: Guidelines and Best Practices to Share Deidentiﬁed Data and Code

In 2022, the Journal of Statistics and Data Science Education (JSDSE) instituted augmented requirements for authors to post deidentified data and code underlying their papers. These changes were prompted by an increased focus on reproducibility and open science (NASEM 2019). A recent review of data availability practices noted that"such policies help increase the reproducibility of the published literature, as well as make a larger body of data available for reuse and re-analysis"(PLOS ONE, 2024). JSDSE values accessibility as it endeavors to share knowledge that can improve educational approaches to teaching statistics and data science. Because institution, environment, and students differ across readers of the journal, it is especially important to facilitate the transfer of a journal article's findings to new contexts. This process may require digging into more of the details, including the deidentified data and code. Our goal is to provide our readers and authors with a review of why the requirements for code and data sharing were instituted, summarize ongoing trends and developments in open science, discuss options for data and code sharing, and share advice for authors.


Introduction
In 2022, the Journal of Statistics and Data Science Education (JSDSE) instituted augmented requirements for authors to post deidentified data and code underlying their papers (Horton et al. 2022).These changes were prompted by an increased focus on reproducibility and open science (National Academies of Sciences, Engineering, and Medicine 2019).A recent review of data availability practices noted that "such policies help increase the reproducibility of the published literature, as well as make a larger body of data available for reuse and re-analysis" (Federer et al. 2018).
JSDSE values accessibility as it endeavors to share knowledge that can improve educational approaches to teaching statistics and data science.Because institution, environment, and students differ across readers of the journal, it is especially important to facilitate the transfer of a journal article's findings to new contexts.This process may require digging into more of the details, including the deidentified data and code used to support the article's findings.
Two years on, we felt that it was valuable to provide our readers and authors with a review of why the requirements for code and data sharing were instituted, summarize ongoing trends and developments in open science, discuss options for data and code sharing, and share advice for authors.

Why is data and code sharing necessary?
Why the push for the sharing of data and code?Data sharing, as well as the sharing of the code for the processing and analysis steps, is a necessity if readers or other researchers are to reproduce findings published in the journal.Beyond building trust in a particular article's findings, the sharing of deidentified data can help others in their own studies.But it's not just deidentified data that are needed.As but one example of the importance of code sharing, Oza (2023) found that when a large group of biologists analyzed the same dataset, their results were widely divergent.

An editorial in PLOS ONE
A NASEM report on Reproducibility and Replicability in Science concluded that "Journal editors should consider ways to ensure reproducibility for publications that make claims based on computations, to the extent ethically and legally possible (page 2)" (National Academies of Sciences, Engineering, and Medicine 2019).
A recent editorial in the Journal of the American Statistical Association (JASA), the flagship journal of the American Statistical Association (ASA) noted that their reproducibility process "fosters a culture of transparency and accountability that is critical to nurture the trust placed by the lay public in scientific research" (Wrobel et al. 2024).
We anticipate that other entities will be moving in this direction in the future.

How have other journals responded?
In the past ten years, many top journals have added requirements to support reproducible research, including policies on data sharing.In March 2014, PLOS ONE started requiring that the supporting data for the results in its publications be shared (Federer et al. 2018).The Harvard Data Science Review released a data sharing requirement in 2021 (Harvard Data Science Review 2021).
In September 2016, Nature started to require that data availability statements be included as part of its publications.Although this policy doesn't require that the data actually be made available, the intention of the data availability statement is to make the authors' decision to share or not share the data more transparent.
The ASA released recommendations in 2018 for their journals that are jointly published with Taylor & Francis (this set of journals includes JSDSE ).These recommendations encourage data sharing and the inclusion of data availability statements (Wasserstein 2018).
In 2016, JASA began to require both code and data at the revision stage of submissions to the "Applications and Case Studies" section (JASA Reproducibility Initiative 2019).JASA editors extended this requirement to the revision stage of submissions to the "Theory and Methods" section in 2021.
4 How best to share deidentified data

What datasets can be shared?
Many journals' data sharing requirements reference a "minimum" or "minimal" dataset that is required to be shared.For example, PLOS ONE considers a minimal dataset as one that has "the data required to replicate all study findings reported in the article, as well as related metadata and methods" (PLOS Journals 2019) while Nature's minimum dataset is one that is "necessary to interpret, verify and extend the research in the article, transparent to readers" (Nature Portfolio 2019).The FAIR Data Principles of data sharing request that these shared datasets are Findable, Accessible, Interoperable , and Reusable with the goal of increasing "knowledge discovery and innovation" (Wilkinson et al. 2016).
At times, investigators have gone back to their Institutional Review Board (IRB) to seek approval for requests to retroactively share deidentified data.At the heart of such a request is the creation of a "minimal" dataset without any direct or indirect identifiers.
As one example, this process was undertaken to make the data from the Health Evaluation and Linkage to Primary Care (HELP) study (Samet et al. 2003) available on the Center for Open Science's Open Science Framework (OSF) (Horton 2024).
However, sharing a "minimum" dataset that just elides identifiers may not be sufficient to avoid reidentification.When creating a "minimum" dataset it may be necessary to only share a proper subset of the attributes recorded on individual subjects to minimize risk of de-anonymization.
Moving forward, researchers need to be thinking about data sharing when they begin to design their studies.The DMP Tool offers useful guidance for investigators designing data management and sharing plans (California Digital Library 2024).
Other approaches may be needed, depending on the study.Consider a hypothetical qualitative research example where students are being interviewed before graduating to find out about their trajectory through a data science program.They are asked a series of questions, and their responses are transcribed.The full text of the interview would reveal a lot of personally identifiable information.However, the transcripts are often coded by researchers as a pre-processing step before analysis.These codes are what are used in the analysis phase of the project, so having access to the coded data rather than the raw transcript data would Many journals' data sharing requirements provide caveats and exceptions for situations where data sharing may be problematic, especially with respect to privacy, and leave room for more controlled access.Similar policies operate for JSDSE (see Section 6.3).

What file formats and licensing?
Making data "interoperable," as the FAIR Data Principles advocate for, includes sharing it in a non-proprietary file format (such as CSV or TXT) so that it is easily accessed across platforms and using a variety of software.Making data "reusable" includes sharing it with a clear usage license for those who want to use it in their own work.Recent papers published in JSDSE are made available under a Creative Commons license, with copyright held by the author.This gives readers the right to build upon the work.The "accessible" piece of FAIR Data principles speaks to making data freely available, which can broaden who can do further research on the topic (Nagaraj et al. 2020).

Which repository to use?
Journals like Nature (Nature 2024) and Proceedings of the National Academy of Sciences (PNAS) (Proceedings of the National Academy of Sciences 2024) provide detailed guidelines on appropriate repositories for sharing data; these can be used as references for exploring repository options.
The FAIR Data Principles also provide a framework for determining what makes a "good" repository.For example, part of the "findable" piece of the FAIR Data Principles involves having a way to link to the data that is permanent.A digital object identifier (DOI) is an example of a permanent, FAIR, way to access a dataset.A GitHub repository is not "findable" in this sense because the repository can be deleted, renamed, or reorganized, such that the link to the repository no longer brings a reader to a dataset as intended (FAIRdata Forum 2021).
JSDSE authors have deposited data in a variety of FAIR repositories.A number have utilized the OSF (Alzen et al. 2024) which facilitates anonymization of authors (that can then be unmasked once a paper has been accepted) (Soderberg 2018).Others have used Zenodo (Mocko et al. 2024) and Mendeley Data (Miller & Pyper 2024), sometimes based on recommendations or requirements of their institutions.

Challenges to sharing of deidentified data
Although broad access is a value, the devil really is in the details.There is a need to balance both transparency and reproducibility with data privacy (Zaslavsky & Horton 1998).Many ethical considerations exist for data that often comes from students.
While many investigations might be deemed exempt from human subjects oversight (involving no or minimal risk), we don't have to look far to come up with a plausible, yet, more complicated scenario that our journal has already faced.
Suppose each row in a dataset is a student.There are columns for performance on a pre-test, performance on a post-test, and a variety of demographic data including race, major, gender, and class year.If this data were shared without any names of the students, that would on the surface look deidentified.However, the combination of demographic variables can reveal individuals, especially those belonging to intersectional, minority groups.For example, if there was only one first-year, Asian, male student in the class, we automatically know how they performed on the pre-test and post-test, which violates this student's privacy.(Even if there were more than one such student, the probability of reidentification is non-zero.) One option is to remove the sensitive demographic columns of the data, and focus on the response and covariates of interest.If we are interested in the relationship between pre-test and post-test scores, the demographic variables are not essential to reproduce the analysis.Removing potentially identifying demographic variables in this way and sharing a deidentified dataset that is a subset of the full data is the approach taken by Evans et al. (2023).
However, if those demographic variables are part of the research question, i.e., we want to investigate how demographic variables may mediate the relationship between pre-test and post-test performance, then these variables would be an important part of the reproducibility process.Another option might be to create synthetic data (Census Bureau 2021) that has similar characteristics to the raw data but where each row does not represent an actual student.
The idea here is that the shared processing and analysis code could be run using the synthetic data, but no student data would be revealed.Such an approach would be indicated if there was substantial risk of disclosure (typically beyond what is considered "exempt from human subjects oversight" by many IRBs).
6 Answers to Frequently Asked Questions

What should a data availability statement look like?
Even in a simple scenario, we can see that each dataset comes with its own nuance.Therefore, in answer to this question, we resort to a statistician's or data scientist's favorite phrase: it depends!Articles published in JSDSE are not intended to adhere to a single, rigid, data availability policy but rather the intent of requiring a data availability statement at all is to start a conversation and a creative problem solving process.The journal wants to work with authors to negotiate a minimal dataset that balances the dueling goals of accessibility and protecting the rights of the data subjects.
That being said, we provide several examples of possible data availability statements that meet the letter and spirit of the guidelines: 1.The deidentified data and code that support the findings of this study are openly available at the Open Science Framework (OSF): provide URL here.
2. The authors confirm that the data supporting the findings of this study are available within the article.
3. Data sharing is not applicable to this article as no new data were created or analyzed in this study.
Student data can come in many forms, their responses to a survey, their performance on an assessment, their demographics, but they can also come in a variety of types including qualitative data.Written responses, audio/video, and interview transcripts are personal and can contain more identifiable information than quantitative data, which can make sharing the raw data challenging.Such a study might adopt the following data availability statement: 4. Due to the nature of the research, the video recordings of the participants are not available.The deidentified coded data and analysis code that support the findings of this study are openly available at the Open Science Framework (OSF): provide URL here.

Is share upon request an acceptable option?
Alternative approaches to data availability statements such as "share upon request" have been notoriously ineffective as a way of fostering data and code sharing (Tedersoo et al. 2021, Gabelica et al. 2022)."Share upon request" is not, in general, allowed for JSDSE papers.

How can a waiver be requested?
Authors seeking an exemption to deidentified data sharing should send a formal request to the editor asking for a waiver.Such a request should include (1) the justification for why a deidentified dataset cannot be made available, (2) background on the study including the original IRB protocol and study documents, and (3) a codebook for the data included in the waiver.

Where should the data availability statement be located?
In terms of placement, the data availability statement should appear just before the references.Information about where deidentified data and code can be found is also provided via metadata input during the submission process.A stub data availability statement is included in the template (LaTeX and Quarto) provided for authors (Journal of Statistics and Data Science Education 2024).

Where can I learn more about computational reproducibility?
This editorial has focused most heavily on the details surrounding the sharing of deidentified data as privacy concerns make additional guidance necessary.However, the sharing of the code used to process and analyze that data is also a necessary but insufficient step to foster computational reproducibility (Project TIER 2024).Sandve et al. (2013) provide ten simple rules for reproducible computational research.While all are important, rule ten describes ways to provide public access to scripts and results.Ball et al. (2022) describe approaches for teaching reproducible science.We encourage authors to take advantage of the Quarto paper template (Quarto Journals Project 2024) that has been made available to facilitate reproducible analysis and reporting.

Closing thoughts
We realize that best practices to foster reproducibility and replicability are fast changing and require additional efforts by authors, reviewers, and editors.This process will require updates to our standard procedures, including improved templates for human subjects oversight.For example, protocols will need to no longer state that "deidentified data will be destroyed at the end of the study".It will be equally important to inform subjects that fully deidentified data will be made available and not just summarized in aggregate.
However, we believe that these changes can and should be undertaken, and that these changes will allow us as a community to balance privacy and sharing in a way that fosters better science.
At present, the deidentified data and code included with submissions are not formally included in the review process (though some reviewers and associate editors may choose to incorporate this information).In the future, the journal may consider formalizing this review and/or starting up a reproducibility process such as that described by Wrobel et al. (2024) or the creative approach undertaken by the journals published by the American Economic Association (Vilhuber et al. 2022).
We hope that this discussion will benefit future authors and readers as they navigate this new territory.We concur with the assessment of Wrobel et al. (2024) which stated that "the journal gains credibility from a system that holds researchers accountable for their work, credibility that can be leveraged to publish controversial and impactful research that has the potential to change a field of study".Nicholas J. Horton ORCID: https://orcid.org/0000-0003-3332-4311Sara Stoudt ORCID: https://orcid.org/0000-0002-1693-8058 (2019) states that: "Data availability allows and facilitates: 1. Validation, replication, reanalysis, new analysis, reinterpretation or inclusion into meta-analyses; 2. Reproducibility of research; 3. Efforts to ensure data are archived, increasing the value of the investment made in funding scientific research; 4. Reduction of the burden on authors in preserving and finding old data, and managing data access requests; 5. Citation and linking of research data and their associated articles, enhancing visibility and ensuring recognition for authors, data producers and curators."Funding bodies have been requiring researchers to share their data as well.In August 2022, the White House Office of Science and Technology Policy directed federal agencies to update their data-sharing policies such that all federally-funded research data would have to be shared freely by 2025 (DuBois et al. 2023).The National Institute of Health's Generalist Repository Ecosystem Initiative has taken a comprehensive approach to foster data and workflow sharing (National Institutes of Health Office of Data Science Strategy 2023).A special theme on "Changing the Culture on Data Management and Data Sharing in Biomedicine" was published in 2022 (Harvard Data Science Review 2022).
be enough to reproduce the work.JSDSE has published papers involving the analysis of qualitative data (Lesser & Santos 2024) where the coded dataset is shared (and not the full transcripts, video recordings, etc.).DuBois et al. (2023) considers a similar question.