ABSTRACT Open data sharing is critical for scientific progress. Yet, many authors refrain from sharing scientific data, even when they have promised to do so. Through a preregistered,

randomized audit experiment (N = 1,634), we tested possible ethnic, gender and status-related bias in scientists’ data-sharing willingness. 814 (54%) authors of papers where data were

indicated to be ‘available upon request’ responded to our data requests, and 226 (14%) either shared or indicated willingness to share all or some data. While our preregistered hypotheses

regarding bias in data-sharing willingness were not confirmed, we observed systematically lower response rates for data requests made by putatively Chinese treatments compared to putatively

Chinese requestors’ trustworthiness and deservingness, impede scientific progress by preventing the free circulation of knowledge. SIMILAR CONTENT BEING VIEWED BY OTHERS DATA SHARING

PRACTICES AND DATA AVAILABILITY UPON REQUEST DIFFER ACROSS SCIENTIFIC DISCIPLINES Article Open access 27 July 2021 PERCEIVED BENEFITS OF OPEN DATA ARE IMPROVING BUT SCIENTISTS STILL LACK

RESOURCES, SKILLS, AND REWARDS Article Open access 20 June 2023 BOUNDED RESEARCH ETHICALITY: RESEARCHERS RATE THEMSELVES AND THEIR FIELD AS BETTER THAN OTHERS AT FOLLOWING GOOD RESEARCH

PRACTICE Article Open access 06 February 2024 INTRODUCTION 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 skepticism1. 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-analysis2,3. This ultimately hinders the cumulative knowledge-building that drives scientific progress4.

Open data improves the credibility of scientific claims, and while journal editors increasingly acknowledge the importance of disclosing data5,6, many authors refrain from sharing their

data, even when they have promised to do so7,8,9,10,11,12,13. 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 gender7,14,15,16,17,18,19. Much less

attention is given to the demand-side issues of data sharing. Ideally, everyone, irrespective of background, should be able to contribute to science1. 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 science20,21,22,23,24,25,26,27 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 deservingness28,29. Such status cues are more

likely to guide people’s judgments in ambiguous situations, where information is scarce30,31. 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 intentions32,33. We conducted a pre-registered (https://osf.io/378qd),

randomized audit experiment to examine possible ethnic, gender and status bias in data sharing. We requested data from authors of papers published in widely recognized journals that commit

to the FAIR data principles (_the Proceedings of the National Academy of Sciences_ [PNAS] and _Nature_-portfolio outlets), where data were indicated to be available upon request. We varied

the identity of the fictitious data requestor on four factors: (i) country of residence (China vs. United States [US]), (ii) institution prestige (high status vs. lower status university),

(iii) ethnicity (putatively Chinese vs. putatively Anglo-Saxon), and (iv) gender (masculine-coded vs. feminine-coded name). Motivated by evidence of gender, ethnic and status bias in trust

games34,35, in correspondence tests of employers, legislators, educators and citizens27,36,37,38,39, and in field- and survey experiments of peer evaluation22,40,41, we hypothesized that

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, which have previously been covered in correspondence tests of scientists42,43, we were interested in the specific disadvantages facing researchers with Chinese names and

university affiliations. China is currently the world’s largest producer of scientific outputs44 and Chinese expatriates by far outnumber any other group of foreign graduate students at US

universities45,46. Considering these figures, a study of the possible bias facing Chinese nationals and descendants in a globalized science system seems timely. MATERIALS AND METHODS

SAMPLING Our data collection included four steps, summarized in Figure S1. The experimental population consisted of authors of scientific papers published in PNAS and Nature-portfolio

journals (between 2017 and 2021), where data were indicated to be ‘available upon request’. We queried journal websites to identify all peer-reviewed papers that included the text-string

“available upon request”. This resulted in an initial sample of 6,798 papers. C.A. and M.W.N. manually checked and coded the data availability section of each paper to identify cases where

data were unambiguously made available upon request (Table S1). If the primary author contact listed in a data statement occurred multiple times in our sample (due to >1 publications), we

only included the author’s most recent publication. We removed all retracted papers and sources with corrections. We matched the authors listed as primary data contacts in the papers with

bibliographic metadata from Clarivate’s Web of Science (WoS) to retrieve up-to-date information on emails and affiliations (Figure S2). Our final sample consists of 1,826 author-paper pairs.

Due to bounced emails and authors withdrawing from the study after receiving a debriefing statement (in total 78 authors decided to withdraw), our analysis sample is further reduced to

1,634 author-paper pairs. According to our registered power analysis, this sample size should be sufficient for detecting a small effect size of Cohen’s f2 = 0.02, with α = 0.01 and a power

of 0.95. PROCEDURES Efforts to measure ethnic, gender and status bias are complicated by observer effects and issues of social desirability47. Audit studies can mitigate such biases by

allowing the experimenter to estimate participants’ responses to the treatment in a realistic setting47,48,49. Our unpaired audit study randomized participants across twelve treatment

conditions (Figure S3). Given our four treatments, a 4 × 4 factorial design would be the typical set-up for our study. Yet, to keep the treatments realistic, we decided not to include a

putatively Anglo-Saxon (male or female) treatment associated with a (higher or lower status) Chinese University (during the COVID-19 pandemic, the number of international students in China

decreased enormously). Thus, we adopted 12, instead of 16, treatment conditions. Previous research indicates that data-sharing practices differ depending on the author’s gender and

disciplinary field11,13,50. Hence, we block-randomized the sample population according to scientific field and gender. Inspired by previous audits on discrimination in science42,43, our

data-sharing requests were emailed from fictitious “about-to-become” PhD students. We created Gmail accounts for all of the gender-ethnicity combinations (in total 4 email addresses). In the

emails, the fictitious data requestors asked the participants to share data related to a specific publication (Figure S4). If participants did not respond to our initial email, we sent out

a follow-up request after one week. Data collection was completed two weeks after the follow-up request. All data were collected in April and May 2022. Email correspondences were managed

through YAMM (https://yamm.com/), a private email service provider (google sheet add-on for Gmail and Google Workspace users). This tool allowed us to track the email delivery metrics and

provided information on whether an email had been received (or bounced) and opened (or left unopened). Our sampling and analysis plan was preregistered at the Open Science Framework. We have

followed all steps presented in the registered plan, with one minor deviation. In the preregistration we planned to run two linear probability models with cluster-robust standard errors at

the field level, while in the results section, we report the outcomes of these models without the cluster-robust standard errors. Cluster robust standard errors proved unsuitable given the

low number of clusters and because randomized treatments were assigned at the individual level as opposed to the group-level51. MANIPULATIONS Our treatment conditions varied on the following

four factors: gender, ethnicity, country of residence and institutional affiliation. We used first and last names to signal the fictitious requestors’ gender and ethnicity (Figure S4).

Emails from putatively Chinese treatments included a masculine- or feminine-coded Anglo-Saxon middle name in a parenthesis to signal the requestor’s gender (e.g., ‘Yadan (Cecilia) Xing’).

This is a widely used name-practice among transnational Chinese students52. We created four email addresses for our treatments. All accounts were opened and used for at least 90 days before

the experiment started. Warm-up activities of email domains (e.g., sending small volumes of emails and slowly increasing) helped build a positive sender reputation for the accounts. Such

standard email activity from the newly created account can reduce the risk of an account being labeled as a fake, hence reducing the email’s likelihood of being filtered as spam. To select

relevant names, we relied on a historical dataset of Olympic athletes53. We limited our focus to Chinese and American athletes participating in the Olympic Games between 1932 and 2016. For

each country treatment (US and China), we randomly selected separate first and last names until finding appropriate combinations. For the putatively Anglo-Saxon requestor conditions, we used

the following names: Jeffrey Killion and Hilary Witmer. For the putatively Chinese requestor conditions, we used the following names: 张嘉实 (Jia Shi Zhang) and 邢雅丹 (Yadan Xing). We used the R

package “rethnicity”54 to ensure that the selected Anglo-Saxon first names were typical Anglo-Saxon names and that the selected Chinese first names were well-known Asian names. In addition,

we manually verified the rethnicity estimates by looking up the first names on Linkedin. To ensure that participants perceived the Chinese names as Chinese, we wrote the signature names (at

the bottom of each email) in both English letters and Chinese characters (‘Yadan (Cecilia) Xing | 邢雅丹’ and ‘Jiashi (Wilson) Zhang | 张嘉实’) (see also Figure S4). According to the US Census of

Frequently Occurring Surnames55, 96% of individuals with the surname “Witmer” identified as white, 87% of individuals with the surname Killion identified as white, 97% of individuals with

the surname Xing identified as Asian, while 98% with the surname Zhang identified as Asian. Previous research suggests that some race and ethnicity-based manipulations also include class

primes56. In our design, we followed Block _et al_.36 and attempted to reduce the potential effect of socio-economic status by holding occupation constant. Specifically, all participants

received an email from an ‘about to become’ Ph.D. student at a Chinese or US research institution. Emails from putatively Chinese treatments (located in the US or China) that targeted

authors outside of China were written in English. Emails from Chinese nationals that targeted authors located in China were written in Mandarin. We used four international university

rankings to identify appropriate university affiliations that varied in institution status (high status [Carnegie Mellon and Zhejiang University] vs. lower status [Baylor University and

Chongqing University) (SI _Appendix_, Table S2). This manipulation also signaled the fictitious requestors’ country of residence (China or the US). Appropriate university affiliations were

identified using a combination of different university rankings (Times Higher Education, Shanghai national and international rankings, QS ranking, and the PP-top 10% indicator from the

Leiden Ranking). To select high status affiliations, we focused on universities that scored consistently high across the rankings. When selecting lower-status affiliations, we focused on

universities that scored consistently low across rankings. We decided to exclude top-ranked (top 10%) and bottom-ranked universities (bottom 10%) to lower the likelihood that participants

would discern the purpose of the experiment. In addition, we restricted our selection to universities with multiple faculties and active Ph.D. programs within each faculty. MEASURES We

measured the experimental treatments using four dichotomous variables: country (US = 0, China = 1), ethnicity (Anglo-Saxon name = 0, Chinese name = 1), institution status (high status

university = 0, lower status university = 1), and gender (masculine coded name = 0 and feminine coded name = 1). In addition, all statistical models included five dichotomous variables to

adjust for scientific field and publication outlet (Table S3). Our preregistered dependent variable is a binary measure of data-sharing willingness. This measure is based on a systematic

coding of participants’ email responses (respondents did not share data nor indicate willingness to share data = 0, respondents shared data or indicated willingness to share data = 1). Two

authors systematically coded all responses. The codebook (Table S4) was first tested by coding 10% of the sample. The pilot phase was repeated, and the codebook was further adjusted, until

coding reliability measures reached a satisfactory level (Kappa coefficient ≥ 0.8). In the manual coding, the coders were blinded to the treatments. As a second outcome (not registered), we

measured whether participants responded to our data requests or not (non-response = 0, response = 1). This outcome measure is widely used in previous email-based correspondence tests57. As

opposed to data-sharing willingness, our measure of email responses did not involve any coding of textual content and can thus be viewed as a more objective indicator of bias in data

sharing. In accordance with our registered analysis plan, we measured and reported the dependent variables in two ways. In one case, we excluded all unopened emails from the analysis. In the

results section, we refer to this sample as the sample of “opened emails”. In the other case, we included unopened emails and coded them as indicating unwillingness to share data or

non-responses. In the results section, we refer to this sample as the “full sample”. Given that all participants in the sample of opened emails have been directly exposed to a treatment, we

would expect the treatment effects to be larger for this sample. In contrast, given that some participants in the full sample were not directly exposed to a treatment, we would expect the

noise-to-signal ratio to be larger and the treatment effects to be smaller for this sample. Thus, while the sample of opened emails gets us closer to the direct effect of the treatments, the

full sample gives us a better sense of the real-world disadvantages associated with a given treatment. Due to a minor error in our email management, responses to one of the treatment emails

were associated with an alias during the first wave of data collection. Specifically, participant responses to emails sent from Yadan Xing were forwarded to a different alias address with a

similar manipulation condition. Only two participants noticed this issue in their responses, but nonetheless positively engaged with our request. The rest of the respondents exposed to this

error either responded to the alias email while addressing their message to Yadan (e.g., “Dear Yadan”) or responded directly to the correct email manipulation. Based on this evidence, we

find it reasonable to assume that the vast majority of respondents exposed to the error either did not notice the issue or perceived it to be irrelevant. This mistake was corrected in the

follow-up email addressed to all recipients who did not respond to the first email. Table S5 presents the response rates across treatments for the first and the second wave. The treatment

including the alias mistake (Yadan Xing) had the highest response rate of the four treatment emails after wave one. Note also that our main findings concerning a bias in responsiveness and

data-sharing willingness (Figs. 1,4) pertain to the male Chinese treatment (Jiashi Zhang) and not the female Chinese treatment (Yadan Xing). STATISTICAL ANALYSES Given that no participants

in our study were exposed to the putatively Anglo-Saxon treatments located in China, we estimated two groups of linear probability models. In one group, we estimated the direct effects of

gender, ethnicity and institution prestige on data sharing and email responses among participants exposed to treatments affiliated with US universities (Fig. 1 and S5). In another group, we

estimated the direct effects of gender, country-location, and institution prestige on data-sharing willingness among putatively Chinese treatments located in the US and China (Fig. 2 and

S6). Despite the unidirectional nature of our hypotheses, we report all outcomes with two-directional 95% and 99% confidence intervals. All analyses were conducted in R58, and we used the

estimator package59 to perform the linear probability models. ETHICS Since the social biases examined in our study may operate subconsciously, a pre-experimental consent process could damage

the validity of the experiment. For this reason, we decided to operate with ex-post consent and information disclosure. Our study was approved by the Ethics Review Board at the Department

of Sociology, University of Copenhagen (UCPH-SAMF-SOC-2022-03). At the end of the experiment, a debriefing email was sent to all participants (non-respondents as well as respondents). In

this debriefing, we explained the general purpose of the study, its experimental manipulations, the potential risks and benefits to participants, and the principles of data management and

anonymization. Moreover, we informed participants about their right to withdraw from the study without penalty. In total, 78 authors (5%) decided to withdraw from our study after receiving

the debriefing statement. RESULTS Eight-hundred-and-eighty-four scientists responded to our data requests, and 226 either shared or indicated willingness to share all or some of their data.

This corresponds to 54% (884 of 1634) and 14% (226 of 1634) of the full sample, and 75% (884 of 1179) and 20% (226 of 1179) of the sample of opened emails. In Fig. 1, we estimate how

institution prestige, gender and ethnicity influence participants’ responsiveness and data-sharing willingness in the sub-sample (that opened emails) exposed to treatments with US

affiliations. As shown in the figure (Panel A), neither university status nor gender affected participants’ likelihood of responding to data requests from treatments with US affiliations.

Yet, treatments with Chinese-sounding names were 7 percentage points less likely to receive a response than treatments with putatively Anglo-Saxon names (β = −0.07, 95% CI: −0.13:−0.01; 99%

CI: −0.15:0.01). This corresponds to an odds ratio of 0.66, or 34% lower odds of obtaining a response for the treatments with Chinese-sounding names compared to putatively Anglo-Saxon

treatments (Table S16). Results are similar, though associated with larger uncertainties, when estimates are based on the full sample as opposed to the sample of opened emails (Figure S5).

Our pre-registered analysis of ethnic, gender and status-related bias in scientists’ willingness to share data with treatments located in the US proved inconclusive in both the full sample

(Figure S5, Panel B) and in the sample of individuals that opened emails (Fig. 1, Panel B). In conflict with one of our hypotheses, participants seemed slightly more willing to share data,

when the request came from a lower status US institution (Baylor University) compared to a higher status US institution (Carnegie Mellon), but the confidence intervals for this effect all

include zero (β = 0.05, 95% CI: −0.01:0.11; 99% CI: −0.02:0.13). In Fig. 2, we estimate how university status, gender and country matter for participants’ responsiveness and data-sharing

willingness in the sub-sample exposed to treatments with Chinese-sounding names, located in China vs. the US. As shown in Fig. 2 (Panel A), we do not find any association between the

treatment conditions and participants’ responsiveness to data-requests in this subsample. All coefficients are close to zero, indicating no discernible effects. Similarly, for data-sharing

willingness (Fig. 2, Panel B), results are weak and inconclusive. The effects of university status and country on data-sharing willingness are close to zero, albeit women, not men

(contrasting our pre-registered hypothesis), are met with slightly higher data sharing willingness in this subsample. Yet, the confidence intervals for this gender effect include zero (β = 

0.05, 95% CI: −0.01:0.11; 99% CI: −0.02:0.13). Again, the results are similar, when models are based on the full sample as opposed to the sample of opened emails (Figure S6). In Fig. 3, we

explore if the ethnicity bias indicated in Fig. 1 (Panel A) is specific to the US treatments with Chinese-sounding names, or whether it affects putatively Chinese data requestors more

generally. In this analysis, which covers both the US-located and China-located treatments, we obtain results comparable to those reported in Fig. 1 (Panel A). The effects for university

status and gender remain inconclusive, but treatments with Chinese-sounding names have a 7 percentage points lower likelihood of receiving a response than treatments with typically

Anglo-Saxon names (β = -0.07, 95% CI: -0.12:-0.02; 99% CI: -0.14:-0.00). This corresponds to an odds ratio of 0.67, or 33% lower odds of response for putatively Chinese treatments compared

to putatively Anglo-Saxon treatments (Table S20). This estimated effect is smaller and associated with larger uncertainties when the analysis is based on the full sample compared to the

sample of opened emails (Figure S7). Given the indication in Fig. 2 (Panel B) that data-sharing willingness is lower for men with Chinese names compared to women with Chinese names, we also

estimated the conditional effects of ethnicity on data-sharing behaviors for male and female treatments. To secure the highest possible statistical power, we ran interaction-analyses on the

combined samples of US-located and China-located treatments. Figure 4 plots the conditional coefficients from four interactions between Ethnicity and Gender in the sample of authors that

opened emails. Male treatments with Chinese-sounding names (compared to male treatments with putatively Anglo-Saxon names) face consistent disadvantages both with respect to 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) when requesting data, while this is not the case for female

treatments with Chinese-sounding names. We obtain comparable results when the conditional coefficients are estimated based on the full sample (Figure S8). DISCUSSION Previous research on

social bias in data sharing is scarce. Several studies document low response rates and low data-sharing willingness among scientists who agreed to make their data available upon

request7,9,11,12,15. Yet, demand-side determinants of data sharing remain largely unexamined. In a small-scale field experiment (N = 200), Krawczyk and Reuben60 tested whether economists’

willingness to share supplementary materials differed depending on the prestige of the requestor’s university and position level (Columbia University vs. University of Warsaw) and found

negligible effects. In our audit experiment, which draws on a larger multi-disciplinary sample of participants, none of the four preregistered hypotheses predicting national, ethnic, gender

and institutional prestige bias in data-sharing willingness were supported. Yet, tests for differences in scientists’ responsiveness to data requests (our most objective measure of

disparities in data-sharing) indicated lower response rates for Chinese treatments compared to Anglo-Saxon treatments. This may indicate that ethnic bias is more likely to occur at the

initial stage of a (potential) data exchange when scientists make rapid and unreflective judgments on whether to engage with a requestor or not. Indeed, previous research into ethnic bias in

pro-sociality also emphasizes the role of implicit attitudes (activated quickly and spontaneously) in discriminatory behaviors61,62. Contrary to our expectations, scientists exposed to

US-located treatments seemed slightly less willing to share data with requestors from prestigious universities compared to requestors from lower status universities (although the 95%

confidence bounds spanned zero for this estimate). One possible explanation for this may be that the perceived career risks associated with data sharing (in terms of lost publication

opportunities and lowered competitive advantages), on average, are higher, when requests are made from prestigious US universities compared to lower-status US universities or Chinese

universities. Indeed, previous research finds that scientists’ data-sharing willingness tends to be lower when perceived competition is high7,50,63,64. Such risks and concerns could

potentially be reduced through the use of data licensing on public repositories like OSF. Importantly, the negative prestige effect was only salient for scientists exposed to treatments from

US universities. This may be because the participants in our sample, which primarily reside in Europe or North America (Table S26), are more familiar with the prestige hierarchy of US

institutions and less knowledgeable about the relative standing of Chinese institutions. Also contrary to our predictions, Chinese treatments with feminine-coded names were met with slightly

higher data-sharing willingness than Chinese treatments with masculine-coded names (although the 95% confidence bounds spanned zero for this estimate). Importantly, this finding reflects an

underlying pattern of male-specific ethnic discrimination. Conditional effects derived from interaction analyses suggested a clear bias in responsiveness and data-sharing willingness

against male Chinese treatments, while results for female Chinese treatments were inconclusive. Given the “double-burden hypothesis” in intersectional theory65, which states that minority

women are most likely to face discrimination, these findings may seem counter-intuitive. Nevertheless, they have some bearing in previous studies on trust and discrimination. Indeed,

evidence suggests that women are stereotypically seen as more trustworthy than men66, and studies on helping behavior also indicate a greater taste for helping women than men in various

social situations67,68,69,70. Further, while field experiments have generally neglected intersectional perspectives on ethnic and gender discrimination71, studies that _do_ cover this aspect

typically find that minority males are subject to larger ethnic penalties than minority females in job-markets, housing markets, and sharing economies72,73,74,75,76,77,78,79. Building on

research on gender and nationality stereotypes80, Arai and colleagues theorize that when stereotypes against specific ethnic groups are negative, they are more likely to disadvantage men

than women, because it is men who are primarily presumed to embody these stereotypes72. Additional research is required to determine the root causes of the observed bias towards Chinese men.

However, we hypothesize that it likely arises from stereotypic beliefs about the group’s trustworthiness and deservingness in data exchange relationships. Such beliefs may have been

particularly salient during 2022, when we collected our data, due to recurring discussions about China’s alleged intellectual property theft in the US and Europe81, but also in the wake of

COVID-19, where prejudice and discriminatory intent against Asians aggravated82,83. While our field-experiment has clear advantages over survey-based approaches to measuring data-sharing

behaviors, it is not without limitations. Our data requests were sent from Gmail accounts, and this may have increased the likelihood that emails ended up in the recipients’ spam filters.

Further, some recipients may have found our data requests more suspicious than they would have, if the same emails were sent via institutional accounts. We have attempted to account for this

issue by tracking opened and unopened emails throughout the study-period and by reporting results for samples that include and exclude unopened emails (SI _Appendix_). Compared to previous

correspondence studies, where data-requests were made from institutional email accounts, the response rate in our study is quite high. Participants ignored 49% of the requests made by

Tedersoo and colleagues11 concerning data for recent papers in _Nature_ and _Science_; while 86% of the data requests made by Gabelica and colleagues to authors publishing in BioMed Central

journals were also ignored12. In comparison 54% of scientists in our sample responded to the data requests. This suggests that the drawbacks of using Gmail compared to an institutional

account have been small. Another limitation of our study design concerns the generic nature of our data request, which may have increased the level of suspicion among some recipients. After

the study was completed, a few authors approached us with concerns that the email request lacked detail and was not tailored to the specific practices of their discipline, which made them

hesitant to respond. While we acknowledge this critique, experiments like ours will always be subject to trade-offs between ecological validity and treatment bias. In this case, we decided

to keep the emails generic to hold constant all other components than our manipulations. A fourth limitation concerns our sampling strategy. Because we only targeted authors of papers in

Nature portfolio journals and PNAS, our results are limited to scientists publishing in these journals. In the future, researchers should examine whether data sharing behaviors differ for

authors publishing in journals that are less committed to open science and the FAIR data-sharing principles. Finally, given that none of our four preregistered hypotheses were directly

confirmed, our results concerning gender-specific ethnic discrimination in data sharing can only be seen as suggestive. Despite these limitations, our paper offers important new insights on

scientific data-sharing practices in science. Compared to unregistered experiments, our preregistered analysis has the advantage of providing a clear record of what ideas our study was

designed to evaluate, how we planned to examine them, and how our most notable finding of ethnic bias in data-sharing relate to these ideas84. Put differently, the preregistered analysis

plan has limited our degrees of freedom as researchers and thereby increased the validity and reliability of our study. Our paper has important implications for open science policies.

Despite clearly indicating intent to make their data available upon request, only around half of the targeted authors responded to our data requests, and only 14% indicated willingness to

share all, or some, of their data. While some participants may have had good reasons not to share, this behavior conflicts with the FAIR principles adopted by PNAS and Nature portfolio

journals, hence demonstrating the drawbacks of enabling researchers to make data available upon request. Our study further complicates this issue by exposing potential inequalities in who

can benefit from data-sharing, when disclosure decisions are left to the discretion of individual scientists7. In accordance with previous work, our study shows that data requests often

require more than trivial efforts from the side of the requestor. These efforts could be reduced if funders and publishers required authors to release all relevant data, whenever

possible11,12. Unfortunately, the reality is that most journals do not incentivize data sharing. In a review of editorial policies for 318 biomedical journals5, only 12% explicitly required

data sharing as a condition for publication, 9% required data sharing without stating it as a condition for publication, while around one third of the journal sample did not mention

data-sharing at all. Under such conditions, we expect that “data availability upon request” will remain a widespread practice in many disciplines. Importantly, disclosure is sometimes also

challenged by practical issues (e.g., data size and propriety rights) or ethical issues (e.g., sensitive information on human subjects), and publishers could do more to help mitigate these

challenges. From our experience, it seems that many authors that cannot share their data for practical or ethical reasons currently opt to indicate data availability upon request to

circumvent a journal’s data-sharing requirements. Assistance from the side of publishers in providing the necessary storage space or easy-to-use methods for making synthetic datasets for

sensitive populations, could help mitigate these problems85. In summary, our field experiment extends on research about scientists’ compliance with open data principles by indicating that

sharing behaviors may differ depending on who is asking for the data. These disparities, which likely arise from stereotypic beliefs about specific requestors’ trustworthiness and

deservingness, hamper the core principle of mutual accountability in science and impede scientific progress, by preventing the free circulation of knowledge. DATA AVAILABILITY Data available

at this link: osf.io/kzrc486 CODE AVAILABILITY Code can be accessed at this link: osf.io/kzrc4 REFERENCES * Merton, R. K. The normative structure of science. in _In_ _Norman W_. _Storer_

_(Ed.)_, _The sociology of science: Theoretical and empirical investigations_. 267–278 (The University of Chicago Press., 1942). * Nosek, B. A. & Bar-Anan, Y. Scientific Utopia: I.

Opening Scientific Communication. _Psychol. Inq._ 23, 217–243 (2012). Article  Google Scholar  * Murray-Rust, P. Open Data in Science. _Nat. Preced._ https://doi.org/10.1038/npre.2008.1526.1

(2008). Article  Google Scholar  * Bird, A. What Is Scientific Progress? _Noûs_ 41, 64–89 (2007). Article  Google Scholar  * Vasilevsky, N. A., Minnier, J., Haendel, M. A. & Champieux,

R. E. Reproducible and reusable research: are journal data sharing policies meeting the mark? _PeerJ_ 5, e3208 (2017). Article  PubMed  PubMed Central  Google Scholar  * Wilkinson, M. D. _et

al_. The FAIR Guiding Principles for scientific data management and stewardship. _Sci. Data_ 3, 160018 (2016). Article  PubMed  PubMed Central  Google Scholar  * Andreoli-Versbach, P. &

Mueller-Langer, F. Open access to data: An ideal professed but not practised. _Res. Policy_ 43 (2014). * Federer, L. M. _et al_. Data sharing in PLOS ONE: An analysis of Data Availability

Statements. _PLoS One_ 13, e0194768 (2018). Article  PubMed  PubMed Central  Google Scholar  * Savage, C. J. & Vickers, A. J. Empirical study of data sharing by authors publishing in

PLoS journals. _PLoS One_ 4 (2009). * Roche, D. G. _et al_. Slow improvement to the archiving quality of open datasets shared by researchers in ecology and evolution. _Proc. R. Soc. B Biol.

Sci._ 289, 20212780 (2022). Article  Google Scholar  * Tedersoo, L. _et al_. Data sharing practices and data availability upon request differ across scientific disciplines. _Sci. Data 2021

81_ 8, 1–11 (2021). Google Scholar  * Gabelica, M., Bojčić, R. & Puljak, L. Many researchers were not compliant with their published data sharing statement: a mixed-methods study. _J.

Clin. Epidemiol._ 150, 33–41 (2022). Article  PubMed  Google Scholar  * Tenopir, C. _et al_. Data sharing by scientists: Practices and perceptions. _PLoS One_ 6, (2011). * Feigenbaum, S.

& Levy, D. M. The market for (ir)reproducible econometrics. _Soc. Epistemol._ 7, 215–232 (1993). Article  Google Scholar  * Campbell, H. A., Micheli-Campbell, M. A. & Udyawer, V.

Early Career Researchers Embrace Data Sharing. _Trends Ecol. Evol._ 34, 95–98 (2019). Article  PubMed  Google Scholar  * Tenopir, C., Christian, L., Allard, S. & Borycz, J. Research Data

Sharing: Practices and Attitudes of Geophysicists. _Earth Sp. Sci._ 5, 891–902 (2018). Article  ADS  Google Scholar  * Stieglitz, S. _et al_. When are researchers willing to share their

data? – Impacts of values and uncertainty on open data in academia. _PLoS One_ 15, e0234172 (2020). Article  CAS  PubMed  PubMed Central  Google Scholar  * Houtkoop, B. L. _et al_. Data

Sharing in Psychology: A Survey on Barriers and Preconditions. _Adv. Methods Pract. Psychol. Sci._ 1, 70–85 (2018). Article  Google Scholar  * Linek, S. B., Fecher, B., Friesike, S. &

Hebing, M. Data sharing as social dilemma: Influence of the researcher’s personality. _PLoS One_ 12, e0183216 (2017). Article  PubMed  PubMed Central  Google Scholar  * Weisshaar, K. Publish

and Perish? An Assessment of Gender Gaps in Promotion to Tenure in Academia. _Soc. Forces_ 96, 529–560 (2017). Article  Google Scholar  * Ross, J. S. _et al_. Effect of blinded peer review

on abstract acceptance. _J. Am. Med. Assoc._ 295, 1675–1680 (2006). Article  CAS  Google Scholar  * Tomkins, A., Zhang, M. & Heavlin, W. D. Reviewer bias in single- versus double-blind

peer review. _Proc. Natl. Acad. Sci. USA_ 114, 12708–12713 (2017). Article  ADS  CAS  PubMed  PubMed Central  Google Scholar  * Krawczyk, M. & Smyk, M. Author’s gender affects rating of

academic articles: Evidence from an incentivized, deception-free laboratory experiment. _Eur. Econ. Rev._ 90, 326–335 (2016). Article  Google Scholar  * Card, D., DellaVigna, S., Funk, P.

& Iriberri, N. Are Referees and Editors in Economics Gender Neutral?*. _Q. J. Econ._ 135, 269–327 (2020). Article  Google Scholar  * Peng, H., Teplitskiy, M. & Jurgens, D. Author

Mentions in Science News Reveal Wide-Spread Ethnic Bias. _ArXiv Prepr_. ABS/2009.0, (2020). * Peng, H., Lakhani, K. & Teplitskiy, M. Acceptance in Top Journals Shows Large Disparities

across Name-inferred Ethnicities. _SocArXiv_ https://doi.org/10.31235/osf.io/mjbxg (2021). Article  Google Scholar  * Milkman, K. L., Akinola, M. & Chugh, D. What happens before? A field

experiment exploring how pay and representation differentially shape bias on the pathway into organizations. _J. Appl. Psychol._ 100, 1678–1712 (2015). Article  PubMed  Google Scholar  *

Ridgeway, C. L. Why Status Matters for Inequality. _Am. Sociol. Rev._ 79, 1–16 (2013). Article  Google Scholar  * Berger, J., Cohen, B. P. & Zelditch, M. Status Characteristics and

Social Interaction. _Am. Sociol. Rev._ 37, 241–255 (1972). Article  Google Scholar  * Correll, S. J., Weisshaar, K. R., Wynn, A. T. & Wehner, J. D. Inside the Black Box of Organizational

Life: The Gendered Language of Performance Assessment. _Am. Sociol. Rev._ 85, 1022–1050 (2020). Article  Google Scholar  * Melamed, D. & Savage, S. V. Status, Numbers and Influence.

_Soc. Forces_ 91, 1085–1104 (2013). Article  Google Scholar  * Hsu, M.-H. & Chang, C.-M. Examining interpersonal trust as a facilitator and uncertainty as an inhibitor of

intra-organisational knowledge sharing. _Inf. Syst. J._ 24, 119–142 (2014). Article  Google Scholar  * Rutten, W., Blaas-Franken, J. & Martin, H. The impact of (low) trust on knowledge

sharing. _J. Knowl. Manag._ 20, 199–214 (2016). Article  Google Scholar  * Fershtman, C. & Gneezy, U. Discrimination in a Segmented Society: An Experimental Approach*. _Q. J. Econ._ 116,

351–377 (2001). Article  MATH  Google Scholar  * Cettolin, E. & Suetens, S. Return on Trust is Lower for Immigrants. _Econ. J._ 129, 1992–2009 (2019). Article  Google Scholar  * Block,

R., Crabtree, C., Holbein, J. B. & Monson, J. Q. Are Americans less likely to reply to emails from Black people relative to White people? _Proc. Natl. Acad. Sci. USA_ 118, (2021). *

Booth, A. L., Leigh, A. & Varganova, E. Does Ethnic Discrimination Vary Across Minority Groups? Evidence from a Field Experiment*. _Oxf. Bull. Econ. Stat._ 74, 547–573 (2012). Article 

Google Scholar  * Baert, S. Hiring Discrimination: An Overview of (Almost) All Correspondence Experiments Since 2005 BT - Audit Studies: Behind the Scenes with Theory, Method, and Nuance. in

(ed. Gaddis, S. M.) 63–77. https://doi.org/10.1007/978-3-319-71153-9_3 (Springer International Publishing, 2018). * Gaddis, S. M. & Ghoshal, R. Searching for a Roommate: A

Correspondence Audit Examining Racial/Ethnic and Immigrant Discrimination among Millennials. _Socius_ 6, 2378023120972287 (2020). Article  Google Scholar  * Ross, J. S. _et al_. Effect of

Blinded Peer Review on Abstract Acceptance. _JAMA_ 295, 1675–1680 (2006). Article  CAS  PubMed  Google Scholar  * Harris, M. _et al_. Explicit bias toward high-income- country research: A

randomized, blinded, crossover experiment of English clinicians. _Health Aff._ 36, 1997–2004 (2017). Article  Google Scholar  * Milkman, K. L., Akinola, M. & Chugh, D. Temporal Distance

and Discrimination: An Audit Study in Academia. _Psychol. Sci._ 23, 710–717 (2012). Article  PubMed  Google Scholar  * Gerhards, J., Hans, S. & Drewski, D. Global inequality in the

academic system: effects of national and university symbolic capital on international academic mobility. _High. Educ._ 76, 669–685 (2018). Article  Google Scholar  * Tollefson, J. China

declared world’s largest producer of scientific articles. _Nature_ 553, 390–391 (2018). Article  ADS  CAS  PubMed  Google Scholar  * Brumfiel, G. Chinese students in the US: Taking a stand.

_Nature_ 438, 278–280 (2005). Article  ADS  CAS  PubMed  Google Scholar  * Bartlett, T., & Fischer, K. The China Conundrum. _The New York Times. (Retrieved October 2022)_ (2011). *

Pager, D. & Quillian, L. Walking the Talk? What Employers Say Versus What They Do: Am. _Sociol._ 70, 355–380 (2005). Google Scholar  * Riach, P. A. & Rich, J. Field Experiments of

Discrimination in the Market Place*. _Econ. J._ 112, F480–F518 (2002). Article  Google Scholar  * Gaddis, S. M. Understanding the “How” and “Why” Aspects of Racial-Ethnic Discrimination: A

Multimethod Approach to Audit Studies: Sociology of Race and Ethnicity 5, 443–455 (2019). * Thursby, J. G., Haeussler, C., Thursby, M. C. & Jiang, L. Prepublication disclosure of

scientific results: Norms, competition, and commercial orientation. _Sci. Adv._ 4, eaar2133 (2022). Article  ADS  Google Scholar  * Abadie, A., Athey, S., Imbens, G. W., & Wooldridge, J.

When should you adjust standard errors for clustering? (_No. w24003)_. _Natl. Bur. Econ. Res_. (2022). * Diao, W. Between Ethnic and English Names: Name Choice for Transnational Chinese

Students in a US Academic Community. _J. Int. Students_ 4, 205–222 (2014). Article  Google Scholar  * Griffin, R. 120 years of Olympic history: athletes and results. * Xie, F. rethnicity: An

R package for predicting ethnicity from names. _SoftwareX_ 17, 100965 (2022). Article  Google Scholar  * U.S. Census Bureau:

https://www.census.gov/topics/population/genealogy/data/2010_surnames.html (2022). * Gaddis, S. M. Signaling class: An experiment examining social class perceptions from names used in

correspondence audit studies. Preprint at: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3350739 (2019). * Crabtree, C. An Introduction to Conducting Email Audit Studies BT - Audit

Studies: Behind the Scenes with Theory, Method, and Nuance. in (ed. Gaddis, S. M.) Ch. 6 (Springer International Publishing, 2018). * R Core Team. _R: A language and environment for

statistical computing_. (2016). * Blair, G., Cooper, J., Coppock, A., Humphreys, M. & Sonnet, L. Estimatr: Fast estimators for design-based inference. _R Packag. version_ (2019). *

Krawczyk, M. & Reuben, E. (Un)Available upon Request: Field Experiment on Researchers’ Willingness to Share Supplementary Materials. _Account. Res_. 19, (2012). * Bhati, A. Does Implicit

Color Bias Reduce Giving? Learnings from Fundraising Survey Using Implicit Association Test (IAT). _Volunt. Int. J. Volunt. Nonprofit Organ._ 32, 340–350 (2021). Article  Google Scholar  *

Stepanikova, I., Triplett, J. & Simpson, B. Implicit racial bias and prosocial behavior. _Soc. Sci. Res._ 40, 1186–1195 (2011). Article  Google Scholar  * Vogeli, C. _et al_. Data

Withholding and the Next Generation of Scientists: Results of a National Survey. _Acad. Med_. 81, (2006). * Kim, Y. & Zhang, P. Understanding data sharing behaviors of STEM researchers:

The roles of attitudes, norms, and data repositories. _Libr. Inf. Sci. Res._ 37, 189–200 (2015). Article  Google Scholar  * Browne, I. & Misra, J. The Intersection of Gender and Race in

the Labor Market. _Annu. Rev. Sociol._ 29, 487–513 (2003). Article  Google Scholar  * Fiske, S. T., Cuddy, A. J. C., Glick, P. & Xu, J. A model of (often mixed) stereotype content:

Competence and warmth respectively follow from perceived status and competition. _Journal of Personality and Social Psychology_ 82, 878–902 (2002). Article  PubMed  Google Scholar  * Eagly,

A. H. & Crowley, M. Gender and helping behavior: A meta-analytic review of the social psychological literature. _Psychol. Bull._ 100, 283–308 (1986). Article  Google Scholar  *

Dufwenberg, M. & Muren, A. Generosity, anonymity, gender. _J. Econ. Behav. Organ._ 61, 42–49 (2006). Article  Google Scholar  * Weber, M., Koehler, C. & Schnauber-Stockmann, A. Why

Should I Help You? Man Up! Bystanders’ Gender Stereotypic Perceptions of a Cyberbullying Incident. _Deviant Behav._ 40, 585–601 (2019). Article  Google Scholar  * Erlandsson, A. _et al_.

Moral preferences in helping dilemmas expressed by matching and forced choice. _Judgm. Decis. Mak._ 15, 452–475 (2020). Article  Google Scholar  * Bursell, M. The Multiple Burdens of

Foreign-Named Men—Evidence from a Field Experiment on Gendered Ethnic Hiring Discrimination in Sweden. _Eur. Sociol. Rev._ 30, 399–409 (2014). Article  Google Scholar  * Arai, M., Bursell,

M. & Nekby, L. The Reverse Gender Gap in Ethnic Discrimination: Employer Stereotypes of Men and Women with Arabic Names. _Int. Migr. Rev._ 50, 385–412 (2016). Article  Google Scholar  *

Carol, S., Eich, D., Keller, M., Steiner, F. & Storz, K. Who can ride along? Discrimination in a German carpooling market. _Popul. Space Place_ 25, e2249 (2019). Article  Google Scholar

  * Dahl, M. & Krog, N. Experimental Evidence of Discrimination in the Labour Market: Intersections between Ethnicity, Gender, and Socio-Economic Status. _Eur. Sociol. Rev._ 34, 402–417

(2018). Article  Google Scholar  * Flage, A. Ethnic and gender discrimination in the rental housing market: Evidence from a meta-analysis of correspondence tests, 2006–2017. _J. Hous. Econ._

41, 251–273 (2018). Article  ADS  Google Scholar  * Midtbøen, A. H. Discrimination of the Second Generation: Evidence from a Field Experiment in Norway. _J. Int. Migr. Integr._ 17, 253–272

(2016). Article  Google Scholar  * Simonovits, B., Shvets, I. & Taylor, H. Discrimination in the sharing economy: evidence from a Hungarian field experiment. _Corvinus J. Sociol. Soc.

Policy_ 9, 55–79 (2018). Article  Google Scholar  * Sidanius, J. & Pratto, F. _Social dominance: An intergroup theory of social hierarchy and oppression. Social dominance: An intergroup

theory of social hierarchy and oppression_. (Cambridge University Press, 1999). * Ert, E., Fleischer, A. & Magen, N. Trust and reputation in the sharing economy: The role of personal

photos in Airbnb. _Tour. Manag._ 55, 62–73 (2016). Article  Google Scholar  * Eagly, A. H. & Kite, M. E. Are stereotypes of nationalities applied to both women and men? _J. Pers. Soc.

Psychol._ 53, 451–462 (1987). Article  Google Scholar  * Guo, E., Aloe, J., & Hao, K. The US crackdown on Chinese economic espionage is a mess. We have the data to show it. _MIT

Technology Review_ (2021). * Lu, Y., Kaushal, N., Huang, X. & Gaddis, S. M. Priming COVID-19 salience increases prejudice and discriminatory intent against Asians and Hispanics. _Proc.

Natl. Acad. Sci._ 118, e2105125118 (2021). Article  CAS  PubMed  PubMed Central  Google Scholar  * Cao, A., Lindo, J. M. & Zhong, J. _Can Social Media Rhetoric Incite Hate Incidents?

Evidence from Trump’s“ Chinese Virus” Tweets_. (2022). * Ryan, T. J. & Krupnikov, Y. Split Feelings: Understanding Implicit and Explicit Political Persuasion. _Am. Polit. Sci. Rev._ 115,

1424–1441 (2021). Article  Google Scholar  * Quintana, D. S. A synthetic dataset primer for the biobehavioural sciences to promote reproducibility and hypothesis generation. _Elife_ 9,

e53275 (2020). Article  PubMed  PubMed Central  Google Scholar  * Acciai, C., Jesper, W. S. & Mathias, W. N. Estimating social bias in data sharing behaviours: an open science

experiment. _Open Science Framework_ https://doi.org/10.17605/OSF.IO/PJC9G (2023). Download references ACKNOWLEDGEMENTS This study was funded by Carlsbergfondet (the Carlsberg foundation) –

Award # CF19-0566. [P.I. M.W.N] The funder had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. AUTHOR INFORMATION AUTHORS AND

AFFILIATIONS * Department of Sociology, University of Copenhagen, Øster Farimagsgade 5, 1353, Copenhagen, Denmark Claudia Acciai & Mathias W. Nielsen * Danish Centre for Studies in

Research and Research Policy, Department of Political Science, Aarhus University, Bartholins Allé 7, 8000, Aarhus C, Denmark Jesper W. Schneider Authors * Claudia Acciai View author

publications You can also search for this author inPubMed Google Scholar * Jesper W. Schneider View author publications You can also search for this author inPubMed Google Scholar * Mathias

W. Nielsen View author publications You can also search for this author inPubMed Google Scholar CONTRIBUTIONS C.A. and M.W.N. designed research with input from J.W.S.; C.A. and M.W.N.

performed research; J.W.S. contributed supplementary data. C.A. and M.W.N. analyzed data; C.A. and M.W.N. and wrote the paper with input from J.W.S. CORRESPONDING AUTHORS Correspondence to

Claudia Acciai or Mathias W. Nielsen. ETHICS DECLARATIONS COMPETING INTERESTS The authors declare no competing interests. ADDITIONAL INFORMATION PUBLISHER’S NOTE Springer Nature remains

neutral with regard to jurisdictional claims in published maps and institutional affiliations. SUPPLEMENTARY INFORMATION SUPPLEMENTARY INFORMATION RIGHTS AND PERMISSIONS OPEN ACCESS This

article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as

you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party

material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s

Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.

To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. Reprints and permissions ABOUT THIS ARTICLE CITE THIS ARTICLE Acciai, C., Schneider, J.W. & Nielsen,

M.W. Estimating social bias in data sharing behaviours: an open science experiment. _Sci Data_ 10, 233 (2023). https://doi.org/10.1038/s41597-023-02129-8 Download citation * Received: 01

February 2023 * Accepted: 31 March 2023 * Published: 21 April 2023 * DOI: https://doi.org/10.1038/s41597-023-02129-8 SHARE THIS ARTICLE Anyone you share the following link with will be able

to read this content: Get shareable link Sorry, a shareable link is not currently available for this article. Copy to clipboard Provided by the Springer Nature SharedIt content-sharing

initiative