Verification checks (P4)
- Clean-state rerun: deleted results.json/table1.csv/figure1.png/exclusions.csv, reran analysis.py;
cmpof results.json against previous run → IDENTICAL. PASS - Exclusion ledger: 1200 − (0+0+12) = 1188 = n_analysed. PASS
- SEs positive: naive_se 0.0146, effect_se 0.0098, skill_index_se 0.078. PASS
- CIs contain estimates: naive [-0.160, -0.103] ∋ -0.132; effect [0.234, 0.273] ∋ 0.254. PASS
- Metrics: model_metric 1.976 < baseline_metric 3.860 (nested model adds regressor, RMSE must not increase). PASS
- CI width ≈ 2×1.96×SE: effect 0.0385 vs 3.92×0.00981=0.0385. PASS
- HC1 vs conventional SE nearly equal (0.00981 vs 0.00984): no major heteroskedasticity concern. PASS
- Conclusion code consistent with CI (low > 0 → positive_effect). PASS
- Sensitivity: including the 12 negative-hours rows changes effect to 0.252; adding skill² gives 0.254; interaction p=0.96; hours² p=0.86 → linear homogeneous spec not rejected. PASS
- Magnitude vs baseline: sign flips (−0.132 → +0.254), explained by positive skill–hours correlation and negative skill–growth relation (omitted-variable bias formula direction negative). PASS (consistent)
- Information rule: only data/firms.csv read (grep of analysis.py for read_csv). PASS