Fix remaining DR fallback text, update test comments for Peduzzi convention

Make all fallback warnings and docstrings method-specific: IPW says
covariates dropped, DR says propensity model unconditional while outcome
regression still uses covariates. Update test comments from old
intercept-inclusive arithmetic to predictor-variable counts.

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

Author: igerber

diff_diff/staggered.py

@@ -205,8 +205,9 @@ class CallawaySantAnna(
         - "error": Raise the exception (default). Ensures the user is
           aware of estimation failures.
         - "unconditional": Fall back to unconditional propensity
-          ``n_treated / (n_treated + n_control)`` with a warning. This
-          effectively drops all covariates for the affected cell.
+          with a warning. For IPW, this drops all covariates. For DR,
+          the propensity model becomes unconditional but outcome
+          regression still uses covariates.
         When ``rank_deficient_action="error"``, errors are always
         re-raised regardless of this setting.
 

diff_diff/staggered_triple_diff.py

@@ -1389,7 +1389,8 @@ def _compute_pscore(
                 warnings.warn(
                     f"Propensity score estimation failed{ctx}. "
                     f"Falling back to unconditional propensity "
-                    f"(all covariates dropped for this cell). "
+                    f"(propensity model ignores covariates; outcome "
+                    f"regression still uses them for DR). "
                     f"Consider estimation_method='reg' to avoid "
                     f"propensity scores entirely.",
                     UserWarning,

diff_diff/triple_diff.py

@@ -398,7 +398,9 @@ class TripleDifference:
         Action when propensity score estimation fails:
         - "error": Raise the exception (default)
         - "unconditional": Fall back to unconditional propensity with
-          a warning (drops all covariates for the affected comparison)
+          a warning. For IPW, drops all covariates. For DR, the
+          propensity model becomes unconditional but outcome regression
+          still uses covariates.
         When ``rank_deficient_action="error"``, errors are always
         re-raised regardless of this setting.
 
@@ -1040,8 +1042,8 @@ def _estimate_ddd_decomposition(
                         ps_estimated = False
                         warnings.warn(
                             f"Propensity score estimation failed for subgroup "
-                            f"{j} vs 4; dropping covariates and using "
-                            f"unconditional probability. "
+                            f"{j} vs 4; using unconditional probability. "
+                            f"For DR, outcome regression still uses covariates. "
                             f"Consider estimation_method='reg' to avoid propensity "
                             f"scores entirely.",
                             UserWarning,

tests/test_linalg.py

@@ -1683,7 +1683,7 @@ def test_epv_warning_below_threshold(self):
         from diff_diff.linalg import solve_logit
 
         rng = np.random.default_rng(42)
-        # 40 events (minority class), 9 params (8 covariates + intercept) → EPV = 4.4
+        # 40 events (minority class), 8 predictor variables → EPV = 5.0
         n = 200
         X = rng.standard_normal((n, 8))
         y = np.concatenate([np.ones(40), np.zeros(n - 40)])
@@ -1695,7 +1695,7 @@ def test_epv_no_warning_above_threshold(self):
         from diff_diff.linalg import solve_logit
 
         rng = np.random.default_rng(42)
-        # 100 events, 3 params (2 covariates + intercept) → EPV = 33.3
+        # 100 events, 2 predictor variables → EPV = 50
         n = 200
         X = rng.standard_normal((n, 2))
         y = np.concatenate([np.ones(100), np.zeros(100)])
@@ -1725,7 +1725,7 @@ def test_epv_threshold_configurable(self):
         rng = np.random.default_rng(42)
         n = 200
         X = rng.standard_normal((n, 2))
-        # 15 events, 3 params → EPV = 5.0
+        # 15 events, 2 predictor variables → EPV = 7.5
         y = np.concatenate([np.ones(15), np.zeros(n - 15)])
 
         # Default threshold 10 → should warn (EPV=5)

tests/test_staggered.py

@@ -3760,7 +3760,7 @@ def test_cs_diagnose_propensity_identifies_critical(self):
             data, outcome="outcome", unit="unit", time="time",
             first_treat="first_treat", covariates=["x1", "x2", "x3"],
         )
-        # With 1 treated unit and 3 covariates: EPV = 1/4 = 0.25 → critical
+        # With 1 treated unit and 3 predictor variables: EPV = 1/3 ≈ 0.33 → critical
         assert any(df["status"] == "critical")
 
     def test_cs_epv_in_summary_output(self):