Which clinical trials share their protocols? What 100,000 completed trials show

A trial’s protocol and statistical analysis plan (SAP) say what the trial set out to do and how it planned to analyse the results. They are what you need to judge a trial properly, design a similar one, or check that a publication reports what was planned. ClinicalTrials.gov lets sponsors post them, and for some trials US law requires it.

So how many do? Of the 102,102 interventional trials that finished between 2018 and 2024, only 27% have posted any document: 26% a protocol, 25% a SAP and 7% a consent form. ClinLit AI counted them by sponsor, phase, disease and country, linked them to PubMed, and read the documents themselves. Each section has a Reproduce this with ClinLit AI box with the question we asked, an excerpt of the answer and how it was produced.

Key findings

• Almost half of completed trials have left no public trace. Of the trials completed in 2018–2022, 46% have posted no results and no documents, and have no paper linked to them in PubMed. For phase 3 trials it’s 21%; for phase 1, 57%.
• One rule explains almost all of it. Sharing jumped from 2.5% to 26% of trials in a single year, 2017, when the US rule requiring a protocol and SAP with posted results took effect. Of trials completed since 2018, every one that has posted results has also posted documents. Trials without results have shared in 4.7% of cases.
• Big pharma shares nearly everything it must; small companies don’t. Among trials that look subject to the US rule, the largest drug companies posted documents for 99% and companies with fewer than five completed trials for 65%.
• Industry rarely shares more than it must. Companies posted documents voluntarily for 1.1% of trials without results, against 5.4% for universities and hospitals.
• Consent forms are an academic and government habit, driven by a separate rule for federally funded trials: NIH and federal sponsors include one in about half their document sets, industry in 4%.
• About one in eight industry documents carries a redaction marker, against 1 in 200 from other sponsors.
• The posted SAP rarely shows when the analysis was fixed. More than half of standalone SAPs are dated on or after the trial’s primary completion, and the three we read didn’t say whether they were finalised before the data were unblinded.

Sharing followed the rule

The numbers behind the chart

Primary completion	Trials	With any document	Protocol	SAP	Consent form
2010	7,471	0.3%	0.3%	0.3%	0.1%
2013	9,207	0.7%	0.6%	0.5%	0.2%
2016	11,962	2.5%	2.4%	2.1%	0.7%
2017	12,467	26.4%	26.2%	25.9%	3.1%
2018	13,319	28.2%	27.8%	27.1%	3.5%
2019	14,353	28.1%	27.4%	26.6%	5.5%
2020	12,769	27.9%	26.9%	26.2%	6.2%
2021	13,842	26.9%	25.8%	25.1%	7.7%
2022	15,549	26.3%	25.2%	24.4%	7.9%
2023	16,181	25.2%	23.7%	23.0%	8.7%
2024	16,089	24.5%	22.8%	22.2%	9.3%
2025	13,853	18.2%	14.9%	14.1%	9.0%

The US rule that implements the FDA Amendments Act (42 CFR Part 11) took effect on 18 January 2017. It makes the full protocol and SAP, with all amendments, part of the results a sponsor must submit, normally within a year of the trial’s primary completion. It covers “applicable clinical trials”: broadly, trials of FDA-regulated drugs, biologics and devices with a US connection, but not phase 1 drug trials or small device feasibility studies (ClinicalTrials.gov final rule webinar, eCFR 42 CFR Part 11, subpart C). NIH applies the same requirement to the trials it funds (University of Colorado summary). Sponsors may redact personal information and confidential commercial information.

In practice the documents go further than the rule: all 23,481 trials in our cohort that have posted results also have posted documents, including trials outside the rule’s reach. ClinLit AI suspected that perfect 100% was a quirk of the data, so we checked a random sample of trials with no US site and no FDA-regulated product on the live registry. All had their results and their protocol and SAP posted on ClinicalTrials.gov. Whatever the legal scope, posting results on ClinicalTrials.gov today comes with the documents.

The flip side: a trial that hasn’t posted results has almost never posted its protocol. Only 4.7% of the 78,621 completed trials without results have any document.

Reproduce this with ClinLit AI: the definitions, the overall shares and the trend

We asked:

Let’s measure how often clinical trials post their documents (protocol, statistical analysis plan, informed consent form) on ClinicalTrials.gov. Cohort: interventional trials with overall status COMPLETED and a primary completion date from 1 January 2018 to 31 December 2024. A trial ‘has documents’ if it has at least one row in provided_documents; use the has_protocol, has_sap and has_icf flags for the document types. ‘Has results’ means calculated_values.were_results_reported is true. Give: the cohort size; the share with any document, with a protocol, with a SAP and with a consent form; and the share with documents among trials with results and among trials without results.

Ask ClinLit AI this question →

	Trials	Any document	Protocol	SAP	Consent form
With results	23,481 (23.0%)	23,481 — 100.0%	23,481 — 100.0%	23,017 — 98.0%	5,454 — 23.2%
Without results	78,621 (77.0%)	3,688 — 4.7%	2,584 — 3.3%	2,325 — 3.0%	1,796 — 2.3%

It then flagged the 100% as “almost certainly an artefact of this snapshot”. That was a reasonable suspicion, but wrong: the live registry shows the same thing. Check the reasoning as well as the citations.

Follow-up in the same conversation:

We asked:

Same definitions, but for all completed interventional trials with a primary completion date from 2010 to 2025: by primary completion year, how many trials are there and what share have documents? Show a table and a line chart.

And, separately, what the rule requires:

We asked:

What does the US FDAAA final rule (42 CFR Part 11) require about uploading the protocol and statistical analysis plan to ClinicalTrials.gov: for which trials, from which date, by when, and is the informed consent form required? Cite official sources.

Ask ClinLit AI this question →

the protocol and the statistical analysis plan (SAP) are part of the results information a responsible party must submit — they are not a registration data element and are not submitted “at registration”. The informed consent form is not required by Part 11 at all.

Who shares

Because documents come with results, sharing is highest where the US rule applies and where trials are large.

Trial	Trials	With documents
FDA-regulated device, US site	4,090	74%
FDA-regulated drug, US site	12,579	65%
FDA-regulated drug, no US site	1,929	48%
Not FDA-regulated, US site	14,454	38%
Not FDA-regulated, no US site	55,655	9%

• By phase: phase 2 (57%) and phase 3 (55%) trials share far more than phase 1 (20%) and trials with no phase, mostly device and behavioural studies (19%). Phase 1 drug trials are outside the rule.
• By number of sites: 20% of single-site trials, 36% with 2 to 10 sites, 61% with 11 to 50, and 89% with more than 50.
• By lead sponsor: NIH 83%, other US federal agencies 78%, industry 41%, universities and hospitals 21%, non-US governments 7%.
• By disease (MeSH categories; a trial can count in several): blood and lymphatic diseases (44%) and immune diseases (42%) share the most, then eye diseases (39%), cancer (37%) and infections (36%). Pregnancy complications (20%), injuries (21%) and musculoskeletal diseases (22%) share the least.

Reproduce this with ClinLit AI: sharing by phase, sponsor, sites, FDA status and disease

Follow-ups in the same conversation as the first question:

We asked:

Same cohort (2018-2024). Show the share with documents by: phase; lead sponsor class (studies.source_class); number of facilities (1, 2-10, 11-50, 51+, from calculated_values); and the combination of is_fda_regulated_drug, is_fda_regulated_device and has_us_facility. For each group give the number of trials, the share with documents and the share with results.

We asked:

Same cohort. Show the share with documents by therapeutic area, using these MeSH ancestor terms in browse_conditions (a trial can be in several): Neoplasms, Nervous System Diseases, Cardiovascular Diseases, Mental Disorders, Nutritional and Metabolic Diseases, Respiratory Tract Diseases, Immune System Diseases, Digestive System Diseases, Infections, Skin and Connective Tissue Diseases, Endocrine System Diseases, Hemic and Lymphatic Diseases, Musculoskeletal Diseases, Wounds and Injuries, Eye Diseases, Pregnancy Complications, Urogenital Diseases. Give trials, share with documents and share with results for each.

ClinLit AI noted that about a quarter of trials (24,774) have no MeSH disease terms at all, so the disease breakdown covers the rest, and that “the document share is only 2.3–4.8 points above the results share” in every area: the ranking is mostly a ranking of which areas’ trials post results.

Like for like: the gap is small companies

Raw rates mix very different trials. A company running mostly phase 1 studies in healthy volunteers will look like a poor sharer even if it posts everything it must. So ClinLit AI compared trials that look subject to the US rule: FDA-regulated drug or device, at least one US site, not phase 1 or phase 1/2, completed 2018–2023. That leaves 9,213 trials, and 86% posted documents.

Industry sponsor’s size (completed trials it led, 2018–2024)	Sponsors	Trials	With documents
100 or more	18	1,083	98.6%
20 to 99	85	1,122	94.7%
5 to 19	409	1,434	82.9%
Fewer than 5	1,179	1,604	65.0%

On this comparison, Pfizer, Merck Sharp & Dohme, AstraZeneca, Bristol-Myers Squibb, Roche, GSK, Gilead, Amgen, Bayer and Takeda posted documents for every one of their trials, and Novartis (99%), AbbVie (99%), Janssen (97%), Eli Lilly (94%) and Novo Nordisk (94%) for nearly all. Universities and hospitals (88%) and NIH (99.5%) are also high.

A missing document doesn’t mean a broken rule. Our test is a rough proxy for “applicable clinical trial”, and the rule lets some trials delay their results, for example while a product awaits approval. But the gradient by company size is steep, and it suggests that small sponsors find the reporting work hardest.

Reproduce this with ClinLit AI: the like-for-like comparison

Follow-up in the same conversation as the first question:

We asked:

Now a like-for-like comparison. Take the cohort trials that look subject to the US results-reporting rule: FDA-regulated drug or device, at least one US facility, phase other than phase 1 or early phase 1, primary completion 2018-2023. Give the share with documents overall and by lead sponsor class. Then, for industry sponsors only, group sponsors by how many completed trials they lead in the whole 2018-2024 cohort (100 or more, 20-99, 5-19, under 5) and give the share for each group. Finally list the industry sponsors with at least 40 such trials and their share.

It checked its own joins first (“sponsors has exactly one lead row per trial … so sponsor grouping doesn’t double-count”) and reported a sensitivity analysis without the trials that have no phase: 91.3%, “Same conclusions”.

Voluntary sharing is rare, but growing

Among completed trials without posted results, the share with documents rose from 3.4% (2018) to 5.8% (2024).

• Industry almost never shares voluntarily: 1.1% of its trials without results, against 5.4% for universities and hospitals.
• A few institutions make it a habit. Wake Forest University Health Sciences has posted documents for 76 of its 98 completed trials without results; Geisinger Clinic (44%) and Ohio State’s cancer centre (32%) follow. Many large academic sponsors post for under 10%.
• Countries differ 28-fold, counting trials with a site in the country: Denmark 8.3%, the United States 7.1% and the United Kingdom 6.6% at the top; France 0.8% and South Korea 0.3% at the bottom.

Reproduce this with ClinLit AI: voluntary sharing

Follow-up in the same conversation as the first question:

We asked:

Voluntary sharing: among cohort trials without results, what share have documents, by primary completion year? Which lead sponsors have the most such trials with documents? Which countries (counting trials with a site in the country, only countries with at least 1,500 cohort trials without results) have the highest and lowest share? And what is the share for each lead sponsor class?

The FED 26.9% is not a real signal. It rests on 268 trials, and its documents are overwhelmingly consent-form-only (23.1% ICF vs 6.3% protocol) — chiefly small US federal sponsors such as the Army research institutes, USDA nutrition centres and VA facilities that upload a consent form and nothing else.

Beyond the registry: what PubMed adds

A protocol on ClinicalTrials.gov is one way a trial becomes visible; a paper is another. ClinLit AI linked the 69,832 trials completed in 2018–2022 (giving each about three years to publish) to PubMed, through the trial number that articles record in PubMed and the publications the registry lists for each trial.

• 40% have a paper in PubMed. Trials that posted results are more likely to have one (51%) than those that didn’t (37%). Results, documents and papers tend to go together.
• Journals add little for protocols. 3% of trials published their protocol as a journal article, for example in BMJ Open or Trials. Counting those as well, the protocol is available somewhere for 29% of trials, against 27% from ClinicalTrials.gov alone.
• 46% have no public trace: no results, no documents and no linked paper.

The numbers behind the chart

Trials completed 2018–2022	Trials	With a linked paper	No public trace
NIH-sponsored	594	59%	10%
US federal	632	49%	16%
Phase 3	4,890	62%	21%
Phase 2	7,248	48%	23%
Industry	16,829	38%	41%
Universities and hospitals	49,835	41%	48%
No phase (mostly device and behavioural)	46,005	39%	51%
Non-US governments	1,525	41%	56%
Phase 1	7,868	31%	57%

“No public trace” means none that we can find by the trial number. To test the links, ClinLit AI took 300 of these trials at random and searched every PubMed title and abstract for their trial numbers: 2 turned up, and only one was a report of the trial. The same search found 78 of 100 trials that are linked, so it works. Missed links move the 46% by about half a point. What no search by trial number can find are papers that never cite it, papers published after late 2025, and papers outside PubMed, so this is a measure of traceability rather than proof that the trials went unpublished. Even so, a trial whose results, protocol and publications can’t be found from its registration number is, for most readers, a trial that left no record.

Reproduce this with ClinLit AI: linking trials to PubMed

We asked:

Let’s link completed trials to PubMed. Cohort: interventional trials with overall status COMPLETED and a primary completion date from 1 January 2018 to 31 December 2022. A trial ‘has a linked paper’ if its NCT ID appears in pubmed.article_databank_accessions (databank_name ‘ClinicalTrials.gov’) or it has an aact.study_references row of type RESULT or DERIVED with a PMID. A trial ‘has a protocol paper’ if one of those linked PMIDs has the publication type ‘Clinical Trial Protocol’ in pubmed.publication_types. ‘Has documents’ means at least one row in aact.provided_documents; ‘has results’ means calculated_values.were_results_reported. Give: the share with a linked paper, overall and split by results and documents; the share with a protocol paper; the share with a protocol available anywhere (documents on ClinicalTrials.gov or a protocol paper); and the share with no public trace (no results, no documents and no linked paper), overall, by phase and by lead sponsor class (studies.source_class).

Ask ClinLit AI this question →

It noted what the links can’t see: “Some linked PMIDs are outside the pubmed schema (which covers 2010+ articles with abstracts): 8,090 of 54,853 distinct linked PMIDs, affecting 4,504 trials. Their publication types are invisible here, so the 3.0% protocol-paper share is also a floor.”

Follow-up in the same conversation:

We asked:

How complete is that linkage? Take a random sample of 300 trials from the no-public-trace group, and for each one check whether its NCT ID appears anywhere in a PubMed title or abstract, using full-text search on pubmed.articles.search (NCT IDs are indexed as lower-case tokens such as nct03580369). Search the IDs in batches with constant queries rather than one join, so each query stays fast. How many of the 300 turn up, and what does that say about the no-trace figure?

The method was validated before I believed it: the same search on 100 known-linked trials turned up 78 … The linkage incompleteness I flagged last time is real but worth about 0.5 percentage points, not the 10–20 points that would change the story.

This was a long answer: it reached ClinLit AI’s per-answer limit, and the final summary was written from the work done by then.

Consent forms: a different rule, a different habit

The US results rule doesn’t ask for consent forms. A separate one does: since 2019, the revised Common Rule requires federally funded clinical trials to post one IRB-approved consent form, no later than 60 days after the last study visit (NIH guidance). It shows in the data. Consent-form sharing tripled from 3.1% of trials completed in 2017 to 9.3% in 2024, and among trials with documents, NIH (49%), federal (48%) and academic (39%) sponsors include a consent form far more often than industry (3.8%).

That matters for anyone writing a consent form or a patient-facing summary: posted examples exist for many academic and NIH-funded trials, but hardly any for industry trials.

Inside the documents

Length. Industry protocols run to a median of 88 pages, against 20 for others; industry SAPs to 37 pages, against 5 (many non-industry “SAPs” are short statistical sections). About 3% of documents have no readable text: scanned images without a text layer, which no search tool can read.

Redaction. The rule lets sponsors redact confidential commercial information, and many mark the gaps with “CCI”. ClinLit AI found the marker in 13% of industry documents and 0.5% of others, and in 5% of industry documents it appears on more than a fifth of the pages. Protocols and SAPs are redacted; consent forms almost never are. Some companies redact silently, without a marker, so these figures are a minimum.

When was the SAP written? 54% of standalone SAPs carry a date on or after the trial’s primary completion date (55% for industry). That isn’t necessarily a problem: sponsors post the latest version, and an SAP can be finalised after the last patient visit but before the data are unblinded. But the posted document rarely lets you check. ClinLit AI read three industry phase 3 SAPs dated after completion. Each showed a version number and a date, and none said when it was finalised relative to database lock or unblinding. ICH E6(R3) expects deviations from the planned statistical analysis to be documented and justified (Annex 1 §3.16.2(e)), which is easier to judge when the SAP records its own timing.

The dates are also widely spread. For industry, the middle half of SAPs are dated between 109 days before and 60 days after primary completion; for other sponsors, between 341 days before and 227 days after.

Reproduce this with ClinLit AI: length, readability, redaction and SAP timing

Follow-up in the same conversation as the first question:

We asked:

What do the documents look like? Among cohort trials with documents, what share include a consent form, by lead sponsor class? Using aact.study_documents and aact.study_document_pages, give the median and 90th percentile page count by document type (protocol only, protocol plus SAP, SAP only, consent form) for industry and non-industry sponsors, and the share of documents with no readable text (no pages, or under 200 characters per page on average). For documents that are SAPs only, what share have a document_date on or after the trial’s primary completion date?

We asked:

How common is redaction in the protocols, SAPs and consent forms posted on ClinicalTrials.gov? Using aact.study_documents and aact.study_document_pages, count documents whose text contains the redaction marker ‘CCI’ (used by sponsors for confidential commercial information) as a whole word, or the word ‘redacted’, separately for industry and non-industry lead sponsors, and count documents where more than 20% of pages contain ‘CCI’. Check a sample of matches to make sure ‘CCI’ is a redaction marker and not an abbreviation such as the Charlson Comorbidity Index, and correct the counts if needed. Which industry sponsors’ documents are most often heavily redacted?

Ask ClinLit AI this question →

It checked what “CCI” meant before counting it, and corrected for false matches:

False positives found and quantified: CCI-779 = temsirolimus (drug code) … Charlson Comorbidity Index (CCI) … I corrected by dropping pages on which CCI occurs only inside parentheses … or only as CCI-####. Effect: industry 3,665 → 3,579 (−2.3%); non-industry 347 → 215 (−38%).

We asked:

For NCT03580369, NCT03557307 and NCT03579459, read the statistical analysis plan posted on ClinicalTrials.gov: what version and date is it, and does it say when it was finalised relative to database lock or unblinding? Cite the pages.

Ask ClinLit AI this question →

None of the three states that the SAP was finalised/signed before database lock or before unblinding — the timing statements present are of a different kind (blinding maintained until the final DBL in NCT03580369; the three-lock schedule in NCT03557307; analysis-population classification before unblinding/database release in NCT03579459).

What you can do with the documents

Posted protocols and SAPs answer questions the registry can’t. For example, we asked ClinLit AI how phase 3 trials of biologics in atopic dermatitis handle rescue medication in their primary analysis. It found eight trials with posted SAPs, read six of them, and laid out the approaches with page references: the lebrikizumab trials count any rescue during induction as treatment failure and handle topical rescue differently in the maintenance phase, while the upadacitinib-versus-dupilumab trial excludes all assessments after rescue and counts those patients as non-responders. It also reported the two it couldn’t use: one with a damaged text layer, and one almost entirely redacted.

That kind of comparison used to mean downloading and reading each PDF. For more examples, see the protocol section of our GLP-1 report, which traces how flagship trials changed their primary endpoints through protocol amendments.

Reproduce this with ClinLit AI: comparing SAPs across trials

We asked:

Which phase 3 trials of biologics in atopic dermatitis have posted their statistical analysis plan, and how do their SAPs handle rescue medication in the primary efficacy analysis? Cite the documents and pages.

Ask ClinLit AI this question →

The four Lilly placebo-controlled/long-term SAPs (ADvocate1/2, ADhere, ADjoin) share one template: a hybrid estimand with rescue initiation as the intercurrent event, analysed by MCMC multiple imputation with tipping-point sensitivity.

How we counted

• Trials: interventional trials on ClinicalTrials.gov with status “completed” and a primary completion date from 2018 to 2024, from the AACT snapshot of 1 October 2026, which leaves out observational studies and phase 4 trials. The trend uses completions from 2010 to 2025.
• Documents: a trial has documents if ClinicalTrials.gov lists at least one protocol, SAP or consent form for it, of any date. Combined “protocol and SAP” files count as both.
• Results: whether the trial has posted results on ClinicalTrials.gov.
• Like-for-like: FDA-regulated drug or device, at least one US site, not phase 1 or phase 1/2, primary completion 2018–2023. This approximates the rule’s “applicable clinical trial” but isn’t the legal test.
• Publications: trials completed 2018–2022, linked to PubMed through the trial number recorded in PubMed and the publications the registry lists. PubMed covers articles with an abstract from 2010, up to late 2025. A protocol paper is a linked article of the type “Clinical Trial Protocol”.
• Document contents: from the text extracted from the 72,734 documents in the snapshot.
• Cross-check: we also counted with our own SQL against the same data. The cohort, overall shares and breakdowns matched ClinLit AI’s to the trial, apart from small differences in the disease breakdown, where it used only the MeSH ancestor terms. The publication counts matched too; our own sample check of 300 other trials found 8 trial numbers in PubMed abstracts, within ClinLit AI’s estimate. The live-registry check of the 100% link is described above.