorship network characteristics are presented in Table 2. Network size increased from 242 nodes in 2008 to 1,523 nodes in 2022. Its growth was most rapid in the first 5 years (2008–2012) after establishing ICTS in 2007. The average degree increased from 1.76 in 2008 to 7.35 in 2022. (This large increase in degree is mainly an indication that more of an ICTS member’s co-authors are themselves becoming ICTS members over time.) Modularity by gender averaged 0.04, ranging from 0.02 in 2013 to 0.07 in 2010. Modularity by discipline averaged 0.11 with a range of 0.08 to 0.14. Modularity scores for both characteristics are close to 0, indicating that the overall structure of the collaboration networks are not explained easily by gender or by discipline.

3.2 Differences in network position by gender

Centrality measures are used in network science to indicate the extent to which network members are centrally (or peripherally) positioned in an overall network. Average centrality scores stratified by gender are presented in Table 3. Degree scores are the number of direct co-authorship ties. Betweenness centrality is a measure of how often a network member sits ‘between’ other members in the network; this measure was normalized so that scores range from 0 to 1. For male researchers, the average degree centrality was 5.6, ranging from 1.7 in 2008 to 8.6 in 2021. In comparison, female researchers had an average degree centrality of 4.7, with a range from 2.0 in 2008 to 6.9 in 2021. The difference in average degree is shown in Figure 8. Average normalized betweenness centrality differed slightly between genders. Male researchers had an average of 0.0028, ranging from 0.0016 to 0.0051, while female researchers averaged 0.0021, with a range from 0.0014 to 0.0038. Taken together, the degree and betweenness scores suggest that there are small, but consistent gender differences in centrality in the co-authorship networks.

Figure 8

3.3 Dyadic patterns of collaboration by gender over time

A co-authorship tie between two ICTS members can have three different gender-to-gender patterns: female–female, male–male, and female–male. The first two are homophilous ties, the latter is a heterophilous tie, with respect to gender. The distribution of collaborations by collaborator gender are shown in Figure 9 (a table of frequencies and percentages for each year is available in Appendix C). In 2008, there were 94 Male–Male collaborations, 88 Male–Female collaborations, and 31 Female–Female collaborations. In 2022, there were 2,386 Male–Male collaborations, 2,350 Male–Female collaborations, and 858 Female–Female collaborations. Until 2016, Male–Male collaborations accounted for over half of all collaborations in a year. This trend declined, converging at 42% with Male–Female collaborations in 2022. The distribution of Female–Female collaborations remained relatively stable at 15% in 2008 and 2022.

Figure 9

3.4 Modeling collaboration ties

The preceding descriptive analyses of ICTS collaboration networks is somewhat equivocal when examining the role of gender. Although there are some stable gender differences, including females being slightly less productive than male researchers, distinctive disciplinary differences by gender, and female ICTS members having lower centrally in the co-authorship networks, these differences are somewhat small. However, we can examine the role of gender in collaboration networks more closely by building statistical network models of collaboration ties.

3.4.1 Model results

ERGM (exponential random graph) models were used to predict a publication co-authorship tie between any two ICTS members. Model summaries for 2009, 2016, and 2022 are presented in Table 4. (These years were selected to illustrate model results near the beginning of ICTS funding, in the middle, and near the end. Model results for all years are available in Appendix D). Models for each year are built up using three blocks of covariates. The first model is a null model with no covariates, which serves as a baseline for subsequent models. The second model adds academic and engagement covariates, along with an endogenous structural term. The third model introduces gender as a differential homophily term, testing whether dyadic gender patterns predict the likelihood of authorship collaboration. Parameter estimates and standard errors are presented within the table. The estimate represents the log-odds of tie formation, more specifically, log-odds of co-authoring a publication (Harris, 2013). A positive estimate indicates ties are more likely to form while a negative estimate indicates ties are less likely to form. The reference group for each categorical variable is listed next to the covariate name. The AIC value for each modeling block is listed below the covariates. AIC is a relative measure of model fit, and lower values indicate better fit of a model to the observed network data.

For the recent 2022 publication network, we see that the AIC gets smaller as covariates are added in the 2nd and 3rd models (going from 68,012 to 66,210.) The main effects for academic and engagement attributes are presented in the second model block. When considering the influence of researcher discipline on collaboration, with clinical discipline as the reference group, we found that social science researchers were more likely to collaborate compared to clinical researchers (0.27; p < 0.001). While both basic science and allied health researchers were less likely to collaborate compared to clinical researchers, the negative relationship was stronger for allied health researchers (−0.18; p < 0.001). Rank was entered as a scale variable with 1 as non-faculty, 2 as assistant professor, 3 as associate professor, and 4 as full professor. The positive Rank estimate indicates that tie formation is more likely as academic rank increases (0.11; p < 0.001). Collaboration is most likely for full professors and is least likely for non-faculty. Engagement measures were entered into the model as binary covariates. Researchers that used a core service in the previous year were much more likely to co-author publications than those that did not (0.35; p < 0.001). Similarly, researchers that received ICTS funding in the previous year were also more likely to co-author publications than those that did not (0.09; p < 0.01). The addition of academic and engagement covariates decreased the AIC to 69,422.19, indicating an improvement in model fitness. The structural GWDegree term was used to model the endogenous effects of the degree distribution in the collaboration network. Use of a structural covariate often improves the robustness and fit of an ERGM model, but it is not substantively interpreted here (Cranmer and Desmarais, 2011).

The third model adds a gender homophily term, with male–male as the reference group, to examine how gender homophily is related to the likelihood of co-authoring publications. In the 2022 collaboration network, tie formation was more likely to occur between two male researchers when compared to collaborations involving a male and female researcher dyad. Specifically, heterophilous collaboration pairs (male and female) were less likely to collaborate than male pairs (−0.21; p < 0.001). Female pairs showed a small negative log-odds, but this was not significant.

As described above, ERGM models were fit for each year from 2008 to 2022. In general, the parameter estimates for all of these models were similar to 2022, especially the heterophilous gender effect as well as the rank, discipline, and ICTS core utilization effects. A heatmap of odds ratios for each predictor in the full model for each year is shown in Figure 10 (odds ratios for the final model of each year are provided in Appendix E). Cross-hatched boxes indicate the odds ratio was not significant (p > 0.05) for that model year. Yellow shaded boxes indicate significant positive relationships (odds-ratios > 1), while blue shaded boxes indicate negative significant relationships (odds-ratios < 1). Basic science and allied health researchers were less likely to co-author publication compared to clinical researchers. This was more pronounced among allied health researchers. Social science researchers were more likely to co-author publications compared to clinical researchers, particularly in 2008–2010. Across all 15 years, co-authoring publications was more by members who had used core during the previous year. Also, more senior researchers had higher likelihood to collaborate with others. Finally, there were strong gender homophily effects on likelihood of co-authorship. Compared to two male co-authors, female and male co-authors were consistently less likely to collaborate.

Figure 10

3.4.2 Model diagnostics and goodness of fit

A strength of ERGM modeling (and Monte Carlo Markov Chain estimation more broadly) is the availability of diagnostic tools and information on how well an estimated model fits the observed data. In general, we want models that converge without problems, the MCMC process shows good mixing, and the final models show good fit to the observed network data (Cranmer and Desmarais, 2011). All the academic engagement and gender homophily models converged with no problems. The MCMC diagnostics show excellent mixing, where the model parameter estimates fit within the simulated quantiles, the lagged correlations approach 0 as the lags increase in the MCMC estimation, and the trace and density plots show that the MCMC chains are well-behaved (i.e., random patterns and posterior distribution centered on 0). The goodness of fit plots show that the individual model parameters (e.g., gender homophily) match the observed data well. However, other structural characteristics of the estimated models are less satisfying. The final models tend to overestimate the average degree of the nodes in the network (while still capturing the long-tail nature of the overall degree distribution) and underestimate the average geodesic distances in the simulated networks. So, while the models appear to be robust and the individual parameters are estimated well, we should be less confident that simulated networks would have general structural characteristics that match the observed networks. This suggests that these models may be missing some other covariates that would help produce more accurate simulated networks. Appendix F contains the detailed diagnostic and goodness of fit information for the final collaboration models.

4 Discussion

Using a rich research collaboration dataset spanning 15 years, we use network analytic modeling to reveal persistent gender differences in collaboration patterns, even when controlling for other individual-level factors. This study demonstrates that gender inequities in team science persist through subtle differences in opportunities within an otherwise integrated network. Longitudinal network models revealed that, regardless of academic rank, discipline, and institutional engagement, collaborations involving a male and a female researcher were significantly less likely to occur than collaborations between two male researchers, while female–female collaborations did not show a corresponding increase. Together, these findings indicate that gender disparities in collaboration are not explained by participation, engagement, or network growth alone. These findings add to the existing debate about structural determinants of equitable (or inequitable) research collaborations (Abramo et al., 2013).

Overall, the increase in collaboration across the entire network is consistent with other translational science programs (Yu et al., 2023). While growth occurred for both males and females, female researchers joined the network at a faster rate. Other studies have highlighted similar disparities in female researcher differences in rank and research discipline, suggesting that the likelihood for males to be full professors is greater than their female counterparts (Jena et al., 2015; Yu et al., 2023). Within this network, males were concentrated in clinical and basic science disciplines while female members had broader representation across all scientific fields. The distribution patterns in discipline remained relatively stable over time, suggesting enduring gender-based preferences or structural influences in discipline choice.

This study offers insights into how team collaboration functions within academia and the structures that contribute to collaboration beyond gender differences (Hall et al., 2018). This is important in how we consider scientific advances and our approaches to solving complex problems in healthcare. Our results complicate this interpretation by showing that gender remains a significant predictor of collaboration likelihood even after accounting for rank, discipline, and institutional engagement. These findings support theories of cumulative advantage and homophily, demonstrating how gendered patterns of tie formation can subtly reinforce inequality over time within ostensibly inclusive scientific networks. This has important implications for how we recruit, retain, and reward researchers, support equitable team science, and shape the research enterprise in ways that match the values of public health and clinical science (Asmal et al., 2022; Misra et al., 2017).

We further explored this concept to understand if individuals sub-grouped based on other factors aside from gender and discovered that individuals within these networks collaborate across discipline and academic rank, illustrating that across the ICTS network people are collaborating regardless of any individual-level factors. This is important, suggesting that a variety of perspectives are included in the development and dissemination of research.

A strength of this study was that it observed collaboration across 15 years, during a time of particularly impressive growth in translational sciences programs broadly (Jena et al., 2015; Llewellyn et al., 2018; Yu et al., 2023). This was demonstrated through overall network increase and the increase in collaboration for both males and females over time. In general, this allows us to understand how teams are growing and generating more science and the influence of team science is increasing how individuals collaborate within academic settings. Except for the first year, males had more collaboration ties within the network; however, some of this is due to the larger number of males in the network. When looking at the model, males were still more likely to collaborate with other males. This suggests that when male researchers enter the ICTS network, they are more likely to establish new network connections with other males in the same network. Females, overall, have fewer numbers of collaborative ties and are similarly less likely to form new ties with either female or male researchers when compared to their male counterparts. However, we do see that in the last 10 years, female involved dyads (female–female and female–male) are increasing relative to male–male co-authorship collaborations (Figure 9).

Over 15 years, the differences in female and male collaboratives have continued to persist, with females having fewer numbers of collaborators as well as fewer publications within the clinical and translational sciences programs. While efforts within these programs have provided opportunities for collaboration, it is necessary to consider potential opportunities to include and empower female scientists to have equitable opportunities for collaboration, publication, and the subsequent career opportunities that follow. Furthermore, our findings revealed that fewer female researchers hold higher academic ranks and are more often represented in scientific fields that exacerbate disparities in collaboration and productivity. These differences must be considered collectively, as they underscore systemic barriers that have long been recognized as challenges for female researchers in science.

However, it is important to recognize that there are differences between males and females with regards to their collaboration that continue over time, even when we control for other individual-level factors. Other work has presented this conceptually and through survey methods, but our structural analysis highlights that over time these differences in collaboration have sustained (Finkelstein et al., 2016; Fox and Nikivincze, 2021). Other literature suggests that female researchers form fewer, deeper collaborations that persist over time and continue to re-occur, which would explain our findings of female researchers having fewer number of collaborators (Bunker Whittington et al., 2024; Ductor et al., 2023). To better understand this, future work should account for instances of collaboration in addition to unique numbers of collaborators. Other research should investigate intersections of collaboration and publications, as well as other intersections of individual factors that influence collaboration activities. Additionally, efforts should be targeted at supporting female researchers, particularly those of lower academic rank, in finding and fostering new collaborations and supporting opportunities for increased engagement and network building. The ICTS networks are well positioned to develop and deliver this support, due to their infrastructure, capacity, and prominence.

This study examined collaboration networks of ICTS members at a time when membership was rapidly increasing. Growth can influence member composition which may complicate the comparison of 1 year to another. Rapid membership growth can also obscure the underlying drivers of collaboration, making it challenging to differentiate between collaborations arising from organic processes or programmatic efforts and those driven by the addition of new members. This study was conducted at a single academic institution in the Midwest, which may limit the generalizability of these findings to other institutions. We also acknowledge that our analysis is limited by the definition of gender using response options associated with sex (male or female). Also, we did not include other social identities such as race, ethnicity, age, and sexual orientation. The data available either lack these constructs altogether or all categories of the variables did not have sufficient representation in the data to be included for analyses. We realize that social identities are not experienced in isolation and recommend that future data collection efforts and analyses find ways to include these important characteristics. As in many academic collaboration studies, the data were limited to what were available in administrative databases. Thus, in future work it would be beneficial to use behavioral and perceptual data that touch on the lived experiences of scientific researchers. Finally, although the broader ICTS membership includes researchers outside of Washington University, because of data access limitations (notably, gender and rank) we were only able to include WashU scholars in the network analyses. This is an obvious limitation for generalizability, and it will be important in the future to look at collaboration data from other institutions.

A few notable strengths contribute to the significance of this study. First, we utilized rich data on researcher attributes collected over 15 years, providing a comprehensive view of the network’s composition and dynamics. This dataset allowed us to contextualize the network structures and statistics identified in our structural analysis, offering deeper insights into collaboration patterns. Additionally, the use of statistical models enabled us to identify significant factors influencing collaboration, facilitating a broader understanding of the determinants of collaborative relationships. Notably, we found that engagement with ICTS through core utilization consistently increased the likelihood of collaboration, underscoring the importance of programmatic support in fostering research partnerships.

Given our findings, as well as the design and analysis limitations discussed above, there are a number of important next steps to outline for this line of research. First, it will be important to replicate the gender homophily findings using similar co-author and research collaboration data from other institutions and initiatives. This will be particularly important for research from smaller, less-well funded academic institutions, as well as from outside the U.S. Second, the analyses described here use a cross-sectional approach to look at annual collaboration data across 15 years. An obvious next step is to treat these network data as longitudinal, which could allow us to identify predictors of changes in network composition or ties over time. New network modeling approaches such as STERGMs (separable temporal exponential random graph models) would prove useful here (Krivitsky and Handcock, 2014).

References

• 1

  AbdallaS.AbdallaM.SaadM.JonesD.PodolskyS.AbdallaM. (2023). Ethnicity and gender trends of UK authors in the British medical journal and the lancet over the past two decades: a comprehensive longitudinal analysis. EClinicalMedicine64:102174. doi: 10.1016/j.eclinm.2023.102174,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 2

  AbramoG.D’AngeloC. A.MurgiaG. (2013). Gender differences in research collaboration. J. Informetr.7, 811–822. doi: 10.1016/J.JOI.2013.07.002

  • CrossRef
  • Google Scholar

• 3

  AsmalL.LampG.TanE. J. (2022). Considerations for improving diversity, equity and inclusivity within research designs and teams. Psychiatry Res.307:114295. doi: 10.1016/J.PSYCHRES.2021.114295,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 4

  AthanasiouT.PatelV.GarasG.AshrafianH.HullL.SevdalisN.et al. (2016). Mentoring perception, scientific collaboration and research performance: is there a ‘gender gap’ in academic medicine? An academic health science Centre perspective. Postgrad. Med. J.92, 581–586. doi: 10.1136/POSTGRADMEDJ-2016-134313,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 5

  BearJ. B.WoolleyA. W. (2011). The role of gender in team collaboration and performance. Interdiscip. Sci. Rev.36, 146–153. doi: 10.1179/030801811X13013181961473

  • CrossRef
  • Google Scholar

• 6

  BennettL. M.GadlinH. (2012). Collaboration and team science: from theory to practice. J. Investig. Med.60, 768–775. doi: 10.2310/JIM.0B013E318250871D,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 7

  BornerK.ContractorN.Falk-KrzesinskiH. J.FioreS. M.HallK. L.KeytonJ.et al. (2010). A multi-level systems perspective for the science of team science. Sci. Transl. Med.2:49cm24-49cm24. doi: 10.1126/scitranslmed.3001399,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 8

  Bunker WhittingtonK.KingM. M.CingolaniI. (2024). Structure, status, and span: gender differences in co-authorship networks across 16 region-subject pairs (2009–2013). Scientometrics129, 147–179. doi: 10.1007/s11192-023-04885-1

  • CrossRef
  • Google Scholar

• 9

  CainC. L.LeaheyE. (2014). Cultural correlates of gender integration in science. Gender Work Organ.21, 516–530. doi: 10.1111/GWAO.12052

  • CrossRef
  • Google Scholar

• 10

  CombsT. B.CarothersB. J.LiuY.EvanoffB.LukeD. A. (2021). If you build it, will they come? Linking researcher engagement and scientific productivity in large infrastructure grants. J. Clin. Transl. Sci.5:e90. doi: 10.1017/CTS.2021.17

  • CrossRef
  • Google Scholar

• 11

  CranmerS. J.DesmaraisB. A. (2011). Inferential network analysis with exponential random graph models. Polit. Anal.19, 66–86. doi: 10.1093/PAN/MPQ037

  • CrossRef
  • Google Scholar

• 12

  DhandA.LukeD. A.CarothersB. J.EvanoffB. A. (2016). Academic cross-pollination: the role of disciplinary affiliation in research collaboration. PLoS One11:e0145916. doi: 10.1371/JOURNAL.PONE.0145916,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 13

  DuctorL.GoyalS.PrummerA. (2023). Gender and collaboration. Rev. Econ. Stat.105, 1366–1378. doi: 10.1162/REST_A_01113

  • CrossRef
  • Google Scholar

• 14

  FinkelsteinM. J.ConleyV. M.SchusterJ. H. (2016). Taking the measure of faculty diversity. Advancing higher education. 1:13.

  • Google Scholar

• 15

  FoxM. F.NikivinczeI. (2021). Being highly prolific in academic science: characteristics of individuals and their departments. High. Educ.81, 1237–1255. doi: 10.1007/s10734-020-00609-z

  • CrossRef
  • Google Scholar

• 16

  GoodreauS. M. (2007). Advances in exponential random graph (p*) models applied to a large social network. Soc. Networks29, 231–248. doi: 10.1016/J.SOCNET.2006.08.001,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 17

  GouldenM.MasonM. A.FraschK. (2011). Keeping women in the science pipeline. Ann. Am. Acad. Polit. Soc. Sci.638, 141–162. doi: 10.1177/0002716211416925

  • CrossRef
  • Google Scholar

• 18

  GrayB. (1994). A feminist critique of collaborating. J. Manage. Inq.3, 286–293. doi: 10.1177/105649269433011

  • CrossRef
  • Google Scholar

• 19

  HallK. L.VogelA. L.HuangG. C.SerranoK. J.RiceE. L.TsakraklidesS. P.et al. (2018). The science of team science: a review of the empirical evidence and research gaps on collaboration in science. Am. Psychol.73, 532–548. doi: 10.1037/AMP0000319,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 20

  HarrisJ. (2013). An introduction to exponential random graph modeling. Vol. 173. Thousand Oaks, CA: Sage Publications.

  • Google Scholar

• 21

  HeidariS.BaborT. F.De CastroP.TortS.CurnoM. (2016). Sex and gender equity in research: rationale for the SAGER guidelines and recommended use. Res. Integr. Peer Rev.1:2. doi: 10.1186/S41073-016-0007-6/TABLES/2,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 22

  HeijstraT.BjarnasonT.RafnsdóttirG. L. (2015). Predictors of gender inequalities in the rank of full professor. Scand. J. Educ. Res.59, 214–230. doi: 10.1080/00313831.2014.904417

  • CrossRef
  • Google Scholar

• 23

  HiltonM.CookeN. (2015). Enhancing the effectiveness of team science. Washington, DC: National Academies Press.

  • Google Scholar

• 24

  HolmanL.MorandinC. (2019). Researchers collaborate with same-gendered colleagues more often than expected across the life sciences. PLoS One14:e0216128. doi: 10.1371/JOURNAL.PONE.0216128,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 25

  JenaA. B.KhullarD.HoO.OlenskiA. R.BlumenthalD. M. (2015). Sex differences in academic rank in US medical schools in 2014. JAMA314, 1149–1158. doi: 10.1001/JAMA.2015.10680,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 26

  KimJ. Y.HowardM.Cox PahnkeE.BoekerW. (2016). Understanding network formation in strategy research: exponential random graph models. Strateg. Manag. J.37, 22–44. doi: 10.1002/SMJ.2454

  • CrossRef
  • Google Scholar

• 27

  KrivitskyP. N.HandcockM. S. (2014). A separable model for dynamic networks. J. R. Stat. Soc. Series B Stat. Methodol.76, 29–46. doi: 10.1111/RSSB.12014

  • CrossRef
  • Google Scholar

• 28

  KutnerM.NachtsheimC.NeterJ.LiW. (2005). Applied linear statistical models. New York: McGraw Hill.

  • Google Scholar

• 29

  LlewellynN.CarterD. R.RollinsL.NehlE. J. (2018). Charting the publication and citation impact of the NIH clinical and translational science awards (CTSA) program from 2006 through 2016. Acad. Med.93, 1162–1170. doi: 10.1097/ACM.0000000000002119,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 30

  LotrecchianoG. R.BennettL. M.VovidesY. (2023). A framework for developing team science expertise using a reflective-reflexive design method (R2DM). Humanit. Soc. Sci. Commun.10:810. doi: 10.1057/s41599-023-02298-2

  • CrossRef
  • Google Scholar

• 31

  LoveH. B.StephensA.FosdickB. K.TofanyE.FisherE. R. (2022). The impact of gender diversity on scientific research teams: a need to broaden and accelerate future research. Humanit. Soc. Sci. Commun.9:386. doi: 10.1057/s41599-022-01389-w

  • CrossRef
  • Google Scholar

• 32

  LukeD. A.CarothersB. J.DhandA.BellR. A.Moreland-RussellS.SarliC. C.et al. (2015). Breaking down silos: mapping growth of cross-disciplinary collaboration in a translational science initiative. Clin. Transl. Sci.8, 143–149. doi: 10.1111/CTS.12248,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 33

  MisraJ.Smith-DoerrL.DasguptaN.WeaverG.NormanlyJ. (2017). Collaboration and gender equity among academic scientists. Soc. Sci.6:25. doi: 10.3390/SOCSCI6010025

  • CrossRef
  • Google Scholar

• 34

  PezzoniM.VisentinF. (2024). Gender bias in team formation: the case of the European Science Foundation’s grants. Sci. Public Policy51, 247–260. doi: 10.1093/SCIPOL/SCAD058

  • CrossRef
  • Google Scholar

• 35

  RawlingsC. M.SmithJ. A.MoodyJ.McFarlandD. A. (2023). Network analysis: integrating social network theory, method, and application with R. Cambridge, UK: Cambridge University Press.

  • Google Scholar

• 36

  RossiterM. W. (1993). The Matthew Matilda effect in science. Soc. Stud. Sci.23, 325–341. doi: 10.1177/030631293023002004

  • CrossRef
  • Google Scholar

• 37

  SantosJ. M.HortaH.FengS. (2024). Homophily and its effects on collaborations and repeated collaborations: a study across scientific fields. Scientometrics129, 1801–1823. doi: 10.1007/s11192-024-04950-3

  • CrossRef
  • Google Scholar

• 38

  SatoS.GygaxP. M.RandallJ.Schmid MastM. (2020). The leaky pipeline in research grant peer review and funding decisions: challenges and future directions. High. Educ.82, 145–162. doi: 10.1007/S10734-020-00626-Y,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 39

  SchmalingK. B.GalloS. A. (2023). Gender differences in peer reviewed grant applications, awards, and amounts: a systematic review and meta-analysis. Res. Integrity Peer Rev.8:2. doi: 10.1186/S41073-023-00127-3,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 40

  Ten HagenK. G.WolinetzC.ClaytonJ. A.BernardM. A. (2022). Community voices: NIH working toward inclusive excellence by promoting and supporting women in science. Nat. Commun.13:1682. doi: 10.1038/s41467-022-28665-2,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 41

  U.S. Department of Health and Human Services (n.d.). Information and Instructions for completing statement of appointment (Form PHS 2271). Available online at: http://aspe.hhs.gov/ (Accessed December 28, 2025)

  • Google Scholar

• 42

  WalkerL. E.HeatonH. A.KohlerK. (2024). Gender dynamics and academic rank in emergency medicine collaboration networks: a social network and gender propensity analysis. Am. J. Emerg. Med.83, 40–46. doi: 10.1016/J.AJEM.2024.06.037,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 43

  WassermanS.FaustK. (1994). Social network analysis: methods and applications. Cambridge, UK: Cambridge University Press.

  • Google Scholar

• 44

  WuchtyS.JonesB. F.UzziB. (2007). The increasing dominance of teams in production of knowledge. Science316, 1036–1039. doi: 10.1126/SCIENCE.1136099/SUPPL_FILE/WUCHTY.SOM.PDF,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 45

  YangY.TianT. Y.WoodruffT. K.JonesB. F.UzziB. (2022). Gender-diverse teams produce more novel and higher-impact scientific ideas. Proc. Natl. Acad. Sci. USA119:e2200841119. doi: 10.1073/PNAS.2200841119,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 46

  YuF.PatelT.CarnegieA.DaveG. (2023). Evaluating the impact of a CTSA program from 2008 to 2021 through bibliometrics, social network analysis, and altmetrics. J. Clin. Transl. Sci.7:e44. doi: 10.1017/CTS.2022.530,

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 47

  ZhangM. Z.WangT. R.LyuP. H.ChenQ. M.LiZ. X.NgaiE. W. T. (2024). Impact of gender composition of academic teams on disruptive output. J. Informetr.18:101520. doi: 10.1016/J.JOI.2024.101520

  • CrossRef
  • Google Scholar