Normalize weights in SRS fallback for scale invariance

Both SRS fallback branches now normalize positive weights to mean=1
before computing variance, ensuring SEs are invariant to constant
weight rescaling (important for replicate designs that preserve raw
weight scale). Add scale-invariance regression test with 5x rescaling.

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

Author: igerber

diff_diff/prep.py

@@ -1373,10 +1373,17 @@ def _cell_mean_variance(
     y_bar = float(np.sum(w_valid * y_clean) / sum_w)
 
     # SRS fallback if below min_n threshold
+    # Normalize positive weights to mean=1 so fallback is scale-invariant
+    # (replicate designs preserve raw weight scale per survey.py:L189-240)
     used_srs = False
     if n_valid < min_n:
-        resid_sq = w_valid * (y_clean - y_bar) ** 2
-        variance = float(np.sum(resid_sq) / (sum_w**2) * n_valid / (n_valid - 1))
+        w_norm = w_valid.copy()
+        w_pos = w_norm[w_norm > 0]
+        if len(w_pos) > 0:
+            w_norm[w_norm > 0] = w_pos / w_pos.mean()
+        sum_wn = float(np.sum(w_norm))
+        resid_sq = w_norm * (y_clean - y_bar) ** 2
+        variance = float(np.sum(resid_sq) / (sum_wn**2) * n_valid / (n_valid - 1))
         return y_bar, max(variance, 0.0), n_valid, True
 
     # Full-design domain estimation: construct full-length psi with zeros
@@ -1396,8 +1403,13 @@ def _cell_mean_variance(
 
     # SRS fallback when design-based variance is unidentifiable
     if np.isnan(variance):
-        resid_sq = w_valid * (y_clean - y_bar) ** 2
-        variance = float(np.sum(resid_sq) / (sum_w**2) * n_valid / (n_valid - 1))
+        w_norm = w_valid.copy()
+        w_pos = w_norm[w_norm > 0]
+        if len(w_pos) > 0:
+            w_norm[w_norm > 0] = w_pos / w_pos.mean()
+        sum_wn = float(np.sum(w_norm))
+        resid_sq = w_norm * (y_clean - y_bar) ** 2
+        variance = float(np.sum(resid_sq) / (sum_wn**2) * n_valid / (n_valid - 1))
         used_srs = True
 
     return y_bar, max(float(variance), 0.0), n_valid, used_srs

tests/test_prep.py

@@ -2855,3 +2855,43 @@ def test_replicate_weight_min_n_fallback(self):
             panel_rep["outcome_se"].values,
             rtol=1e-6,
         )
+
+    def test_srs_fallback_scale_invariant(self):
+        """SRS fallback SEs are invariant to constant weight rescaling."""
+        rng = np.random.RandomState(55)
+        n = 60
+        data = pd.DataFrame(
+            {
+                "geo": np.repeat(["A", "B", "C"], n // 3),
+                "time": np.ones(n, dtype=int),
+                "wt": rng.uniform(0.5, 2.0, n),
+                "y": rng.normal(10, 2, n),
+            }
+        )
+        design1 = SurveyDesign(weights="wt")
+
+        # Force SRS fallback with high min_n
+        with pytest.warns(UserWarning, match="SRS fallback"):
+            panel1, _ = aggregate_survey(
+                data,
+                by=["geo", "time"],
+                outcomes="y",
+                survey_design=design1,
+                min_n=9999,
+            )
+
+        # Rescale weights by 5x → should give identical SEs
+        data2 = data.copy()
+        data2["wt"] = data2["wt"] * 5.0
+        design2 = SurveyDesign(weights="wt")
+        with pytest.warns(UserWarning, match="SRS fallback"):
+            panel2, _ = aggregate_survey(
+                data2,
+                by=["geo", "time"],
+                outcomes="y",
+                survey_design=design2,
+                min_n=9999,
+            )
+
+        np.testing.assert_allclose(panel1["y_se"].values, panel2["y_se"].values, rtol=1e-10)
+        np.testing.assert_allclose(panel1["y_mean"].values, panel2["y_mean"].values, rtol=1e-10)