Geng Integrity Report: Data Inconsistencies in "The Impact of Parental Absence on Adolescents' Health Risk Behaviors: Evidence from the China Education Panel Survey"

Academic fraud report · Geng Detector

Summary

This report assesses the article by Nan Wang et al. in Emerging Markets Finance and Trade (DOI: 10.1080/1540896X.2026.2630740, wait: 10.1080/1540496X.2026.2630740), Vol. 62, No. 14, pp. 5693–5710. Overall verdict: highly suspicious, based on internal data-consistency checks. Two findings are confirmed as hard inconsistencies: (1) Table 2's School ranking group statistics are identical to three decimal places across subsamples (0.809, SD 0.393) despite different sample sizes, suggesting direct copying from full-sample statistics; (2) Radius and Kernel PSM matching produce exactly identical coefficients to the third decimal (0.049/0.049, 0.036/0.036, 2.188/2.188), although other estimates do differ, implying the authors knew the methods should diverge. Three further findings carry plausible benign explanations (last-digit mismatches between text and Table 3; implausibly large control-variable coefficients; garbled Table 6 layout; asymmetric factor-score distribution) and require original output for confirmation. Automated pixel-level and Benford/last-digit forensics were deemed inapplicable and excluded. Confidence is moderate-to-high for Findings 1–2; final determination requires institutional investigation.

Verdict

Highly suspicious (high-confidence internal inconsistency findings). The paper's problems are confined to internal data self-consistency; image-forensics and digit-distribution tests were judged inapplicable to this empirical social science paper and excluded from the assessment. Findings 1 and 2 are confirmed as hard inconsistencies after adversarial review; Findings 3–6 remain flagged with plausible benign explanations pending original data/output from the authors.

Key findings

Finding 1 (confirmed, orange): Table 2 (p. 5700, Panel B) reports identical School ranking means and SDs — 0.809 (0.393) — for divorced Yes/No groups (Diff −0.001) and migrant Yes/No groups (Diff −0.001), matching Table 1 full-sample statistics exactly, despite differing subsample sizes (N1=5940/N2=446; N1=5678/N2=708) and stated exclusion of missing-value cases. Suggests copy-paste from full-sample statistics.
Finding 2 (confirmed, orange): Radius vs. Kernel matching yield coefficients identical to the third decimal: Table 4 Panel A (0.049 (0.020) / 0.049 (0.020); 0.036 (0.016) / 0.036 (0.016)) and Table 3 migration column (2.188* (0.037) / 2.188* (0.033)). Other pairs (divorce: 0.156 vs 0.186; 0.162 vs 0.161) do differ, indicating the two methods should not coincide exactly.
Finding 3 (flagged, yellow): Text states β = 0.155 and β = 0.185 for parental divorce (Columns 1–2), while Table 3 shows 0.156* and 0.186 — last-digit mismatch, possibly transcription error.

Finding 4 (flagged, yellow): Table 3 shows extreme binary-control coefficients (School type 7.385, SE 0.097; School ranking 6.591, SE 0.095; School region −7.590, SE 0.173; constant −6.929) alongside modest adj_R² (0.363–0.264); possibly explained by rare categories and school fixed effects, but the coefficient–SE–R² relationship appears mismatched.

Finding 5 (flagged, yellow): Table 6 (p. 5704) contains garbled, concatenated numbers (e.g., "6,3866,3866,3866,386"; "0.0650.065 0.1300.1270.035 0.035"), preventing verification of stated column correspondences; likely typesetting/PDF extraction artifact.

Finding 6 (flagged, yellow): Health risk behaviors factor score (Mean 0, SD 1, Min −0.419, Max 9.612, N=6386) is highly asymmetric; principal factor scores need not be symmetric, so a benign explanation exists.

Evidence highlights

Table 2 Panel B: 0.809 (0.393) repeated four times across subsamples of differing sizes (446 vs. 708 in migrant split), with Diff −0.001 in both panels.

Table 4 Panel A: paired ATT estimates identical to three decimals across two distinct matching algorithms; Table 3: 2.188* identical across methods with only SEs differing (0.037 vs 0.033).
Text vs Table 3: 0.155/0.185 in prose vs 0.156/0.186 in table.
Table 1: Min −0.419, Max 9.612, adj_R² 0.363 cited in Finding 4 verified verbatim in source.
Automated Bayesian synthesis (not a determination of misconduct) assigned posterior ≥99.9% (BF = 4.46e+11), driven mainly by Finding 2 and PRNU image-pair comparisons of table screenshots (NCC=1.000, PCE=128.0), the latter considered layout-consistent and not independently probative.

Notes

Benford/last-digit tests applied to regression coefficients and factor scores were excluded for unmet test assumptions; page-level copy-move hits were regular table/axis elements and not treated as evidence of fabrication.
Recommended actions (from report): request CEPS raw data, matching code, and original Stata/R output (especially Table 2 School ranking row, Tables 3–4 coefficients); post findings 1–2 on PubPeer; notify the journal (Emerging Markets Finance and Trade / Taylor & Francis) and the authors' institution.
AI-assisted report; final misconduct determination requires professional investigation. Findings 3–6 require original regression output and factor-score distributions to resolve.
One automated synthesis caveat: evidence treated as conditionally independent, which may overstate joint strength of related evidence.