Case Studies: What Statistical Forensics Would Have Flagged
All cases below have official investigation findings or formal retraction records. Our goal is not to pass judgment on anyone — those conclusions were already reached by investigating institutions — but to dissect: if statistical forensics screening had been applied at the time, which signals could have been flagged earlier?
⚠️ Each case's final determination rests with its official investigation report; this page cites only public records. The role of statistical fingerprints in these cases is “reproducible signals in hindsight” — not proof that would have stood alone at the time.
Jan Hendrik Schön (condensed-matter physics)
Official finding: data fabricationPublished 15+ papers in Science, Nature, and other top journals on superconductivity and molecular electronics in 2000-2001. In 2002, Bell Labs' official investigating committee found large-scale data falsification and fabrication; 16+ papers were retracted. Statistical post-mortems found the “experimental” data suspiciously smooth, with noise residuals from different experiments matching far too closely — something real physical measurements virtually never produce.
Reproducible screening signals: variance anomaly (data “too clean”) + residual/terminal-digit fingerprints.Try the variance screen →
Source: Beasley Report (official Bell Labs investigation committee, 2002).
Diederik Stapel (social psychology)
Official finding: data fabricationIn 2011, Tilburg University's official investigation (the Levelt Committee) determined that he fabricated experimental data in dozens of papers; 50+ papers were retracted. The report noted his data was “too beautiful”: large effect sizes with perfectly patterned between-group variances, lacking the statistical “imperfections” common to real data. The case became a landmark of psychology's credibility crisis.
Reproducible screening signals: variance anomaly (low CV) + GRIM/SPRITE consistency + terminal-digit preference.About SPRITE →
Source: Flawed Science (official Levelt Committee report, 2012).
Brian Wansink (food behavior research)
Official investigation + mass retractionsThe GRIM test's proving ground: Brown & Heathers (2017) applied GRIM to the summary statistics of his famous “all-you-can-eat pizza” paper and found several reported means were mathematically impossible given the sample sizes and integer scoring — triggering a mass re-examination of his work on PubPeer. Cornell's official 2018-2019 investigation found academic misconduct (including misreporting of research data); he resigned in 2019, with 17+ retractions to date.
Reproducible screening signals: GRIM (30 seconds, just n + mean) — this case proved the lightest method can move a major investigation.Try GRIM →
Sources: Brown NJL & Heathers JAJ (2017, The GRIM Test); Cornell University official statements (2019).
Paper Mills
Systematic retraction waveIn 2022-2024, major publishers retracted over ten thousand papers as “products of paper mills” (a single large open-access publisher retracted 8,000+). Typical flaws of mass-produced papers: references to non-existent works, templated tortured phrases, and suspicious consistency in terminal digits and baseline data. Publishers' B2B screening (e.g., STM Integrity Hub) is built on exactly these fingerprints.
Reproducible screening signals: citation verification (Crossref “not found” references) + terminal digit analysis.Try citation verification →
Sources: publisher retraction notices (2022-2024); public STM Integrity Hub materials.
The common pattern
- The signals were in the data all along — GRIM needs only the two numbers printed in the paper: n and the mean.
- Discovery came from bystanders — the breakthroughs above came from peers and independent researchers checking the work, not from the review system.
- No single method suffices — every case combined statistical fingerprints with requests for raw data and official investigations.
⚠️ Important: This page discusses historical cases with official findings only. For any paper not yet investigated, every method on this site can only produce “statistical anomaly” signals and must never be treated as a determination of misconduct; concerns should be raised through due process via PubPeer, journal editors, or institutional channels.