Integrity Review Report: 'Can Agricultural Water Rights Trading Promote Green Development in Agriculture?'

Academic fraud report · Geng Detector

Summary

This report assesses a 2023 Chinese economics paper (DOI: 10.20077/j.cnki.11-1262/f.2023.02.010) by Yao Peng and Li Huizhao on agricultural water rights trading. The overall verdict is highly suspicious (orange/high). Pixel-level image analysis was not possible because original high-resolution figures were not provided. However, multiple severe issues were identified in data logic and econometric methodology. The authors interpret a baseline DID coefficient of 0.0865 as an 8.65% increase, which is arithmetically incorrect; the proper calculation against a dependent-variable mean of 0.485 yields approximately 17.83%, suggesting either fabrication or fundamental miscalculation. In 2SLS regressions, the coefficient on the treatment variable explodes to 7.5789, against a theoretical maximum of 0.938 for the dependent variable—a roughly 88-fold inflation consistent with a weak instrument or severe misspecification. The constructed instrument (1984 regional telecommunications × national internet ports) is time-invariant cross-sectionally and time-varying but region-invariant, and would be fully absorbed by two-way fixed effects that the authors claim to include. Finally, subgroup sample counts in Tables 7–8 fail to sum to the stated 1,356 total observations without explanation, raising p-hacking concerns. Image-based checks remain pending pending raw files.

Verdict

Highly suspicious (orange/high). The paper contains multiple severe mathematical and econometric inconsistencies that strongly suggest data manipulation, fabrication, or fundamental methodological incompetence. Pixel-level image analysis was not performed because the report was based on text/tables only.

Key findings

Arithmetic misrepresentation of baseline coefficient: The authors claim the DID coefficient of 0.0865 represents an 8.65% increase in green development. Given a dependent-variable mean of 0.485, the correct percentage change is approximately 17.83% (0.0865 / 0.485). Reporting 8.65% amounts to misplacing the decimal point by two positions.
Implausible IV coefficient magnitude: In Table 4 (2SLS), the coefficient on the water-rights-trading dummy is 7.5789. The dependent variable (composite green development index) has a maximum of 0.938 (from Table 2 descriptive statistics). This is a roughly 88-fold inflation over the OLS/DID estimate (0.0865) and is mechanically impossible within the variable's range.
Collapsed instrument under two-way fixed effects: The instrument—1984 per-capita regional postal/telecom volume × 2015–2020 national internet ports—is constructed from a time-invariant cross-sectional variable multiplied by a time-varying but region-invariant variable. If the model includes both region fixed effects (σᵢ) and year fixed effects (λₜ) as Equation 1 specifies, this instrument is fully absorbed and Stata should drop it. The reported first-stage results therefore imply the fixed effects were not actually estimated, or the instrument description was copy-pasted.
Disappearing observations in subgroup regressions: Sub-sample counts do not reconcile with the 1,356 total observations: marketization subgroups (684 + 660 = 1,344, –12); economic development (636 + 684 = 1,320, –36); primary-industry share (624 + 690 = 1,314, –42); only the grain-producing-region split (828 + 528 = 1,356) reconciles exactly. This pattern is consistent with selective sample trimming (p-hacking).

Evidence highlights

Table 3, column (3): treatment coefficient = 0.0865; Table 2: mean of dependent variable = 0.485.
Table 4: 2SLS coefficient on did = 7.5789; Table 2: max of dependent variable = 0.938.
Equation 1 specifies region FE (σᵢ) and year FE (λₜ); instrument = (1984 telecom per million people) × (national internet ports 2015–2020).
Table 7/8 sub-sample counts vs. total N = 1,356 as documented above.
DOI: 10.20077/j.cnki.11-1262/f.2023.02.010 (2023).

Notes

Scope limitation: First-form (image reuse) and third-form (image splicing) pixel-level checks were not possible without the original high-resolution figures.
The arithmetic error (8.65% vs. 17.83%) and the implausible 2SLS coefficient could in principle arise from extreme methodological sloppiness rather than intentional fabrication; however, the combination of all four issues substantially raises the probability of misconduct.
Confidence: high on the arithmetic and IV-absorbing-FE findings (deterministic from the published numbers); medium on the p-hacking inference (sample-count discrepancies could in principle reflect undocumented missing-data rules).
Recommended actions include PubPeer comment, editorial notification requesting raw data and Stata do-files, and institutional ethics review.