PaperlayerView exact revision on arXiv

Source-linked AI summary

Historical comparison of gender inequality in scientific careers across countries and disciplines

Junming Huang, Alexander J. Gates, Roberta Sinatra, Albert-Laszlo Barabasi

arXiv:1907.04103v1cs.DLphysics.soc-ph

TL;DR

Evidence that men publish more over their careers leaves open how career length relates to productivity gaps. Measuring career length from first to last publication, the study finds productivity and impact gender gaps increased over time.

  • Problem

    Evidence that men publish more over their careers leaves unresolved how productivity differences relate to career length.

    Source: pp. 3, 7

  • Method

    The study measures career length as the time between an author's first and last publication to examine gender gaps in productivity.

    Source: pp. 7–8

  • Results

    Productivity and impact gender gaps increased steadily over time.

    Source: p. 6

  • Takeaways & Limitations

    Understanding gender gaps requires distinguishing productivity differences from differences in career length.

    Source: p. 7

    The analysis does not include currently active careers, so further work is needed to assess the impact of recent efforts.

    Source: p. 14

Abstract

from arXiv · show

There is extensive, yet fragmented, evidence of gender differences in academia suggesting that women are under-represented in most scientific disciplines, publish fewer articles throughout a career, and their work acquires fewer citations. Here, we offer a comprehensive picture of longitudinal gender discrepancies in performance through a bibliometric analysis of academic careers by reconstructing the complete publication history of over 1.5 million gender-identified authors whose publishing career ended between 1955 and 2010, covering 83 countries and 13 disciplines. We find that, paradoxically, the increase of participation of women in science over the past 60 years was accompanied by an increase of gender differences in both productivity and impact. Most surprisingly though, we uncover two gender invariants, finding that men and women publish at a comparable annual rate and have equivalent career-wise impact for the same size body of work. Finally, we demonstrate that differences in dropout rates and career length explain a large portion of the reported career-wise differences in productivity and impact. This comprehensive picture of gender inequality in academia can help rephrase the conversation around the sustainability of women's careers in academia, with important consequences for institutions and policy makers.

Supplementary Materials

The supplementary figures compare gender differences in scientific careers across disciplines, countries, affiliation ranks, and decades, and examine whether career length accounts for productivity and impact gaps.

Supplementary Materials

  • The career measure definitions distinguish total productivity and impact from annual productivity and career length, with dropout marked by an author's last article.Annual productivity is average yearly productivity; career length spans first to last publication.Source: Figure 1 · p. 20
  • Annual productivity is nearly identical for male and female authors across percentiles, disciplines, countries, affiliation ranks, and decades.Source: Figure 2 · p. 21
  • The productivity gender gap increased from the 1950s to the 2000s across decade groupings.Source: Figure 2 · p. 21
  • 27.4% in the population versus 12.4% in matched samples: controlling for discipline, country, and career length reduces the productivity gap.The matching experiment also compares average impact in matched samples with the original unmatched sample, but the supplied caption gives no impact result.Source: Figure 3 · p. 22
  • The career-length gender gap correlates with the productivity gap across disciplines (Pearson correlation 0.80) and countries (0.56).Source: Figure 3 · p. 22

C D E

The analyses examine gender gaps in productivity and impact in relation to age-dependent dropout, and report that the total impact gap is eliminated in a sample matched on total productivity.

E

  • The total impact gap is eliminated in the sample matched on total productivity.Figure 4E presents the matched-sample comparison.Source: Figure 4 · p. 23

C D

  • Controlling for age-dependent dropout examines its effect on gender gaps in total productivity and impact.Figure 4C examines total productivity, while Figure 4D examines impact.Source: Figure 4 · p. 23

C D E

  • The figure compares male and female authors’ dropout rates and cumulative survival across academic ages.These comparisons are shown in panels A and B, respectively.Source: Figure 4 · p. 23

S1 Data sets · S1.1 Web of Science

The study’s publication data came from the Web of Science Core Collection, which provided publication histories, citation relationships, and a hierarchical discipline classification.

S1.1 Web of Science

  • The source was the Web of Science Core Collection, including the Science Citation Index Expanded and Social Sciences Citation Index.Source: S1.1 Web of Science · p. 26
  • The dataset included articles, reviews, and letters published from 1900 to 2016, excluding other document types generally not peer-reviewed.Source: S1.1 Web of Science · p. 26
  • The Web of Science dataset covers 7,863,861 authors, 101,961,318 authorships, and 53,788,499 publications.The study also extracted 694,439,758 citation relationships.Source: S1.1 Web of Science · p. 26
  • Web of Science assigns each article to at least one discipline in a three-layer hierarchy of 153 disciplines.The layers include a top-level category, a middle-level category, and a leaf-level category.Source: S1.1 Web of Science · p. 26
  • Discipline assignments rely primarily on journal information, while selected multidisciplinary journals provide article-specific categories.Source: S1.1 Web of Science · p. 26
  • Because leaf categories were too fine-grained and the two higher layers insufficiently detailed, the study grouped leaf categories into a coarser partition.The grouping is described in Section S2.7.Source: S1.1 Web of Science · p. 26

S1.2 Microsoft Academic Graph

The study uses the Microsoft Academic Graph (MAG), a comprehensive index of journal and conference publications. The authors downloaded 77,642,549 publications in November 2017 through its authorized API.

  • MAG indexes scientific publications in both journals and conferences and supplied 77,642,549 publications downloaded in November 2017 through its authorized API.Source: S1.2 Microsoft Academic Graph · p. 26

S1.3 DBLP

The DBLP bibliography provides the computer-science publication corpus for this analysis, with records spanning 1970–2010 and contributions from over two million authors.

S1.3 DBLP

  • 4,181,940 publications from computer science journals and conference proceedings make up the DBLP bibliography corpus.The bibliography was downloaded June 5, 2018.Source: S1.3 DBLP · p. 27
  • Articles, review articles, proceedings, book chapters, and dissertations published from 1970 to 2010 were included, while other document types were excluded.Webpages and notes were excluded because they are generally not peer-reviewed.Source: S1.3 DBLP · p. 27
  • 2,129,492 authors produced the included publications, contributing 12,090,783 authorships.Source: S1.3 DBLP · p. 27

S2 Data pre-processing · S2.1 Identifying scientific careers

To test robustness to database bias and author-disambiguation errors, the study replicated its analysis in WoS, MAG, and DBLP, which use different author-identification procedures. Because disambiguation and gender inference can be difficult for Asian names, the researchers excluded specified Asian countries.

S2.1 Identifying scientific careers

S2.2 Career selection criteria

The analysis selected authors using publication-count, career-duration, annual-rate, and final-publication-date criteria, and its main conclusions were robust to stricter or modified filters.

S2.3 Country label

Authors were assigned countries using the most frequent country in their known affiliation records; the country-specific analysis retained 83 countries after applying minimum sample-size criteria.

S2.3 Country label

  • 1,876,950 authors received country labels based on the most frequently occurring country among their known affiliations.Source: S2.3 Country label · p. 29
  • 58,576 authors (3.12%) received a different country under the earliest-affiliation method, which did not qualitatively affect results.Source: S2.3 Country label · p. 29
  • The country-specific analysis retained 83 countries after excluding countries with fewer than 100 male or 100 female authors.Source: S2.3 Country label · p. 29

S2.4 Affiliation rank

Authors were assigned an affiliation rank based on their highest-ranked institute during their career; 1,296,995 of 1,876,950 authors with recorded affiliations were matched to an institute rank.

Affiliation rank

  • Affiliations were matched to the Times Higher Education World University Rankings 2019, covering over 1,250 universities, with Google Maps used to disambiguate name variations.Source: S2.4 Affiliation rank · p. 30
  • Each author was assigned the rank of their highest-ranked affiliated institute over the course of their career.Source: S2.4 Affiliation rank · p. 30
  • 1,296,995 of 1,876,950 authors with at least one recorded affiliation were aligned to an institute rank.Source: S2.4 Affiliation rank · p. 30

S2.5 Gender assignment … S7 Tables and Figures

This supplementary section documents how author gender, citations, disciplines, career indicators, and the study population were constructed and evaluated, including exclusions, metadata alignment, and robustness checks. It presents supplementary analyses of performance and career patterns across disciplines, countries, cohorts, and affiliations, alongside matching and cross-database analyses of productivity, impact, career length, and dropout.

S2.5 Gender assignment

  • Genderize.io inferred binary gender from country-specific first names; low accuracy for Asian names prompted exclusions covering China, both Koreas, Japan, Malaysia, Singapore, and Brazil.China exclusions included mainland China, Hong Kong, Macau, and Taiwan.Source: S2.5 Gender assignment · p. 30

S2.5.1 WoS and MAG authorship alignment

  • By aligning WoS and MAG papers published in the same year, with identical author-last-name sets and titles differing by no more than 25%, the authors completed first names for 1,334,886 additional authors.Source: S2.5.1 WoS and MAG authorship alignment · p. 31

S2.5.2 Gender label inference

S2.5.3 Gender label accuracy

  • 94.25% of 1,512 overlapping author names had inferred gender labels agreeing with a manually collected ground-truth benchmark.The benchmark contained 2,000 male and female full names collected in Lariviere et al. (2013).Source: S2.5.3 Gender label accuracy · p. 32

S2.6.2 Removing self-citations

S2.6.3 Citation normalization

  • Impact used each paper’s citations within 10 years (c10), normalized by the average c10 of papers published in the same year and multiplied by 12.Source: S2.6.3 Citation normalization · p. 33

S2.7 Discipline hierarchy

  • Using the Wikipedia3 scientific-field classification, the study reorganized 153 WoS categories into 75 disciplines, then focused its analyses on the 12 disciplines with the largest total populations, which cover 90.3%.Source: S2.7 Discipline hierarchy · p. 33

S2.8 Data summary

  • 1,523,002 authors (1,110,194 male and 412,808 female) contributed 18,750,502 authorships to 13,081,184 papers across 13 disciplines and 83 countries.Source: S2.8 Data summary · p. 34

S4.2 Career length matching

  • Matching annual productivity slightly increased total-productivity and total-impact gaps by 1.6% and 0%, respectively, emphasizing career length as the dominant factor.The authors report that the change in the total-productivity gap was not significant.Source: S4.2 Career length matching · p. 37

S4.3 Total productivity matching

  • Among authors of approximately the same affiliation rank, career impact favored male authors by 50.7%; after additionally matching on total productivity, the gap flipped, with female scientists receiving 1.9% more citations.Source: S4.3 Total productivity matching · p. 37

S4.4 Relationship between productivity and number of collaborators

S4.5 Controlling for the dropout rate

  • Female scientists had higher age-dependent dropout rates, and the authors simulated equal rates by randomly terminating additional male careers across 100 trials.The reported mean gender gaps had standard deviations described as near zero; the supplied passages do not give the resulting gap values.Source: S4.5 Controlling for the dropout rate · pp. 38–39

S5.1 Distributions of measurements

  • Men averaged more publications and citations and longer careers, with the largest inequalities among top scientists; annual productivity and citation rates looked similar by gender.These comparisons describe rank distributions across four major indicators.Source: S5.1 Distributions of measurements · p. 39

S5.2 Statistics and gender gaps in each discipline, country, and year

  • Supplementary tables report male and female population sizes and mean performance by discipline and country, and by career start and finish year; figures label the country gaps and career-length relationship.The supplied passages describe the scope of Tables S3–S6 and fully labeled Figures S4 and S5, without giving their cell values.Source: S5.2 Statistics and gender gaps in… · p. 40

S6.1 Microsoft Academic Graph

  • In the Microsoft Academic Graph, men showed large advantages in total productivity and career length but only slight differences in annual productivity; women had higher dropout rates.The passages identify the dropout difference as producing a separation in survival.Source: S6.1 Microsoft Academic Graph · p. 40

S6.2 DBLP

  • In the DBLP analysis, because affiliation information was largely absent and location could not assist gender assignment, the authors identified 301,150 male and 69,473 female authors.Source: S6.2 DBLP · p. 41

S7 Tables and Figures

  • The supplementary discipline hierarchy maps scientific fields into broader categories.Table S1 is titled “The discipline hierarchy.”Source: Table S1 · p. 46
  • Supplementary tables report gender-specific academic-performance averages and Welch tests overall, by discipline and country, and by career-start and career-end decade.The measures include total productivity, total impact, career length, and annual productivity.Source: Table S2 · p. 47, Table S3 · p. 49, Table S4 · p. 60
  • Further tables report gender-specific academic-performance averages and Welch tests by primary-affiliation rank and number of unique collaborators.The reported measures include total productivity, total impact, career length, and annual productivity.Source: Table S7 · p. 67, Table S8 · p. 70
  • Controlling for annual productivity does not reduce gender gaps in total productivity or total impact.The comparison is shown before and after the annual-productivity control.Source: Figure S2 · p. 72
  • Across countries, the gender gap in career length is highly correlated with the productivity gap.Figure S5 is described as a full version of Figure 3B.Source: Figure S5 · p. 75
  • Microsoft Academic Graph reproduces WoS gaps in total productivity, annual productivity, and career length; a separate figure reports dropout and cumulative survival rates.The supplied caption does not state the direction of the dropout or survival differences.Source: Figure S6 · p. 76, Figure S7 · p. 76
  • In DBLP, women’s annual productivity is nearly identical to men’s, while career length and dropout differences accompany a growing productivity gap that matching eliminates, mirroring WoS results.The caption reports higher female dropout at all career stages.Source: Figure S8 · p. 77

Related papers

Loading 1907.04103v1…