What Are Zombie Trials and Why Do They Matter for Biomedical Evidence?

The term “zombie trial” was introduced by John Carlisle, editor of Anaesthesia, to describe RCTs whose data are so problematic that the study should have been retracted—but was not. These are trials that look alive on PubMed, carry DOIs, get cited in Cochrane reviews, yet contain data that is fabricated, falsified, or so deeply flawed that no reliable conclusion can be drawn from them (PMID:33040331).

The distinction matters because zombie trials occupy the top of the evidence hierarchy. Randomized controlled trials are the gold standard for causal inference in medicine. When they are fabricated, the distortion propagates upward through meta-analyses, systematic reviews, and clinical practice guidelines—ultimately reaching the patient.

Ioannidis estimated that hundreds of thousands of zombie randomized trials circulate in the literature, arguing that the problem extends far beyond the anesthesiology field where Carlisle’s original analysis was conducted (PMID:33124075).

Prevalence of Fabricated and False Data in Randomized Controlled Trials

The most granular prevalence data come from Carlisle’s analysis of 526 RCTs submitted to Anaesthesia between February 2017 and March 2020. Among the 153 trials where individual participant data (IPD) were available, 44% contained false data and 26% met the threshold for “zombie” classification—meaning the flaws were so severe that the trial was deemed irrecoverable. Among the 373 trials without IPD access, only 2% were flagged as false and 1% as zombie, strongly suggesting that aggregated summary data conceal fabrication (PMID:33040331).

The geographic distribution was uneven. China and Egypt had the highest proportions of false data submissions: 48% and 90% of their respective IPD-available trials contained false data. South Korea, India, and Japan followed at 32%, 62%, and 18% (PMID:33040331).

Self-reported data provide a complementary but lower bound. Fanelli’s meta-analysis of 18 surveys found that 1.97% of scientists admitted to fabricating, falsifying, or modifying data at least once, while 14.12% reported personal knowledge of a colleague committing such misconduct. Medical and pharmacological researchers reported higher rates than scientists in other fields (PMID:19478950).

How Zombie Trials Distort Meta-Analyses and Systematic Reviews

The downstream consequences of zombie data in evidence synthesis are severe and quantifiable. O’Connell et al. systematically traced a cohort of untrustworthy spinal-pain trials through 32 unique systematic reviews and 10 clinical practice guidelines. Across 55 meta-analytic comparisons, removing the untrustworthy “index trials” reduced treatment effect sizes by a median of 58% (IQR 40–74). In 85% of comparisons, the impact was classified as “high” (PMID:37453533).

The statistical significance of pooled estimates is also affected. In 12 of 40 comparisons (30%) that were originally statistically significant (p < 0.05), the effect became non-significant once the zombie trials were removed. Nine out of 10 reviews conducting narrative synthesis had drawn positive conclusions about the intervention; nine out of 10 CPGs had made positive treatment recommendations (PMID:37453533).

The Self-Correction Failure: Why Zombie Data Persists in Clinical Guidelines

The biomedical literature’s capacity for self-correction is, in practice, minimal. Kataoka et al. conducted a meta-epidemiological study of 587 articles (525 systematic reviews and 62 CPGs) that cited retracted RCTs. Among the 127 articles published after retraction that included the retracted trial in their evidence synthesis, none corrected themselves. Among the 239 articles published before retraction where the RCT was later retracted, only 5% of systematic reviews and 5% of CPGs corrected or retracted their results (PMID:35779825).

The citation pattern compounds the problem. Approximately 90% of systematic reviews published after an RCT had been retracted cited the trial without caution or acknowledgment of its retracted status. The retracted data continues to function as if it were valid, feeding into new syntheses and guideline updates (PMID:35779825).

Investigation timelines are part of the problem. Misconduct investigations in fields such as women’s health have been shown to take a median of over 11 years. During this interval, the suspect data continues to accumulate citations and influence downstream evidence products (PMID:39628711).

Superretractors: Concentrated Sources of Fabricated RCT Data

A 2026 JAMA Network Open cohort study linked an openly available dataset of retracted RCTs (VITALITY) to lists of authors with the most retractions. Among 1,330 retracted RCTs, just 6 superretractors—concentrated in anesthesiology and endocrinology/metabolism—coauthored 290 (22%) of all retracted trials. Expanding the list to 18 career-long top-cited scientists with 10 or more retractions captured 327 trials (25%), with 84% of these overlapping with the superretractor set (PMID:41984475).

The lag between publication and retraction was striking. Superretractor-authored RCTs had a median publication-to-retraction interval of 5,111 days—approximately 14 years. During that window, these papers accrued a median of 21 citations (IQR 12–42), compared with 5 citations (IQR 1–19) for non-superretractor retractions. In multivariable regression, time to retraction was the only variable significantly and positively associated with total citations (PMID:41984475).

The implication is that a small number of prolific fabricators can poison an entire evidence ecosystem. Their work, often published in high-impact journals and cited by dozens of systematic reviews, has an outsized influence on the conclusions that shape clinical practice.

Statistical Methods for Detecting Fabricated Data in Clinical Trials

Detection of zombie trials depends on a hierarchy of statistical forensics, from baseline plausibility checks through machine-learning anomaly detection to full individual-participant-data audit.

The Carlisle method. The most widely cited approach tests whether the distribution of baseline characteristics across treatment groups is mathematically consistent with genuine randomization. Key indicators include perfectly balanced groups across many variables, distributions that are too narrow or too wide, and means/SDs that are inconsistent with the reported sample size. In one cohort, this method flagged 86% of 35 RCTs from a single author group as having implausible baselines (PMID:39628711).

Central statistical monitoring (CSM). CSM applies thousands of statistical tests across multiple variables to calculate a “data inconsistency score” for each clinical site in a multicentre trial. Pogue et al. developed and validated risk scores that discriminated between centres with and without fabricated data, achieving area-under-the-curve values of 0.90–0.95 (PMID:23283577).

IPD forensics. When raw datasets are available, investigators look for repeating blocks of data (identical participant rows duplicated), calculation errors where figures do not reconcile with tables, and excessive similarity or difference between groups that would not arise naturally. The Cochrane ivermectin review discovered fabricated spreadsheets through exactly this kind of row-level inspection (PMID:36054583).

Funnel-plot and sensitivity analysis. At the meta-analytic level, zombie trials often appear as extreme outliers reporting implausibly large effects. Integrity-based sensitivity analysis—removing the suspect trials and measuring the shift in the pooled estimate—quantifies the degree to which conclusions depend on untrustworthy sources (PMID:37453533).

Emerging Integrity Frameworks: RIGID, TRACT, and INSPECT-SR

The recognition that conventional risk-of-bias tools assume data authenticity has driven the development of dedicated integrity screening frameworks.

RIGID (Research Integrity in Guidelines and evIDence synthesis) is a six-step framework co-developed with 80 international experts. Its steps are: (1) standard systematic-review processes, (2) exclusion of retracted studies, (3) integrity assessment using the TRACT checklist, (4) committee discussion and voting on risk ratings, (5) formal author engagement including IPD requests, and (6) re-assessment and pathway allocation—included, awaiting classification, or not included. In its pilot deployment for the international PCOS guideline (endorsed by 39 societies across six continents), 45 of 101 originally identified studies (45%) were excluded due to integrity concerns (PMID:39628711).

TRACT (Trustworthiness in RAndomised Clinical Trials) is a 19-item checklist organized into seven domains: governance, author group, plausibility of intervention usage, timeframe, drop-out rates, baseline characteristics, and outcomes. Each item is rated as no concerns, some concerns, or major concerns. When a majority of items are rated as major concern, the framework triggers a more thorough investigation including IPD assessment (PMID:37337220).

INSPECT-SR (INveStigating ProblEmatic Clinical Trials in Systematic Reviews) is a Cochrane-led project developing a tool specifically for use within systematic reviews. It combines expert consensus from a Delphi survey with empirical evidence, and includes a planned extension (INSPECT-IPD) for forensic checks of underlying datasets (PMID:38471680).

RIA (Research Integrity Assessment), developed by the Cochrane COVID-19 ivermectin review team, assesses six criteria: retraction status, prospective registration, ethics approval, author group, plausibility of methods, and plausibility of results. It was used at the eligibility-check stage, resulting in the exclusion of more than 40% of studies in the first update of the ivermectin review (PMID:36054583).

Who Needs to Act on the Zombie Trial Problem?

Frequently Asked Questions About Zombie Trials