P3 R3 — Complete revision package ready for MIR submission

outputs/r3/audit/D1_diagnosis.json

@@ -0,0 +1,134 @@
+{
+  "target": {
+    "TCI": 0.187,
+    "fsts_c2_DAI": 2.972,
+    "adj_rsq": 0.192
+  },
+  "candidates": {
+    "C0_canonical (Table 2 M8)": {
+      "n": 617,
+      "TCI": 0.15282307108044393,
+      "DAI": 0.01876816338509209,
+      "fsts_c": 2.408923999224807,
+      "fsts_c2": -2.5426209374013347,
+      "fsts_c_DAI": -1.1666305870701332,
+      "fsts_c2_DAI": 3.118509354618096,
+      "rsq": 0.21060011168872506,
+      "adj_rsq": 0.19624738644670192
+    },
+    "C1a_DAI_thin_drop_k38": {
+      "n": 617,
+      "TCI": 0.15861554170001224,
+      "DAI": 0.008479481907202824,
+      "fsts_c": 2.5975854723773915,
+      "fsts_c2": -2.7531245984435566,
+      "fsts_c_DAI": -1.28498819624694,
+      "fsts_c2_DAI": 3.0922643284141547,
+      "rsq": 0.20950753766039276,
+      "adj_rsq": 0.19513494743603632
+    },
+    "C1b_DAI_thin_drop_k33": {
+      "n": 617,
+      "TCI": 0.1645477490187506,
+      "DAI": -0.013059716011683726,
+      "fsts_c": 2.231405718780488,
+      "fsts_c2": -1.5707464704683172,
+      "fsts_c_DAI": -0.7340104392100292,
+      "fsts_c2_DAI": 1.6191483683450174,
+      "rsq": 0.2008414357858288,
+      "adj_rsq": 0.18631128007284392
+    },
+    "C1c_DAI_thin_drop_c22b": {
+      "n": 617,
+      "TCI": 0.15373599136284502,
+      "DAI": 0.07010918490892645,
+      "fsts_c": 2.16139417274263,
+      "fsts_c2": -1.684150616632968,
+      "fsts_c_DAI": -0.6318527981029047,
+      "fsts_c2_DAI": 1.6562886210137013,
+      "rsq": 0.21016108503120345,
+      "adj_rsq": 0.1958003774863163
+    },
+    "C2a_TCI_thin_drop_h8": {
+      "n": 617,
+      "TCI": 0.1422183914950359,
+      "DAI": 0.02445551244246606,
+      "fsts_c": 2.4776853929272105,
+      "fsts_c2": -2.58671910902815,
+      "fsts_c_DAI": -1.1246997216921406,
+      "fsts_c2_DAI": 3.0147641864048897,
+      "rsq": 0.20902024523838514,
+      "adj_rsq": 0.19463879515181037
+    },
+    "C2b_TCI_thin_drop_h1": {
+      "n": 617,
+      "TCI": 0.16981767965791494,
+      "DAI": 0.019533892146092555,
+      "fsts_c": 2.2913682645292943,
+      "fsts_c2": -2.43244987819773,
+      "fsts_c_DAI": -1.1124207429032902,
+      "fsts_c2_DAI": 3.0793424970607717,
+      "rsq": 0.2147293679377481,
+      "adj_rsq": 0.2004517200820709
+    },
+    "C2c_TCI_thin_drop_e6": {
+      "n": 617,
+      "TCI": 0.14741399333234664,
+      "DAI": 0.01635189443854478,
+      "fsts_c": 2.5867681625242747,
+      "fsts_c2": -2.821870711641307,
+      "fsts_c_DAI": -1.2685230743042986,
+      "fsts_c2_DAI": 3.3250968361991293,
+      "rsq": 0.2096788042460085,
+      "adj_rsq": 0.19530932795957234
+    },
+    "C2d_TCI_thin_drop_b8": {
+      "n": 617,
+      "TCI": 0.0828870585521267,
+      "DAI": 0.03455671681008279,
+      "fsts_c": 2.5675529267842103,
+      "fsts_c2": -2.6942803193003795,
+      "fsts_c_DAI": -1.146727198099981,
+      "fsts_c2_DAI": 3.07065655794518,
+      "rsq": 0.1984443520258241,
+      "adj_rsq": 0.18387061297174823
+    },
+    "C4_HC3_SE": {
+      "n": 617,
+      "TCI": 0.15282307108044393,
+      "DAI": 0.01876816338509209,
+      "fsts_c": 2.408923999224807,
+      "fsts_c2": -2.5426209374013347,
+      "fsts_c_DAI": -1.1666305870701332,
+      "fsts_c2_DAI": 3.118509354618096,
+      "rsq": 0.21060011168872506,
+      "adj_rsq": 0.19624738644670192
+    },
+    "C5_HC0_SE": {
+      "n": 617,
+      "TCI": 0.15282307108044393,
+      "DAI": 0.01876816338509209,
+      "fsts_c": 2.408923999224807,
+      "fsts_c2": -2.5426209374013347,
+      "fsts_c_DAI": -1.1666305870701332,
+      "fsts_c2_DAI": 3.118509354618096,
+      "rsq": 0.21060011168872506,
+      "adj_rsq": 0.19624738644670192
+    }
+  },
+  "best": {
+    "name": "C2a_TCI_thin_drop_h8",
+    "result": {
+      "n": 617,
+      "TCI": 0.1422183914950359,
+      "DAI": 0.02445551244246606,
+      "fsts_c": 2.4776853929272105,
+      "fsts_c2": -2.58671910902815,
+      "fsts_c_DAI": -1.1246997216921406,
+      "fsts_c2_DAI": 3.0147641864048897,
+      "rsq": 0.20902024523838514,
+      "adj_rsq": 0.19463879515181037
+    },
+    "l1": 0.09018459006166424
+  }
+}

outputs/r3/audit/D1_root_cause.md

@@ -0,0 +1,140 @@
+# D1 Root-Cause Diagnosis — Table 3 Mismatch
+
+**Date:** 2026-05-02
+**Reviewer issue:** Manuscript Table 3 baseline row reports values that do
+not match Manuscript Table 2 Model M8 despite both being labeled as the
+same specification on the same N=617 sample.
+
+---
+
+## What the reviewer flagged
+
+Manuscript Table 3, baseline row:
+- TCI β_z = **0.187***
+- FSTS² × DAI = **2.972***
+- Adj. R² = **0.192**
+
+Manuscript Table 2, Model M8:
+- TCI β_z = **0.153***
+- FSTS² × DAI = **3.119***
+- Adj. R² = **0.196**
+
+Both rows claim N = 617 and the same TCI_full + DAI_rich specification.
+
+---
+
+## Audit findings
+
+### 1. Canonical M8 (locked spec) reproduces Table 2 M8 exactly.
+
+`tools/r3/00_canonical_m8.py` re-estimates M8 from the raw `.dta` using
+the locked specification (4-component TCI, 3-component DAI with k33/k38
+single-missing imputation, HC1 robust SE, broad-sector NACE buckets,
+1/99 winsorized ln-labor-productivity). Result:
+
+| Term | Coefficient | SE | p |
+|---|---|---|---|
+| TCI | 0.1528 | 0.0412 | <0.001 |
+| DAI | 0.0188 | 0.0454 | 0.679 |
+| FSTS² × DAI | 3.1185 | 1.1236 | 0.006 |
+| **N** | **617** | | |
+| **Adj. R²** | **0.196** | | |
+
+→ Matches Table 2 M8 exactly. Canonical M8 is therefore the single source
+  of truth for all subsequent R3 work.
+
+### 2. No alternative specification reproduces the Table 3 baseline row.
+
+`tools/r3/01_diagnose_table3_baseline.py` and
+`tools/r3/03_diagnose_tci_spec.py` test ten plausible alternative specs:
+
+- DAI variants (drop k33, drop k38, drop c22b)
+- TCI variants (drop one of b8/e6/h1/h8)
+- TCI standardization variants (z-then-avg-then-z; simple avg then z;
+  z-then-avg only; sum-of-z then z)
+- Robust-SE variants (HC0, HC1, HC3)
+
+Closest match (V2_simple_avg_then_z) yields TCI=0.159 and FSTS²×DAI=3.110,
+which still disagrees with Table 3 baseline (0.187 / 2.972). No spec
+reproduces the joint triple (0.187, 2.972, 0.192).
+
+### 3. All six rows of Table 3 contain transcription errors.
+
+`tools/r3/02_audit_table3.py` re-estimates each Table 3 row from raw data:
+
+| Row | Spec | N (rep/man) | TCI (rep/man) | FSTS²×DAI (rep/man) | AdjR² (rep/man) |
+|---|---|---|---|---|---|
+| Baseline | TCI_full + DAI_rich | 617/617 ✓ | 0.153/0.187 ✗ | 3.119/2.972 ✗ | 0.196/0.192 ✓ |
+| R1 | DAI_thin (c22b only) | 623/623 ✓ | 0.180/0.211 ✗ | 1.552/1.563 ✓ | 0.188/0.184 ✓ |
+| R2 | TCI_thin (e6 + b8) | 617/617 ✓ | 0.159/0.171 ✗ | 2.930/2.749 ✗ | 0.199/0.188 ✗ |
+| R3 | Excl micro-firms (<10 empl) | 464/464 ✓ | 0.170/0.218 ✗ | 3.521/3.535 ✓ | 0.219/0.226 ✗ |
+| R4 | SMEs only (≤200 empl) | 595/595 ✓ | 0.156/0.180 ✗ | 3.505/3.483 ✓ | 0.199/0.197 ✓ |
+| R5 | Exporters only (FSTS > 0) | 84/84 ✓ | 0.130/0.143 ✗ | **2.821**/1.308 ✗ | 0.165/0.140 ✗ |
+
+Ns are correct in all six rows. Coefficients deviate systematically:
+- TCI is consistently **higher** in the manuscript than canonical (0.012 to 0.048).
+- The R5 (exporters-only) FSTS²×DAI is off by **+1.513** (manuscript reports 1.308 but canonical is 2.821) — this is the single largest deviation.
+
+### 4. Most plausible explanation
+
+The Table 3 numbers appear to be a **stale snapshot from an earlier
+manuscript draft** that used a slightly different TCI or sample-build
+pipeline. The subsequent revision updated Table 2 M8 to the locked spec
+but did not propagate the change to Table 3, leaving Table 3 with stale
+coefficients across all rows.
+
+This matches the reviewer's observation that "the baseline row in Table
+3 appears intended to represent the same full specification as Model
+M8" yet does not match.
+
+---
+
+## Recommended fix
+
+**Replace the entire Table 3 with the canonical re-estimation
+(corrected six-row block produced by `02_audit_table3.py`).**
+
+Corrected Table 3:
+
+| Specification | N | TCI β_z | FSTS² × DAI | Joint F (p) | Adj. R² |
+|---|---|---|---|---|---|
+| Baseline (TCI_full + DAI_rich) | 617 | +0.153*** | +3.119** | 4.56 (.011) | 0.196 |
+| R1: DAI_thin (website c22b only) | 623 | +0.180*** | +1.552 | 4.01 (.019) | 0.188 |
+| R2: TCI_thin (e6 + b8) | 617 | +0.159*** | +2.930** | 4.27 (.014) | 0.199 |
+| R3: Excl micro-firms (<10 empl) | 464 | +0.170*** | +3.521** | 4.61 (.010) | 0.219 |
+| R4: SMEs only (≤200 empl) | 595 | +0.156*** | +3.505** | 5.30 (.005) | 0.199 |
+| R5: Exporters only (FSTS > 0) | 84 | +0.130 | +2.821 | 6.32 (.003) | 0.165 |
+
+Significance stars: † p<.10, * p<.05, ** p<.01, *** p<.001.
+
+### Implication for narrative (R5 exporters-only row in particular)
+
+The corrected R5 row (FSTS²×DAI = +2.821, joint F = 6.32, p = .003)
+**strengthens** rather than weakens the moderation finding within the
+exporter subsample. The previously reported R5 (FSTS²×DAI = +1.308,
+joint F = 3.18, p = .048) suggested the result was attenuated when
+restricting to exporters. The canonical re-estimation shows the
+opposite: the moderation pattern is stronger and more significant
+within the exporter subsample.
+
+This affects how Section 7 caveats should be worded. The reviewer's
+warning that R5 "illustrates how the dynamics change when non-exporters
+are excluded" relied on the (incorrect) attenuation. With the corrected
+numbers, the dynamic is **reinforcement, not attenuation**.
+
+This finding will also be central to Phase 1's extensive–intensive
+split design: the intensive-margin (exporter-only) test now has a
+stronger basis.
+
+---
+
+## Provenance
+
+- Raw data: `data/raw/Singapore2023fulldata.dta` (WBES Singapore 2023, 623 firms)
+- Locked spec: `tools/r3/00_canonical_m8.py`
+- Diagnosis: `tools/r3/01_diagnose_table3_baseline.py`,
+  `tools/r3/03_diagnose_tci_spec.py`
+- Full audit: `tools/r3/02_audit_table3.py`
+- Audit JSON: `outputs/r3/audit/m8_canonical.json`,
+  `outputs/r3/audit/D1_diagnosis.json`,
+  `outputs/r3/audit/table3_audit.json`

outputs/r3/audit/bootstrap_tp.json

@@ -0,0 +1,17 @@
+{
+  "n_reps": 5000,
+  "n_valid": 5000,
+  "n_after_clip": 4833,
+  "full_sample_tp_pct": 82.42734943255799,
+  "boot_ci_95": [
+    52.783447184437435,
+    252.88850840165713
+  ],
+  "boot_iqr": [
+    68.34763515648574,
+    101.48642800926018
+  ],
+  "boot_median": 79.98398279880973,
+  "pct_inverted_U_shape": 96.26,
+  "pct_TP_in_0_100_range": 72.3980964204428
+}

outputs/r3/audit/extensive_margin.json

@@ -0,0 +1,60 @@
+{
+  "n": 617,
+  "n_exporters": 84,
+  "n_domestic": 533,
+  "pseudo_rsq": 0.18588284144786205,
+  "llf": -199.86874012076993,
+  "ll_null": -245.5036575770315,
+  "lr_chi2": 91.26983491252315,
+  "lr_p": 6.785563111143895e-17,
+  "params": {
+    "Intercept": -4.426522310624799,
+    "C(broad_sector)[T.manufacturing]": 2.060908687966435,
+    "C(broad_sector)[T.retail_services]": 1.8344999585877233,
+    "TCI": 0.498604602591217,
+    "DAI": 0.5823037154349072,
+    "ln_empl": -0.1532903849791616,
+    "firm_age": 0.03827272490857476,
+    "foreign": 0.9469228747992615
+  },
+  "bse": {
+    "Intercept": 0.7818548560166404,
+    "C(broad_sector)[T.manufacturing]": 0.6436399672390086,
+    "C(broad_sector)[T.retail_services]": 0.6283907155093299,
+    "TCI": 0.12395138741848782,
+    "DAI": 0.17713834216464655,
+    "ln_empl": 0.1293901370523867,
+    "firm_age": 0.010237936660509615,
+    "foreign": 0.281220425069438
+  },
+  "pvalues": {
+    "Intercept": 1.4999841915915947e-08,
+    "C(broad_sector)[T.manufacturing]": 0.0013649649617863816,
+    "C(broad_sector)[T.retail_services]": 0.0035074868430648473,
+    "TCI": 5.7563637646633504e-05,
+    "DAI": 0.0010115941350844339,
+    "ln_empl": 0.23613030632319265,
+    "firm_age": 0.00018525106840965825,
+    "foreign": 0.0007593808345014743
+  },
+  "odds_ratios": {
+    "Intercept": 0.011955996699375046,
+    "C(broad_sector)[T.manufacturing]": 7.853102588914332,
+    "C(broad_sector)[T.retail_services]": 6.262002102050488,
+    "TCI": 1.6464222537054076,
+    "DAI": 1.7901576980383642,
+    "ln_empl": 0.8578805700067292,
+    "firm_age": 1.039014559392361,
+    "foreign": 2.577765336068539
+  },
+  "ame": {
+    "C(broad_sector)[T.manufacturing]": 0.20329813750792605,
+    "C(broad_sector)[T.retail_services]": 0.1809640703719163,
+    "TCI": 0.04918480263174326,
+    "DAI": 0.05744129349499352,
+    "ln_empl": -0.015121315149040734,
+    "firm_age": 0.003775409234139334,
+    "foreign": 0.09340911508325701
+  },
+  "ame_weight_phi": 0.09864490294741146
+}

outputs/r3/audit/extensive_margin.txt

@@ -0,0 +1,21 @@
+                           Logit Regression Results                           
+==============================================================================
+Dep. Variable:               exporter   No. Observations:                  617
+Model:                          Logit   Df Residuals:                      609
+Method:                           MLE   Df Model:                            7
+Date:                Sat, 02 May 2026   Pseudo R-squ.:                  0.1859
+Time:                        08:52:05   Log-Likelihood:                -199.87
+converged:                       True   LL-Null:                       -245.50
+Covariance Type:            nonrobust   LLR p-value:                 6.786e-17
+======================================================================================================
+                                         coef    std err          z      P>|z|      [0.025      0.975]
+------------------------------------------------------------------------------------------------------
+Intercept                             -4.4265      0.782     -5.662      0.000      -5.959      -2.894
+C(broad_sector)[T.manufacturing]       2.0609      0.644      3.202      0.001       0.799       3.322
+C(broad_sector)[T.retail_services]     1.8345      0.628      2.919      0.004       0.603       3.066
+TCI                                    0.4986      0.124      4.023      0.000       0.256       0.742
+DAI                                    0.5823      0.177      3.287      0.001       0.235       0.929
+ln_empl                               -0.1533      0.129     -1.185      0.236      -0.407       0.100
+firm_age                               0.0383      0.010      3.738      0.000       0.018       0.058
+foreign                                0.9469      0.281      3.367      0.001       0.396       1.498
+======================================================================================================