Add LPDiD R-parity validation harness (Dube et al. 2025), Phase B2

CHANGELOG.md

@@ -7,6 +7,21 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+- **`LPDiD` R-parity validation (absorbing).** `tests/test_methodology_lpdid.py` pins the
+  estimator against the method authors' own R recipes (`danielegirardi/lpdid` event-study /
+  reweight / premean / pooled `fixest::feols` specifications) with an `alexCardazzi/lpdid`
+  cross-check gate on the event-study variants, generated by
+  `benchmarks/R/generate_lpdid_golden.R`. Variance-weighted, reweighted, premean,
+  direct-covariate, pooled, and regression-adjustment-point estimates match to ~1e-12
+  (pooled is anchored to the authors' fixed-composition recipe only — `alexCardazzi`'s pooled
+  uses a laxer clean-control window and is recorded for transparency, not gated). The
+  regression-adjustment standard error (canonically Stata `teffects ra` only — no R-package
+  analogue) is pinned and its calibration validated by an ungated Monte-Carlo coverage study
+  (`benchmarks/python/coverage_lpdid_ra.py`, ≈0.95 coverage across cluster counts). Confirms
+  the two previously-provisional REGISTRY notes (RA influence-function variance convention;
+  pooled fixed-composition estimand) resolve in the library's favour with no estimator change.
+
 ## [3.6.0] - 2026-06-29
 
 ### Added

TODO.md

@@ -116,6 +116,7 @@ exists but parity can't be verified without a local toolchain.
 | `CallawaySantAnna` bootstrap: align p-value computation with R `did`'s symmetric-percentile method (former "CallawaySantAnna Bootstrap Improvements" section). | `staggered.py` | — | Low |
 | **`bias_corrected_local_linear` (lprobust) Phase-1c follow-ups:** extend golden parity to `kernel ∈ {triangular, uniform}` (epa-only today); expose `vce ∈ {hc0,hc1,hc2,hc3}` on the public wrapper once R goldens exist (port supports all four; needs a per-mode generator + a hc2/hc3 q-fit-leverage decision); clustered-DGP auto-bandwidth parity is **blocked upstream** on an nprobust singleton-cluster bug in `lpbwselect.mse.dpi` (Phase-1c DGP 4 uses manual `h=b=0.3`). | `_nprobust_port.py`, `local_linear.py`, `generate_nprobust_lprobust_golden.R` | Phase 1c | Low-Med |
 | `HeterogeneousAdoptionDiD` Stute-family Stata-bridge parity: no public R `Stutetest` package exists; would add `benchmarks/stata/generate_stute_golden.do` + a Stata dependency. | `benchmarks/stata/`, `tests/test_stute_test_parity.py` | follow-up | Low |
+| **`LPDiD` regression-adjustment SE — no runnable R reference.** The RA influence-function cluster SE is canonically Stata `teffects ra ... atet vce(cluster)` only; no R package computes it (`alexCardazzi/lpdid` does direct covariate inclusion, not RA). Today the RA *point* is R-anchored (~1e-12), the SE is pinned + MC-coverage-validated (`coverage_lpdid_ra.py`). Follow-up: contribute the RA path to `alexCardazzi/lpdid` so a runnable R RA reference exists — only a *trusted* anchor once cross-checked vs Stata `teffects` (else circular). | `tests/test_methodology_lpdid.py`, `benchmarks/python/coverage_lpdid_ra.py` | #B2 follow-up | Low |
 | `HeterogeneousAdoptionDiD` Phase-3 R-parity: ships coverage-rate validation on synthetic DGPs, not tight point parity vs `chaisemartin::stute_test` / `yatchew_test` (needs bootstrap-seed-semantics + `B` alignment across numpy/R). | `tests/test_had_pretests.py` | Phase 3 | Low |
 
 ### Parked — pending user demand / out of scope

benchmarks/data/lpdid_golden.json

@@ -0,0 +1,76 @@
+{
+  "meta": {
+    "estimator": "LPDiD (Dube, Girardi, Jorda & Taylor 2025) - absorbing",
+    "r_version": "R version 4.5.2 (2025-10-31)",
+    "fixest_version": "0.13.2",
+    "lpdid_alexcardazzi_version": "0.6.1",
+    "lpdid_alexcardazzi_commit": "ba64563983861be5e616f6020842c1a1cdf17a27",
+    "seed": 20260629,
+    "pre_window": 3,
+    "post_window": 4,
+    "se_convention": "Default/weighted/direct/pmd/pooled SEs: feols cluster-robust at unit with the reghdfe-style finite-sample factor (G/(G-1))*((n-1)/(n-k)) via setFixest_ssc(adj=TRUE, cluster.adj=TRUE) - the authors' convention.",
+    "ra_se_note": "ra_cov records the RA POINT estimate (full-interaction == teffects ra atet). The canonical RA SE is Stata teffects (unconditional IF, NO finite-sample factor) - no R package computes it; the library influence-function SE is pinned on the Python side and calibration-validated by benchmarks/python/coverage_lpdid_ra.py. ra_cov[h] = c(att, conditional_CR0_se_for_reference_only).",
+    "pooled_note": "Pooled uses the authors' fixed-composition window-mean recipe (mean(y over [t,t+POST]) - y_{t-1}, clean through F.POST). alexCardazzi's pooled uses a laxer clean-control window, so its pooled value is recorded here for transparency but NOT used as the cross-check gate.",
+    "nocomp_note": "no_composition is NOT a golden parity variant: the library fixes the realized sample across all post horizons (paper Section 3.6) and excludes cohorts p_g > T-H, more faithful to the paper than alexCardazzi's looser per-horizon version (recorded below). The library path is validated by the pure-Python B1 tests.",
+    "alex_pooled_post": [3.487463576882, 0.3495904202394],
+    "alex_nocomp_looser": {
+      "-3": [-0.7532927692809, 0.5003193796022],
+      "-2": [-0.250198606317, 0.6160670067758],
+      "-1": [0, 0],
+      "0": [1.46387943226, 0.545454853302],
+      "1": [2.335727820716, 0.5511865769279],
+      "2": [2.865671038762, 0.5566693486612],
+      "3": [3.245223911323, 0.4728816808867],
+      "4": [4.100734546643, 0.4370902112454]
+    }
+  },
+  "vw_es": {
+    "0": [2.891454184758, 0.4972838592697],
+    "1": [3.215287698688, 0.4242105863868],
+    "2": [3.570283061763, 0.4502488079078],
+    "3": [3.968564586288, 0.3904079437881],
+    "4": [4.269176444588, 0.396524220757],
+    "-2": [0.1183626967809, 0.4413440724331],
+    "-3": [-0.1254433057345, 0.4292726043484]
+  },
+  "ew_es": {
+    "0": [2.903628272087, 0.4973282590126],
+    "1": [3.221973963653, 0.4235473199649],
+    "2": [3.57527972142, 0.4499773773075],
+    "3": [3.971884855509, 0.390089371568],
+    "4": [4.265997983221, 0.3934421813065],
+    "-2": [0.1195811966608, 0.4402860720602],
+    "-3": [-0.1219233071656, 0.4298940947632]
+  },
+  "pmd_es": {
+    "0": [2.832272836367, 0.4337549369753],
+    "1": [3.156106350297, 0.3522180898809],
+    "2": [3.511101713373, 0.394076563941],
+    "3": [3.902081106111, 0.3654320104818],
+    "4": [4.269986848274, 0.3537196151693],
+    "-2": [0.05918134839047, 0.2206720362165],
+    "-3": [-0.184624654125, 0.3713701578016]
+  },
+  "direct_es": {
+    "0": [2.876086144279, 0.4110153204559],
+    "1": [3.217671119032, 0.433940624349],
+    "2": [3.571778263524, 0.4571207641365],
+    "3": [3.969558831006, 0.3912243590417],
+    "4": [4.268878161094, 0.3995261701574],
+    "-2": [0.1185563879358, 0.442300060411],
+    "-3": [-0.1239894706087, 0.4328879644821]
+  },
+  "ra_cov": {
+    "0": [2.888871838355, 0.2774049784919, null],
+    "1": [3.22526891068, 0.3671167840304, null],
+    "2": [3.577950221628, 0.3920693801281, null],
+    "3": [3.973278128237, 0.363703531293, null],
+    "4": [4.265103781445, 0.3614251956151, null],
+    "-2": [0.1204444745696, 0.4103497220604, null],
+    "-3": [-0.1197046094519, 0.4026535247538, null]
+  },
+  "pooled": {
+    "post": [3.532265685012, 0.3521082333958],
+    "pre": [-0.003540304476795, 0.377308359478]
+  }
+}

benchmarks/data/lpdid_ra_coverage.json

@@ -0,0 +1,76 @@
+{
+  "study": "LPDiD RA influence-function SE coverage",
+  "nominal": 0.95,
+  "tau": {
+    "0": 1.0,
+    "1": 1.5,
+    "2": 2.0,
+    "3": 2.5
+  },
+  "note": "RA IF variance is asymptotic (teffects convention, no finite-sample factor); empirical coverage of the true effect holds near nominal across G for both the event-study horizons and the headline pooled-row RA CI - the t(G-1) reference keeps it well-calibrated even at modest G. Underwrites REGISTRY ## LPDiD Deviation 2.",
+  "sweep": [
+    {
+      "n_units": 30,
+      "reps": 500,
+      "per_horizon": {
+        "0": 0.954,
+        "1": 0.952,
+        "2": 0.94,
+        "3": 0.968
+      },
+      "mean_event_coverage": 0.9535,
+      "pooled_att_coverage": 0.968,
+      "n_valid": {
+        "0": 500,
+        "1": 500,
+        "2": 500,
+        "3": 500
+      },
+      "n_pooled_valid": 500,
+      "n_fit_errors": 0,
+      "min_valid_share": 1.0
+    },
+    {
+      "n_units": 100,
+      "reps": 500,
+      "per_horizon": {
+        "0": 0.942,
+        "1": 0.932,
+        "2": 0.944,
+        "3": 0.942
+      },
+      "mean_event_coverage": 0.94,
+      "pooled_att_coverage": 0.946,
+      "n_valid": {
+        "0": 500,
+        "1": 500,
+        "2": 500,
+        "3": 500
+      },
+      "n_pooled_valid": 500,
+      "n_fit_errors": 0,
+      "min_valid_share": 1.0
+    },
+    {
+      "n_units": 300,
+      "reps": 500,
+      "per_horizon": {
+        "0": 0.95,
+        "1": 0.948,
+        "2": 0.946,
+        "3": 0.936
+      },
+      "mean_event_coverage": 0.945,
+      "pooled_att_coverage": 0.958,
+      "n_valid": {
+        "0": 500,
+        "1": 500,
+        "2": 500,
+        "3": 500
+      },
+      "n_pooled_valid": 500,
+      "n_fit_errors": 0,
+      "min_valid_share": 1.0
+    }
+  ]
+}