Why Clinical Trials Fail: An Analysis of 50,541 Terminated and Withdrawn Studies

When a clinical trial collapses, the public explanation is usually scientific: the drug didn't work, or it wasn't safe. We tested that assumption against the registry itself. Every sponsor who terminates or withdraws a study on ClinicalTrials.gov must state why, in their own words. We retrieved all 50,541 of those studies and classified the 44,908 stated reasons. The result contradicts the common story. Science is one of the rarest causes of death for a clinical trial. The most common cause, by a factor of four, is that not enough people walked through the door.

What kills a clinical trial

ClinicalTrials.gov records 594,309 registered studies. Of those, 33,973 are marked TERMINATED — stopped after enrolling at least one participant — and 16,568 are marked WITHDRAWN, meaning the study closed before enrolling anyone at all. Together they represent 50,541 abandoned research efforts, each one a question that was asked and never answered.

Sponsors supply a free-text whyStopped explanation. 44,908 studies (88.9%) provided one. We classified each into twelve categories using keyword rules. The distribution:

Percentages are of studies that stated a reason (n=44,908 all; 30,452 terminated; 14,456 withdrawn), not of all 50,541.

The headline number is the first row. Recruitment failure accounts for 36.7% of all trial deaths — 40.9% among terminated studies. The next largest cause is a distant 9.4%. If you add together every reason that relates to the treatment itself — safety at 4.9%, futility at 3.6% — you reach 8.5%. Trials are roughly four times more likely to die because patients never arrived than because the science failed.

Trials die from money before they open, and from patients after

Splitting terminated from withdrawn studies produces the sharpest result in the dataset, and it is one we did not anticipate. The two groups fail in fundamentally different ways.

Among withdrawn studies — those that closed before enrolling a single participant — funding is the cause in 15.5% of cases, more than double its weight among terminated trials (6.6%). Meanwhile safety collapses to 1.5%, and futility all but vanishes at 0.4%.

Among terminated studies — those that enrolled someone and then stopped — safety rises to 6.6% and futility to 5.2%.

The logic is almost tautological once stated, but the magnitude is not obvious in advance: before a single patient is enrolled, there is no data that can kill your trial. Only money, staffing and logistics can. A withdrawn trial with a futility reason is nearly a contradiction in terms, and the data confirms it — 0.4%, effectively noise. Safety and efficacy only become lethal once human beings are in the study.

This has a practical consequence for how trial failures are read. A withdrawn study tells you almost nothing about the treatment. A terminated study tells you slightly more — but even then, 87% of terminations had nothing to do with safety or efficacy.

The academic recruitment gap

Splitting by sponsor type produces the finding with the largest practical implications.

Academic sponsors fail on recruitment at nearly twice the rate of industry (41.8% vs 24.3%), and run out of money almost five times as often (12.5% vs 2.6%). At the level of individual institutions the gap is starker still:

M.D. Anderson — the largest cancer centre in the United States — loses half of its failed trials to recruitment. Pfizer loses 15.5%. Mayo Clinic sits at 49.0%, Duke at 47.8%, Massachusetts General at 46.8%. These are among the most prestigious research institutions in the world, with the deepest patient populations and the most experienced investigators available anywhere.

They still cannot fill their studies at anything close to the rate that a pharmaceutical company can.

The dataset does not explain the gap, and we will not pretend it does. But it constrains the plausible explanations. It is not scientific quality — these institutions produce the field's best work. The more likely candidates are structural: industry trials come with dedicated recruitment budgets, professional coordinators, contract research organisations, multi-site networks and advertising spend. Academic trials frequently rely on a principal investigator referring their own patients, alongside a clinical workload. Industry also has an option academia largely lacks — closing a trial for strategic reasons and calling it a business decision (7.5% of industry-heavy terminations) before recruitment failure is ever recorded.

One further data point sharpens this. Trial networks — consortia built specifically to pool patients across institutions — have the worst recruitment failure rate in the entire dataset at 54.4%. The structures designed to solve the recruitment problem fail at it more than anyone.

Most terminated trials barely started

Among the 30,353 terminated studies reporting an enrollment figure, the median number of participants at termination was 20. Nearly a third — 31.6%, or 9,580 studies — stopped with fewer than ten participants enrolled.

This reframes what "terminated" means. These are not studies that ran their course and fell short at the finish line. The typical terminated trial recruited a couple of dozen people and then stopped — too few to answer anything, but enough for those participants to have undergone screening, consent, procedures and follow-up. Roughly ten thousand studies exposed fewer than ten people each to the burden of research participation and produced no usable answer in return.

Phase 4 is where recruitment goes to die

The expected pattern would be that early-phase trials — asking healthy volunteers or heavily pre-treated patients to take an unproven compound — struggle most to recruit. The data shows the opposite. Phase 1 has the best recruitment record of any phase at 26.9%. Phase 4 has the worst at 46.4%.

Phase 4 studies take place after a drug is approved and on the market. There is no access argument left to make: a patient who wants the drug can be prescribed it. The trial offers no earlier access, no unavailable treatment, and no experimental hope — only extra visits, extra paperwork and extra monitoring for a medication already obtainable at a pharmacy. Nearly half of failed Phase 4 studies died for want of participants, which suggests post-marketing surveillance is structurally the hardest research to staff in medicine.

COVID-19 hit the 2019 cohort hardest

Attributing failures to COVID-19 by the year the trial started produces a clean curve. Trials that began in 2019 were the most likely to be killed by the pandemic — 11.9% of their failures name COVID — compared with 9.6% for 2020 starts, 4.7% for 2021 and 1.4% by 2023.

The 2019 cohort was mid-flight when the world closed: enrolled, running, dependent on in-person visits, and unable to pause. Trials that started in 2020 and later were designed in the presence of the disruption and could plan around it. The pandemic's cost fell hardest not on the research that was being planned, but on the research that was already underway.

Methodology

We retrieved every study on ClinicalTrials.gov with an overall status of TERMINATED (33,973) or WITHDRAWN (16,568) via the public ClinicalTrials.gov API v2 on 16 July 2026 — 50,541 studies in total, retrieved in full rather than sampled.

Of those, 44,908 (88.9%) contained a non-empty whyStopped field. Each free-text reason was matched against ordered keyword patterns and assigned to the first matching category of twelve. Because the rules are ordered, a reason mentioning both funding and recruitment is counted once, under recruitment.

Limitations, stated plainly:

• 24.6% of stated reasons could not be classified. The residue is long-tail free text — "Production halt of FUDR in China", "Standard of care changed" — much of it genuinely one-off. If those reasons are not randomly distributed across categories, every percentage here shifts.
• These are stated reasons, not audited ones. A sponsor who cancels a programme for commercial reasons may write "slow recruitment"; one whose drug underperformed may write "business decision". The dataset records what sponsors chose to say. The industry–academia gap in particular should be read with this in mind.
• 5,633 studies (11.1%) stated no reason at all and are excluded from all percentages.
• Sponsor-level rates are denominated on failed trials only. A 50.0% recruitment-failure rate at M.D. Anderson means half of its failed trials died that way — not half of all its trials. This analysis cannot compute an institution's overall failure rate, and does not claim to.

Category counts and the classification rules underlying every figure above are reproducible against the live API by any reader who wishes to check them.

What this means for patients

Three things follow from the data.

A trial that stopped is usually not a warning about the treatment. Only 8.5% of failures were attributed to safety or futility combined. If a study you were following was terminated, the overwhelmingly likely explanation is that it could not fill its slots or fund its staff.

Enrolling is more consequential than it appears. The single largest cause of research failure in medicine is that people do not join. This is not a rhetorical point — it is the top line of a 44,908-study dataset. The median terminated trial needed roughly twenty more participants than it found.

Where a trial is run affects whether it finishes. An academic trial is more likely to close for reasons unrelated to the science than an industry-sponsored one. That is not an argument against academic research — it produces work industry will not fund — but it is a real difference in the odds that a study reaches an answer, and it is visible in the registry for anyone who cares to look.