Use positive-weight count for HC1/classical df with zero-weight rows

When pweight/aweight fits have zero-weight rows (from subpopulation),
use np.count_nonzero(weights > 0) for HC1 df adjustments instead of
total n. Zero-weight rows contribute nothing to the sandwich and
should not inflate df.

Also fix clustered SEs to exclude zero-total-weight clusters from
G/(G-1) small-sample correction, and DEFF effective_n to use
positive-weight count.

Survey design df (n_PSU - n_strata) unchanged — preserves design
structure matching R's survey::degf() convention.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: igerber

diff_diff/linalg.py

@@ -1046,10 +1046,16 @@ def _compute_robust_vcov_numpy(
     else:
         bread_matrix = X.T @ X
 
-    # Effective n for df computation (fweights use sum(w))
+    # Effective n for df computation
+    # fweights: sum(w) (frequency expansion)
+    # pweight/aweight with zeros: positive-weight count (zero-weight rows
+    # contribute nothing to the sandwich and should not inflate df)
     n_eff = n
-    if weights is not None and weight_type == "fweight":
-        n_eff = int(round(np.sum(weights)))
+    if weights is not None:
+        if weight_type == "fweight":
+            n_eff = int(round(np.sum(weights)))
+        elif np.any(weights == 0):
+            n_eff = int(np.count_nonzero(weights > 0))
 
     # Compute weighted scores for cluster-robust meat (outer product of sums).
     # pweight/fweight multiply by w; aweight and unweighted use raw residuals.
@@ -1075,6 +1081,11 @@ def _compute_robust_vcov_numpy(
         unique_clusters = np.unique(cluster_ids)
         n_clusters = len(unique_clusters)
 
+        # Exclude clusters with zero total weight (subpopulation-zeroed)
+        if weights is not None and weight_type != "fweight" and np.any(weights == 0):
+            cluster_weights = pd.Series(weights).groupby(cluster_ids).sum()
+            n_clusters = int((cluster_weights > 0).sum())
+
         if n_clusters < 2:
             raise ValueError(f"Need at least 2 clusters for cluster-robust SEs, got {n_clusters}")
 
@@ -1741,10 +1752,14 @@ def fit(
             nan_mask = np.isnan(coefficients)
             k_effective = k - np.sum(nan_mask)  # Number of identified coefficients
 
-            # For fweights, df uses sum(w) - k (effective sample size)
+            # Effective n for df: fweights use sum(w), pweight/aweight with
+            # zeros use positive-weight count (zero-weight rows don't contribute)
             n_eff_df = n
-            if self.weights is not None and self.weight_type == "fweight":
-                n_eff_df = int(round(np.sum(self.weights)))
+            if self.weights is not None:
+                if self.weight_type == "fweight":
+                    n_eff_df = int(round(np.sum(self.weights)))
+                elif np.any(self.weights == 0):
+                    n_eff_df = int(np.count_nonzero(self.weights > 0))
 
             if k_effective == 0:
                 # All coefficients dropped - no valid inference

diff_diff/survey.py

@@ -849,6 +849,9 @@ def compute_deff_diagnostics(
     from diff_diff.linalg import compute_robust_vcov
 
     n = X.shape[0]
+    # Use positive-weight count for effective n (zero-weight rows from
+    # subpopulation don't contribute to the effective sample)
+    n_eff = int(np.count_nonzero(weights > 0)) if np.any(weights == 0) else n
 
     # SRS baseline: HC1 weighted sandwich ignoring design structure
     srs_vcov = compute_robust_vcov(
@@ -861,7 +864,7 @@ def compute_deff_diagnostics(
     # DEFF = survey_var / srs_var
     with np.errstate(divide="ignore", invalid="ignore"):
         deff = np.where(srs_var > 0, survey_var / srs_var, np.nan)
-        eff_n = np.where(deff > 0, n / deff, np.nan)
+        eff_n = np.where(deff > 0, n_eff / deff, np.nan)
 
     survey_se = np.sqrt(np.maximum(survey_var, 0.0))
     srs_se = np.sqrt(np.maximum(srs_var, 0.0))

docs/methodology/REGISTRY.md

@@ -2093,11 +2093,12 @@ Domain estimation preserving full design structure.
 - **Note:** Weight validation relaxed from "strictly positive" to
   "non-negative" to support zero-weight observations. Negative weights
   still rejected. All-zero weight vectors rejected at solver level.
-- **Note:** Survey df and variance adjustment factors use total n
-  (including zero-weight rows), matching R's `survey::degf()` convention
-  after `subset()`. This preserves the design structure for correct
-  variance estimation. Zero-weight rows contribute zero-valued scores
-  to the sandwich meat but are counted in df = n_PSU - n_strata.
+- **Note:** Survey design df (`n_PSU - n_strata`) uses total design
+  structure (including zero-weight rows), matching R's `survey::degf()`
+  convention after `subset()`. The generic HC1/classical inference paths
+  use positive-weight count for df adjustments, ensuring zero-weight
+  padding is inference-invariant outside the survey vcov path. DEFF
+  effective-n also uses positive-weight count.
 - **Note:** For replicate-weight designs, `subpopulation()` zeros out both
   full-sample and replicate weight columns for excluded observations,
   preserving all replicate metadata.