Honor rank_deficient_action="error" in propensity score paths

Re-raise ValueError in IPW/DR except blocks when rank_deficient_action
is "error" instead of silently falling back to unconditional estimation.
Applies to CallawaySantAnna and TripleDifference PS paths.

Add estimator-level regression tests asserting ValueError propagation
with collinear covariates under error mode.

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

Author: igerber

diff_diff/staggered.py

@@ -1489,6 +1489,8 @@ def _ipw_estimation(
                     if pscore_cache is not None and pscore_key is not None:
                         pscore_cache[pscore_key] = beta_logistic
                 except (np.linalg.LinAlgError, ValueError):
+                    if self.rank_deficient_action == "error":
+                        raise
                     # Fallback to unconditional if logistic regression fails
                     warnings.warn(
                         "Propensity score estimation failed. "
@@ -1662,6 +1664,8 @@ def _doubly_robust(
                     if pscore_cache is not None and pscore_key is not None:
                         pscore_cache[pscore_key] = beta_logistic
                 except (np.linalg.LinAlgError, ValueError):
+                    if self.rank_deficient_action == "error":
+                        raise
                     # Fallback to unconditional if logistic regression fails
                     warnings.warn(
                         "Propensity score estimation failed. "

diff_diff/triple_diff.py

@@ -772,6 +772,8 @@ def _estimate_ddd_decomposition(
                             rank_deficient_action=self.rank_deficient_action,
                         )
                     except Exception:
+                        if self.rank_deficient_action == "error":
+                            raise
                         pscore_sub = np.full(n_sub, np.mean(PA4))
                         ps_estimated = False
                         warnings.warn(

tests/test_methodology_triple_diff.py

@@ -1556,3 +1556,24 @@ def test_r_squared_respects_rank_deficient_action(self):
         # Both should produce finite results regardless
         assert np.isfinite(result_silent.att)
         assert np.isfinite(result_warn.att)
+
+    @pytest.mark.parametrize("method", ["ipw", "dr"])
+    def test_rank_deficient_action_error_raises_in_ps_path(self, method):
+        """rank_deficient_action='error' raises ValueError in PS-based paths with collinear covariates."""
+        data = generate_ddd_data(n_per_cell=50, seed=42, add_covariates=True)
+        data["age_dup"] = data["age"].copy()
+
+        ddd = TripleDifference(
+            estimation_method=method,
+            rank_deficient_action="error",
+        )
+
+        with pytest.raises((ValueError, RuntimeError)):
+            ddd.fit(
+                data,
+                outcome="outcome",
+                group="group",
+                partition="partition",
+                time="time",
+                covariates=["age", "age_dup"],
+            )

tests/test_staggered.py

@@ -1275,6 +1275,46 @@ def test_reg_nyt_rank_deficient_action_warn(self):
         assert results.overall_att is not None
         assert results.overall_se > 0
 
+    def test_ipw_rank_deficient_action_error_raises(self):
+        """IPW path raises ValueError with rank_deficient_action='error' and collinear covariates."""
+        data = generate_staggered_data_with_covariates(seed=42)
+        data["x1_dup"] = data["x1"].copy()
+
+        cs = CallawaySantAnna(
+            estimation_method="ipw",
+            rank_deficient_action="error",
+        )
+
+        with pytest.raises(ValueError, match="[Rr]ank"):
+            cs.fit(
+                data,
+                outcome="outcome",
+                unit="unit",
+                time="time",
+                first_treat="first_treat",
+                covariates=["x1", "x1_dup"],
+            )
+
+    def test_dr_rank_deficient_action_error_raises(self):
+        """DR path raises ValueError with rank_deficient_action='error' and collinear covariates."""
+        data = generate_staggered_data_with_covariates(seed=42)
+        data["x1_dup"] = data["x1"].copy()
+
+        cs = CallawaySantAnna(
+            estimation_method="dr",
+            rank_deficient_action="error",
+        )
+
+        with pytest.raises(ValueError, match="[Rr]ank"):
+            cs.fit(
+                data,
+                outcome="outcome",
+                unit="unit",
+                time="time",
+                first_treat="first_treat",
+                covariates=["x1", "x1_dup"],
+            )
+
     def test_bootstrap_single_unit_cohort_handles_gracefully(self, ci_params):
         """Test that bootstrap handles cohort with 1 treated unit without crashing."""
         # Build small dataset where one cohort has exactly 1 unit