Racial and ethnic imbalance in neuroscience reference lists and intersections with gender
Maxwell A. BertoleroiD, Jordan D. DworkiniD, Sophia U. David, Claudia López LloredaiD, Pragya SrivastavaiD, Jennifer StisoiD, Dale ZhouiD, Kafui DzirasaiD, Damien A. FairiD, Antonia N. KaczkurkiniD, Bianca Jones MarliniD, Daphna Shohamy AiD, Lucina Q. UddiniD, Perry ZurniD, Danielle S. BassettiD
Abstract
Discrimination against racial and ethnic minority groups exists in the academy, and the associated biases impact hiring and promotion, publication rates, grant funding, and awards. Precisely how racial and ethnic bias impacts the manner in which the scientific community engages with the ideas of academics in minority groups has yet to be fully elucidated. Citations are a marker of such community engagement, as well as a currency used to attain career milestones. Here we assess the extent and drivers of racial and ethnic imbalance in the reference lists of papers published in five top neuroscience journals over the last 25 years. We find that reference lists tend to include more papers with a White person as first and last author than would be expected if race and ethnicity were unrelated to referencing. We show that this imbalance is driven largely by the citation practices of White authors, and is increasing over time even as the field diversifies. To further explain our findings, we examine co-authorship networks and find that while the network has become markedly more integrated in general, the current degree of segregation by race/ethnicity is greater now than it has been in the past. Citing further from oneself on the network is associated with greater balance, but White authors preferential citation of White authors remains even at high levels of network exploration. We also quantify the effects of intersecting identities, determining the relative costs of gender and race/ethnicity, and their combination in women of color. Our findings represent a call to scientists and journal editors of all disciplines to consider the ethics of citation practices, and actions to be taken in support of an equitable future.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 186
- Organism
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (12 Oct 2020)
In five top neuroscience journals, papers with name-inferred first and last authors of color are under-cited, mostly by White-led teams, and the gap is growing.
Citations function as career currency, and prior studies found women under-cited in neuroscience and other fields, but how racial and ethnic bias shapes engagement with minority scholars' work was unresolved. The work quantifies race/ethnicity-linked reference-list imbalance against several null models, ties it to citer identity and co-authorship network segregation, and estimates compounded gender-by-race citation costs.
Key results
ShownNeuroscience reference lists over-cite papers with White first and last authors (WW) and under-cite those with an author of color first (CW) or in both positions (CC), both against field composition and after adjusting for paper characteristics.
Of 365,859 citations from 2009 to 2019, WW papers received 50.4% vs 46.7% expected (7.9% over) and CC papers 11.7% vs 14.1% (17.2% under) by random draws; adjusting for paper characteristics gave WW 5.4% over, CW 7.8% under, CC 9.3% under. Census-based inference and a 30555-article Journal of Neuroscience subfield check agreed. · Data collection and author race modeling; Citation imbalance relative to overall…; Citation imbalance after accounting for papers’…; Figure legends
ShownThe imbalance is concentrated in White-led (WW) reference lists; teams with an author of color as first or last author over-cite WW papers less and do not over-cite CC papers, which the authors read as evidence against a simple outgroup-bias account.
Paper-characteristics model: in 206,729 citations from WW teams, WW papers were 7.1% over and CC papers 14.3% under expectation; in 159,130 citations from teams with an author of color, WW 3.2% over and CC 3.2% under. Random-draw CC under-citation was 24.1% in WW lists vs 8.0% in the others. · The effect of authors’ race on citation behavior; Citation imbalance after accounting for papers’…
ShownAlthough authorship diversified, WW over-citation and CC under-citation grew over time among both White citers and citers of color, contrary to the hypothesized decline; observed citation shares stayed flat while expected shares for author-of-color-led papers rose.
Probability of a first or last author of color rose from 47% in 1995 to 59% in 2019 (mean 0.49% per year, 95 CI 0.48% to 0.51%, 1000 bootstraps). Year-vs-imbalance slopes over 10,000 bootstraps were significant against the paper-characteristics null (slopes not reported numerically). · Trends in authorship; Temporal trends in over/under-citation; Hypothesis 3: Under-citation of C∪C papers will…
ShownCo-authorship networks grew less clustered overall yet more segregated by race/ethnicity; authors cite network-near researchers, and among White-led teams, citing farther afield is tied to less WW over-citation, though the preference persists.
InfoMap partition modularity Q peaked in 1995-2000 then fell, while race-based partition Q rose from near zero in 1995. Mean distance to cited authors was 3.5 vs 4.95 to all authors (paired t = 418, df=18,381). In WW teams, path length β = −0.98% (p < 0.0001); teams two standard deviations farther still over-cite WW papers by roughly 8.8%. · Temporal dynamics of co-authorship and its…; Co-authorship network model of…; Time-evolving segregation of co-authorship…
ShownGender is the dominant axis of imbalance: White men are most over-cited, men of color fall between White men and White women, and women of every racial/ethnic category are under-cited (Black women most, per the authors); race gaps among women are uncertain.
Random-draw gaps in percentage points (95% CI): White men vs White women 54.41 (52.61, 56.18); White men vs men CC 31.84 (11.13, 49.01); men CC vs women CC 31.06 (7.73, 60.15); White women vs women CC 8.47 (-8.49, 23.06). Paper-characteristics model: White men vs White women 29.06 (27.08, 31.0). · The intersection of gender and race/ethnicity; Discussion; Hypothesis 5: Under-citation of C∪C papers will…
Methods, data and reproducibility
First- and last-author names were converted to Asian, Black, Hispanic and White probabilities (ethnicolr classifier, Florida voter data) and binary gender; citation shares in reference lists of papers from 2009 to 2019 were compared with bootstrapped expectations from random-draw, ridge-regression paper-characteristics and co-authorship-network null models.
- Web of Science records of research articles, reviews and proceedings papers in Brain, Journal of Neuroscience, Nature Neuroscience, NeuroImage and Neuron, 1995 to 2019 · n = 63,677 articles
- Citing papers published 2009 to 2019 and their references to earlier in-corpus papers, with self-citations removed · n = 33,934 citing papers; 365,859 citations
- Florida Voter Registration File (February 2017) first and last names, training data for the main-text race/ethnicity model · human (Florida voter registry)
- US Census 2000 and 2010 surname data (last names occurring 100 or more times) for sensitivity analyses · human (US Census population)
- Social Security Administration baby-name data plus Gender API lookups for binary first-name gender assignment · human · n = 56,144 papers (88%) with first- and last-author gender
- Journal of Neuroscience articles with subfield labels, used for a subfield-adjusted sensitivity analysis · n = 30555 articles
- Methods & tools
- Web of Science (Science Citation Index Expanded), Crossref API (first-name retrieval for papers up to 2006), Secure Open Enterprise Master Patient Index nickname list, ethnicolr (Python) long short-term memory (LSTM) name classifiers, R 'gender' package with SSA baby-name data, Gender API (gender-api.com), sklearn RidgeCV ridge regression, InfoMap community detection with modularity Q, Shortest-path distance and random-walker co-authorship null models, Bootstrap resampling of citing papers, Holm-Bonferroni correction, Generalized additive model with cubic spline basis
- Software & environment
- ethnicolr Python package (LSTM models on Florida voter and Census names), sklearn RidgeCV, R 'gender' package with SSA names, Gender API and InfoMap are named; no software versions, containers or dependency files are reported.
- Reproducibility gaps
- Authors' browser tool and Python scripts lack a printed URL (only the third-party ethnicolr repository is linked); no repository or accession for the Web of Science-derived dataset; some gender labels came from Gender API, a paid service.
Limitations (6)
- Race/ethnicity is inferred probabilistically from names, not self-identification; the Census-based sensitivity model often mislabels Black and Hispanic people as White and uses decade-old data, and the authors expect misclassification to bias results toward the null. · stated by the authors
- Source name data lack American Indian/Native Alaskan and mixed-race categories, so Indigenous and mixed-race authors are not explicitly modeled, and the four coarse categories pool Asian, Black and Hispanic authors as 'authors of color'. · stated by the authors
- Gender is a binary label inferred from first names (probability threshold 0.70) and cannot represent intersex, transgender or nonbinary authors or guarantee a match to gender identity. · stated by the authors
- Name disambiguation errors can drop initials-only papers, leave some self-citations in place (judged negligible by sensitivity analysis), or wrongly merge authors and remove legitimate citations. · stated by the authors
- Intersectional race contrasts are imprecise: the White women vs women CC gap has 95% CIs spanning zero in every model reported, in the paper-characteristics model only the White men vs White women gap excludes zero, and one CI is printed for two different contrasts. · noted in this summary
- The name classifiers were trained and validated on Florida voter and US Census records (Florida model precision 0.83), but accuracy was not assessed on this multi-continent author pool, where name-to-race associations may differ by country. · noted in this summary
Who should read:Read this if you study citation equity, bibliometric bias, or name-based demographic inference and want null-model designs that separate race/ethnicity and gender effects from paper characteristics and co-authorship structure.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- The extent and drivers of gender imbalance in neuroscience reference lists
J. D. Dworkin, K. A. Linn, … D. S. Bassett·Nature NeurosciencePeer-reviewed·Jun 19, 2020
- Gender imbalance in citations most pronounced in high-impact neuroscience journals
Kat Hefter, Erin G. Teich, Danielle S. Bassett·bioRxiv·Feb 19, 2025
- Gender (im)balance in citation practices in cognitive neuroscience
J. M. Fulvio, I. Akinnola, B. R. Postle·bioRxiv·Aug 19, 2020
- Trends in Self-citation Rates in High-impact Neurology, Neuroscience, and Psychiatry Journals
Matthew Rosenblatt, Saloni Mehta, … Dustin Scheinost·eLifePeer-reviewed·May 14, 2025
- Factors that Influence Career Choice Among Different Populations of Neuroscience Trainees
L. E. Ullrich, J. R. Ogawa, M. D. Jones-London·bioRxiv·Apr 12, 2021
- Analysis of science journalism reveals gender and regional disparities in coverage
N. R. Davidson, C. S. Greene·bioRxiv·Jun 21, 2021