The puzzle of sharing scientific data

ABSTRACT

Government funding entities have placed data sharing at the centre of scientific policy. While there is widespread consensus that scientific data sharing benefits scientific progress, there are significant barriers to its wider adoption. We seek a deeper understanding of how researchers from different fields share their data and the barriers and facilitators of such sharing. We draw upon the notions of epistemic cultures and collective action theory to consider the enablers and deterrents that scientists encounter when contributing to the collective good of data sharing. Our study employs a mixed-methods design by combining survey data collected in 2016 and 2018 with qualitative data from two case studies sampled within two scientific communities: high-energy physics and molecular biology. We describe how scientific communities with different epistemic cultures can employ modularity, time delay, and boundary organisations to overcome barriers to data sharing.

KEYWORDS:

• Open science
• data commons
• collective action theory
• epistemic cultures

1. Introduction

In September 2011, researchers in the Oscillation Project with Emulsion-tRacking Apparatus (OPERA) fired a 730-km beam of muon neutrinos from European Organisation for Nuclear Research (CERN) in Geneva, Switzerland to the Gran Sasso National Laboratory in central Italy at what appeared to be a velocity faster than the speed of light. Puzzled by this result, researchers uploaded the data, with unprecedented granularity, to the open access archive arXiv.org. The data included all the necessary procedural descriptions to enable other scientists to search for an explanation of this surprising violation of a physical law. Subsequently, more than 200 papers were posted on arXiv.org attempting to explain the anomalous result. With ruthless external scrutiny, the mystery was resolved within a year – the OPERA team announced the identification of two potential sources of timing error that had corrupted the measurements (Royal Society 2012). More recently, the COVID-19 pandemic has exemplified the value of scientific data sharing, as data sharing was critical for understanding the methods of transmission and infection of the SARS-CoV-2 virus, as well as the symptoms. Within a short time, the extensive and timely sharing of COVID-19-related data informed the rapid development of vaccines (e.g. EMBL-EBI COVID-19 data portal)¹ (Fegan and Cheah 2021). There are numerous additional examples of the value of data sharing among scientific communities, although researchers across disciplines engage with data sharing in vastly different ways (Tenopir et al. 2015).

The relevance of data sharing has become prominent in recent years, as scientific research is generating ever-increasing volumes of data (Hey 2009). Some disciplines have a long tradition of working with high volumes of data, particularly the big science research infrastructures (Weinberg 1961) in physics and astronomy (Atkins et al. 2003; Borgman 2012, 2015; Carillo and Papagni 2014), whereas other scientific fields have just recently grown more data-intensive (EIROforum IT working group 2013). These disciplines include computational social science (Lazer 2009), digital humanities (Kaplan 2015), social media data (Plantin et al. 2018), citizen science research projects (Hochachka et al. 2012), and political science and public policy (Lee, Almirall, and Wareham 2016).

With the increased quantity of scientific data, perspectives on data sharing have evolved, leading to an increase in the role and status of data. Scientific data are now recognised as a scholarly object in their own right, with dedicated journals such as Nature-Scientific Data. With this shift in perspective, the increase in data-intensive methods has been labelled the ‘fourth paradigm’ in science (Atkins et al. 2003; Hey 2009), augmenting ‘the existing paradigms of experimental, theoretical, and computational science’ (Edwards et al. 2011, 670). As the potential of scientific big data grows, so too does the expectation to share data and allow others to mine, aggregate, and recombine them with other data for novel applications: ‘If the rewards of the data deluge are to be reaped, then researchers who produce those data must share them, and do so in such a way that the data are interpretable and reusable by others’ (Borgman 2012, 1059). Data reuse can be facilitated by making data Findable, Accessible, Interoperable, and Reusable (FAIR)² (Wilkinson et al. 2016). Recent studies have estimated the annual financial cost of not sharing FAIR data to be at least €10.2bn for the European economy, while the impact of FAIR on potential economic annual growth is estimated to be €16bn annually (European Commission 2019).

The importance of sharing FAIR data comes as part of a more general ‘open’ movement, embracing greater transparency in science (Edwards 2019). Starting with open access publishing, the open movement extends to open scientific data, open standards, open repositories, open bibliography, open lab-notebooks, open-source software and hardware – a virtually endless list of ‘open’ qualifiers to all activities in the scientific realm (Friesike et al. 2015). The urgency of sharing FAIR data is not only based on concerns of reproducibility (Baker 2015) or scientific fraud (Kupferschmidt 2018), but also in recognition of the novel technological and scientific innovations that can result from data sharing (Borgman 2010). As such, government funding entities, particularly in Western Europe and the United States, have placed open data at the crux of scientific policy. As European Union Commissioner for Research, Science, and Innovation, Carlos Moedas made open research data one of the EU’s priorities in 2015. This led to the formation of several expert working groups (e.g. High-level expert group on FAIR data, the Open Science Policy Platform, Expert group on altmetrics) to provide advice on how to foster and promote research data sharing in Europe. In 2016, the EU launched the Open Science Cloud initiative, which is a federated data infrastructure with cloud-based services to provide the scientific community with an open environment for storing, sharing, and reusing scientific data. In parallel, many funding agencies now require that scientific data be publicly available: for example, the US National Institutes of Health (NIH) has required this since 2003 for grants over $500,000 (NIH 2003), the National Science Foundation (NSF) since 2010 (Borgman 2012), and the European Commission for the Horizon 2020 programme since 2014 (European Commission 2014). Accompanying policy, new private and public entities have emerged to facilitate the aggregation and publication of research data. Examples include the Research Data Alliance and the National Data Service, as well as for-profit publishers who attempt to build on existing structures (e.g. Mendeley Data) (Borgman 2015). Platforms such as Dataverse (King 2007), FigShare (Thelwall and Kousha 2016), Zenodo (Peters et al. 2017), DataHub (Bhardwaj et al. 2014), EUDat (Lecarpentier et al. 2013), and other data repositories³ that offer scholars new venues to archive and share their data have also emerged (Cragin et al. 2010).

Although scientific data sharing has been positively promoted for some time, several challenges that inhibit data sharing have become apparent. Critics have pointed out that data sharing imposes increased costs on scientists and their institutions without commensurate professional benefits (Borgman 2015; Edwards 2019; Edwards et al. 2011; Tenopir et al. 2015; Wallis et al. 2013). More importantly, while consistent with the Mertonian norm of sharing to build cumulative academic knowledge (Merton 1973), data sharing clashes with a system of scientific rewards where ‘the first person to discover a result gets the “prize” associated with discovery’ (Haeussler et al. 2014, 465; Dasgupta and David 1987; Stephan 1996). There is therefore a tension between Mertonian ideals and the actual incentives of scientists who perceive data sharing as costly without commensurate professional recognition (Hagstrom 1974; Dasgupta and David 1994; Murray and O’Mahony 2007). While the prevalence of limited data sharing is known (Blumenthal et al. 1996; Campbell and Bendavid 2002; Haeussler et al. 2014; Mukherjee and Stern 2009), there is a limited understanding of the complexities and intricacies of how – and where – actual data sharing occurs.

To understand the enablers and deterrents of scientific data sharing practices, we draw upon both cultural and economic perspectives. From one side, we draw upon the notion of ‘epistemic cultures’, which originates from the sociology of science and has been applied in organisational studies and Information Systems (IS) to understand information and knowledge sharing across communities (e.g. Kellogg, Orlikowski, and Yates 2006; Mørk et al. 2008). We follow anthropologist Knorr Cetina (1999) to understand the challenges and processes involved in scientists’ epistemic work. This perspective predicts that researchers from different scientific communities will share relatively more – or less – compared to other scientific communities due to differences in disciplines’ shared norms. On the other side, taking an economic perspective, we employ collective action theory (Hess and Ostrom 2003; Olson 1965; Ostrom 1990) to understand the disincentives and deterrents that scientists face when considering data contributions to common information infrastructures for the collective good (Constantinides 2012; Constantinides and Barrett 2015; Vassilakopoulou, Espen, and Aanestad 2016). While both sociological and economic perspectives potentially offer theoretical explanations of differences in scientists’ data sharing practices, the processes that lead researchers from different scientific fields to share – or not share – are not well understood. Using a mixed-methods design (Venkatesh, Brown, and Bala 2013) (Figure 1), this study seeks to address two key research questions:

1.1. RQ1: why and how do researchers from different scientific fields share their data?

1.1.1. RQ2: what mechanisms mitigate deterrents to researchers’ data sharing?

To address RQ1, we employ survey data collected in 2016 (n = 1,162) and 2018 (n = 1,029) to explore why and how scientists share their data, considering contextual differences across scientific communities as well as data-sharing behaviour at an individual level. For RQ2 we take a case study approach to identify the mechanisms that overcome data-sharing deterrents by comparing two distinct scientific communities: high-energy physics and molecular biology (Knorr Cetina 1999). High-energy physics is a field with a communitarian culture, whereas molecular biology is relatively individualistic (Knorr Cetina 1999), and have both achieved high data sharing levels, but in different ways. Substantial differences in the scientific technologies, processes, and norms between these two disciplines provide the variance needed to identify and qualify the levers and incentives that underpin data sharing practices.

Figure 1. The research design: a mixed-methods approach to address RQ1and RQ2.

The contributions of the paper are twofold. First, we provide evidence and develop theory about the enablers and deterrents of scientific data sharing behaviour. Specifically, we provide a nuanced understanding of how both individual and discipline-level factors interact to determine researchers’ data sharing attitudes and behaviours. Consistent with predictions from epistemic culture theory, we find that researchers from different fields vary in data sharing practices due to their experienced norms. In agreement with collective action theory, we also find that researchers will respond to their perceived individual incentives regarding data sharing. Additionally, our analysis identifies three concrete mechanisms – modularity, time delay, and boundary organisations – that overcome scientists’ limited data sharing as a collective resource of the scientific corpus. Elaboration of these mechanisms can inform open science policies struggling to shape scientists’ willingness to embrace scientific data sharing.

In the remainder of the paper, first we contextualise the research setting by reviewing the background concepts of data sharing from the organisational studies, IS, and Science and Technology Studies (STS) literature and identify the incentives and reasons for sharing – or not sharing – scientific data. We then review the theoretical foundations of our research study and sequentially describe our methods and results. Methods and results from the survey data are presented first and then the methods and results from the case studies (Venkatesh, Brown, and Bala 2013). Finally, we synthesise the findings and discuss the theoretical and practical implications of the study, its limitations, and future directions.

2. Conceptual background

‘Data are representations of observations, objects, or other entities used as evidence of phenomena for research or scholarship’ (Borgman 2015, 18). A more operational definition from Open Archival Information System (OAIS) defines data as ‘a reinterpretable representation of information in a formalized manner suitable for communication, interpretation, or processing’. Examples of data include: ‘sequence of bits, a table of numbers, the characters on a page, the recording of sounds made by a person speaking, or a moon rock specimen’ (Consultative Committee for Space Data Systems 2012, 10).

An important qualification is that the research context determines what becomes data and how data are processed (Kallinikos and Constantiou 2015). As such, it becomes paramount that all relevant contextual information is gathered in the description of the data; that is, the correct and complete metadata are critical to optimising the utility of data across disciplines, time, geographies, or application domains (Edwards et al. 2011). The genesis of data may also affect an operational decision of whether to preserve the data and for how long (National Science Board 2005). For instance, observational data are considered essential to preserve, because they are the most difficult to replicate. Nevertheless, the question remains as to who makes such decisions: who has the authority to decide whether to destroy, share, or withhold data?

Historically, the dominant perspective is that the ‘producers’ of scientific data hold the authority to disclose or withhold it. Yet, with the increasing volumes of data generated by publicly funded programmes and computationally intensive environments, the ‘scientist-as-owner’ paradigm obscures a rather complex and more subtle picture of the different stakeholders involved in data production. Whether it is the scientist, team, lab, repository, or other organisational layer, the allocation of data ownership requires a careful examination of how data-sharing incentives and deterrents interact across diverse institutional levels.

2.1. Enablers and deterrents for data sharing: two theoretical perspectives

The tenor of Isaac Newton’s famous quote – ‘If I have seen further than others, it is by standing upon the shoulders of giants’ (Kuhn 1962) – is still prevalent in scientific practice. The progression of science relies on the accumulation of knowledge and, thus, the sharing of the results of prior research. Yet even though the economic costs and needed efforts of data processing and storage are falling (Romasanta and Wareham 2021), how the accumulation and reuse of knowledge occurs within and across disciplines is often unclear, confounding scientists’ motivations for data sharing.

Prior literature suggests that while authors are initially the copyright holders of their academic publications, the jurisdiction over the data is more ambiguous: uncertainty around ownership, control, and access to the data generates tensions amongst stakeholders (Borgman 2015). ‘Even when individuals and groups assign authority for data, the rights and responsibilities may remain unclear’ (Bowker and Leigh Star 1999, 646). As a consequence, policymakers, funders, and academic institutions are working to increase awareness that, while the publications and knowledge derived from research data pertain to the authors, research data needs to be considered a public good so that its potential social and scientific value can be realised (European Commission 2014; OECD 2015; Järvenpää and Markus 2018; Vassilakopoulou, Espen, and Aanestad 2016).

2.2. Epistemic cultures

Knorr Cetina (1999) coined the notion of epistemic cultures to describe ‘those amalgams of arrangements and mechanisms – bonded through affinity, necessity and historical coincidence – which, in a given field, make up how we know what we know’ (p. 1). The notion of epistemic culture claims that the nature of scientific activities, types of reasoning, and practices of establishing evidence are variable across scientific fields. The epistemic cultural approach disputes the ‘unity of science’ associated with the Vienna Circle (Knorr Cetina 1999) and ‘reveals the fragmentation of contemporary science’ (Mørk et al. 2008, 15).

The concept of different scholarly cultures can be drawn back to the idea of ‘styles of thought’ shared by ‘thought collectives’ (Knorr Cetina 1999); it also relates to a concept of ‘thought worlds’ (Dougherty 1992) or the idea of ‘communities of knowing’ (Boland and Tenkasi 1995). Haas (1992) used the notion of ‘epistemic communities’ to define groups of people engaged in knowledge production. The general and universal idea across such notions is that knowledge is situated and local. ‘There is no “view from nowhere” – knowledge is always situated in a place, time, conditions, practices, and understandings’(Borgman 2012 p.37).

In the context of our study, Knorr Cetina’s ideas are useful because the definition of ‘culture’ is rooted in practice. The ‘epistemic machinery’ defines the shared tools, techniques, instruments, methods, and architectures of shared empirical practices that the epistemic subjects use to produce and distribute knowledge. She describes the making of science through interiorised processes where scientists, organisations, and collectives operate (e.g. labs and experiments) (Knorr Cetina 2007).

Employing Knorr Cetina’s lens, data sharing practices would be expected to be community-bound and largely determined by epistemic culture. This is logical if we consider the long cycles through which new members are trained; the specificities in the technology tools; the commonly accepted methods, funding sources, collaborative norms; and the ways in which responsibility and authorship are assigned. Differences in data sharing practices would also depend on whether the scientific community is more communitarian or individualistic: a communitarian culture should be predisposed to share, with fewer concerns about individual incentives and rewards; whereas an individualistic culture would be more driven by individual motivations that give weight to the costs of sharing data and the lack of clear individual compensations.

2.3. Collective action theory

To examine the mechanisms by which self-interested researchers would contribute to a data as a public good, we employ collective action theory and the idea of a data as commons. The term ‘commons’ designates a ‘resource shared by a group of people that is subject to social dilemmas’ (Hess and Ostrom 2003 p.3).

Collective action theory has been widely used in sociology and economics to understand individuals’ motivation to engage in collective action (Fulk et al. 2004; Monge et al. 1998). Collective action research originated with Olson’s work in the classic Logic of Collective Action (Olson 1965), from which Hardin (1968) developed his ‘tragedy of the commons’ thesis, suggesting that uncontrolled individual self-interest pursuits can undermine common public resources (Greco and Floridi 2004). In other words, the tragedy of the commons can be viewed as a prisoner’s dilemma (with n-people) where the rational pursuit of self-interest results in suboptimal management of public resources and social goods (e.g. forests, fisheries) (Greco and Floridi 2004; Fletcher and Zwick 2000; Ostrom 1986).

As Hardin (1982) described, communities benefit when the individual perceives gains from making contributions to the commons. If such benefits are not perceived, the shared pool of resources is ‘latent’ and can deteriorate without external intervention. What makes collective action useful in understanding scientific data sharing is its focus on how the appropriation of individual gains is determined by adjusting the costs and benefits that accrue with contributions to a common resource (Fulk et al. 2004; Ostrom 1990; Vitali, Mathiassen, and Rai 2018; Weill and Ross 2004).

Table 1 summarises common reasons cited to justify the benefits of data sharing along with frequent explanations for deterrents. Interestingly, whereas arguments for data sharing reflect benefits to the scientific community, reasons not to share data are predominantly based on concerns that individual costs exceed individual benefits. While not exclusively so, the theory of epistemic cultures largely emphasises norms that favour data sharing, where collective action theory is predominantly concerned with self-interest and individual costs.

Table 1. Enablers and deterrents for data sharing in science according to prior literature and implications from collective action and epistemic cultures perspectives

Reasons for sharing data from policy and community perspective	Description	Literature	Implications from Epistemic cultures perspective
Reproducibility	Sharing research data and making them easier to peer review increases transparency and opportunities for the reproduction of research findings. It also increases the potential for publishing negative results and enables accurate verifications of research findings.	(Baker 2015; Fecher et al. 2015; Lyon 2016; OECD 2015; Pujol Priego and Wareham 2019; Tenopir et al. 2015)	Data sharing practices may be community-bound because of the epistemic culture of the community. Differences in data sharing practices across scientific communities depend on how community values shape the norms towards integrating the importance of scientific reproducibility, acceleration of scientific progress, scientific quality, efficiency, and innovation into their collective behaviour.
Higher scientific efficiency and progress	The availability of the Gene Expression Omnibus (GEO) database at the U.S. National Center for Biotechnology Information led to more than 1,150 published articles by third-party contributors by the end of 2010.	(Borgman 2015; Lyon 2016; OECD 2015; Pasquetto, Randles, and Borgman 2017; Piwowar, Vision, and Whitlock 2011;Whyte and Pryor 2011)
Higher scientific quality and transparency	Sharing research data is related to the strength of the evidence supporting the results, the quality of the statistical results reporting, and enables transparency and greater scrutiny of research.	(Kupferschmidt 2018; Wicherts et al. 2011)
Reasons for sharing/not sharing data by the individual scientists	Description	Literature	Implications from a Collective Action perspective
Personal credit and rewards	There is a lack of consistency in the way data is cited. The scholarly system is heavily biased towards publications; secondary products such as data or code are rendered far less credit. Relatedly, for scholars driven by credit, sharing data offers little benefit, particularly in light of intentions to try to publish future articles out of the same data, or aggregating it with complementary datasets.	(Borgman 2015; Piwowar, Vision, and Whitlock 2011) (Harley, Acord, and Earl-Novell 2010; Howison et al. 2015; Meijer et al. 2017; Plantin et al. 2018)	Employing a collective action perspective, we would expect that scientists would share the data if they perceive gains from its contribution to the commons by adjusting the values and costs associated with the resource contribution. Data sharing practices would depend on how the community adjusts the costs (e.g. lack of credit, rewards, costs of input metadata) with the potential value of the data contribution.
Misuse, misinterpretation, liability concerns	Uncertainty over who is going to reuse the data and for what purposes and lack of understanding of the data and thus misuse.	(Meijer et al. 2017; Tenopir et al. 2015; Wallis et al. 2013)
Lack of skills	Lack of expertise and knowledge of tools to make their data available.	(Borgman 2015; European Commission 2019; OECD 2015)
Costs to prepare data	The effort and time-consuming activity of providing contextual information and detailed descriptions of the data. Costs could span an estimated two to three weeks from an average of a two-year research grant application.	(Edwards 2010; Holzner, Igo-Kemenes, and Mele 2009; OpenAire 2019)

3. Methods and results

We first present the methods and results of the analysis of the survey data. We then follow with the case studies and synthesise the findings from both (Venkatesh, Brown, and Bala 2013). The synthesis of the results in the discussion is a ‘bridging’ process (Creswell 2018), where we leverage the complementarities across the findings to enrich our empirical and theoretical understanding of scientific data sharing practices.

3.1. Survey

3.1.1. Methods and data

We developed a large-scale global online survey in the framework of the Open Science Monitor for the European Commission⁴ in collaboration with a consortium of scholars including a major academic publisher, and responses were collected in 2016 and 2018. The survey data allow us to address RQ1- why and how do researchers share their data? The survey of 2016 was sent in June–July 2016 by the publisher and generated 1,162 responses, which represented a 2.3% response rate. The margin of error for 1,162 responses was estimated ± 2.87% at 95% confidence levels (see prior analysis of the survey and full dataset in Meijer et al. 2017). The survey of 2018 was sent in October–November 2018 to 40,991 individuals randomly selected from the Scopus author database, weighted to be representative of the researcher population (UNESCO counts of researchers, 2013), to which 1,029 researchers responded (2.5% response rate). The responses in both surveys were anonymous, not containing any personal identifying information. There were several differences between the 2016 and 2018 survey questionnaires: four additional questions were added in 2018 to assess the consequences of data sharing for scientists in their future collaborations with for-profit entities. Additional minor modifications were introduced in the 2018 questionnaire to improve semantic clarity.

By using random sampling methods to ensure that the individuals selected were representative of the researcher population at large, the study sought to mitigate potential selection bias. Nonetheless, we can compare UNESCO’s count of researchers across fields with the mix we received from respondents in the survey. As we report in the Appendix, we have overrepresentation in the natural sciences and slight underrepresentation in engineering.

In the first phase of the analysis, we explore which factors promote or inhibit data sharing among individual researchers. Specifically, we examine two aspects: 1) their willingness to let others access their research data, and 2) actual experience in sharing their data. By analysing both attitudes and behaviours, we obtain a more holistic picture of data sharing. These two variables were measured through a 5-point Likert scale.

The independent variables were the factors related to data sharing enumerated in the literature review. For each of these factors (Table 2), we identified relevant survey questions as proxies to represent each enabler or deterrent, which were either ordinal or binary. The ordinal variables were measured through a 5-point Likert scale of agreement with statements provided. The binary variables were from a question asking respondents about the benefits of data sharing in their field.

Table 2. Factors related to data sharing

Variable	Survey question	Relevant factor represented	Type
Data sharing (Dependent variables)
Willingness	I am willing to allow others to access my research data	Data sharing	5-point scale (Strongly disagree to Strongly agree)
Experience	I have previously shared my research data with others	Data sharing	5-point scale (Strongly disagree to Strongly agree)
Enablers: Epistemic culture lense
Reproducibility of research	Reproducibility of research	Reproducibility	Yes/No
Data reuse	Data reuse	Higher scientific efficiency and progress	Yes/No
Research aggregation	Research aggregation	Higher scientific efficiency and progress	Yes/No
Importance of data sharing in field	Sharing research data is important for doing research in my field	Higher scientific quality	5-point scale (Strongly disagree to Strongly agree)
Collaboration	More possibilities for collaboration	Higher scientific quality	Binary
Deterrents: Collective action incentives
Data sharing rewarded in field	Sharing research data is associated with credit or reward in my field	Personal credit/rewards	5-point scale (Strongly disagree to Strongly agree)
Higher paper acceptance	Article more likely to be accepted for publication	Personal credit/rewards	Yes/No
Higher citation	Article more likely to be cited	Personal credit/rewards	Yes/No
Compliance to funding body	Compliance with funding body mandates	Misuse and liability concerns	Yes/No
Compliance with journal/publisher	Compliance with journal or publisher requirements	Misuse and liability concerns	Yes/No
Training	I have received sufficient training in research data sharing	Lack of skills	5-point scale (Strongly disagree to Strongly agree)
Costs	Effort required prior to sharing data	Costs to prepare data	5-point scale (Strongly disagree to Strongly agree)

Respondents with missing data or a response of N/A were removed from the analysis. This left us with 491 respondents with complete data across the different variables we wanted to test in Table 2. Due to the dominance of non-continuous variables, we used diagonally weighted least squares (DWLS) to estimate the model parameters through the lavaan package in R (Rosseel 2012). DWLS is robust against non-normality and works better with ordinal data than maximum likelihood methods (Mîndrilă 2010) and performs well with smaller sample sizes (Flora and Curran 2004).

In the second phase of the analysis, we explored the differences in data sharing practices across scientific communities. We looked at the same two aspects of researchers’ data sharing: willingness to let others access their data and actual experience sharing data. Considering the ordinal nature of these variables, we used the Kruskal-Wallis test to determine whether different research fields have distinct willingness and experiences in sharing data. Kruskal–Wallis is a nonparametric statistical test that assesses the differences among three or more groups (Kruskal and Allen Wallis 1952; McKight and Najab 2010). In contrast to one-way Analysis of Variance (ANOVA), which is used for normally distributed continuous variables, Kruskal–Wallis is more appropriate for ranked data such as those from our survey. To identify which specific fields share more – or less – than other fields, we followed up with Dunn’s test, the pairwise multiple-comparison procedure used when a Kruskal–Wallis test is rejected (Dunn 1961; Dinno 2015). To correct for the increased false positives that arise from conducting more statistical tests, we also present the results after Holm’s sequential adjustment (Holm 1979).

3.1.2. Findings of the survey

We began the analysis with a focus on data sharing for individual researchers. To do this, we used the various factors identified under epistemic cultures and collective action and related these to both researchers’ willingness and previous experience in data sharing.

We found support for epistemic norms stimulating data sharing. Our results suggest that researchers had positive data sharing outcomes in both attitude and behaviour, if they believed that the data are crucial to the advancement of their research field. Moreover, the push for data reproducibility, data reuse, and collaborations in their field also seemed to lead researchers to engage with data sharing (Table 3). Furthermore, supporting individual incentives to overcome collective action problems was related to positive data sharing outcomes. Compliance with funding bodies was related to increased willingness to share data. Receiving necessary skill training also seemed to lead to data sharing engagement. In summary, various individual-level factors seemed to motivate researchers’ sharing of data.

Table 3. Researcher willingness to share data for various reasons and experience of it

		Willingness	Experience
Epistemic culture	Reproducibility of research	0.287 (0.131)*	0.181 (0.119)
	Data reuse	0.205 (0.118)	0.366 (0.117)**
	Research aggregation	0.111 (0.122)	−0.017 (0.114)
	Importance of data sharing in field	0.565 (0.08)***	0.568 (0.072)***
	Collaboration	0.251 (0.118)*	0.181 (0.113)
Collective action	Data sharing rewarded in field	0.094 (0.053)	0.029 (0.047)
	Higher paper acceptance	−0.199 (0.113)	−0.023 (0.115)
	Higher citation	−0.179 (0.109)	0.007 (0.106)
	Compliance with funding body	0.360 (0.166)*	0.293 (0.172)
	Compliance with journal/publisher	0.034 (0.130)	0.027 (0.130)
	Training	0.091 (0.047)	0.136 (0.046)**
	Costs	−0.048 (0.071)	0.072 (0.070)
	N	491	489
	Test statistic	0.000	0.000
	R^2	0.278	0.143

p-values: * <0.05, ** < 0.01, ***<0.001

Analysis performed in survey data 2018

We present varying attitudes and behaviour towards data sharing among scientific disciplines in Figure 2. Visual examination shows a mixed finding, where many disciplines are similar in data sharing, and some disciplines in the extremes share more or share less compared to the average. To make a better judgement of these disciplines’ differences, we used the Kruskal-Wallis test, which confirmed that certain disciplines differ in their data-sharing activity. By their willingness to allow others access to research data, we found differences across disciplines (statistic = 26.9, p-value = 0.003). Moreover, in their experience sharing of data, we also found significant differences (statistic = 25.7, p-value = 0.004).

Figure 2. Data sharing by discipline.

The fields are ordered by their researchers’ average experience sharing data. The fields included are Maths (MATHS), Medicine and Allied Health (MED), Engineering (ENG), Materials Science (MAT SCI), Computer Science (COMP SCI), Social Sciences, Humanities & Economics (SOC HUM ECON), Chemistry (CHEM), Physics & Astronomy (PHY), Life Sciences (LIFE SCI) and Earth & Environmental Science (ENV). The responses of ‘N/A’ and ‘Neither agree nor disagree’ were left out for clarity. Width relates to sample size. + denote significant difference to at least one other field after posthoc Dunn test without adjustment (p-value<0.05). * denote fields that continue to have significant differences to at least one other field after posthoc dunn test with Holm adjustment (p-value <0.05).

To identify which disciplines differed in data sharing, we used post-hoc pairwise Dunn test with Holm correction. Researchers in environmental sciences tend to be the most willing to allow others access to their research data. This is relative to the disciplines of medicine and engineering, where many respondents mentioned that they are not as engaged in data sharing. As for the remainder of the disciplines, we do not find significant differences in their sharing of data. In continuation, we explored the various outcomes of data sharing: with whom data is shared, where it is shared, and what type of data is shared. This information is summarised in Figure 3.

Figure 3. Data sharing outcomes across all fields.

Plots show percent responses across researchers from all fields. Asterisks denote p-value after Kruskal–Wallis test (* <0.05, ** < 0.01, ***<0.001) comparing different fields in their responses. The type of data shared is less than 100% due to researchers saying that it is not very important in their field.

Regarding the question of with whom researchers tend to share their data, we find that they are highly discriminatory. Although they may grant non-collaborators access to their data through personal communication, we see that researchers tend to be most comfortable sharing with collaborators they know personally. This was consistent across all fields. The most variance we see across fields is whether they share data with their research partners such as funders (Figure 3).

Location of data storage differed significantly among the disciplines. Certain fields prefer to publish data (1) in the appendix, (2) as stand-alone peer-reviewed data publications, (3) in data repositories, or (4) through other avenues altogether. Finally, fields also vary across the types of data they generate and, logically, the types of data they are willing to share. Across all fields, the primary type of data that researchers generate is numerical, followed by textual data, although fields do not vary in their rates of sharing such data.

What becomes clear from the survey results is that data sharing is a practice that varies not only across disciplines, but also between researchers from the same discipline. In other words, the results of the questionnaire thus far reveal a complex picture where discipline-level and individual-level factors are highly interlinked. Consistent with an epistemic culture explanation, we see that certain disciplines are more open to sharing than others. Additionally, consistent with a collective action perspective, we find that individual researchers within the same discipline can vary in their data sharing.

To contextualise our subsequent cases studies, we highlight the specific results for physics and astronomy and life science as they compare to all other fields across the survey items. Both the disciplines of Life Science and Astronomy/Physics share data at high levels as compared to the other fields surveyed. As seen in Figure 4, they share the same motivators and deterrents in data sharing despite the theoretical differences previously emphasised in the literature. Comparing the two would provide useful cross-discipline variance for addressing RQ2.

Figure 4. Comparison of data sharing among disciplines, highlighting life sciences and physics/astronomy.

For reference, we also show the mean response across all fields. For questions with 5-point Likert scale coded from 0 to 4, we divided the response by 4 for the plot. All the statistical analysis was done with the original (non-transformed) data through Kruskal–Wallis test. Asterisks denote p-value after Kruskal–Wallis test comparing the two fields of Life Science and Astrophysics in their responses. (* <0.05, ** < 0.01, ***<0.001).

4. Case studies

4.1. Methods and data

To capture the data sharing practices of life science and physics, we thematically sampled (Creswell 2018) high-energy physics (HEP) and molecular biology (MB) communities’ practices anchored in two information infrastructures. Information infrastructures have been defined as ‘a digital library system based on commonly shared standards and containing information of both local and/or widespread interest’ (Kahn and Cerf 1988, 3), designed or intended ‘to augment our ability to search for, correlate, analyse and synthesize available information’ (Kahn and Cerf 1988, 11). Our decision to focus on information infrastructures (as opposed to less-institutionalised data sharing practices) is because the highest data sharing levels are in communities that actively use information infrastructures (Edwards et al. 2009; Ribes and Lee 2010). Both cases represent infrastructures based upon data commons, i.e. infrastructures that co-locate data, storage, and computing facilities with commonly used services and tools for analysing and sharing data for a defined base of users (Grossman et al. 2016).

4.1.1. Empirical context 1: molecular biology and open targets

The sequencing of the human genome (Human Genome Project, HGP) is recognised as ‘the largest undertaking in the history of biological science’ (Chaguturu, Murad, and Murad 2014, 35). Not only did it transform biology into a data-driven discipline with a deluge of new data and computational techniques, but it also opened the debate about research data sharing. Celera, a private undertaking, initially announced their intention to patent fully-characterised important structures amounting to 100–300 genetic targets, which prompted considerable outcry (Leonelli 2012). In March 2000, U.S. President Bill Clinton announced that the data on the Human Genome Project (HGP) sequence should be made freely available to the entire research community. The HGP propelled discourse and debates on open research data to the forefront of molecular biology research (Leonelli 2012) and spawned a new generation of information infrastructures to generate, integrate, and curate the growing data pools with commonly used tools and analytical methods (Grossman et al. 2016; Vamathevan et al. 2019). As a result, the discipline has been very active in developing infrastructures based upon data commons, one of them being Open Targets (Pujol Priego and Wareham 2018).

Open Targets (OT) is a MB consortium created in 2015 by the European Molecular Biology Laboratory-European Bioinformatics Institute (EMBL- EBI), Europe’s flagship laboratory for life science, with the Wellcome Sanger Institute, and five pharmaceutical companies (i.e. Biogen, Celgene, GSK, Sanofi, Takeda); it aims to accelerate knowledge about the links between genetic targets and disease development. The OT consortium includes publicly funded, non-profit, and for-profit organisations with vastly divergent institutional objectives and stakeholders. The infrastructure started as a result of pharmaceutical companies searching to scale requisite R&D capabilities by generating, integrating, and curating large data pools with commonly used tools and analytical methods about the early phases of drug discovery for the research community (Grossman et al. 2016; Vamathevan et al. 2019). The architecture, data policies, and procedures from researchers participating in OT provide insights about the mechanisms that effectively foster data sharing across the MB research community. As of this writing, OT infrastructure contains more than 27,717 genetic targets, 7,999,050 associations, 13,445 diseases, and 20 data sources (Open Targets 2020).

4.1.2. Empirical context 2: high-energy physics and reana

Big scientific research infrastructures within HEP, such as CERN, have a long tradition of embracing open data. Large volumes of data generated via expensive, unique, and elaborate experiments make data preservation and reuse important. Reana is a reusable and reproducible research data analysis infrastructure that was created at CERN in 2018 to facilitate data and code reuse. The infrastructure was initiated to combat the reproducibility crisis in the particle physics field. CERN built Reana as a data infrastructure to allow the different HEP experiments to adhere to FAIR principles and facilitate data sharing and reuse in the community. Reana allows the reuse and reinterpretation of the data shared by helping HEP scientists to structure their input data, analysis code, containerised environments, and computational workflows to run the analysis on remote clouds (Pujol Priego and Wareham 2019). What makes Reana attractive is that the infrastructure helps to generalise computational practices employed by HEP scientists, thereby systematising reproducibility. The infrastructure supports a plurality of ‘container technologies (Docker), workflow engines (CWL, Yadage), shared storage systems (Ceph, EOS) and compute cloud infrastructures (Ku- Kubernetes/OpenStack, HTCondor)’ used by the HEP scientific community (Simko et al. 2018, 1). The infrastructure sits on extant platforms and services provided by CERN to the HEP community; these include Zenodo, a free and open data repository, and CERN open data portal, both of which are precedents to Reana infrastructure (Pujol Priego and Wareham 2019).

4.1.3. Data sources and analysis

The study of both cases relies on the diverse primary and secondary data sources described in Table 4. Numerous discussions with managers from Open Targets and Reana were an integral part of the Open Science Monitor, published by the European Commission in separate reports (Pujol Priego and Wareham 2018, 2019). Primary data included 18 semi-structured interviews and direct observations from a study visit at the Wellcome Genome Campus for the OT open days (June 2019) and recurrent study visits at CERN from 2018–2020. As part of participation in the two additional EU H2020 funded projects, the authors benefited from extensive conversations with policymakers, research infrastructure managers, data architects, and programmers, in which they discussed data sharing practices and future open research data initiatives (CS3MESH4EOSC part of European Open Science Cloud and ATTRACT funded by Research Infrastructure Innovation H2020-INFRAINNOV). The interview process was concluded when no significant additional insights were obtained from the data, and theoretical saturation was achieved.

Table 4. Data collection sources for both Molecular Biology (MB) and High-Energy Physics (HEP)

	MB – Open Targets	HEP- Reana and related platforms
Primary data sources	13 interviews with scientists and managerial team of OT	5 interviews with scientists and managerial team of Reana and related platforms; and 4 interviews with CERN programmers and data architects.
Observations	Study visit to Genome Campus OT Open Days – workshop, working groups and social event (June 2019)	Study visits to CERN (2018, 2019, 2020). Partner in H2020 funded CS3MESH4EOSC, a constituent project of the European Open Science Cloud https://cordis.europa.eu/project/id/863353, and ATTRACT https://attract-eu.com/. Interviews and discussions with open data related services at CERN (Zenodo, Open Data Portal, CS3-ScienceMesh).
Secondary data sources	41 publications 1 tutorial on OT infrastructure 3 outreach posts; 19 release notes; 6 posts; 7 websites	Experimental data policy and guidelines: CMS, ALICE, ATLAS, LHCb, OPERA data policy; CERN open data terms of use; 22 guidelines in CERN open data portal; CERN Analysis Preservation Portal; Joint declaration and Taskforce documentation on HEP data preservation; Reana workshop presentations June 2018; 12 runnable examples of Reana; 6 publications, 6 release notes, 2 blog posts.

Secondary sources included data from media outlets, with 47 publications resulting from OT and Reana, 1 tutorial and 12 runnable examples about how to use the infrastructures. Additionally, 8 blog posts, 25 release notes, 2 webinars, several workshop presentations, and information available in the different HEP experiments and OT organisations’ websites were used. These secondary sources were very useful in the first stages of the analytical processes, enabling us to have more technically informed conversations in both the infrastructures and data sharing practices. The combination of our primary data with secondary data allowed us to crosscheck findings and build our theoretical inferences from the cases.

Data were analysed by performing a two-stage inductive analysis, relying on established procedures for inductive research (Miles and Huberman 1994). The first stage was devoted to reading the abundant material available online about OT, HEP experiments, and Reana. We produced brief summaries that moved from technical descriptions of the infrastructure to managerial inferences. In-depth interviews were then conducted to understand how scientists use the infrastructures. We performed the interviews and analysis in several iterations, and thus earlier transcripts informed and incorporated information emerging from later interviews. In addition, we contrasted the transcripts from the interviews with our analysis of secondary sources. We generated research memos that synthesised the emergent themes identified in the analysis and compared them with prior research. Finally, to validate our findings, we applied respondent validation (Miles and Huberman 1994) by sharing our initial findings with the study participants.

4.2. Findings of the case studies

Preliminary observations about HEP and MB communities suggest two different epistemic cultures consistent with Knorr Cetina’s thesis, with HEP being more communitarian and MB more individualistic. When looking at how HEP data flows are organised, we first realised the importance of the institutional entity of ‘the experiment’. In HEP, a limited number of capital-intensive experiments have been designed and constructed over 20 years. For example, CERN currently hosts seven large experiments on the Large Hadron Collider, four of which are elaborate international collaborations (ATLAS, CMS, ALICE, LHCb). By contrast, MB is organised around the ‘laboratory’ – or even teams within a single institution. Very often, the molecular biologists are shaped by the conviction that they need to compete ‘for the priority of important findings’ (Knorr Cetina 1999), generating competition within – and across – laboratories.

When comparing how HEP and MB ascribe contributions to an individual scientist, we soon realised that HEP publications list a vast number of authors, as the construction and operation of HEP experiments often depends on many people; the record being over 5,000 authors on one article from CERN (Aad et al. 2015). In MB, although there are also challenges in ascribing results to individual scientists, the experiments are typically far less capital-intensive and permit differentiation in contributions within smaller teams. Finally, it is worth noting that some MB research is closer to commercial organisations (life sciences and pharma), whereas HEP is traditionally considered basic research with a more extended pathway towards any commercial outcome (Wareham et al. 2021; Romasanta et al. 2021). Accordingly, we would expect a more competitive culture with less data sharing in MB than HEP.

4.2.1. Open targets

The architecture of the OT data infrastructure is modular, containing different layers of access rights and data standards that employ a variety of mechanisms for researchers to be able to share their data (compliant with post-HGP norms). The stratified architecture grants different access rights to the data, where data generators are awarded complete access to a hidden layer, augmented by a public data layer (with different rights) that is accessible to any researcher willing to reuse the data. This modularity simultaneously allows researchers to grasp any individual or competitive benefits of being the generators of the data, while also being compliant with the collective norms of data transparency and sharing.

The modular architecture, with different access rights, also engages a time delay between the generation of the data and the publication of the data in the infrastructure that spans, on average, two years. As an informant explains: ‘Everybody understands that until there is a formal publication after the project, there is no disclosure.’

Finally, the information infrastructure acts as a ‘boundary organisation’ (O’Mahony and Bechky 2008); that is, ‘structures capable of effectively mediating between disparate constituencies and establishing common ground among the differing interests in the play’ (Perkmann and Schildt 2015, 1134). An interviewee explains: ‘There is a need to coordinate the integration of data into OT, both from the projects that generate data but also with the data providers such as Chembl and Uniprot and all the data that goes into the platform to keep it up to date. We also work with the developer team that creates some of the features that users will use to visualize the data coming through.’

Modularity and time delay are coordinated by the boundary organisation: normative governance on data access and reuse is embedded in the infrastructure, where the ownership and responsibilities over the data are explicit. These three mechanisms fit in a ‘logic of exchange’ that seeks to maximise benefits for the researchers (that is, the potential of data reuse and the commercial interests of data generators), while minimising the costs of sharing data (e.g. loss of potential commercial value, publication rights, recognition). This is achieved through protocols and data standards. The fact that for-profit companies form a significant part of the OT consortium suggests that the mechanisms developed are effective in balancing incentives to scientists while mitigating the risks of a competitive loss to other re-users of their data.

4.2.2. Reana

CERN built Reana upon data access and preservation policies agreed within the main experiments. Although the data policies may differ slightly across experiments, they all stratify the data generated by the HEP community in four main layers: (a) data directly related to the publications, which include the complete documentation for the published results; (b) simplified data formats devoted to training exercises within the physics community; (c) reconstructed data, simulations, and software analysis to facilitate research analysis; and finally, (d) the raw data and associated software, allowing access to the full potential of the experimental data’s reuse (Pujol Priego and Wareham 2019). Data sharing is concentrated in data layers (b) and (c). Raw data (d) are not made available to other researchers to reuse for pragmatic reasons. For instance, one of the core CERN experiments, CMS (Compact Muon Solenoid), produces on average 1 petabyte (100 gigabytes) of ‘raw’ data per second; similar data volumes characterise other experiments. As the Large Hadron Collider (LHC) data policy explains⁵ ‘It is practically impossible to make the full raw data-set from scientific endeavours of the scale of high-energy physics easily usable in a meaningful way outside of the collaboration […]It should be noted that, for these reasons, direct access to the raw data is not even permitted to individuals within the collaboration, and that instead the production of reconstructed data is performed centrally.’

Experiments also employ a time delay between the generation of the experimental data and the time of sharing with the external research community. These periods are also referred to as embargo periods that allow the data generators within the experiment to publish the results. As explained in the LHC experiment data policy: ‘In general data will be retained for the sole use of the collaboration for a period commensurate with the substantial investment in the effort needed to record, reconstruct and analyse those data. After this period, some portion of the data will then be made available externally, with this proportion rising with time …. The portion of the data which LHCb would normally make available is 50% after five years, rising to 100% after ten years.’

The main idea behind Reana’s infrastructure is to preserve software and data workflows so that they can enhance collaborative scientific work and diffuse knowledge of the experimental procedures (Dphep Study Group 2009). Such data sharing protocols and preservation techniques are embedded in the Reana framework and reinforce the need for quality metadata: ‘Our own experience from opening up vast volumes of data is that openness cannot simply be tacked on as an afterthought at the end of the scientific endeavour. Besides, openness alone does not guarantee reproducibility or reusability, so it should not be pursued as a goal in itself. Focusing on data is also not enough: it needs to be accompanied by software, workflow, and explanations, all of which need to be captured throughout the usual iterative and closed research lifecycle, ready for a timely open release with the results’ (Chen et al. 2018).

Reana acts as a boundary organisation or ‘interface’ to the experiment’s knowhow, so that other researchers outside the experiment can reuse it. While normative governance defining data access rights and responsibilities exists, it applied aapplied at the experiment level, not the infrastructure level. As such, the infrastructure is required to respect distinct data policies.

Table 5 provides a detailed description of the progression of our empirical analysis towards the three theoretical constructs: modularity, time delay, and boundary organisations. Table 6 summarises the similarities and dissimilarities identified from the Open Targets (MB) and Reana (HEP) analyses. While both scientific communities employ similar mechanisms to overcome scientists’ deterrents to share data, they differ in how such mechanisms are used in their respective infrastructures and scientific communities.

Table 5. Theoretical progression of our analysis

Empirical observations from data sources		Identification of theoretical constructs
MB- Open Targets	HEP- Reana
‘So, we have a platform that is public and open to everybody. Then, for the experimental projects, the partners share the data while they are creating it in Google buckets.’ ‘We have other features that are private, that we do not share with others. Those hidden features allow me, for instance, to work with my compound library on the platform, which I do not share with other OT partners.’	‘Open access to its data by people outside the collaboration can be considered at four levels of increasing complexity.’ CMS experiment preserves ‘the reconstructed data and simulations by keeping available a copy of the data reconstructed with the best available knowledge of the detector performance and conditions for each period of data-taking a virtualised computing environment, compatible with the software version with which the original data can be analysed’ (Dphep Study Group 2009: 7).	Modularity (Mechanism 1)
‘We have an internal internet that we use to say here’s what this data is and you can request it and I will send it to you in a password protected encrypted format. They get it sent and then I send them passwords separately, and then they take it from there. Lots of our partners, they prefer sometimes to use the raw data, so they all go through their own pipeline.’ ‘Once the project themselves have published the data in their own time, once that’s publicly available, then we can link to it on our public platform. In the meantime, though, it’s all very confidential and we don’t share anything outside of the internal platform that we have. It’s up to the project to have that publication before we start sending it out to the world.’	‘New data will enter the portal once the embargo periods for them are over.’ (CERN Open Data Portal) ‘The first data release of 2010 data took place in 2014.’ (R1) ‘The first data release was followed by a full analysis of the procedure, which was endorsed by the Collaboration Board in 2015, and regular data releases, accompanied by appropriate simulated data, each approved by the Collaboration Board, are now taking place.’ (CMS April 2018)	Time delay (Mechanism 2)
‘There is a need to coordinate the integration of data into OT, both from the projects that generate data but also with the data providers such as Chembl and Uniprot and all the data that goes into the platform to keep it up to date. We also work with the developer team that creates some of the features that users will use to visualize the data coming through.’	‘The data preservation process should follow well-defined policies, defined as soon as possible during the lifetime of the collaborations, and possibly embedded in a global HEP data preservation initiative.’	Boundary organisation (Mechanism 3)

Table 6. Similarities and differences observed between Reana (HEP) and open targets (MB)

Similarities (What mechanism)	Differences (How the mechanism is represented)
	HEP – Reana	MB – Open Targets
Modularity (Mechanism 1)	HEP establishes four layers of data: raw data is not released, while more curated versions of data are opened (level 2 in open data portal and reused in Reana; level 1 from publications through HEP library systems).	In MB, raw data from target associations with metadata is released in OT. However, the aggregations with data related to the next steps of the drug discovery process (e.g. proprietary compound libraries) remain closed.
Time Delay (Mechanism 2)	The embargo period of HEP is around 5 to 10 years, depending on the experiments. After the embargo period in HEP, only a % of the data is agreed to be released.	In MB, the time delay between the generation of the data and release in OT is of 18–24 months. In MB, all the data generated is shared in OT infrastructure.
Boundary organisation (Mechanism 3)	The boundary organisation and what makes the interface that mediates the data flows between researchers and establishes the rules, responsibilities, and drivers in data policies varies in the two cases. In HEP, the prominent role is played by the experiment, which decides rights and responsibilities across data. These rules prevail across infrastructures, including Reana. The competition over the data is not between scientists but between experiments.	In MB, the different experimental projects need to comply with the data governance and rules of OT, which establishes the protocols to avoid unintended spillovers and regulates the process to release the data.

5. Discussion

Our study aimed to answer two research questions: (1) Why and how do researchers from different scientific fields share their data? and (2) What mechanisms enable researchers to share their data? Regarding RQ1, we find that data sharing varies significantly across certain disciplines, although data sharing attitude and experience can be similar across fields. Most of the data sharing is carried out between collaborators on the same projects, suggesting that researchers adopt a discriminatory approach by sharing data with selected partners. Addressing RQ2, we find that communitarian and individualistic scientific communities employ three mechanisms (with some variation) to enable data sharing: (1) modularity, (2) time delay, and (3) boundary organisations. These mechanisms serve to establish transparent data governance and facilitate the identification of the ‘bona fide’ researcher.

Scientists are often professionally competitive. In this sense, individual incentives, professional recognition, and status are important components of a scientist’s career. However, nothing precludes collective norms and values from coexisting alongside individual motives; most scientists care about advancing science as a social good yet seek recognition for their contributions to it. Our survey evidence suggests that data sharing perceptions and practices are highly variable among academic disciplines. We infer that differences in the balance between individual and collective orientations explains, at least partially, some of the larger variance of data sharing across academic communities (Fulk et al. 2004; Hardin 1982; Ostrom 1990; Vitali, Mathiassen, and Rai 2018).

The nature of, and manner in which, science is conducted across disciplines is also highly determinative of data sharing (Borgman 2012; Knorr Cetina 1999, 2007; Gläser et al. 2015). HEP primarily conducts fundamental research with few immediate applications in industry; that is, while physics research has informed industrial development in a multitude of ways, the path to commercial applications is a longer one. MB, by contrast, is often more proximate to the life sciences and pharma industries. In fact, much MB research is funded by big pharma (Contreras and Vertinsky 2016; Vertinsky 2014; Cain 2012; Mittleman, Neil, and Cutcher-Gershenfeld 2013). So where HEP researchers have little reason to refrain from disclosing research data once academic credit is recognised, MB’s proximity to industries premised on finite periods of IP protection makes the calculus of disclosure far more complex.

With this background, our case analysis contrasted two information infrastructures that deployed mechanisms to align scientists’ professional incentives with data sharing practices (Figure 5).

Figure 5. Mechanisms that enable HEP and MB researchers to share data. The (+) show significant motivator/deterrent of data sharing from our study.

Our framework shows the tension between the community epistemic norms and the individual costs and benefits of data sharing. Our analysis identified three mechanisms to accommodate these tensions:

Data modularity enables data governance that acknowledges that research data are heterogeneous, as are the producers, audiences, and applications of such data. In HEP, more pragmatic considerations of the size and usability of data are determinative, while in MB, data modularity is conditioned by the applications of the data by its generators and consumers. Specifically, where most HEP research is publicly funded, MB research is funded by constellations of public and private sources. Consequently, demands for public disclosure need to be balanced with potential commercial appropriation for the private entities that have funded the research (Cain 2012; Mittleman, Neil, and Cutcher-Gershenfeld 2013).

Time delay also serves to balance any conflicting interests between the generators and consumers of data. In HEP, practical uncertainties about how data should be structured, analysed, or interpreted can require delays in its disclosure. In parallel, the requirements of the principal research teams who need sufficient time for data analysis and publication are also determinative. In MB, time delays serve a similar function: they permit the generators and funders of data to develop research leads towards commercial appropriability before releasing data into public platforms (Contreras 2010; Contreras and Vertinsky 2016).

Boundary organisations, finally, are of particular significance in their responsibility for data governance. These are defined by the different infrastructures required to conduct science in HEP and MB: the experiment or the information infrastructure. While both disciplines are data intensive, HEP requires particle accelerators, detection and imaging technologies of vast size, energies, and economic investment, that bind their operation to very large teams of scientists working on centrally coordinated and internationally funded experiments. This generates a highly collective culture with commensurate communal recognition and norms. Data collection, storage, and analysis are governed centrally, and are publicly transparent to a broad contingency of stakeholders whenever feasible. Most importantly, centralised data governance exists at the point of data genesis: for HEP, ‘data openness cannot simply be tacked on as an afterthought’ (Chen et al. 2018).

MB, by contrast, does not require the same level of public investment in infrastructure (it should, however, be noted that a significant amount of cell biology does transpire at larger synchrotron, free-electron laser, and neutron scattering facilities.) Many diagnostic and analytical instruments are owned and operated by individual organisations and laboratories; data generation is de-centralised. The functional units in MB are often smaller teams of researchers where the contributions of individual researchers are more transparent. Given the high status of much MB research, this fragmented structure can lead to competitive dynamics across research teams that inhibit data sharing. Additionally, a critical difference from HEP is that any decision to disclose MB data to a centrally governed information infrastructure is discretionary and most often occurs after the data genesis.

The appropriate governance of this subtle yet fundamental difference (mandatory data governance at data genesis, or discretionary data governance after data genesis) is supported by the time delays and modularity that are adequately adjustable to accommodate the divergent incentives and objectives of the researchers, institutions, and funders.

5.1. Implications for policy and practice

Our study reveals a complex interpretation of where a community’s norms mesh with individual incentives to share data for the collective benefit. The research community by and large is optimistic about the scientific benefits of data sharing. From our data, we find that 74% of researchers say that having access to other data would benefit them. There is a consensus that scientific data should be a public good: replicability and transparency are beneficial to science; FAIR data practices are desirable in principle; cooperation makes science more efficient and reduces scientific fraud. However, a closer examination of what scientists proclaim ideally, versus what they actually practice, reveals a more ambiguous situation. Scientists need assurance of recognition of their work, and private entities that fund research desire reasonable safeguards for a fair return on their financial investments. This implies that policies seeking to effectively boost scientific data sharing should aim to be tailored to meet the needs of discipline-specific practices and norms at both individual and institutional levels.

Accordingly, beyond the mechanisms identified in this study (data modularity, time delay, and boundary organisations), funding institutions and policymakers can consider additional levers that allow scientists to receive commensurate rewards for data cultivation and publication. Fundamentally, this means elevating the status of data curation from a necessary input to the scientific process, to a valid, high-status outcome in its own right. Inspiration can be found in the practices of patenting and software licencing which ensure that inventors share their knowledge in exchange for various rights. This implies that public agencies, academic institutions, and other arbiters of scientific merit, award data curation and publication greater status in funding decisions, recruitment and promotion processes, or other professional accolades. Such structural changes could potentially have shorter-term effects in the individual cost-benefit calculus of collective action (Fulk et al. 2004; Hardin 1982; Ostrom 1990; Vitali, Mathiassen, and Rai 2018), as well as longer-term effects in the evolution of the epistemic cultures (Borgman 2012; Knorr Cetina 1999, 2007; Gläser et al. 2015).

It is also important to highlight that there are additional considerations that determine data sharing attitudes and behaviours across disciplines. Scientists are often legitimately concerned with the potential misuse or erroneous interpretations of their results. The recent Covid-19 pandemic evidences the fact that data publication is not an end in itself, but must be tempered with qualified interpretation to inform public health policy appropriately. By extension, FAIR data practices entail costs of documenting metadata and scientific procedures in a manner facilitate appropriate interpretation and communication, which, for fields such as public health or environmental policy, are increasingly vital.

Many of these additional ‘costs’ (arbitration, interpretation, communication) are currently assumed by scientific organisations such as CERN, NASA, CDC, EMBL, etc. As such, our study informs the potential efforts of other scientific communities currently less reliant on the cultivation of large data quantities, but increasingly so. In addition to appropriate application of modularity, time delay, and boundary organisations, information infrastructures can be designed with complementary mechanisms that foresee data sharing implications of beyond immediate scientific communities, but policy makers and the public at large.

6. Limitations and future research

Our findings are subject to limitations that warrant further investigation. Although the sample sizes in our survey were large, given the relatively short interval between 2016 and 2018, this sampling is likely insufficient to detect long-term patterns of data sharing behaviour. Additional surveys in the coming years can enrich our current data to uncover confounding relationships in scientists’ willingness to share data. Research that purposefully examines heterogeneity in data sharing practices across disciplines can benefit from in-depth comparisons of high- and low- intensity data sharing communities. Furthermore, while HEP and MB represent disciplines that are very capital-intensive, a research design focusing on scientific contexts with different economic dynamics would be useful for extending our understanding of data sharing practices.

By using random selection methods to identify respondents representative of the researcher population at large, our survey sought to mitigate selection bias. However, we acknowledge a potential bias of researchers: those more inclined towards sharing scientific data could have a greater propensity to respond to the survey invitation. In addition, while the sponsorship of a major publisher in the survey might have also influenced survey response bias, we estimate that the additional involvement of an academic consortium and the European Commission could have partially counterbalanced any respondent bias.

Finally, regarding the case study analysis, we acknowledge that the challenge of the case method is to generalise the findings. Nevertheless, it is worth mentioning that there is a trade-off between internal and external validity. Our results are deeply grounded in the studied contexts, and we employed established procedures in inductive research to maximise the internal validity of our results. Consequently, we should be prudent in extrapolating our results to other contexts and scientific communities that do not display the same institutional and economic characteristics. We encourage additional in-depth research across other epistemic cultures and academic disciplines to better inform our understanding of how data sharing can be governed.

7. Conclusion

Data sharing is a practice intended for the collective benefit of scientific progress. Yet, reasons for its gradual and disparate adoption are less obvious. Scientific communities are far from united and display heterogeneous practices and norms in the way science is produced and how merit and status are allocated. Consequently, a delicate system of mechanisms needs to be established to align individual and collective incentives. The use of modularity, time delay, and boundary organisations are pivotal in the information infrastructures created by the scientific disciplines currently at the forefront of scientific data sharing. Other academic communities that seek to follow these examples can apply these mechanisms in a manner consistent with their own epistemic cultures and professional practices.

Acknowledgments

This study was funded by the Open Science Monitor (2017- 2019), a service contract with European Commission- DG RTD (Contract number PP-05622-2017) and implemented in collaboration with Elsevier, ESADE, Leiden University, and Lisbon Council.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the European Commission [Contract number PP-05622-2017].

Notes

¹ EMBL-EBI COVID-19 data portal: https://www.covid19dataportal.org/support-data-sharing-covid19

² The term FAIR was launched in the Lorentz workshop in 2014. The resulting FAIR principles were published in 2016. See https://www.go-fair.org/fair-principles/

³ re3data.org

⁴ Open Science Monitor: https://ec.europa.eu/info/research-and-innovation/strategy/goals-research-and-innovation-policy/open-science/open-science-monitor_en

⁵ Large Hadron Collider (LHC) data policy: https://twiki.cern.ch/twiki/pub/LHCb/LHCbDataPreservation/130321-LHCbDataAccessPolicy.pdf Retrieved 20^th October 2021:

Appendix 1. Analysis of survey sample

Table

Survey in this study			UNESCO data			Difference
Fields	Percent among repondents	Grouped	Fields	Percent across all researchers	Grouped
SocSci + Arts Hum + Economics	15.3%	15.3%	Social sciences	14.7%	20.8%	−5.5%
			Humanities	6.1%
Life Sciences	13.8%	38.1%	Natural sciences	18.3%	24.2%	13.9%
Earth & Env. Science	12.9%
Physics & Astronomy	7.0%		Agricultural and veterinary sciences	5.9%
Chemistry	4.4%
Computer Science	7.7%	31.9%	Engineering and tech	41.6%	41.6%	−9.7%
Engineering	15.9%
Material Science	4.1%
Maths	4.3%
Medicine and Allied Health	12.1%	12.1%	Medical and health sciences	13.4%	13.4%	−1.3%
Other	2.6%	2.6%				2.6%
Total	100.0%	100.0%	Total	100.0%	100.0%