Estimating social bias in data sharing behaviours: an open science experiment

Scientific discoveries are made through cumulative and collective efforts, ideally based on full and open communication. For science to work, published claims must be subject to organized skepticism. Yet, science's ethos of rigorous, structured scrutiny is contingent on data sharing. Lack of data prevents results from being reexamined with new techniques, and samples from being pooled for meta-analysis. This ultimately hinders the cumulative knowledge-building that drives scientific progress. Open data improves the credibility of scientific claims, and while journal editors increasingly acknowledge the importance of disclosing data, many authors refrain from sharing their data, even when they have promised to do so. Previous research has focused on the supply-side determinants of data-sharing. Surveys find that scientists' decisions to share data depend on (i) contextual factors such as journal requirements, funding incentives and disciplinary competition, (ii) individual factors such as perceived risks of data misuse, lost publication opportunities, and efforts associated with making data available, and (iii) demographic factors such as experience level, tenure-status and gender. Much less attention is given to the demand-side issues of data sharing. Ideally, everyone, irrespective of background, should be able to contribute to science. As such, data access should not differ depending on who is asking for the data. Yet, research indicates persistent gender, ethnic and status-related bias in science that likely also affects data-sharing practices. Social bias in data-sharing may arise from scientists' stereotypic beliefs about data requestors. According to status characteristics theory, nationality, ethnicity, gender and institution prestige are diffuse cues that, when salient, may influence scientists' impressions of requestors' trustworthiness, competence or deservingness. Such status cues are more likely to guide people's judgments in ambiguous situations, where information is scarce. Further, status cues may be critical for data sharing, as knowledge transfer is generally more likely in high-trust situations, here including the potential data sharer's trust in the requestor's competences and intentions. The study preregistered and tested four hypotheses: scientists would be less willing to share data when a requestor (i) was from China (compared to the US); (ii) was affiliated with a lower-status university (compared to a higher-status university); (iii) had a Chinese-sounding name (compared to a typical Anglo-Saxon name); and (iv) had a feminine-coded name (compared to a masculine-coded name). In addition to gender and institution status, the study examined specific disadvantages facing researchers with Chinese names and university affiliations, given China’s prominence in global scientific output and the large number of Chinese expatriate graduate students in the US.

Prior work on data sharing has predominantly examined supply-side determinants, documenting that sharing decisions are shaped by journal policies, funding incentives, disciplinary competition, perceived risks of misuse and lost opportunities, the effort required to share data, and researcher demographics such as experience, tenure status, and gender. Empirical studies also show that despite stated policies or declarations that data are available upon request, actual data provision is often limited across disciplines. On the demand side, research indicates persistent gender, ethnic, and status-related bias in science that could extend to data-sharing interactions. Status characteristics theory suggests that diffuse status cues (nationality, ethnicity, gender, institutional prestige) can affect perceptions of trustworthiness, competence, and deservingness, especially under ambiguity. Evidence from trust games, correspondence tests, and evaluation experiments shows differential treatment based on these cues in various academic and societal contexts, motivating hypotheses about potential disparities in data access conditioned on who requests the data.

Design: A preregistered, randomized, unpaired audit experiment tested ethnic, gender, national, and status bias in data-sharing behaviors. Treatments manipulated four factors of the fictitious requester: gender (masculine-coded vs. feminine-coded), ethnicity (putatively Chinese vs. putatively Anglo-Saxon), country of residence (China vs. United States), and institutional prestige (higher-status vs. lower-status university). To maintain realism, Anglo-Saxon requesters affiliated with Chinese universities were not included; thus, 12 treatment conditions were used rather than a full 16-cell factorial. Sampling and participants: The experimental population comprised authors of peer-reviewed papers published between 2017 and 2021 in PNAS and Nature-portfolio journals that indicated data were available upon request. Journal sites were queried for the string “available upon request,” yielding 6,798 papers. Authors listed as primary data contacts were deduplicated to the most recent paper when appearing multiple times, retractions/corrections were excluded, and authors were matched to Clarivate Web of Science metadata for current emails/affiliations. The final emailing sample was 1,826 author–paper pairs; after bounced emails and post-debrief withdrawals (78), the analysis sample was 1,634. A registered power analysis indicated sufficient power (detecting Cohen’s f=0.02 at α=0.01, power=0.95). Procedures: Fictitious “about-to-become” PhD students sent data requests referencing specific target publications. Four Gmail accounts represented the four gender-ethnicity combinations; accounts were warmed up to avoid spam filtering. Nonresponders received a follow-up email one week later; data collection occurred April–May 2022. Email delivery and opens were tracked via YAMM. An alias-forwarding error affected early responses to one Chinese-female treatment but was corrected in follow-ups; analyses suggest minimal impact. Manipulations: Names signaled ethnicity and gender. Anglo-Saxon: Jeffrey Killion (male), Hilary Witmer (female). Chinese: 张嘉实 (Jia Shi Zhang; male, presented as Jiashi (Wilson) Zhang | 张嘉实) and 邢雅丹 (Yadan Xing; female, presented as Yadan (Cecilia) Xing | 邢雅丹). Chinese treatments included an Anglo-Saxon middle name in parentheses to cue gender, a practice common among transnational Chinese students. Name typicality was validated using the rethnicity R package and manual checks. Request language matched recipient context (English for non-China recipients; Mandarin for Chinese recipients). Institutional affiliation and country were varied via four universities identified through multiple rankings: high-status (Carnegie Mellon University; Zhejiang University) vs. lower-status (Baylor University; Chongqing University). Extremely top/bottom ranked universities were avoided. Measures: Treatments were coded as binary indicators: Country (US=0, China=1), Ethnicity (Anglo-Saxon=0, Chinese=1), Institution status (High=0, Low=1), Gender (Masculine=0, Feminine=1). Models adjusted for field and publication outlet via five dichotomous controls. Outcomes: (1) Preregistered primary outcome: data-sharing willingness (1 if participant shared or indicated willingness to share any/all data; 0 otherwise). Responses were double-coded via a codebook (Kappa>0.8), with coders blinded to treatments. (2) Secondary outcome (not preregistered): responsiveness (1 if any reply; 0 if no reply), an objective measure requiring no textual coding. Analytic samples: Two samples were analyzed—“opened emails” (excluding unopened emails) and “full sample” (unopened emails coded as nonresponse/unwillingness). Statistical analysis: Linear probability models (LPMs) estimated direct effects. Because no Anglo-Saxon treatments were located in China, two analysis sets were used: (a) participants exposed to US-affiliated treatments (estimating effects of ethnicity, gender, and institution status) and (b) participants exposed to Chinese-named treatments located in the US or China (estimating effects of gender, country, and institution status). Two-sided 95% and 99% CIs were reported. Analyses were conducted in R using the estimatr package. Ethics: Ex-post consent and debriefing were used to preserve validity; the study was approved by the University of Copenhagen Ethics Review Board (UCPH-SAMF-SOC-2022-03). Participants were informed post hoc and could withdraw without penalty (78 withdrawals).

- Response and sharing rates: Of 1,634 authors, 884 responded (54%) and 226 (14%) shared or indicated willingness to share some/all data. Among opened emails (N=1,179), response was 75% (884) and sharing/willingness was 20% (226). - US-affiliated treatments (opened emails, N=770): Institution status and requester gender did not affect responsiveness. Chinese-sounding names received 7 percentage points fewer responses than Anglo-Saxon names (β≈-0.07; 95% CI: -0.13:-0.01; 99% CI: -0.15:0.01), corresponding to OR≈0.66 (34% lower odds). For data-sharing willingness, preregistered hypotheses were not supported; a slight tendency toward higher willingness for lower-status US institutions vs. higher-status (β≈0.05; 95% CI: -0.01:0.11; 99% CI: -0.02:0.13) was not statistically conclusive. - Chinese-named treatments located in US vs. China (opened emails, N=802): No discernible effects of country, institution status, or gender on responsiveness; effects close to zero. For willingness, women (not men) showed a slight, non-significant advantage (β=0.05; 95% CI: -0.01:0.11; 99% CI: -0.02:0.13). - Pooled opened emails (N=1,179): Chinese-sounding names had a 7 percentage point lower likelihood of response vs. Anglo-Saxon names (β=-0.07; 95% CI: -0.12:-0.02; 99% CI: -0.14:-0.00), OR≈0.67 (33% lower odds). Institution status and gender main effects remained inconclusive. - Intersectional effects (Ethnicity×Gender, opened emails): Male Chinese-named treatments faced consistent disadvantages: responsiveness β=-0.10 (95% CI: -0.17:-0.03; 99% CI: -0.19:-0.01) and willingness β=-0.07 (95% CI: -0.15:-0.00; 99% CI: -0.17:0.02). Female Chinese-named treatments did not show comparable penalties. Results were similar in full-sample analyses but with smaller effect sizes and greater uncertainty. - Overall, none of the four preregistered hypotheses predicting bias in willingness to share data were supported; the most robust disparity emerged for responsiveness against Chinese-sounding (especially male) requesters.

The study set out to test whether scientists' data-sharing behaviors differ depending on who is asking for data, focusing on nationality, ethnicity, gender, and institutional prestige cues. While preregistered hypotheses about disparities in willingness to share data were not confirmed, analyses of responsiveness—a more objective precursor to data exchange—revealed a consistent ethnic penalty: requests from Chinese-sounding names were less likely to receive a reply than those from Anglo-Saxon names. This suggests that bias may operate at the very initial stage of interaction, consistent with theories of implicit attitudes guiding quick judgments under ambiguity. Interaction analyses further indicated that the ethnic penalty was concentrated among male Chinese-named requesters, aligning with evidence that minority men can face larger ethnic penalties than minority women, possibly due to stereotype content (trustworthiness, deservingness) and socio-political contexts (e.g., heightened concerns about intellectual property and pandemic-related prejudice during 2022). Contrary to expectations, prestige effects were weak; if anything, US-based requests from higher-status institutions were met with slightly lower willingness to share (not statistically conclusive), potentially reflecting perceived competitive risks of sharing with elite institutions. The findings underscore challenges for the FAIR data principles when data are available only upon request: not only is overall compliance modest, but discretionary sharing may introduce inequities in access. Policy implications include strengthening mandates for proactive data deposition and repository use (with licensing where needed), and improving support for ethically or practically constrained datasets (e.g., storage solutions, synthetic data options).

A large, preregistered audit experiment across authors in PNAS and Nature-portfolio journals found limited support for hypothesized disparities in willingness to share data based on nationality, ethnicity, gender, or institutional prestige. However, a robust and practically meaningful disparity emerged in responsiveness: requests from Chinese-sounding names—especially male—were less likely to receive replies. These results suggest that biases may act at the gateway to data exchange, potentially impeding the equitable circulation of knowledge. The study contributes new demand-side evidence to the open science literature, highlighting the limitations of “available upon request” policies and the need for stronger, standardized data-sharing requirements and supports. Future research should examine broader journal contexts beyond elite outlets, probe mechanisms behind male-specific ethnic penalties, and test interventions (e.g., blinded request workflows or repository-mediated access) to reduce bias and improve compliance.

- Email channel: Requests were sent from Gmail rather than institutional accounts, potentially affecting spam filtering and perceived legitimacy. The study tracked opens vs. unopened to mitigate this concern and observed relatively high response rates compared to prior institutional-account studies. - Generic request content: Standardized, non-discipline-specific emails may have raised suspicion among some recipients; this reflects a trade-off between treatment control and ecological validity. - Sampling frame: Authors were limited to PNAS and Nature-portfolio journals; findings may not generalize to other outlets or disciplines with different data policies. - Design constraints: To maintain realism, Anglo-Saxon treatments affiliated with Chinese universities were omitted, precluding a fully crossed 4×4 design; an alias-forwarding error occurred early for one treatment but was corrected and appears to have had minimal impact. - Statistical uncertainty: Many estimated effects, including prestige and gender main effects on willingness, had confidence intervals spanning zero; none of the preregistered hypotheses about willingness were confirmed, and intersectional findings should be interpreted as suggestive. - Potential period effects: Data collection in 2022 may coincide with heightened geopolitical and pandemic-related prejudices, which could influence observed biases.