Address PR #95 follow-up: fix bootstrap p-value floor and document policy

- Fix bootstrap p-value floor to use valid sample count (not n_bootstrap)
- Remove unused _mask_nonfinite_samples method (dead code)
- Document non-finite inference handling in Methodology Registry
- Fix misleading test docstring that claimed undocumented behavior

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Author: igerber

diff_diff/staggered_bootstrap.py

@@ -688,36 +688,17 @@ def _compute_effect_bootstrap_stats(
 
         # Use only valid samples
         valid_dist = boot_dist[finite_mask]
+        n_valid_bootstrap = len(valid_dist)
 
         se = float(np.std(valid_dist, ddof=1))
         ci = self._compute_percentile_ci(valid_dist, self.alpha)
-        p_value = self._compute_bootstrap_pvalue(original_effect, valid_dist)
-        return se, ci, p_value
 
-    def _mask_nonfinite_samples(self, arr: np.ndarray, context: str) -> np.ndarray:
-        """Return boolean mask of finite samples, warning if any dropped.
-
-        Parameters
-        ----------
-        arr : np.ndarray
-            Array to check (1D bootstrap distribution).
-        context : str
-            Description of where this check is happening (for warning message).
+        # Compute p-value inline with correct floor based on valid sample count
+        if original_effect >= 0:
+            p_one_sided = np.mean(valid_dist <= 0)
+        else:
+            p_one_sided = np.mean(valid_dist >= 0)
+        p_value = min(2 * p_one_sided, 1.0)
+        p_value = max(p_value, 1 / (n_valid_bootstrap + 1))  # Floor uses valid count
 
-        Returns
-        -------
-        np.ndarray
-            Boolean mask where True indicates finite (valid) samples.
-        """
-        finite_mask = np.isfinite(arr)
-        if not np.all(finite_mask):
-            import warnings
-            n_nonfinite = np.sum(~finite_mask)
-            warnings.warn(
-                f"Dropping {n_nonfinite}/{arr.size} non-finite bootstrap samples in {context}. "
-                "This may occur with very small samples or extreme weights. "
-                "Bootstrap estimates based on remaining valid samples.",
-                RuntimeWarning,
-                stacklevel=3
-            )
-        return finite_mask
+        return se, ci, float(p_value)

docs/methodology/REGISTRY.md

@@ -204,6 +204,10 @@ Aggregations:
   - Detection: Pivoted QR decomposition with tolerance `1e-07` (R's `qr()` default)
   - Handling: Warns and drops linearly dependent columns, sets NA for dropped coefficients (R-style, matches `lm()`)
   - Parameter: `rank_deficient_action` controls behavior: "warn" (default), "error", or "silent"
+- Non-finite inference values:
+  - Analytic SE: Returns NaN to signal invalid inference (not biased via zeroing)
+  - Bootstrap: Drops non-finite samples, warns, and adjusts p-value floor accordingly
+  - Threshold: Returns NaN if <50% of bootstrap samples are valid
 
 **Reference implementation(s):**
 - R: `did::att_gt()` (Callaway & Sant'Anna's official package)

tests/test_staggered.py

@@ -684,10 +684,10 @@ def test_extreme_propensity_scores(self):
     def test_extreme_weights_warning(self):
         """Test that extreme weights produce warnings and methodology-aligned behavior.
 
-        Per the Methodology Registry (docs/methodology/REGISTRY.md):
-        - Missing group-time cells: ATT(g,t) set to NaN
-        - Analytic SE: returns NaN to signal invalid inference (not biased via zeroing)
-        - Bootstrap: drops invalid samples and warns, preserving valid distribution
+        Tests that:
+        - ATT point estimates remain finite
+        - SE is finite (valid) or NaN (signals invalid inference), never biased
+        - Bootstrap drops invalid samples and adjusts inference accordingly
         """
         import warnings
         np.random.seed(42)