Geng Academic Fraud Detection Report: Psychological Effects of Air Pollution: Evidence from House Viewings (DOI: 10.1080/1540496X.2026.2735333)

Academic fraud report · Geng Detector

Summary

This report assesses an empirical economics paper (Fang, Deng & Yang, Emerging Markets Finance and Trade, 2026, DOI: 10.1080/1540496X.2026.2735333). The reviewer's overall verdict is 'highly suspicious' (orange tier), while explicitly stating that no core evidence of data fabrication was found. Key retained concerns: (1) Table 3 shows a statistically significant positive effect for AQI_L2 (0.0173***, SE 0.0065) in the low-pollution range (51-100), contradicting the paper's narrative that only the two highest pollution categories show significant effects; (2) Figure 2 error bars and Figure 3 placebo distributions appear 'too smooth,' though Figures 1 and 3 have benign explanations supported by the paper's own text; (3) Table 7 reports incomplete R-squared and first-stage statistics (0.942 and 0.203 verified in-text). Numerical forensics (Benford, last-digit, inter-column difference tests) were downgraded as statistical preconditions are not met for regression coefficient tables. Image forensics signals (e.g., NCC=1.000, PCE=128.0 pairs; noise variance mean 923.7, CV=0.600) likely reflect repeated embedding of the same chart files in the PDF; retained but not treated as fabrication evidence. Confidence is limited: no raw microdata access, and automated Bayesian synthesis is noted as potentially overstating correlated evidence.

Verdict

Highly suspicious (orange tier). The paper contains no biological imagery, and after review, no core evidence of data fabrication was established. However, several independent, mutually consistent concerns remain unresolved, sufficient for a 'highly suspicious' rating. This is not a finding of misconduct; formal determination requires institutional investigation.

Key findings

Table 3 internal inconsistency (verified in-text): AQI_L2 coefficient is 0.0173* (0.0065) for the low-pollution band (51-100), significantly positive where the paper's own narrative ('only the two highest pollution categories … are associated with statistically significant increases') predicts no effect. AQI_L3 = 0.0083 and AQI_L4 = 0.0203 (not significant). EPM2.5_L2 = 0.0071 (0.0067, not significant) shows a different pattern despite the high correlation of AQI and PM2.5. Baseline coefficient 0.0178 (Table 2, col. 4) verified. Alternative benign explanations exist (different binning definitions), but authors should clarify.

Figure aesthetics 'too perfect' (low confidence): Figure 2 error bars of near-equal length at low pollution levels; Figure 3 placebo histogram unusually regular. Figures 1 and 3 have benign explanations supported by the text (AQI is a composite index including PM2.5; placebo coefficients 'centered around zero' across 500 random draws is the expected, self-declared result).

Table 7 reporting incomplete: only one R-squared (0.942) and one value of 0.203 reported despite columns (2)-(3) being first-stage estimates. Anderson canon. corr. LM statistic 988.692 and Cragg-Donald Wald F statistic 457.855 verified; instrument strength appears adequate. Likely a reporting oversight, not misconduct.

Numerical forensics downgraded: Benford (χ²=118.1), last-digit (χ²=1163.3), and inter-column difference tests do not satisfy statistical preconditions for regression coefficient tables; tool-generated statistics, not fabrication evidence.

Image forensics downgraded but retained: copy-move matches at identical offsets (e.g., img-006/012/016/020 at offset (32,0), 102 matched blocks) consistent with repeated chart elements; img-013/017 share identical noise statistics (variance mean 923.7, CV=0.600), suggesting the same file embedded multiple times. No positive statement of a unified export pipeline exists in the paper, so this explanation remains inferential.

Evidence highlights

Table 3: AQI_L2 = 0.0173* (0.0065); EPM2.5_L2 = 0.0071 (0.0067) — all values verified in the original text.
Table 7: R² 0.942, 0.203; Anderson LM 988.692; Cragg-Donald Wald F 457.855 — verified in-text.
PRNU pairs: NCC=1.000, PCE=128.0 (img-007 vs img-013/017/021); NCC=0.150, PCE=53.2 (img-007 vs img-016/020) — forensic tool outputs, not paper content.
Automated Bayesian synthesis (BF = 1.24e+6) treats evidence as conditionally independent and may overstate the joint strength of correlated evidence; it is not a misconduct judgment.

Notes

Recommended follow-up: authors should clarify the Table 3 AQI_L2 result vs. the paper's narrative, release placebo/randomization code, and complete Table 7 column-level statistics. Independent replication of the baseline coefficient 0.0178 with raw transaction data would resolve remaining doubts.
Machine forensics tools misfired on statistical charts and regression tables; 53 cues were excluded after review for benign explanations or unmet test preconditions.
AI-assisted report for academic discussion only; final misconduct determinations rest with professional institutions.