Remove dead Conley scaffolding from panel estimators; align paper review

Address two P3 findings from CI Codex review of PR #411:

P3 (Maintainability) — DiD/MultiPeriodDiD/TWFE all reject vcov_type="conley"
unconditionally at fit-time, but each fit() still materialized
`_conley_coords_array` from data and forwarded `conley_coords`,
`conley_cutoff_km`, `conley_metric`, `conley_kernel` to LinearRegression /
solve_ols. Those code paths were unreachable behind the unconditional
NotImplementedError raise. Removed the dead extraction + arg-passes from
all three estimators. The constructor still accepts the conley_* kwargs
for sklearn-style API symmetry (set_params/get_params round-trip works);
they have no effect on the panel paths.

P3 (Documentation) — `docs/methodology/papers/conley-1999-review.md`
Requirements checklist and Tuning Parameters table still said the
Bartlett kernel is "PSD by construction" and only flagged uniform as
needing the negative-eigenvalue warning. Updated both surfaces to spell
out the radial 1-D pairwise specialization vs Conley's explicit 2-D
separable PSD lattice formula (Eq 3.14) and to apply the warning to
both kernels — matching the registry and the runtime contract.

271 targeted regression tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: igerber

diff_diff/estimators.py

@@ -439,13 +439,6 @@ def fit(
         # For wild bootstrap, we don't need cluster SEs from the initial fit
         cluster_ids = data[self.cluster].values if self.cluster is not None else None
 
-        # Extract Conley coords array (n×2 float64) from the user's data.
-        # Validation of the column existence and the 2-tuple shape happened
-        # at the top of fit(); here we only need to materialize the array.
-        _conley_coords_array = None
-        if self.vcov_type == "conley" and self.conley_coords is not None:
-            _conley_coords_array = data[list(self.conley_coords)].to_numpy(dtype=np.float64)
-
         # When survey PSU is present, it overrides cluster for variance estimation
         effective_cluster_ids = _resolve_effective_cluster(
             resolved_survey, cluster_ids, self.cluster
@@ -487,10 +480,6 @@ def fit(
             weight_type=survey_weight_type,
             survey_design=_lr_survey,
             vcov_type=_fit_vcov_type,
-            conley_coords=_conley_coords_array,
-            conley_cutoff_km=self.conley_cutoff_km,
-            conley_metric=self.conley_metric,
-            conley_kernel=self.conley_kernel,
         ).fit(X, y, df_adjustment=n_absorbed_effects)
 
         coefficients = reg.coefficients_
@@ -1538,13 +1527,6 @@ def fit(  # type: ignore[override]
         # Remap implicit "classical" + cluster to CR1 (legacy backward compat).
         _fit_vcov_type = self._resolve_effective_vcov_type(effective_cluster_ids)
 
-        # Extract Conley coords array (only when vcov_type='conley'; the
-        # estimator-level guards above already validated the column-name
-        # tuple against `data`).
-        _conley_coords_array_mp = None
-        if _fit_vcov_type == "conley" and self.conley_coords is not None:
-            _conley_coords_array_mp = data[list(self.conley_coords)].to_numpy(dtype=np.float64)
-
         # Note: Wild bootstrap for multi-period effects is complex (multiple coefficients)
         # For now, we use analytical inference even if inference="wild_bootstrap"
         coefficients, residuals, fitted, vcov = solve_ols(
@@ -1558,10 +1540,6 @@ def fit(  # type: ignore[override]
             weights=survey_weights,
             weight_type=survey_weight_type,
             vcov_type=_fit_vcov_type,
-            conley_coords=_conley_coords_array_mp,
-            conley_cutoff_km=self.conley_cutoff_km,
-            conley_metric=self.conley_metric,
-            conley_kernel=self.conley_kernel,
         )
 
         # Compute survey vcov if applicable

diff_diff/twfe.py

@@ -325,15 +325,6 @@ def fit(  # type: ignore[override]
         # single source of truth.
         _fit_vcov_type = self._resolve_effective_vcov_type(survey_cluster_ids)
 
-        # Materialize Conley coords from data (validated above; this is just
-        # array extraction). NOTE: data passed to LinearRegression is the
-        # within-transformed matrix, but coords are still in the ORIGINAL
-        # row order — within-transformation preserves row ordering, so the
-        # coords align with the demeaned X 1:1.
-        _twfe_conley_coords = None
-        if _fit_vcov_type == "conley" and self.conley_coords is not None:
-            _twfe_conley_coords = data[list(self.conley_coords)].to_numpy(dtype=np.float64)
-
         if self.rank_deficient_action == "error":
             reg = LinearRegression(
                 include_intercept=False,
@@ -344,10 +335,6 @@ def fit(  # type: ignore[override]
                 weight_type=survey_weight_type,
                 survey_design=_lr_survey_twfe,
                 vcov_type=_fit_vcov_type,
-                conley_coords=_twfe_conley_coords,
-                conley_cutoff_km=self.conley_cutoff_km,
-                conley_metric=self.conley_metric,
-                conley_kernel=self.conley_kernel,
             ).fit(X, y, df_adjustment=df_adjustment)
         else:
             # Suppress generic warning, TWFE provides context-specific messages below
@@ -364,10 +351,6 @@ def fit(  # type: ignore[override]
                     weight_type=survey_weight_type,
                     survey_design=_lr_survey_twfe,
                     vcov_type=_fit_vcov_type,
-                    conley_coords=_twfe_conley_coords,
-                    conley_cutoff_km=self.conley_cutoff_km,
-                    conley_metric=self.conley_metric,
-                    conley_kernel=self.conley_kernel,
                 ).fit(X, y, df_adjustment=df_adjustment)
 
         coefficients = reg.coefficients_

docs/methodology/papers/conley-1999-review.md

@@ -214,10 +214,10 @@ The paper itself does NOT distribute code. Conley's Section 5 empirical example
 - [ ] Coordinates supplied as two columns (lat, lon) or `(x, y)` projected.
 - [ ] Distance metric configured (haversine for lat/lon; euclidean for projected; callable for custom).
 - [ ] Cutoff `conley_cutoff_km > 0` (or unitless `conley_cutoff` for euclidean). Document that `h = 0` reduces to HC0.
-- [ ] Kernel choice `conley_kernel ∈ {"bartlett", "uniform"}`. Bartlett is PSD by construction; uniform is not in general (warn).
+- [ ] Kernel choice `conley_kernel ∈ {"bartlett", "uniform"}`. Conley's explicit PSD Bartlett (Eq 3.14) is the 2-D separable lattice product window; the radial 1-D pairwise Bartlett that diff-diff and R `conleyreg` implement is a practitioner specialization that is **not** formally PSD-guaranteed. Uniform is also not PSD in general. Apply the negative-eigenvalue warning to **both** kernels.
 - [ ] Score outer products `x_i ε̂_i` computed identically to HC0 path.
 - [ ] Robustness sweep: document that practitioners should report estimates at multiple cutoffs (Conley Section 5 standard).
-- [ ] If `conley_kernel="uniform"` and the resulting variance has any negative eigenvalues, warn or fall back to Bartlett.
+- [ ] If the resulting Conley meat / variance has any materially negative eigenvalues (under either Bartlett or uniform), warn the user (the implementation does this for both kernels).
 
 ---
 
@@ -242,7 +242,7 @@ The paper itself does NOT distribute code. Conley's Section 5 empirical example
 | `vcov_method` | str | `"hc0"` | Set to `"conley"` to activate. |
 | `conley_coords` | tuple of 2 str | `None` | User specifies the two column names for lat/lon (or projected x/y). Required when `vcov_method="conley"`. |
 | `conley_cutoff_km` | float | `None` (no default) | User-supplied. Conley does not provide a plug-in selector. Recommend a robustness sweep (3-5 values spanning the relevant economic-distance range). For Phase 1, error if not supplied. |
-| `conley_kernel` | str | `"bartlett"` | `"bartlett"` is PSD by construction (Conley Eq 3.14 page 12) and is the practitioner default. `"uniform"` matches Conley's "truncated window" (page 11) but may fail PSD; emit warning. |
+| `conley_kernel` | str | `"bartlett"` | `"bartlett"` evaluated on pairwise distance `d_ij/h` is the practitioner default, matching R `conleyreg` and Stata `acreg`; this radial 1-D form is a specialization of Conley's explicit 2-D separable PSD-guaranteed Bartlett (Eq 3.14, page 12) and is not formally PSD-guaranteed itself. `"uniform"` matches Conley's "truncated window" (page 11) and is also not PSD in general (footnote 11). Emit a warning under either kernel when the resulting meat has a materially negative eigenvalue. |
 | `conley_metric` | str or callable | `"haversine"` | `"haversine"` for lat/lon (km); `"euclidean"` for projected coords (units = whatever the coord units are - so if coords are degrees, cutoff is in degrees); a callable `(coord_i, coord_j) -> float` for custom metrics (e.g., travel time, network distance). |
 
 ### Relation to Existing diff-diff Estimators