Fifteen ideas about data validation (and peer review)

Many open issues drift around data publication, but validation is both the biggest and the haziest. Some form of validation at some stage in a data publication process is essential; data users need to know that they can trust the data they want to use, data creators need a stamp of approval to get credit for their work, and the publication process must avoid getting clogged with unusable junk. However, the scientific literature’s validation mechanisms don’t translate as directly to data as its mechanism for, say, citation.

This post is in part a very late response to a data publication workshop I attended last February at the International Digital Curation Conference (IDCC). In a breakout discussion of models for data peer review, there were far more ideas about data review than time to discuss them. Here, for reference purposes, is a longish list of non-parallel, sometimes-overlapping ideas about how data review, validation, or quality assessment could or should work. I’ve tried to stay away from deeper consideration of what data quality means (which I’ll discuss in a future post) and from the broader issues of peer review associated with the literature, but they inevitably pop up anyway.

1. Data validation is like peer review of the literature: Peer review is an integral part of science; even when they resent the process, scientists understand and respect it. If we are to ask them to start reviewing data, it behooves us to slip data into existing structures. Data reviewed in conjunction with a paper fits this approach. Nature publishing group’s Scientific Data publishes data papers through a traditional review process that considers the data as well as the paper. Peer review at F1000Research follows a literature-descended (although decidedly non-traditional) process that asks reviewers to examine underlying data together with the paper.
2. Data validation is not like peer review of the literature: Data is fundamentally different from literature, and shouldn’t be treated as such. As Mark Parsons put it at the workshop, “literature is an argument; data is a fact.” The fundamental question in peer review of an article is “did the authors actually demonstrate what they claim?” This involves evaluation of the data, but in the context of a particular question and conclusion. Without a question, there is no context, and no way to meaningfully evaluate the data.
3. Divide the concerns: Separate out aspects of data quality and consider them independently. For example, Sarah Callaghan divides data quality into technical and scientific quality. Technical quality demands complete data and metadata and appropriate file formats; scientific quality requires appropriate collection methods and high overall believability.
4. Divvy up the roles: Separate concerns need not be evaluated by the same person or even the same organization. For instance, GigaScience assigns a separate data reviewer for technical review. Data paper publishers generally coordinate scientific review and leave at least some portion of the technical review to the repository that houses the data. Third party peer-review services like LIBRE or Rubriq could conceivably take up data review.
5. Review data and metadata together: A reviewer must assess data in conjunction with its documentation and metadata. Assessing data quality without considering documentation is both impossible and pointless; it’s impossible to know that data is “good” without knowing exactly what it is and, even if one could, it would be pointless because no one will ever be able to use it. This idea is at least implicit any data review scheme. In particular, data paper journals explicitly raise evaluation of the documentation to the same level as evaluation of the data. Biodiversity Data Journal’s peer review guidelines are not unusual in addressing not only the quality of the data and the quality of the documentation, but the consistency between them.
6. Experts should review the data: Like a journal article, a dataset should pass review by experts in the field. Datasets are especially prone to cross-disciplinary use, in which case the user may not have the background to evaluate the data themselves. Sarah Callaghan illustrated how peer review might work– even without a data paper– by reviewing a pair of (already published) datasets.
7. The community should review the data: Like a journal article, the real value of a dataset emerges over time as a result of community engagement. After a slow start, post-publication commenting on journal articles (e.g. through PubMed Commons) seems to be gaining momentum.
8. Users should review the data: Data review can be a byproduct of use. A researcher using a dataset interrogates it more thoroughly than someone just reviewing it. And, because they were doing it anyway, the only “cost” is the effort of capturing their opinion. In a pilot study, the Dutch Data Archiving and Networked Services repository solicited feedback by emailing a link to an online form to researchers who had downloaded their data.
9. Use is review: “Indeed, data use in its own right provides a form of review.” Even without explicit feedback, evidence of successful use is itself evidence of quality. Such evidence could be presented by collecting a list of papers that cite to the dataset.
10. Forget quality, consider fitness for purpose: A dataset may be good enough for one purpose but not another. Trying to assess the general “quality” of a dataset is hopeless; consider instead whether the dataset is suited to a particular use. Extending the previous idea, documentation of how and in what contexts a dataset has been used may be more informative than an assessment of abstract quality.
11. Rate data with multiple levels of quality: The binary accept/reject of traditional peer review (or, for that matter, fit/unfit for purpose) is overly reductive. A one-to-five (or one-to-ten) scale, familiar from pretty much the entire internet, affords a more nuanced view. The Public Library of Science (PLOS) Open Evaluation Tool applies a five-point scale to journal articles, and DANS users rated datasets on an Amazon-style five-star scale.
12. Offer users multiple levels of assurance: Not all data, even in one place, needs be reviewed to the same extent. It may be sensible to invest limited resources to most thoroughly validate those datasets which are most likely to be used. For example, Open Context offers five different levels of assurance, ranging from “demonstration, minimal editorial acceptance” to “peer-reviewed.” This idea could also be framed as levels of service ranging (as Mark Parsons put it at the workshop) from “just thrown out there” to “someone answers the phone.”
13. Rate data along multiple facets : Data can be validated or rated along multiple facets or axes. DANS datasets are rated on quality, completeness, consistency, and structure; two additional facets address documentation quality and usefulness of file formats. This is arguably a different framing of  divided concerns, with a difference in application: there, independent assessments are ultimately synthesized into a single verdict; here, the facets are presented separately.
14. Dynamic datasets need ongoing review: Datasets can change over time, either through addition of new data or revision and correction of existing data. Additions and changes to datasets may necessitate a new (perhaps less extensive) review. Lawrence (2011) asserts that any change to a dataset should trigger a new review.
15. Unknown users will put the data to unknown uses: Whereas the audience for, and findings of, a journal article are fairly well understood by the author, a dataset may be used by a researcher from a distant field for an unimaginable purpose. Such a person is both the most important to provide validation for– because they lack the expertise to evaluate the data themselves– and the most difficult– because no one can guess who they will be or what they will want to do.

Have an idea about data review that I left out? Let us know in the comments!