Who Gets to See the Raw Trial Data? IPD Sharing Explained

Since July 1, 2018, journals that follow the International Committee of Medical Journal Editors (ICMJE) have required every clinical trial to publish a data-sharing statement declaring whether its individual participant data (IPD) will be shared, with whom, and by what mechanism. That statement discloses intent; it does not guarantee that anyone can obtain the dataset. IPD is the layer beneath the published result: the row-by-row record of every enrolled patient, with baseline characteristics, treatment assignment, doses, outcomes, and dropouts. When outside researchers do get the raw numbers, a reanalysis can either reproduce the trial's conclusion or overturn it.

What counts as the raw data

A published paper reports summary results: means, event counts, hazard ratios, p-values, and the tables and figures built from them. IPD is the layer below that, the de-identified participant-level dataset paired with a data dictionary that explains what each variable means. Serious sharing also includes the supporting documents that make the numbers interpretable, namely the study protocol, the statistical analysis plan, and increasingly the analytic code and the blank consent form. Without the protocol and analysis plan, a spreadsheet of numbers is hard to check, because you cannot tell which analyses were prespecified and which were chosen after the results came in.

What the ICMJE statement requires

In June 2017, the ICMJE finalized a rule with two deadlines. Manuscripts submitted on or after July 1, 2018 must contain a data-sharing statement, and clinical trials that began enrolling on or after January 1, 2019 must register a data-sharing plan when they register the trial itself. The statement has to address a fixed set of questions: whether de-identified IPD will be shared, exactly which data, what supporting documents come with it, when the data becomes available and for how long, and by what access mechanism and criteria, meaning with whom and for what kinds of analysis. The committee is blunt that "'undecided' is not an acceptable answer." Authors must take a position.

One detail is easy to miss. An earlier 2016 ICMJE proposal would have made sharing de-identified IPD a condition of publication. The final 2017 policy stepped back from that. As the committee explained in its editorial in the New England Journal of Medicine, the requirement is transparency about intent, not a mandate to share. A trial can satisfy the rule by clearly declaring that it will not share its data, as long as it says so.

Who is allowed to look

"Sharing" covers a wide range. At one end, a de-identified dataset is deposited openly for anyone to download. Far more common for trials is managed access, where an independent platform holds the data and an outside researcher must submit a proposal, sign a data-use agreement, and pass a review before receiving a gated copy. Repositories such as Vivli and the Yale YODA Project operate this way, acting as neutral intermediaries between sponsors and requesters. This is what the ICMJE means by describing the mechanism and the people who may request access. It is a controlled process, calibrated to protect participant privacy and to record who analyzed what, rather than an open door.

That design has a consequence. The group best positioned to check a trial is often an independent one with no stake in the original conclusion, and managed access is what lets them in without exposing patient identities.

Why the raw data can confirm or undermine a claim

Reanalysis is the reason any of this matters. When outside investigators re-run a trial from its IPD, they sometimes reach a different answer. A 2014 study in JAMA by Ebrahim and colleagues examined 37 published reanalyses of randomized trials and found that 13 of them, about a third, reached conclusions different from the original about which patients should be treated. The dataset did not change. The analytic choices did, and those choices moved the bottom line.

Sharing can also work the other way, by reinforcing a result. A 2018 survey in The BMJ by Naudet and colleagues examined trials in The BMJ and PLOS Medicine, two journals with strong sharing policies. Every trial carried a data-sharing statement, yet the authors could obtain usable data for only 17 of 37, about 46 percent. Where they did get the data, most reanalyses reproduced the original findings. Two lessons sit inside that one study. A statement is not the same as accessible data, and when the data is genuinely available, published conclusions often hold up.

Reading a statement without overreading it

For anyone evaluating a trial, the data-sharing statement is a short paragraph that rewards a careful look. A confident "yes, de-identified IPD will be available through a named repository, with the protocol and analysis plan" signals a team prepared to be checked. A vague or negative statement is not proof that anything is wrong, since legitimate constraints exist, but it does mean the published analysis is the only one anyone can see. And a pledge to share is a starting point, not evidence that the data ever moved. The value of open trial data is realized only when someone requests it, receives it, and looks. This article is educational and not medical advice.