Uri Simonsohn's data-forensics method flags fabricated research data via 'excessive similarity' inconsistent with random sampling

Uri Simonsohn (co-founder of the Data Colada research-integrity blog) has publicly detailed cases where fabricated psychology data was caught purely from its summary statistics, without access to raw data or any admission from the researcher. The diagnostic signature is not noise but its absence: reported means, standard deviations, or cell counts across conditions or studies show "excessive similarity" — the numbers agree with each other, or with a theoretical prediction, more tightly than independent random sampling would ever produce, making the reported pattern "inconsistent with random sampling."

This operationalizes, with modern statistical tooling, the same inference Fisher applied informally to Mendel's peas in 1936: real measurement carries sampling noise, so a dataset with too little noise is not simply lucky — it is evidence the reported numbers did not arise from the process the researcher described. Simonsohn's method sits at the applied, quantitative end of a spectrum that runs from ancient legal doctrine (claim-sanhedrin-unanimous-guilty-verdict-acquits-the-defendant) through Fisher's informal chi-squared suspicion to a named, repeatable forensic technique used to force real retractions — compare claim-hirsch-forensics-drove-dias-superconductivity-retraction, where an implausibly smooth susceptibility curve played the same evidentiary role. See observation-suspicious-perfection-independence-absence-signals-defect for the general law these instances share.