Female authorship of covid-19 research in manuscripts submitted to 11 biomedical journals: cross sectional study

Study design and settings

This cross sectional study is ancillary to a broader research project, the ATHENA study, assessing whether gender bias exists in the editorial decision making process. Both studies are being conducted under a confidentiality agreement between BMJ Publishing Group and the medical school of Geneva University represented by the Department of Health and Community Medicine.

Editors of all BMJ Publishing Group journals were invited to take part in the ATHENA study. We selected a subsample of 11 journals of varying size and impact factor from a range of biomedical disciplines, including two large general medical journals (The BMJ and BMJ Open) and nine specialist journals (BMJ Global Health; Occupational and Environmental Medicine; Heart; Gut; Thorax; Journal of Neurology, Neurosurgery and Psychiatry; Postgraduate Medical Journal; Sexually Transmitted Infections; and Tobacco Control). As the focus was on research submissions, participating journals needed to have clear submission categories for original research, protocols, and/or systematic reviews on ScholarOne (the manuscript submission system) to be eligible to take part, as well as being published by BMJ Publishing Group for at least eight years to enable comparisons over time.

Participants and data source

Using the ScholarOne article types selected by authors on submission, we selected all research manuscripts (original research, systematic reviews and/or meta-analysis), including research study protocols, that were submitted for publication to one of the 11 participating journals between 1 January 2018 and 31 May 2021. We excluded all other article types on the basis of the ScholarOne submission categories indicated by submitting authors. We identified manuscripts that were submitted to one of the 11 journals and then transferred to another of these journals in the study period by using “manuscript transfer ID,” “transferring journal name,” “receiving journal name,” and “manuscript title” and included only the first submission. We extracted data from the ScholarOne manuscript submission systems.

Factors and outcomes

We defined two submission periods: the “pre-pandemic period” (1 January 2018 to 31 December 2019) and the “pandemic period” (1 January 2020 to 31 May 2021). We further defined four time periods during the pandemic period corresponding roughly to the evolution of the pandemic worldwide: Jan-May 2020, Jun-Sep 2020, Oct 2020-Jan 2021, and Feb-May 2021. In total, we had five time-periods and nine different groups (including covid-19 versus non-COVID-19 manuscripts in the pandemic period).

To identify manuscripts describing studies on patients with covid-19 or studies on the impact of covid-19 on health services and policies, we further categorised manuscripts submitted in the pandemic period as “covid-19” manuscripts or “pandemic non-covid-19” manuscripts on the basis of presence/absence of the following keywords in the title and/or abstract variables: “covid,” “covid-19,” “SARS-CoV-2,” “coronavirus,” “2019-nCov,” “coronavirinae,” “coronaviridae,” “hCoV-19,” or “SARS and virus.” To estimate the accuracy of our search strategy for covid-19 relevant manuscripts, two independent raters (AGA and SS) assessed the titles and abstracts of a random sample of 300 manuscripts identified as related to covid-19 and indicated their agreement.

For each manuscript, we collected two levels of information. At the manuscript level, we recorded the following variables: journal name, manuscript ID, title, abstract, date of original submission, number of co-authors on the byline, date of transfer to another journal where applicable, name of the receiving journal, and editorial decision by 11 August 2021. At the author level, we recorded the following variables: author ID; author’s first, middle, and last names; author’s salutation; author’s country; rank authorship order on the article byline (several co-authors could hold the first authorship position, but we used only first declared author), and corresponding authorship status.

Our primary outcome was the first author’s gender. Secondary outcomes were the last and corresponding author’s gender and the percentage of women among all co-authors on the article byline. We evaluated primary and secondary outcomes among nine groups: pre-pandemic and either covid-19 or pandemic non-covid-19 manuscripts in Jan-May 2020, Jun-Sep 2020, Oct 2020-Jan 2021, and Feb-May 2021.

Gender determination

We determined gender by using a sequential four step procedure. Firstly, we used both first name and country to search in Gender API (https://gender-api.com/en) website. Gender API provides gender determination with an accuracy probability from 50% to 100% (under 50% an unknown status is attributed), gives the number of samples retrieved to search for the information, ensures that the level of misclassification is around 5%, and has the highest coverage on the variability in name origins.²² We set gender accuracy above 80% as the minimum criterion for inclusion in the primary gender determination. Secondly, for all authors with undetermined or unknown gender based on the above, we used the authors’ given middle names and country to search in Gender API. Thirdly, we used the online service genderize.io (http://genderize.io) to determine gender on the basis of first names and middle names, also using an accuracy above 80% as the minimum criterion. Fourthly, for still undetermined gender, we used the authors’ salutation and attributed with 100% accuracy male gender to “Mr” or “M” and female gender to “Miss,” “Mrs,” and “Ms.” Where none of these methods determined the gender, we recorded the gender as undetermined.

Statistical analysis

Our sample size was fixed (n=63 259 manuscripts, including 34 222 pre-pandemic manuscripts, 4447 covid-19 manuscripts, and 24 590 pandemic non-covid-19 manuscripts); these numbers allowed us to detect at least a 3.93% decrease in the proportion of female first authors between pre-pandemic and covid-19 related submissions with 90% study power and 5% type I error (two sided), taking into account that the data were clustered within the country of affiliation of first authors nested within journals (we took mean cluster size at 30 and intraclass correlation at 0.05, leading to a design effect of 2.45).

We described continuous variables by mean (standard deviation) or median (interquartile range) and presented categorical variables as frequencies and relative proportions. We plotted the daily number of manuscripts submitted across time by using time series smoothed curves with a moving average by seven days.

We described outcomes and manuscripts’ characteristics in the nine groups. We compared proportions of manuscripts submitted to generalist versus specialised journals among time groups by using a χ² test. We used a Kruskal-Wallis non-parametric test for comparison of total number of co-authors among time groups. For outcomes, we reported P values assessing outcomes across the nine groups from the univariate models described below.

We used bivariate mixed effects logistic regression models to test associations between groups (covid-19 and pandemic non-covid-19 versus pre-pandemic manuscripts) and first author’s gender by journal specialty. We assessed the probability of the first author being a woman across the nine groups, independently of last author’s gender, total number of co-authors, and categories of journal (generalist versus specialist). We further adjusted for the level of missing gender determination at country level (<25%, 25-50%, and ≥50% versus none missing). We also compared the chance of being a female first author on covid-19 and pandemic non-covid-19 manuscripts versus pre-pandemic manuscripts. We graphically verified the linearity assumption of logit for continuous variables. We used mixed effects logistic regression models with first author’s gender as the dependent variable (model 1). To account for some clustering in the data at the journal and country of affiliation levels, we included nested random factors on the intercept by using the country of affiliation within journals. To increase the precision of estimates, we restricted the analyses to countries with at least three submissions per reported group. We used similar models to assess the probability of the last author (model 2) and corresponding author (model 3) being a woman. For model 3, we also adjusted for the position of the corresponding author on the article byline (last, first, or other position). We interpreted global P values as significant at the adjusted cut-off 0.0125, using the Bonferroni method to take into account multiple comparisons (0.05/4 outcomes). We ran these regression models on complete cases and reported adjusted odds ratios, 98.75% confidence intervals, and P values. We did sensitivity analyses with accuracy of gender determination above 95%. We also did post hoc sensitivity analyses to quantify the potential of missing gender determination (for model 1, model 2, and model 3) by using multiple imputation for missing at random,²³ with M=20 imputations and adjustment of the imputation model for author’s country, nine time groups, and journal specialty.

For the assessment of the percentage of women among all co-authors across groups (model 4), we reported results only for manuscripts that had gender determined for at least 80% of all co-authors. Firstly, we plotted the number of women co-authors among the last author’s region of affiliation (as a proxy of research origin), to visualise distribution (supplementary figure C). We observed two different patterns between authors with affiliation in Europe, North America, Latin America, or Oceania (region 1) and those with affiliation in Africa, China, or the rest of Asia (region 2). For manuscripts from region 1, we further used a mixed effects Poisson regression model with total number of co-authors on the byline and with determined gender as offset, nested random factors on the intercept using the last author’s country of affiliation within journals, and we restricted the analyses to countries with at least three submissions per reported group (model 4A). For manuscripts from region 2, we used a second mixed effects Poisson regression model with nested random factors on the intercept and using last author’s geographical region (and not country) of affiliation within journals (model 4B). We adjusted both models for the type of journal (generalist versus specialist) and for the level of missing gender determination at country level. We verified the goodness of fit of the Poisson models by a visual inspection of Anscombe’s residuals. We reported incidence rate ratios, 98.75% confidence intervals, and P values.

We interpreted global P values as significant at the adjusted cut-off of 0.0125, with only P<0.0125 yielding a confidence interval not crossing the null value. We used Stata 16 intercooled and R software (version 3.3.1) for analyses.

Patient and public involvement

We partnered and co-authored with the chair of the BMJ’s LGBTQ+ network (MR). MR contributed to almost every aspect of the study (data collection, gender determination, manuscript revision, final approval).

Studied population

A total of 63 259 manuscripts were included in our study among the 75 731 manuscripts submitted for publication to the 11 journals (supplementary figure A), with 551 341 authors across all included manuscripts (supplementary table A). Corresponding authorship overlapped with first and last authorship in 30 703 (48.5%) and 25 039 (39.6%) manuscripts. Gender was determined for 445 052 (80.7%) of authors (supplementary table A). Supplementary figure B shows the gender determination process. We found a high level of agreement (κ=0.95) between raters on the identification of covid-19 manuscripts.

Descriptive analysis

Manuscript submissions during 2020 increased by 18.6% compared with 2019 and by 23.3% compared with 2018 (fig 1); 15.3% (4447/29 037) of submissions during the pandemic were covid-19 related manuscripts that were submitted as a priority to generalist journals (table 1). We observed similar proportions for editorial decisions (we did not consider manuscripts with no decision yet here to calculate proportions) on manuscripts submitted before the pandemic (n=34 202) and on pandemic non-covid-19 manuscripts (n=20 668): respectively, 24 964 (73.0%) versus 15 468 (74.8%) were rejected, 8884 (26.0%) versus 4855 (23.5%) were accepted, and 354 (1.0%) versus 345 (1.7%) were still under revision by 11 August 2021 (table 1; supplementary table A). Among covid-19 manuscripts (n=3740), we observed a higher proportion of rejected manuscripts (83.8%; n=3134) compared with pre-pandemic and pandemic non-covid-19 manuscripts, as well as a lower proportion of accepted manuscripts (15.2%; n=568) and manuscripts under revision (1.0%; n=38) (table 1; supplementary table A). Total numbers of co-authors on the manuscript byline were significantly higher by a mean of 2 additional co-authors in covid-19 manuscripts compared with pre-pandemic and pandemic non-covid-19 manuscripts.

Overall, we observed an underrepresentation of women in prominent authorship positions (first 22 620 (44.9%), last 16 037 (31.1%), and corresponding 19 373 (37.7%)), and among all co-authors (median percentage of female co-authors 35.7%). The number of female first, last, and corresponding authors respectively were 1313 (37.1%), 996 (27.9%), and 1119 (31.1%) for covid-19 manuscripts (lowest values in Jan-May 2020: 230 (29.4%), 165 (21.1%), and 185 (22.9%)), compared with 8583 (44.9%), 6118 (31.2%), and 7273 (37.3%) for pandemic non-covid-19 manuscripts and 12 724 (46.0%), 8923 (31.4%), and 10 981 (38.9%) for pre-pandemic manuscripts (supplementary table A). The proportions of female first, last, and corresponding authors were significantly lower in covid-19 manuscripts (−8.9%, −3.5%, and −7.8%, respectively) in comparison with pre-pandemic manuscripts (supplementary table A). The drop in proportions of female first, last, and corresponding authors in comparison with pre-pandemic was the greatest in Jan-May 2020 (−16.6%, −10.3%, and −16.0%, respectively; table 1), the deepest being in March-April 2020 (fig 2). These gender differences narrowed in the most recent period compared with the early pandemic period and were very similar to values observed for pre-pandemic manuscripts; for last authorship, the proportion of women was even higher in the two last periods of 2021 than pre-pandemic. The proportions of female first, last, and corresponding authors in pandemic non-covid-19 manuscripts were very similar to pre-pandemic manuscripts (table 1; supplementary table A; fig 2). The percentage of female authors on the byline was lower in covid-19 manuscripts in comparison with pre-pandemic manuscripts (supplementary table A); the drop was the greatest in Jan-May 2020 (median 28.6% female co-authors compared with 36.4% in pre-pandemic; table 1). The odds of the first author being a woman were significantly lower in covid-19 manuscripts compared with pre-pandemic manuscripts for most specialties (supplementary table B). For generalist journals, representing the majority of the dataset, the odds of the first author being a woman was 0.74-fold (98.75% confidence interval 0.68 to 0.81) lower in covid-19 manuscripts compared with pre-pandemic manuscripts; it was significantly higher by 1.05-fold (1.00 to 1.11) in non-covid-19 manuscripts compared with pre-pandemic manuscripts.

In all groups, most authors had a European affiliation, with some variations among first, last, and corresponding authors (supplementary table C), except for covid-19 manuscripts submitted in Jan-May 2020, in which the proportions of authors with a Chinese affiliation were the highest (474 (43.0%), 451 (40.9%), and 460 (41.7%) of first, last, and corresponding authors compared with 6479 (18.9%), 6192 (18.1%), and 6334 (18.5%) in pre-pandemic manuscripts). By June-Sept 2020, the distribution of authors’ affiliations returned to a pattern similar to that seen in the pre-pandemic period, with European affiliations being the most frequent. The proportions of missing gender for first, last, and corresponding authors were the highest for authors with a Chinese affiliation (supplementary table D). Among authors with determined gender, we observed gender bias against women in most countries, except in Latin America and Oceania where first authors were women in respectively 646 (54.4%) and 1644 (57.9%) (supplementary table D). Men predominantly occupied the last author and the corresponding author position in all regions.

Multivariable analysis

Comparison with other studies

Our findings are broadly in line with the three previous studies of gender disparities in research submissions during the early months of the pandemic (up to May 2020). These studies also showed that women submitted proportionately fewer articles overall (from −0.9% up to −24%) than men in the early months of the pandemic and authored fewer covid-19 related articles overall (−7.6% fewer female than male co-authors in health and medicine).^8102021 Bell et al’s smaller study of submissions to American Public Health showed the proportions of first and corresponding authorships to be significantly higher for men (36% and 42% respectively) than women (12% and 11% respectively) between April and May 2020 compared with April to May 2019, but with variation by country. Similarly to us, Bell et al also reported that less than a third of covid-19 related submissions had female corresponding authors.²⁰ Williams et al’s analysis of the increased volume of submissions to the American Journal of Pediatrics during January to May 2020 found that the greatest increase was in international male corresponding authors.¹⁰ Like Squazzoni et al, who looked at the number of manuscripts submitted by female and male authors (regardless of authorship position) in the opening months of the pandemic, we found a lower proportion of female than male co-authors in the covid-19 manuscripts (median 33.3%), which was even lower in January to May 2020 (median 28.6%), compared with the pre-pandemic period (median 36.4%).²¹ We found similar results to a smaller study of articles, news, and opinion pieces on covid-19 published before 1 May 2020, reporting that 31% (95% confidence interval 29% to 33%) of articles were co-authored by women during the first semester of 2020.⁸ In this study, Pinho-Gomes reported even more pronounced gender disparities than in our study, with 29% of first and 26% of last authors being women. However, these values may be lower than ours owing to the inclusion of a large range of article types other than research, use of a different method to distinguish covid-19 related articles, and a focus on publications in the opening months of the pandemic when working patterns will have been most disrupted. Lerchenmüller et al’s longitudinal study of covid-19 related articles published up to 31 January 2021 found that the proportion of women in first and last authorship positions started to trend backwards to values observed before the pandemic.¹⁴ We observed a similar trend regarding first, last, and, additionally, corresponding authorship positions over a longer study period (up to 31 May 2021).

Similarly to others, we found some association between authorship position on the byline and the geographical affiliation of authors.^{3142024252627} Bell et al reported that the increase in submission rates between the pre-pandemic and pandemic periods was greater for male than for female first and corresponding authors affiliated to a non-US country versus the US.²⁰ Pinho-Gomes et al reported that the percentage of female co-authors on published research was lowest in Africa and highest in Oceania.⁸ Lerchenmüller et al outlined deeper gender disparities in first authorship positions at the beginning of the pandemic with some variations according to geographical area, with a greater gender gap in Africa and lesser one in Asia compared with our findings.¹⁴ Beyond scientific production, in many countries very few women researchers work in biomedicine and even fewer in key positions, as women’s place in society, and in science, depends on the socio-cultural context of the country.²⁵ Therefore, meta-research is needed to explore these countries’ academic systems and research practices with the goal of assessing policies that might explain these differences.

Strengths and limitations of study

Although our study included a self-selected sample of journals from one publisher, we selected 11 journals from a range of biomedical specialties including two large general medical journals attracting submissions from around the world, with a combination of hybrid and open access journals with varying journal impact factors. Earlier studies looking at the effects of the pandemic on gender inequalities focused on the opening months and captured only the first international lockdowns; however, we have included submissions across the whole of 2020 until 31 May 2021, as multiple lockdowns occurred and international differences existed in policies and stipulations that were difficult to account for. We included a longer pandemic period and further divided the pandemic into four time periods to try to account for seasonal differences. We also tried to adjust for disparities between countries regarding lockdown policies and their impact on scientific production by including country as a random factor in our statistical analyses. We found the deepest fall in women’s attainment of prominent authorship positions at the beginning of the pandemic (between 1 January and 31 May 2020), when most countries applied strict lockdown measures, including school and university closures. The urgent need for evidence on covid-19 during the pandemic led to a need for an expedited research and publication process. By identifying submissions related to covid-19, we were able to assess the real current impact of the pandemic and lockdown measures on gender disparities in research production and authorship on a prominent and competitive topic. However, our search strategy, similarly to others,²¹ to identify covid-19 relevant manuscripts was over-inclusive, as authors tended to refer to the pandemic to make their manuscripts more topical.

To minimise missing data and misclassification bias, we determined authors’ gender by using a four step sequential procedure and set a high threshold for data accuracy within the software used. Although we were unable to determine gender for 19.3% of all authors across all submissions, we thought that using a high threshold was more important than getting more genders determined with a lower threshold for accuracy. We adjusted the models on gender missingness and did sensitivity analyses using accuracy of gender determination above 95% and multiple imputation to replace missing gender determination data; both sensitivity analyses have confirmed our original results. However, by using software that determines gender in a binary form—namely, “male” and “female” based solely on name and country—we were unable to take into account the gender identity of individuals and may have misrepresented people with diverse gender identities. Further research in this area should try to integrate a broader definition of gender; however, for this to be conducted in large samples, these data will need to be self-reported routinely on journal submission systems.²⁸ More generally, a greater need exists to evaluate the role of gender, ethnicity, and cultural diversity on many aspects of the production of science.²⁹ Finally, we excluded duplicate manuscript submissions by identifying those that were transferred to another journal and removed those with an identical title and a submission date later than the original. We acknowledge that our method to identify duplicates may be imperfect, but it probably covers the vast majority of manuscripts, as automatic transfers use exactly the same manuscript title.

Policy implications

Academic promotion and awarding of research grants are heavily influenced by the publication records of candidates, and competition for prominent authorship positions (first, last, corresponding) is high.³⁰ Studies of published research have already shown gender disparities,^131323334 but our demonstration of even wider disparities at the submission stage during the pandemic suggests that we may observe even wider gender disparities in published research over the next few years once these manuscripts have reached publication stage. This may have important implications for career development, awarding of grant applications, and job opportunities for women.

During the covid-19 pandemic, women, particularly younger women, seem to have paid a heavier price than men from the lockdown measures and the need to reorganise work and research environments while assuming childcare, home schooling, housekeeping, and domestic work.^57353637 Female researchers have been less involved in research related to covid-19, and subsequently they have produced proportionately fewer research manuscripts, reflected by lower submission rates among female first authors compared with male first authors.⁷ In a survey conducted among 3345 Brazilian academics and research centres between April and May 2020, women with children, especially younger children, submitted significantly fewer research manuscripts than did men with children, with an even bigger difference compared with men without any children. This survey additionally showed that the intersection between race, gender, and parenthood has influenced academic productivity, with black female researchers being the most effected by covid-19.³⁸ Furthermore, in general, teaching responsibilities are shouldered more by women faculty members than men,³⁹ and women may have had even less time to write and submit research manuscripts during the pandemic as a result of the increased demands from the rapid shift to online teaching and an increase in mentorship duties and the need to support the mental health of students.³⁶⁴⁰ This slowdown in the race for scientific production will undoubtedly have practical implications for the near future. The consequences will be worse for young women researchers if adjustment measures are not rapidly implemented at academic institutions and competitive funding agencies.

People have proposed the introduction of voluntary disclosure of gender as part of the submission process to overcome the systematically low representation of women in authorship of scientific manuscripts and the application of gender quotas for the appointment to prominent academic positions to correct for gender disparities in the research environment.⁸⁴¹ The systematic participation of all women researchers in mentorship programmes might also be an important area to develop because professional social networks seem to be more prominent among men than women.⁴²⁴³ This might explain why male dominant environments lead to an amplification of homophily and why, despite the support of young female researchers by senior female researchers, gender inequalities in scientific production still persist.⁴⁴ We found that when a woman was last author, women were more likely to hold a prominent authorship position. The benefits of more senior women (reflected by last authorship position) supporting other women as first authors or more generally the promotion of female participation in research by other women has already been discussed and demonstrated by others.^45464748 This mechanism of sisterhood promotion, which we observed in our study, has also been shown by others through a high correlation between the percentage of tenured female professors and the percentage of female full professors at universities.⁴⁹ Implicit bias training may also help assessors in academia to evaluate women more equally for academic promotion.⁴⁷ Finally, journals can also try to reduce gender bias by implementing policies to encourage gender balance in the editorial staff and advisory boards of journals, as well as among those invited to peer review manuscripts and author commissioned articles.⁵⁰⁵¹

Conclusions

Patterns in authorship need to be routinely scrutinised to draw attention to the continued imbalance in female representation until parity can be reached. The low rates of female first, last, and corresponding authorship on submissions may fall further as the pandemic continues to cause disruption to working patterns and may be compounded once the new research that was conceived and designed during the pandemic flows through to publication stage. Thus, our findings should give early warning of inequity to those who rely on authorship of published articles as a metric of productivity to inform academic promotion and award research grants.

Ethical approval

Ethical approval for the ATHENA project was received from the Institutional Review Board of the Geneva Canton on 16 April 2018 (No 2019-00540).