In late 2023, the Cognitive Neuroscience Laboratory at the University of Northwood received a notice from the National Institute for Behavioral Science (NIBS): all data supporting published papers must be deposited in an open, machine-readable format, with a persistent identifier, within six months of acceptance. The lab, which had spent years storing results in proprietary spreadsheet files and a smattering of PDFs, faced a sudden, unglamorous overhaul. The mandate was part of NIBS's broader push for data citation, but its immediate effect was far more concrete: it forced a reckoning with file formats. What followed, documented in the lab's internal logs and interviews with its staff, offers a case study in how research policy lands on the ground, reshaping not just one lab's workflow but the two repositories that received its data.

A Mandate Lands in One Lab

The lab, which I'll refer to as the Northwood Lab, had been collecting behavioral and imaging data for years. Its standard practice was to save raw outputs in Excel, with analysis files in SPSS's native format. Published papers included summary statistics and figures, but the underlying data lived on a shared drive, accessible only to lab members. The mandate, part of NIBS's pilot program on data citation, required that all data be deposited in a repository with a DOI, in a format that could be read by any standard tool.

The immediate reaction among the lab's five postdocs and three graduate students was, by all accounts, resistance. One postdoc, who asked not to be named, described the mandate as "a bureaucratic hurdle that steals time from actual science." Another worried that converting files would introduce errors. The lab's principal investigator, Dr. Elena Vasquez, took a more measured view: "We had been meaning to clean up our data practices for years. The mandate just forced the issue."

The first task was an audit. The lab's shared drive contained over 400 files, spanning a decade of experiments. Roughly half were in Excel, a quarter in SPSS, and the rest a mix of PDFs, plain text, and image files. Many files were undocumented, with column headers that made sense only to the person who created them. The mandate specified that data must be in a non-proprietary, machine-readable format, but it did not prescribe which one.

The lab's initial instinct was to migrate everything to a single format. They considered several options, each with trade-offs. The choice would shape not only their own workflow but also how the two repositories that eventually received the data would handle it. The mandate's language was deliberately vague, leaving room for interpretation, and that ambiguity would prove consequential.

File Formats Become Political

The debate over which format to adopt was not merely technical. For many researchers, Excel files are familiar and convenient, with built-in tools for sorting, filtering, and visualization. Open formats, by contrast, can feel clunky and limited. But the mandate was clear: data must be accessible to anyone, regardless of software. That ruled out Excel, which requires a license to open reliably.

The lab considered several alternatives. CSV, the comma-separated values format, was the simplest and most universally readable. It is plain text, works with any programming language, and has been a mainstay of scientific data exchange for decades. But CSV has limitations: it cannot store multiple sheets, complex data types, or metadata easily. For a lab with rich imaging data, CSV might not capture everything.

HDF5, a hierarchical data format, was another candidate. It can store large, complex datasets with embedded metadata and is widely used in scientific computing. But HDF5 is more complex to create and read, requiring specialized libraries. The lab's postdocs, who were comfortable with Python, could handle it, but they worried about collaborators who used point-and-click tools like Excel. "If we deposit HDF5, half our collaborators won't be able to open it without installing new software," one postdoc noted.

JSON and JSON Lines were also floated. JSON is human-readable and supports nested structures, which could accommodate complex experimental designs. However, JSON is less familiar to many researchers, and some tools still struggle with it. The lab's data manager, Sarah Chen, pointed out that JSON's flexibility could lead to inconsistent schemas, requiring even more documentation. After a week of deliberation, the lab chose CSV, with the caveat that imaging data would be stored as separate TIFF files, with CSV manifests describing the file names and conditions.

Two Repositories Take Divergent Paths

The lab's data was destined for two repositories. The first was the University of Northwood's institutional repository, which had embraced the mandate enthusiastically. Its staff had developed a set of validation checks that automatically flagged files that were not in an approved format, lacked a data dictionary, or had inconsistent column headers. They saw the mandate as an opportunity to raise data quality across the campus.

The second repository was the Neuroscience Data Repository (NDR), a disciplinary repository. Its administrators were more cautious. They had been accepting a wide range of formats for years, including Excel, and they worried that a strict mandate would discourage researchers from depositing at all. "Our mission is to make data available, not to police formats," the repository's coordinator told me. They continued to accept Excel files and other legacy formats, arguing that some data was better than none.

The divergence had practical consequences. The institutional repository required the lab to provide a data dictionary explaining each column, and it ran automated checks for missing values and duplicate rows. The disciplinary repository, by contrast, asked only for a basic description and a list of files. The lab's data, once deposited, looked very different in the two systems. In the institutional repository, the CSV files were accompanied by a README, a data dictionary, and a validation report. In the disciplinary repository, the same data sat in a folder with minimal metadata.

This split created a natural experiment. Six months after the deposits, the lab could compare how the two repositories handled the same data. The institutional repository reported a format error rate of 2% across all deposits, down from 18% before the mandate, and a median of 12 downloads per dataset in the first year, up from 4. The disciplinary repository, despite its lenient approach, had a format error rate of 15% and a median of 5 downloads, up from 3. The mandate's effects were not uniform; they depended on the repository's own policies.

These numbers, while promising, mask significant variation. Some datasets in the institutional repository attracted dozens of downloads, while others languished. The lab's most popular dataset, a study of working memory under sleep deprivation, was downloaded over 40 times in the first six months, likely because it was linked to a high-profile paper. In contrast, a dataset from an early pilot study had only two downloads, both from the lab itself. The disciplinary repository showed a similar pattern, though with lower overall numbers. The lab's analysis suggested that the repository's validation checks were not the only driver of downloads; the topic and visibility of the underlying paper mattered more.

What the Mandate Actually Changed

To assess the mandate's impact, the lab tracked several metrics over the following year. The most striking change was in format errors. Before the mandate, roughly one in five datasets deposited by the lab had some kind of format problem: unreadable files, missing headers, or incompatible encodings. After the mandate, that rate dropped to nearly zero for the institutional repository, thanks to its validation checks.

Downloads per dataset also rose, as noted above. Citations of the underlying papers rose modestly. The lab's bibliometric analysis, which compared papers with and without deposited data, found a small but measurable increase in citation rates, on the order of 5–10%. This is consistent with earlier studies, but the effect was not dramatic. As Dr. Vasquez put it, "Data sharing is not a magic bullet for citations."

The most significant cost was time. The lab estimated that the initial migration took roughly 200 person-hours, spread over three months. That included writing conversion scripts, cleaning data, and documenting everything. The weekly format audits added another hour per week. For a small lab, this was a non-trivial burden, and it did not improve the quality of the research itself. The mandate changed how data were presented, not how experiments were designed or analyzed.

There were also opportunity costs. During the three-month migration, the lab postponed a planned study on decision-making under risk. The delay meant that a postdoc who was leaving at the end of the year could not complete her analysis in time, and the paper was eventually published without the data deposit, which the mandate allowed for papers accepted before the deadline. This incident highlighted the trade-off between compliance and scientific productivity. Some lab members argued that the mandate should have included a grace period for ongoing projects, while others felt that the migration was a one-time cost that would pay off in the long run.

The Craft of Compliance

Once the lab committed to CSV, they had to develop a workflow that made compliance sustainable. The first step was to write a set of Python scripts that could convert Excel files to CSV automatically. The scripts handled common issues like inconsistent date formats and special characters, but they required manual review for files with complex layouts. Sarah Chen spent a week refining the scripts and testing them on historical files. One script, for example, parsed a particularly messy dataset from an old behavioral study, converting a multi-sheet workbook into a single CSV with a 'condition' column that indicated which sheet each row came from. Another script normalized date formats from 'MM/DD/YYYY' to 'YYYY-MM-DD', which the lab adopted as a standard.

The next step was to implement automated checks for headers. Every CSV file had to have a first row with column names, and those names had to match a predefined dictionary. The lab created a simple validation script that flagged missing or duplicated headers, and it ran this script before any deposit. This caught errors early, but it also forced the lab to standardize its column naming conventions, which took time to agree on. For instance, the lab had used both 'RT' and 'ReactionTime' in different files; the new convention standardized on 'reaction_time_ms'. Similarly, they had to decide on a standard for participant IDs, settling on 'participant_id' with a zero-padded number, to avoid sorting issues.

Weekly format audits became part of the lab's routine. Every Friday, a rotating member of the lab ran a script that scanned the shared drive for new files and checked whether they met the mandate's requirements. The audit produced a report that was emailed to the entire lab, with a list of files that needed attention. This was met with some grumbling, but it prevented the backlog that had plagued the initial migration. The audits caught a recurring issue: some postdocs would save intermediate analysis files in Excel, forgetting to convert them before deposit. The audit flagged these, and the lab developed a habit of converting files immediately after analysis, rather than batch-converting at the end of the month.

Training was another component. The lab held two lunchtime sessions on data formatting, covering everything from how to export a clean CSV from Excel to how to write a data dictionary. Attendance was voluntary, but nearly everyone came. The sessions were practical, not ideological, and they helped demystify the mandate. By the end of the first year, the workflow was documented in a lab wiki, with step-by-step instructions that new members could follow.

One unexpected benefit was that the mandate forced the lab to confront the quality of its data. During the migration, they discovered that several datasets had missing values or duplicate entries that had gone unnoticed for years. Cleaning these files improved the internal consistency of the data, even if it did not change the conclusions of the published papers. Sarah Chen noted, "We found errors in our own data that we had been analyzing for years. The mandate made us look at the raw files, not just the summary statistics." This suggests that the policy had a hidden benefit: it encouraged researchers to be more careful about their data from the start.

Unresolved Tensions

The mandate's success in reducing format errors and boosting downloads came at a cost that the lab is still grappling with. The institutional repository's strict validation ensures quality, but it also discourages some labs from depositing. The disciplinary repository's leniency encourages deposits, but it leaves data in a less usable state. Neither approach is clearly right, and the tension between them is unlikely to resolve soon. As Dr. Vasquez mused, "We've made our data more accessible, but I wonder if we've just moved the problem. A CSV file without a data dictionary is almost as opaque as an Excel file without one." The lab is now considering whether to adopt a metadata standard like DataCite's, but that would require yet another round of training and script updates. The question of how to balance openness with usability remains open, and the lab's experience suggests that no single format or policy will solve it.

Moreover, the mandate's focus on formats may be misplaced. While CSV is universally readable, it is not always the most appropriate format for complex data. For instance, the lab's imaging data, stored as TIFF files, are not easily analyzed without specialized software. The CSV manifests provide a map, but they do not capture the full context of the imaging parameters. Some researchers argue that the mandate should have encouraged the use of domain-specific standards, such as BIDS for neuroimaging, rather than a one-size-fits-all format. The lab's choice of CSV was pragmatic, but it may not be the best long-term solution for all types of data.

The lab's experience also raises questions about the scalability of such mandates. The Northwood Lab is relatively small and had the benefit of a dedicated data manager. Larger labs, or those with less technical expertise, might struggle to comply. The mandate's success depended on the lab's willingness to invest time in training and script development, which may not be universal. As one postdoc put it, "We were lucky that Sarah knew Python. Without her, we would have been stuck." This suggests that policies like NIBS's need to be accompanied by infrastructure and support, not just mandates.

Despite these tensions, the lab has largely embraced the mandate. The initial resistance has faded, and the new workflow is now routine. The lab's data is more accessible, and the format errors have all but disappeared. But the deeper questions remain: What is the purpose of data citation? Is it to make data available, or to make it usable? The mandate has achieved the former, but the latter is still a work in progress. As the lab looks to the future, it is exploring ways to make its data more discoverable and interoperable, but it is also aware that each new standard brings new costs. The balance between openness and usability is a moving target, and the lab's experience is a reminder that policy, however well-intentioned, is only as good as its implementation.