Why Clinical Trials Stop Early: Census and Risk Matrix

Clinical trials stop early because the scientific question has been answered, continuing is unlikely to answer it, participant protection requires a stop, or the study becomes operationally or financially unviable. These explanations require different decisions. Termination is not synonymous with drug efficacy failure, and a temporary suspension is different from permanent termination or withdrawal before enrollment. (^[1]) (^[2]) (^[3])

This report adds a deliberately small ClinicalTrials.gov stopped-record census, reported as observed on October 17, 2026: all interventional records first posted during January 1 through 7, 2024, whose current status was terminated, withdrawn or suspended. The complete eligible export contains 23 unique records: 11 terminated, 8 withdrawn and 4 suspended. Single-pass analytical coding assigns 7 records to business/funding, 7 to logistics/design, 4 to recruitment, 4 to unknown and 1 to mixed reasons. Every record contains reason text, but text presence does not guarantee an informative explanation. These are counts within a selected registration-week cohort, not population termination rates. (^[4])

Published studies reinforce the need to preserve denominators. The NIH-led results-database analysis found insufficient accrual in 350 of 905 terminated trials, or 39%, within its restricted results-posting population. The frequently repeated larger accrual percentage refers to a narrower subgroup. A later Nature Genetics analysis combined terminated, withdrawn and suspended records and allowed overlapping reasons, so its proportions answer a different question. Neither study supplies a universal forecast for an individual development program. (^[5]) (^[6]) (^[7])

For development heads, feasibility teams and contract research organizations, the useful output is a reason-specific decision gate. Recruitment signals should trigger an eligibility and site-capacity reassessment; funding signals require runway and commitment checks; design or supply signals require executable remedies. Safety, futility and early efficacy stopping belong within prospectively planned scientific oversight, rather than being treated as interchangeable operational outcomes. (^[8]) (^[9]) (^[10]) The identifiers and coding rules appear below, but no dated frozen export or complete record-level coding appendix accompanies this article. The new coding has no human adjudication or inter-rater reliability estimate. Treat it as an illustrative analytical exercise for hypothesis formation and planning questions, with limits on historical reproducibility and causal inference.

A trial can end before its planned finish without demonstrating that its intervention lacks efficacy. Funding can change, eligible participants can be difficult to find, the treatment context can change, or interim evidence can justify a scientifically planned stop. Reporting guidance therefore asks investigators to explain why a trial ended and who made the decision, rather than leaving readers to infer cause from its status label. (^[11]) (^[12])

This distinction matters when a sponsor is deciding whether to revise recruitment assumptions, add sites, change supply commitments or release another funding tranche. A budget reserve may address a financing interruption; it cannot resolve a statistical futility boundary. Adding sites may improve access, but only if those sites can identify eligible participants and deliver the protocol. Evidence-based feasibility and site selection are central recommendations of the Clinical Trials Transformation Initiative, or CTTI. (^[13])

The IntuitionLabs perspective in this report is advisory: connect traceable registry evidence with practical clinical-development decisions. Its central contribution is a bounded census with inspectable identifiers and conservative reason codes, alongside published research that provides broader context.

The answer to “why clinical trials stop early” is therefore a taxonomy, followed by a denominator question. Which kinds of studies were eligible? Were they observed long enough to stop? Did the analysis include withdrawals before enrollment? Did a missing explanation become an assumed scientific failure? A results-posting sample cannot automatically stand in for all registered trials; the original NIH-led analysis explicitly acknowledged that limitation. (^[7])

The report separates recorded explanation, analyst classification and verified cause. Only the first is submitted registry information. The second is a reproducible interpretation. The third generally requires additional evidence about the decision process. Published classification research likewise cautions that a submitted reason may describe only part of the explanation. (^[14])

ClinicalTrials.gov distinguishes the following statuses. These definitions determine which records can legitimately be combined:

• Terminated: the study stopped early and will not restart. (^[1])
• Withdrawn: the study stopped before enrolling its first participant. (^[15])
• Suspended: the study stopped early but may restart. (^[2])
• Completed: the study ended normally, with participants no longer being examined or treated under the study. (^[16])

“Stopped early” is a useful umbrella for this report's selected statuses, but permanent discontinuation is inappropriate for a cohort that includes suspension. Registry labels also require source-specific interpretation. The World Health Organization, or WHO, describes suspended recruitment as temporary; its registration dataset's “complete” recruitment label concerns cessation of recruitment and enrollment. Analysts should read definitions before mapping labels between registries. (^[17]) (^[18])

Scientific stopping also needs separation. Futility means that the trial is unlikely to demonstrate efficacy under the relevant design and assumptions. It is not a universal proof that an intervention has no biological activity. The US Food and Drug Administration, or FDA, distinguishes nonbinding futility guidance from binding rules, so the protocol's actual decision framework matters. (^[10]) (^[19])

Early efficacy stopping can mean a prespecified benefit criterion has been met. Interpretation still requires the stopping design: the FDA notes that early efficacy termination reduces the sample available and that conventional estimates in group-sequential settings can be biased toward larger effects. A positive stopping decision is therefore not a reason to ignore precision, secondary outcomes or the remaining safety information needed. (^[20]) (^[21])

Safety stopping concerns participant protection and the emerging risk-benefit assessment. It should not be inferred solely from the word “paused” or an unexplained status change. Independent data monitoring committees can recommend continuation, modification or termination; their remit is different from an operational team's response to slow recruitment. (^[22])

Finally, registry information is submitted by a responsible party. National Library of Medicine, or NLM, review does not verify the scientific validity or relevance of every submission. A reason field is evidence of what was reported, not an independently established causal finding. This report uses that distinction throughout its coding and recommendations. (^[23]) (^[24])

The new analysis is a registration-week census of currently stopped records, not a census of every clinical trial. The window was selected for a compact, inspectable analysis; it is not a random or representative sample. First posting is the date a record became available in the registry, rather than the trial's start date. Consequently, studies in the same registration window can have very different operational histories. (^[25])

The inclusion and handling rules are:

• Population: interventional studies, using the returned structured study-type field.
• Registration window: use the inclusive first-posted dates encoded in the reproducible query below.
• Eligible current statuses: TERMINATED, WITHDRAWN or SUSPENDED.
• Unit: one unique National Clinical Trial identifier, or NCT ID, rather than a study version or publication.
• Snapshot: use current status at retrieval; this is not a reconstructed historical database.
• Extraction: all eligible rows returned by the projected application programming interface, or API, query; pagination is checked.
• Denominator: all eligible stopped records in that query, including uninformative reasons.
• Coding: explicit reason text takes priority; missing causal detail remains unknown.

The complete export returned 23 rows and 23 unique identifiers, agreeing with the API total count, without a next-page token. All rows met the structured study-type and registration-date filters. The latest posted updates in these records ranged from January 3, 2024, through September 9, 2026. No dated frozen export or coding metadata is linked here. The query below provides a live refresh route, rather than reader access to the original October 17, 2026 extract. Reproducing the reported historical classifications requires that extract, the original reason text and the record-level codes and ambiguity flags. (^[4])

The exact query expression is:

AREA[StudyType]INTERVENTIONAL AND AREA[StudyFirstPostDate]RANGE[2024-01-01,2024-01-07]

Use the API's filter.overallStatus parameter with TERMINATED,WITHDRAWN,SUSPENDED; request StudyType,NCTId,OverallStatus,WhyStopped,Phase,Condition,LeadSponsorClass,StudyFirstPostDate,LastUpdatePostDate; and set pageSize=1000, countTotal=true and format=json. Follow any nextPageToken in larger reproductions. These are the parameters of the fetched census, rather than an assertion that every registry field has been exhaustively documented here. (^[4]) NLM describes the modernized service as a representational state transfer, or REST, API using OpenAPI and JavaScript Object Notation, or JSON, responses. (^[26]) (^[27])

Enrollment requires historical care. Current enrollment can be estimated or actual; it does not necessarily preserve the original target. WHO separately defines planned enrollment, while the earlier NIH-led study reconstructed targets from archived registrations. No historical target-attainment ratios are calculated here. (^[28]) (^[29])

The new coding is an automated analytical pass with a separate arithmetic recomputation from the returned rows. It has no human double-coding, independent causal adjudication or measured inter-rater agreement. Published precedents show what a stronger validation layer could look like: the NIH-led analysis reported Cohen's kappa of 0.86, and the Nature Genetics study reported differing agreement across annotator pairs. Those results validate neither this codebook nor these new assignments. (^[30]) (^[31])

The reason taxonomy is intentionally conservative. Recruitment, business/funding and logistics/design are separated because they imply different remedies. Safety, futility and efficacy require explicit wording; they are not inferred from status alone. Mixed reasons preserve a visible combination instead of forcing an arbitrary primary cause. Published large-scale classification work allowed overlapping categories, whereas this census assigns mutually exclusive codes and retains a mixed category. (^[6])

Table 1 combines the analytical codebook with recommended decision questions. The definitions are authored for this census; the planning questions apply evidence-based feasibility and prospectively planned stopping principles. They are decision prompts, not estimated probabilities or clinical recommendations. (^[8]) (^[32])

Code	Assignment rule	Decision question before committing resources
Recruitment	Explicit insufficient recruitment, eligible participants or accrual; no explicit second causal category.	Can sites document the eligible pool, screening conversion and achievable pace?
Business/funding	Explicit funding, funder or business decision. An unspecified sponsor decision alone remains unknown.	Is runway committed through the next interpretable milestone?
Logistics/design	Explicit supply, staffing, site operation, protocol/device design or changed treatment context.	Is a practical remedy available, costed and compatible with the study question?
Safety	Explicit safety rationale only.	What does the monitoring framework require before continuation or modification?
Futility	Explicit futility or inadequate prospect of demonstrating efficacy.	Does the prespecified analysis support continuation, and under what rule?
Early efficacy	Explicit benefit or success as the reason for an early stop.	Were the boundary, estimation method and information needs addressed?
Mixed	Explicit reasons span substantive categories.	Which dependencies must be resolved together?
Unknown	Missing, nonspecific or insufficient text.	What documentation is needed before any causal or resource conclusion?

The matrix discourages a common planning error: applying a recruitment intervention to every stopped study. A supply interruption calls for supply planning; an underfunded program calls for financing decisions; a scientifically planned stop calls for the relevant monitoring process. CTTI's recommendations link recruitment milestones to activation, screening and enrollment trajectories, rather than treating site count as a sufficient capacity measure. (^[33])

The reported census identifiers, grouped by assigned code, are listed below. These lists expose the assignments, but do not provide the full projected fields, verbatim reason text or record-level ambiguity flags needed to reproduce the historical coding. The following citation is the live census query; no separate metadata or frozen-export link is supplied. (^[4])

• Business/funding: NCT06189131, NCT06188637, NCT06191900, NCT06191796, NCT06191289, NCT06187701, NCT06190912.
• Recruitment: NCT06193174, NCT06192329, NCT06190028, NCT06192966.
• Logistics/design: NCT06189313, NCT06191250, NCT06187376, NCT06188871, NCT06191692, NCT06188104, NCT06186505.
• Unknown: NCT06188507, NCT06189560, NCT06188208, NCT06189482.
• Mixed: NCT06190249.

Ambiguity is part of the result. The mixed record combines support withdrawal with the time required to meet accrual. Assigning it to recruitment alone would increase that category from 4 to 5; assigning it to funding alone would increase that category from 7 to 8. A stricter treatment of the terse funder-decision record would instead reduce business/funding to 6 and increase unknown to 5. These are alternative coding rules, not changes in the underlying studies. (^[4])

A separate recruitment explanation mentions interim performance, but the explicit stopping rationale remains recruitment. It is not coded as early efficacy. Other vague text remains unknown. Structured status and narrative timing can also disagree, and a record can retain an interventional field while describing a replacement observational protocol. The analysis uses the structured inclusion fields. These discrepancies are discussed here, but a record-level flag file is not supplied. Published classification research similarly cautions that a brief submitted explanation can omit other contributing factors. (^[14])

Table 2 reports counts by phase, sponsor class and a simple condition grouping. The numeric cells are reported for the same October 17, 2026 observation; the underlying dated export is not supplied here. The denominator is the segment's eligible stopped records, not all trials in that segment. The code columns are business/funding, recruitment, logistics/design, unknown and mixed, respectively. (^[4])

Segment	Stopped records	Business/funding	Recruitment	Logistics/design	Unknown	Mixed
All eligible records	23	7	4	7	4	1
Phase not applicable	10	3	3	2	2	0
Phase 1	6	2	1	0	2	1
Phase 2	3	1	0	2	0	0
Phase 3	1	0	0	1	0	0
Phase 4	2	1	0	1	0	0
Combined phases 2 and 3	1	0	0	1	0	0
INDUSTRY lead sponsor	7	2	2	1	2	0
OTHER lead sponsor	15	5	2	5	2	1
OTHER_GOV lead sponsor	1	0	0	1	0	0
Oncology condition grouping	7	2	1	1	2	1
Other conditions	16	5	3	6	2	0

Each block partitions the cohort independently. Adding phase rows to sponsor rows would count studies repeatedly. Combined phases are retained as their own group, and phase-not-applicable records remain visible rather than being forced into drug-development phases. Lead-sponsor class is the registry field, not a reconstruction of every collaborator or funding relationship.

Interpretation should follow several constraints:

• Phase: small segment denominators do not support a ranking of scientific or operational risk.
• Sponsor: INDUSTRY, OTHER and OTHER_GOV describe the recorded lead sponsor; OTHER is not automatically equivalent to an academically funded study.
• Indication: oncology means that a listed condition explicitly denotes cancer, malignancy, tumor, carcinoma, glioma or melanoma. This is an analyst rule, not a validated therapeutic ontology.
• Precision: these are exact counts of the selected snapshot, but code interpretation and representativeness remain uncertain.
• Censoring: studies still recruiting or later changing status are outside this stopped-only extract.

There are no blank reason fields, but 4 of 23 explanations are insufficiently specific for a substantive cause. Missing text and uninformative text therefore need separate reporting. No record meets the explicit safety, futility or early-efficacy coding rule. The appropriate conclusion is zero explicitly classified cases in this cohort, with no inference about their prevalence across clinical development. (^[4])

Two sensitivities illustrate dependence on the population definition. Keeping only INDUSTRY lead sponsors leaves 7 records: 2 business/funding, 2 recruitment, 1 logistics/design and 2 unknown. Excluding recorded start dates during 2020 through 2022 leaves 21 records: 7 business/funding, 4 recruitment, 7 logistics/design, 2 unknown and 1 mixed. The latter removes records by date, not verified pandemic exposure; start-date actual/estimated flags were not preserved in the supplemental projection. No listed condition contains COVID. (^[34]) (^[4])

Sponsor comparisons in broader research can differ because funding definitions and eligibility differ. A rare-disease study compared industry-funded and academically funded cohorts, while a separate ethics-protocol analysis distinguished randomized and nonrandomized designs. Neither relationship can be transferred mechanically to these registry lead-sponsor classes or small cells. (^[35]) (^[36])

There is no defensible universal clinical trial termination rate without a population, status definition and follow-up rule. A share among stopped trials measures reason composition. A proportion discontinued among all eligible trials measures a different outcome. A time-to-event estimate additionally handles when stopping occurs and how ongoing trials are censored.

Table 3 contrasts original research populations. These estimates are not pooled, averaged or used as priors for the new census.

Original analysis	Population and denominator	Reported finding and interpretation
NIH-led results-database study	7,646 results-posting trials, observed February 19, 2013	905 terminated, about 12%; accrual 350/905, or 39% of terminated trials. Restricted results population. ([37]) ([5])
Phase III analysis	9,828 registered interventional phase III trials, registration window January 2013 through August 2017	320 prematurely concluded after recruitment began; 70% stopped for reasons outside interim safety, efficacy or futility. ([38]) ([39]) ([40])
Nature Genetics analysis	28,561 stopped studies in a November 27, 2021 registry snapshot	Reported insufficient enrollment: 36.67%. Includes withdrawn, suspended and terminated studies with overlapping reasons. ([41]) ([42])
Rare-disease randomized trials	659 eligible studies registered during 2010 through 2012, completed or discontinued by December 31, 2014	199 discontinued, or 30.2%; main-results insufficient accrual 62/199. Multiple stopped statuses included. ([43]) ([44]) ([45])
Ethics-protocol cohort	1,017 randomized controlled trials, or RCTs, from ethics-approved protocols	253 discontinued; recruitment-related discontinuation 101/1,017, or 9.9% of the full cohort. ([46]) ([47])
Childhood-obesity registry analysis	1,312 interventional trials	57 stopped, comprising 5 suspended, 34 terminated and 18 withdrawn; 11/57 lacked reasons. ([48]) ([49])

The rare-disease row uses the paper's main-results accrual count; its abstract gives a different count. That discrepancy is retained as a source limitation rather than silently reconciled. (^[45])

The table shows why similar-sounding statistics can differ without contradiction. The NIH-led analysis's 57% accrual figure applies within the subgroup stopped for reasons other than accumulated trial data; its full terminated-trial figure is 39%. Preserving that denominator is more useful than repeating a context-free recruitment-failure percentage. (^[5])

Specialty findings require the same restraint. A cardiovascular analysis reported 141 missing reasons and separately reviewed a sample of 85 terminated studies, finding poor accrual in 47%. That manual-sample percentage is neither a registry-wide termination rate nor an estimate for every therapeutic area. (^[50]) (^[51]) A restricted breast-cancer treatment-optimization study reported 45 prematurely discontinued studies among 240, with recruitment accounting for 18 of those discontinuations. Its treatment-optimization eligibility defines its applicability. (^[52]) (^[53])

A registry comparison of 163 discontinued and 163 completed studies is especially easy to misread. Because the investigators deliberately selected equal groups, the design cannot establish a population discontinuation prevalence. Its reason analysis, including recruitment and funding, can still help refine a codebook or formulate feasibility questions. (^[54]) (^[55])

For a sponsor-facing evidence pack, the minimum useful items are:

• Population: registration dates, study type, disease rules and geography where applicable.
• Outcome: exactly which statuses are counted and whether a suspension can reverse.
• Time: snapshot date, follow-up opportunity and handling of ongoing studies.
• Reason quality: missing, uninformative, mixed and adjudication-disagreement categories.
• Provenance: identifiers, source fields, coding version and a preserved extract.
• Comparability: explicit limits before transferring a percentage between populations.

These requirements turn a headline percentage into evidence a feasibility team can interrogate. Reporting standards also ask for the reason a trial ended and its stopping framework, which helps connect registry summaries with trial reports. (^[11]) (^[32])

The operational response should begin with the mechanism most plausibly limiting execution. CTTI recommends evidence-based feasibility, reducing unnecessary procedures and reviewing eligibility restrictions. NIHR's site checklist asks about realistic timelines, participant-identification databases and visit workload. Those are concrete inputs to a recruitment model, rather than generic optimism about the number of activated centers. (^[8]) (^[56]) (^[57]) (^[58]) (^[59]) (^[60])

Feasibility Calculator (Hypothetical Example)

The following inputs are assumptions, not benchmarks or user-supplied forecasts. They illustrate how a sponsor could replace each value with its own observed evidence. The model assumes constant productivity after activation, no recruitment competition, sufficient supply and no loss of already counted participants.

• Target: assume 120 enrolled participants.
• Screen failure: assume 40%, implying 120 / 0.60 = 200 screenings.
• Screening capacity: assume 8 active sites, each screening 2.5 candidates per month. Capacity is 20 screenings and 12 enrollments per month.
• Recruitment duration: 120 / 12 = 10 months after full activation; a hypothetical uniform 2-month activation delay gives 12 elapsed months.

The arithmetic separates screening productivity from enrollment productivity. If the assumed pace were already enrolled participants per site-month, applying the screen-failure discount again would double-count attrition. NCI guidance distinguishes screened people from enrolled participants, and registry enrollment definitions need to be checked before importing counts into a model. (^[61]) (^[62])

A second hypothetical sensitivity changes only screening conversion: at 60% screen failure, the same screening capacity produces 8 enrollments monthly and requires 15 recruitment months. This is calculated from the stated assumptions, not evidence that a particular protocol will behave that way. Real recruitment models should sum active-site months across staggered openings and incorporate observed conversion, closures and competing capacity. CTTI's milestone guidance explicitly recognizes activation, screening and enrollment trajectories. (^[33])

Convert the Matrix into Funding and Operating Gates

Recommended governance questions are:

• Recruitment gate: do site-level eligible pools and recent conversion data support the recruitment curve?
• Site-mix gate: can proposed additions actually access relevant participants and provide the required visit capacity?
• Supply gate: can available stock and replenishment cover the revised site and participant schedule?
• Funding gate: is committed runway sufficient to reach the next decision-quality dataset under a slower scenario?
• Scientific gate: are interim analyses and decision rules prospectively specified and reviewed by the appropriate monitoring body?
• Evidence gate: can an unexplained registry reason be resolved before it influences a program-level forecast?
• Revision gate: which assumptions, approvals and resource commitments change when observed recruitment departs from plan?

The gates are recommendations, not empirically validated risk scores. Recruitment remediation should protect the study's ability to answer its question; International Council for Harmonisation, or ICH, guidance discusses identifying shortfalls and remedial action to protect statistical power. Scientific interim decisions still require their own prespecified framework. (^[63]) (^[9])

Regulatory timing can also matter to feasibility. For applicable UK clinical trials of investigational medicinal products, or CTIMPs, HRA guidance describes a two-year first-recruitment expectation and extension provisions; applications submitted before April 28, 2026 are outside that new expectation. Sponsors should confirm applicability with their regulatory team rather than treating a generic recruitment forecast as sufficient. (^[64]) (^[65])

Reproduction and Revision Policy

Full historical reproduction requires publishing a dated frozen extract and coding metadata, including the projected fields, verbatim reason text, record-level codes and ambiguity flags. A live-query link alone does not supply those materials. A future refresh should preserve the distinction between a frozen observation and an updated registry:

• Save: retain returned projected records, retrieval date and query parameters.
• Paginate: collect every page before deduplicating by NCT ID.
• Check: reconcile row count, uniqueness, eligibility and mutually exclusive reason totals.
• Review: retain ambiguous assignments and obtain independent human coding before high-stakes reuse.
• Compare: distinguish new records, status transitions, edited explanations and codebook revisions.
• Archive: recover historical enrollment targets from prior versions before calculating target attainment.

Version history is essential when a reason or enrollment field changes. ISRCTN describes retaining dated updates, and the NIH-led enrollment analysis used original archived targets. Neither practice is reproduced by simply downloading a current enrollment number. (^[66]) (^[29]) An expanded study could use a prespecified start cohort, sufficiently long follow-up, validated condition mapping and independently adjudicated reasons. That would address questions this compact stopped-record census deliberately leaves unanswered.

Clinical trials stop early for scientific, participant-protection, operational and financial reasons. The status label alone cannot distinguish them. Terminated, withdrawn and suspended records should be separated before analysts compare counts, interpret a reason or translate a finding into a development decision. Registry explanations remain submitted descriptions, and published classification research cautions that they may capture only part of the story. (^[14])

The census illustrates an approach: define the population, list identifiers, retain unknown explanations, discuss mixed cases and show how alternative coding changes the result. Its small registration window, stopped-only denominator and absence of human adjudication limit its use. The query and identifier lists support further checking, but full historical reproduction requires a reader-accessible dated extract and complete coding metadata. These counts do not provide a universal termination probability or a validated prediction of program success.

For sponsor feasibility teams and CRO leaders, the practical task is to connect each uncertainty with a decision. Recruitment concerns call for eligible-pool evidence and credible site capacity. Logistics concerns call for executable supply or design remedies. Financing concerns call for explicit runway commitments. Scientific stopping belongs within its planned monitoring and analysis framework. Evidence-based feasibility and transparent reporting provide complementary foundations for those decisions. (^[13]) (^[11])

A stronger future census would expand the population while preserving these distinctions and improving adjudication. Until then, the most useful question is specific: what evidence would change the enrollment assumption, site mix, supply commitment or funding gate for this study? An auditable reason taxonomy makes that discussion clearer without pretending that every early stop means the same thing.