Reversals: correct entries that misreport their sources, incl. the replicated soup-bowl effect
Description
Cross-checking the Replications and Reversals page against FLoRA and the cited papers turned up entries that misreport their own sources. For example, the page lists the bottomless soup bowl as "not replicated", although a preregistered direct replication (Lopez et al., 2024, n = 464) reproduced the effect at about half the original size. In another entry, the hurricane-names percentages are male/female ratios mislabelled as decreases. This PR corrects only the errors confirmed against the primary sources. The evidence for each change, with short quotes and page numbers, is in the corrections ledger.
Status changes
Two rules: reversed needs a statistically supported effect in the opposite direction, and a null original has no direction to reverse.
- Bottomless soup bowl: not replicated → replicated (about half the original effect size). Adds Lopez et al. 2024 and keeps the critical blog posts.
- Three entries whose original finding was null move from reversed to not replicated (…): age of acquisition (AoA) in recognition, the AoA verification-task entry, and numerical ratio effects in elephants (no ratio effect in Asian elephants; a later ratio effect in African elephants). The sibling AoA object-decision entry already uses this wording.
- Three entries whose follow-ups found only non-significant opposite-sign estimates move from reversed to not replicated (…): hurricane names (Smith 2016, p = .31 and .50), elderly priming (Doyen et al. 2012: no effect with automated timing unless experimenters expected slower walking) and scarcity resource allocation (O'Donnell et al. 2021: F(1, 530) = 3.59, p = .059).
- Depressive realism: reversed → mixed (much smaller effect in the original direction). The meta-analysis reports d = −0.07 in the same direction.
- Object distance: reversed → NA. The only cited follow-up used a different task, so it cannot reverse the original effect.
Labels checked because FLoRA disagreed
Where FLoRA's replication codes disagreed with a page label, both were checked against the papers. The FLoRA code alone never decided a label.
- Abstract rule learning, induced-compliance dissonance, choosing vs rejecting, facial feedback, moral priming (cleanliness) and men's preference for competition become mixed. Each now cites the replication the page had omitted: Geambașu et al. 2022, Vaidis et al. 2024, Many Labs 2 and Shafir & Cheek 2024, Many Smiles, Huang 2014 and Carlsson et al. 2020. With the original materials, Many Labs 2 found a significant opposite effect for choosing vs rejecting.
- Scarcity brand attitudes ("replicated (NS)…") and neonate imitation (NA) become not replicated (…). Neonate imitation now also cites Koepke et al. 1983.
- Last place aversion: replicated → mixed. Two direct replications found no last-place aversion: Camerer et al. 2016 (the redistribution game) and Martinangeli & Windsteiger 2021. A modified design and a queueing study found support. The description now covers both of the original's experiments, not just the gambles. The original's DOI pointed to an unrelated chemistry cover image and is fixed.
- Prosocial spending: replicated → replicated (smaller effects; one of three experiments was null).
- Less-is-better: fixes "d = 76" and a confidence interval.
- FLoRA's own errors are filed as forrtproject/fred-data#175.
Description, sample and study-type fixes
- Elephants: the original tested five elephants in Experiment 1 and four in Experiment 2, not three; effect lines say which condition the correlations come from.
- Hurricane names: describes Smith's data correctly (expanded Atlantic set and a new Pacific set, not "same data") and replaces the impossible "264%/103% decrease" with the reported means and p values.
- Object colour: describes Connell's original mismatch advantage.
- Facial feedback: separates the Coles et al. 2019 meta-analysis (286 effect sizes, 138 studies) from the Many Smiles study (n = 3,878).
- Loss aversion: removes a misplaced unconscious-thought meta-analysis.
- Slow blinking: corrects the sample to n = 18.
- Colour approach/avoidance: marks the 2013 poster as the same sample as Steele 2014.
- Peptic ulcers: identifies Levenstein et al. 1999 as a review; the page had attributed an n = 4,595 sample to it.
- 5-HTT × stress: Chabris et al. 2012 (intelligence) and Farrell et al. 2015 (schizophrenia) test other genes, so they move from the replication line to a labelled context note. Serretti et al. 2007 (HTR2A, unlinked) is removed.
Links
- 22 citations (25 links) pointed to unrelated works. 19 now point to the intended work, each checked on Crossref against the cited authors, year and title. Three are unlinked because the intended identifier could not be verified.
The remaining six reversed labels (Romeo and Juliet, monolingual creativity, object colour, handedness and dyslexia, mate guarding, personality and political attitudes) are kept; the ledger lists how each was checked. No unrelated wording or citation formatting changed.
Related to #612.
Type of Change
- Content/documentation update
- New feature
- Bug fix
- Breaking change
Testing
- Tested locally:
hugobuild,scripts/check_seo_html.pyandscripts/check_sitemap_lastmod.pypass; the edited entries render correctly - Previewed on staging.forrt.org
- Not tested yet
Checklist
- Self-reviewed my changes
- Verified links and formatting are correct
- No new warnings or errors
Notes
git merge-tree shows no conflict with rebrand-doors.
Created with Claude Opus 5.5 in Claude Code (T3 Code); GPT-6.1 Sol independently reviewed the age-of-acquisition verification-task and elephants corrections, and the FLoRA disagreements (four reviews and a second review of the labels where we disagreed).
🤖 Generated with Claude Code