Gender diversity of research consortia contributes to funding decisions in a multi-stage grant peer-review process

The study investigates whether and how the gender composition of research consortia influences outcomes in a multi-stage grant peer-review process. Motivated by longstanding evidence that gender stereotypes can shape perceptions of competence and leadership in science, prior literature on gender bias in grant allocation has yielded mixed results across countries, disciplines, and agencies. This paper focuses on group-level gender representation (proportion of female PIs in a consortium) rather than solely individual PIs, and examines its association with evaluations at all stages of the EUROCORES scheme (2003–2015) across all scientific domains. The authors aim to (1) test whether gender bias extends to groups (consortia), (2) compare effects across decision-making by expert panels versus individual external reviewers and trace propagation across stages, and (3) assess consistency between reviewers’ textual assessments and quantitative scores via sentiment analysis. The work addresses the broader concern of women’s under-representation in scientific leadership and potential barriers in access to research funding.

The paper synthesizes a broad literature on gender bias in research funding, noting mixed evidence. Several studies find gender disparities disadvantaging women in success rates or scores (e.g., Wenneras and Wold 1997; Ginther et al. 2011; van der Lee and Ellemers 2015; Severin et al. 2020; Witteman et al. 2019; Bol et al. 2022), while others report no or small effects in certain contexts (e.g., Bazeley 1998; Sandström and Hällsten 2008; Cañibano et al. 2009; Yip et al. 2020). Only a few works analyze each evaluation step or reviewer texts alongside scores (e.g., Magua et al. 2017; van der Lee and Ellemers 2015). The authors situate their contribution as examining group-level gender composition across all stages with both quantitative outcomes and textual sentiment, within a large, multidisciplinary, pan-European program.

Data source and setting: EUROCORES (2003–2015), a pan-European, multi-stage grant scheme supporting multinational collaborative research across disciplines. Applications were organized as Collaborative Research Projects (CRPs), each comprising multiple Individual Projects led by Principal Investigators (PIs). Evaluation process: Three sequential stages: (1) Expert panels evaluated Outline Proposals (OP) via in-person meetings and consensus decisions. (2) Selected consortia submitted Final Proposals (FP) for written assessment by at least three anonymous external reviewers who provided section-specific scores on 5-point Likert scales and narrative comments. (3) Panels met again to make funding recommendations informed by the reviewers’ reports and applicant rebuttals. Sample and preprocessing: Raw data included 10,533 applicants, 1642 OPs (886 accepted, 756 rejected) and 886 FPs (223 accepted, 663 rejected), 2182 external reviewers and 491 panel members. Missing person-level attributes (e.g., gender, age, affiliation) were manually completed from public sources when possible. The analytic sample includes 1347 CRPs (from 9158 unique applicants), 467 panel members, and 1862 reviewer reports, restricted to proposals with complete information across stages. Measures: Key independent variable is the proportion of female PIs in each consortium (female ratio). Outcomes: Stage 1 and Stage 3 are binary selections (passed vs. not passed); Stage 2 outcome is the average reviewer score focusing on scientific quality of the project and PI qualifications (stable items over time). Text analysis outcomes include sentiment polarity and intensity (VADER; robustness with Syuzhet and sentimentR) and the presence of evaluative bi-grams referring to PIs/teams (derived via a Word2vec Skip-Gram model trained on the review corpus); a binary indicator marks whether at least one of the 30 most frequent evaluative terms appears. Covariates: Consortium productivity (publications per PI, log), presence of highly cited scientists, research diversity (Blau index), cognitive proximity (overlap in WoS subject categories between consortium and evaluators), network proximity (pre-existing coauthorship ties), average age (log), size, institutional prestige (Top 100 Shanghai affiliation present), private sector partnership, requested budget, prior EUROCORES grant experience, number of participating countries. Evaluator characteristics include gender, productivity, age, institutional prestige; panel-level factors include workload and panel size. Statistical analysis: Stage 1 and Stage 3 modeled via probit with marginal effects; Stage 2 via OLS. Standard errors clustered at the research programme level. The sequential nature was addressed using Heckman two-step selection correction for Stage 2 (selection equation from Stage 1). Fixed effects for year, domain, and programme included. Sentiment analyses: VADER primary; Syuzhet and sentimentR as robustness checks. Word2vec model used 512-dimensional embeddings, context window of 7, negative sampling with k=15; bi-grams constructed after preprocessing (stop word removal, frequency thresholds) to identify evaluative terms related to PI/consortium/team.

• Descriptive patterns: Female participation among applicants declined across stages: 19.8% at OP, 17.5% at FP, 16.7% after final panel decisions. Female evaluators constituted ~20%. In the analytic window, Stage 1 (OP to FP) success rate was ~38% (511 projects; 3579 applicants) and Stage 3 funding recommendation rate was ~60% (306 projects; 2200 applicants); overall scheme-wide success ~13%.
• Stage 1 (first panel selection): Higher female ratio in consortia reduced the probability of advancing. In the preferred model (Table 2, Col. 3), the female ratio had a marginal effect of −0.213 (SE 0.081; p<0.01), implying a 1% increase in female share lowers the advancement likelihood by ~0.2 percentage points. Other significant factors: positive—consortium size, number of participating countries, cognitive proximity, prior EUROCORES success, and higher requested budget; negative—research diversity and network proximity to panel members; panel size and workload negatively associated with selection likelihood.
• Stage 2 (external reviewer scores): A higher female ratio was associated with lower reviewer scores. In the preferred specification (Table 3, Col. 6), coefficient −0.356 (SE 0.175; p<0.05), indicating that a 1% increase in female share predicts a 0.356% decrease in average score. Productivity increased scores; cognitive proximity decreased scores. Heckman selection correction did not materially change results.
• Text–score discrepancy: Sentiment analysis of reviewer narratives showed no significant association between consortium female ratio and sentiment polarity or presence of positive evaluative terms (e.g., VADER estimate 0.024, SE 0.071, p>0.10; evaluative terms 0.042, SE 0.091, p>0.10). Thus, while narratives did not devalue female competence, numeric scores were lower for consortia with higher female representation.
• Stage 3 (final panel recommendation): No direct effect of female ratio on final funding recommendations (e.g., −0.089, SE 0.191, p>0.10), but panel decisions were strongly and positively related to reviewers’ mean scores (e.g., ~0.69–0.71; p<0.001), implying an indirect gender effect transmitted via reviewer scoring. Network proximity had a strong positive association with success (consistent with old-boy network patterns).

The findings demonstrate that gender composition at the group level matters in grant peer review: consortia with a higher proportion of female PIs are disadvantaged in early panel triage and in external reviewer scoring. Despite reviewers’ narratives not explicitly signaling lower perceived competence of female scientists, their quantitative scores penalize higher female representation, suggesting implicit bias operating outside conscious reporting. Because panel funding recommendations heavily weight reviewer scores, biases at the reviewer stage propagate through the sequential process to influence final outcomes. Beyond gender, the positive role of network proximity indicates potential favoritism consistent with established network effects. Collectively, these patterns highlight systemic mechanisms by which women’s representation and leadership in funded collaborative science may be suppressed, with implications for equity and scientific resource allocation.

This study contributes evidence that gender bias in grant peer review extends from individuals to groups: a higher share of female PIs within consortia is associated with lower advancement odds and reviewer scores, and these reviewer biases indirectly shape final panel outcomes in a sequential evaluation system. The observed misalignment between reviewers’ textual assessments and scores suggests implicit bias in scoring practices. Given the similarity of EUROCORES procedures to many national and international funding schemes (e.g., ERC, NIH, NSF, ANR, DFG), these findings are relevant for research funders and policymakers. Potential remedies include increasing awareness by providing evaluators ex-post feedback on detected biases, carefully managing conflicts of interest and proximity (cognitive, social, network), and recognizing that anonymization alone is unlikely to suffice when bibliometric verification is integral to review. Future research should address identified limitations by measuring proposal text quality and style, accounting for applicants’ time allocation and caregiving responsibilities, examining heterogeneity by scientific areas, and investigating self-selection into applying.

• Proposal text quality and writing/editing style were not measured and may correlate with gender composition, potentially affecting evaluations.
• Lack of data on applicants’ time allocation (e.g., teaching, administrative load) and family responsibilities (e.g., childcare), which could influence productivity and perceptions.
• High interdisciplinarity of projects constrained analysis of heterogeneity in gender effects across macro-disciplines.
• Potential self-selection bias in who applies to EUROCORES (gender-differential application patterns cannot be ruled out).