Data publication consensus and controversies

Why publish data?

The underlying goals of data publication are to enable research to be reproduced and data to be reused. Hidden primary data exacerbates science’s very public “reproducibility crisis”^5–9, most recently illustrated by the collapse of a pair of irreproducible Nature articles describing a simple method to transform somatic cells into pluripotent stem cells^10,11. Widespread publication of the data underlying research papers could help expose both honest errors and fraud¹². The leaders of the US National Institutes of Health (NIH) recently cited “provid[ing] greater transparency of the data that are the basis of published manuscripts” as one way to improve scientific reproducibility¹³.

Journals already frequently require authors to supply underlying data on request. In 2011, Alsheikh-Ali et al.¹⁴ found that 88% of high-impact journals required a statement regarding the availability of underlying data; half of those made willingness to provide data a condition of publication. However, the authors of 59% of papers examined in the study failed to adhere to the availability instructions. Vines et al. (2014)¹⁵ could only obtain underlying data from 101 of 516 papers published from 1991 to 2011. Availability dropped off sharply with time; data could be obtained from only two of the 62 oldest papers. Now, some journals require that underlying data be published simultaneously with the article.

In 2010, a coalition of Ecology and Evolutionary Biology journals began to require that the data underlying articles be archived with a maximum embargo of one year^16,17. F1000Research has had a similar policy (without an embargo period) since its inception, and the Public Library of Science (PLOS) journals followed suit earlier this year¹⁸. Although there can be no substitute for funding new experiments and data collection, appropriate data reuse lowers costs and accelerates research. Documenting, publishing, and archiving data is time consuming and costly, but usually far less so than repeating the data collection. For example, Open Context published archaeological data from a site in eastern Turkey at the substantial cost of $10,000–15,0000, but this publication expense was minor compared to $800,000 spent to collect the data¹⁹. Piwowar (2011) contrasted the impact of $100,000 in National Science Foundation (NSF) grants, which generates an average of three to four papers, with an estimate that the same investment in curating, archiving, and publishing data could contribute to over 1,000 publications²⁰. Furthermore, while some data is merely expensive to recreate, time-dependent or ephemeral data, (e.g. climate records or observations of unique astronomical events) should be published because it can never be recreated for any price²¹.

Data that supplements a paper

The most familiar kind of data publication is a traditional journal article accompanied by underlying data. That data can be hosted by the journal as supplementary material or deposited in a third-party repository. The trend is away from supplemental material because repositories are considered to be better suited to ensure long-term preservation and access to the data. For instance, The Journal of Neuroscience stopped publishing supplemental material in 2010; the announcement promotes disciplinary repositories as “vastly superior to supplemental material as a mechanism for disseminating data”²². Data underlying any peer-reviewed or otherwise “reputable” publication can be deposited in the Dryad repository. Dryad makes data available and citable, but the publisher of the article must manage any assessment of scientific validity. Other third-party repositories include Figshare, Zenodo, institutional repositories (e.g. the Purdue Research Repository), and discipline-specific repositories (e.g. DNA sequences are deposited in GenBank²³ and protein structures in the Protein Data Bank²⁴).

Data as the subject of a paper

A data paper describes a dataset with thoroughly detailed rationale and collection methods, but lacks any analysis or conclusions²⁵. Data papers are flourishing as a new article type in journals such as F1000Research, Internet Archaeology, and GigaScience, as well as in dedicated journals like Geoscience Data Journal, Nature Publishing Group’s Scientific Data, and a trio of “metajournals” from Ubiquity Press.

Data paper length and structure varies between journals, but the tendency is toward a short, tightly structured format. All journals require an abstract, collection methods, and a description of the dataset; a few encourage authors to suggest potential uses for the data (e.g. Internet Archaeology, and Open Health Data). Some journals supplement this general framework with field-specific sections. (e.g. Internet Archaeology and the Journal of Open Archaeology Data each include a section for temporal and geographic scope.) Data papers are most sharply defined not by the presence of any particular information, but by the absence of analysis or conclusions. A crisp distinction from other article types is important because many journals do not consider a data paper to be prior publication if the authors seek to publish an analysis of the same dataset (e.g. Nature-titled journals, Science, and others listed by F1000Research).

Data journals generally limit themselves to publishing the description of the dataset; a trusted repository publishes the data itself. For instance, Scientific Data and Geoscience Data Journal each direct authors to a list of approved repositories. As an exception, GigaScience hosts data in an integrated repository named GigaDB. An early implementer of data papers, The International Journal of Robotics Research²⁵ is unusual in that they permit authors to host datasets on their own websites.

Data independent of any paper

To be useful or reproducible, a dataset must be accompanied by descriptive information (i.e. metadata)²¹, but this need not take the form of a journal article. Instead, some repositories publish rich, structured and/or freeform description together with the data. The distinction between a data repository and a data publisher is often indistinct. Repositories provide access and citability, but the degree of validation varies widely and few are equipped to provide peer review. For instance, to make data publication as easy as possible for authors, Figshare and Zenodo publish datasets from any field with minimal validation.

Simple case

The present consensus is that a dataset should be cited using, at a minimum, five elements largely familiar from article citations: creator(s), title, year, publisher and identifier. This format agrees with CODATA’s recommendation³² and conveys all the information required to obtain a DataCite DOI³³ or be listed in the Thomson-Reuters Data Citation Index. However, this article-derived format fails to address some of the complications unique to datasets, described below.

Deep citation

The first major complication that data citation faces is the need for deep citation. When supporting an assertion in writing, it usually suffices to cite the entirety of a journal article and leave it to the inquisitive reader to find the relevant passage. But, to reproduce an analysis performed on a subset of a larger dataset, the reader needs to know exactly what subset was used (e.g. a limited range of dates, only the adult subjects, wind speed but not direction). Datasets vary so widely in structure that there may not be a good general solution for describing subsets. The most common suggestion is to cite the entire dataset in the reference list and describe the subset in the text of the paper³⁴. In straightforward cases, the Federation of Earth Science Information Partners (ESIP) and the National Snow and Ice Data Center (NSIDC) both recommend including a list of variables or range of dates in the formal citation.

Dynamic datasets

The second major complication arises when datasets change. In the past, the printing process cemented one version of an article as the version of record. Even for traditional scholarly literature, web-based publishing and preprint servers (e.g. arXiv.org) are complicating the situation, but datasets are especially prone to be dynamic. Two kinds of dynamic datasets warrant consideration: growing datasets that add new data while never changing or deleting existing data, and revisable datasets where data may by added, deleted, or changed.

Consider USC00046336, a weather station at the Oakland Museum. Each day, the high temperature, low temperature and amount of precipitation recorded at the Museum³⁵ flow, together with data from more than 20,000 other stations, into the swelling Global Historical Climate Network (GHCN)-Daily³⁶ dataset. Or, consider WormBase³⁷, a genome database used by the Caenorhabditis elegans research community. WormBase encompasses genomic sequences of C. elegans and 20 related species massively annotated with gene structures, protein sequences, expression patterns, and a host of other information from empirical data and computational predictions. Every two months, WormBase responds to new data and better computational models by issuing a revised version with new material added and inaccurate material deleted or corrected.

Additions and updates to published datasets are extremely valuable, but a researcher seeking to reproduce an analysis of a dynamic dataset needs access to a particular version. To enable that access, previous versions must be preserved and citable. Growing datasets can be cited with an access date or a date range in the citation, as recommended by ESIP and NSIDC. Revisable datasets are more difficult; the most common approach is to accumulate revisions and periodically publish a new version with a citable version number. For example, WormBase identifies each release with a citable version number and makes all of the previous versions available.

Controversy persists around the specific issue of identifiers for dynamic datasets. DataCite recommends, but does not insist, that their DOIs refer to immutable objects. NSICD and ESIP instruct researchers to use a single identifier for growing datasets and include the access date in the citation; each major version of a revisable datasets gets a new identifier, but minor versions do not. In contrast, the DCC, Dataverse, and the UK Natural Environment Research Council (NERC) insist that any change to a dataset should trigger a new identifier^4,34,38. To handle the difficulties with dynamic data that this policy creates, the DCC recommends periodically issuing growing datasets a new identifier that refers to the “time-slice” of new records and freezing versions of revisable datasets as individually-identified “snapshots”.

Just-in-time identifiers

The difficulties surrounding deep citation and dynamic data could potentially be solved by turning the identifier-issuing process on its head. Instead of the dataset publisher issuing identifiers for data at the level that researchers seem likely to cite, researchers could issue identifiers for precisely the part of the dataset that they want to cite. The Research Data Alliance (RDA) Data Citation Working Group recently put forth a sophisticated proposal applicable to data in (or convertible to) databases. Identifiers created under this scheme would wrap together identification of a database, a query to return the cited dataset, the version of the database queried for this analysis, and a number of other useful components. Although promising, many technical and policy issues must be resolved before this approach can be widely adopted.

Data paper peer review

Peer review guarantees that journal articles entering the scholarly record reach some level of validity (although the aforementioned reproducibility crisis calls into question exactly what that level is). In many fields, peer-reviewed publications enjoy a much higher status than any other literature. Any effort to apply the prestige of “publication” to datasets cascades naturally into an effort to apply the prestige of “peer review”. But as data validation seeks to model itself on literature peer review, literature peer review itself is in flux^39–41. Open peer review at F1000Research and post-publication commenting at PubMed Commons are just two of many ongoing web-enabled experiments in article evaluation.

Journal article reviewers traditionally consider whether the methods used are appropriate for the questions asked and the data collected support the conclusions drawn. In the absence of particular questions and conclusions, it is not obvious what peer review of data should certify. A dataset may be suitable for some purposes, but not for others⁴². In addition, while a reviewer can be expected to read an entire article, they cannot inspect every point in a large dataset. Finally, researchers are already over-whelmed by peer review of articles⁴³ and may find any increased workload unreasonable. Despite all these difficulties, venues for peer-reviewed data papers are opening rapidly.

Data paper journals wrap scientific peer review of the paper and the dataset together into a single process. GigaScience, an exception, assigns technical review of the dataset to a separate data reviewer. The standards that various data journals provide to reviewers are fairly uniform, with the exception that about half of consider novelty or potential impact, while the rest only require that the dataset be scientifically sound. While review standards are similar, processes differ widely.

As an example, compare Biodiversity Journal and Scientific Data. Both journals divide reviewer guidelines into three sections along similar lines, which Biodiversity Journal calls “quality of the data”, “quality of the description”, and “consistency between manuscript and data”. Scientific Data follows a traditional peer-review process: an editor appoints reviewers who are encouraged to remain anonymous. In contrast, review at Biodiversity Journal follows a flexible and open process featuring entirely optional anonymity and multiple types of reviewer. There, an editor appoints two or three “nominated” reviewers who must report back and several “panel” reviewers who read the paper and only comment at their discretion. Additionally, the authors may choose to open the paper to public comment during the review process.

Independent data validation

Data journals all model their data validation more or less faithfully on literature peer review, but independent data validation practices and proposals are considerably more varied. On the conservative end of the spectrum, Lawrence et al. (2011) propose a set of criteria for independent data peer review⁴⁴. The Planetary Data System (PDS) peer-reviews datasets through the unusual process of holding an in-person meeting with representatives of the repository, the dataset creators, and the reviewers.

Two examples from archaeology, Open Context and the Digital Archaeological Record (tDAR), illustrate the diversity of approaches to data validation. Open Context provides multiple validation processes that incorporate peer review in a way that goes beyond the simple accept/reject binary¹⁹. Each Open Context dataset is rated from one to five based not on quality per se, but on the thoroughness of the validation; a one comes with no guarantees, a three has passed a technical review, and a five has passed external peer review. Whereas Open Context is a boutique publisher, focusing on data presentation and reuse, tDAR is a large repository primarily concerned with with collecting and preserving archaeology data for future use. tDAR is able to operate at scale by performing only technical validation and streamlining data deposition with a minimum of mandatory description. However, tDAR also serves as a platform for high-quality data publication. The repository accommodates contributors who wish to provide more information, and much of the content is deposited by digital curators who can be relied on to supply rich descriptions. Furthermore, two data paper journals, Internet Archaeology and Journal of Open Archaeological Data, recommend tDAR as a repository for their peer-reviewed data. Thus, data validation depends not only on discipline and data type, but on a host of external factors, including the goals of the organizations and researchers involved.

Pre-publication validation can be supplemented or replaced by post-publication feedback from successful or unsuccessful reusers. Parsons et al. (2010) suggest that “data use in its own right provides a form of review”, and go on to point out that the context of reuse demonstrates that the data is not generically “good”, but fit for some particular purpose⁴². The DANS repository solicits feedback from researchers who use its datasets: users are asked to rate the dataset on a one to five scale in each of six criteria (e.g., data quality, quality of the documentation, structure of the dataset)^45,46.