Address P2/P3: method-specific fallback warnings, docstring wording

DR fallback warnings now say propensity model is unconditional while
outcome regression still uses covariates, instead of misleading "all
covariates dropped" text. IPW warnings unchanged.

Update REGISTRY.md fallback description to distinguish IPW vs DR
behavior. Fix docstrings to say "predictor variables (excluding
intercept)" consistently.

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

Author: igerber

diff_diff/linalg.py

@@ -1170,7 +1170,8 @@ def solve_logit(
         regression.
     epv_threshold : float, default 10
         Events Per Variable threshold. When the ratio of minority-class
-        observations to parameters falls below this value, a warning is
+        observations to predictor variables (excluding intercept) falls
+        below this value, a warning is
         emitted (or ValueError raised if ``rank_deficient_action="error"``).
         Based on Peduzzi et al. (1996).
     context_label : str, default ""

diff_diff/staggered.py

@@ -192,10 +192,11 @@ class CallawaySantAnna(
         functions.
     epv_threshold : float, default=10
         Events Per Variable threshold for propensity score logit.
-        When the ratio of minority-class observations to parameters
-        falls below this value, a warning is emitted (or ``ValueError``
-        raised if ``rank_deficient_action="error"``). Based on Peduzzi
-        et al. (1996). Only applies to IPW and DR estimation methods.
+        When the ratio of minority-class observations to predictor
+        variables (excluding intercept) falls below this value, a
+        warning is emitted (or ``ValueError`` raised if
+        ``rank_deficient_action="error"``). Based on Peduzzi et al.
+        (1996). Only applies to IPW and DR estimation methods.
         Use ``diagnose_propensity()`` for a pre-estimation check across
         all cohorts.
     pscore_fallback : str, default="error"
@@ -2443,7 +2444,8 @@ def _doubly_robust(
                     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). "
                         f"Consider estimation_method='reg' to avoid "
                         f"propensity scores entirely.",
                         UserWarning,
@@ -3519,7 +3521,8 @@ def _doubly_robust_rc(
             warnings.warn(
                 f"Propensity score estimation failed{ctx} (RCS DR). "
                 f"Falling back to unconditional propensity "
-                f"(all covariates dropped for this cell). "
+                f"(propensity model ignores covariates; outcome "
+                f"regression still uses them). "
                 f"Consider estimation_method='reg' to avoid "
                 f"propensity scores entirely.",
                 UserWarning,

diff_diff/triple_diff.py

@@ -389,10 +389,11 @@ class TripleDifference:
         - "silent": Drop columns silently without warning
     epv_threshold : float, default=10
         Events Per Variable threshold for propensity score logit.
-        When the ratio of minority-class observations to parameters
-        falls below this value, a warning is emitted (or ``ValueError``
-        raised if ``rank_deficient_action="error"``). Based on Peduzzi
-        et al. (1996). Only applies to IPW and DR estimation methods.
+        When the ratio of minority-class observations to predictor
+        variables (excluding intercept) falls below this value, a
+        warning is emitted (or ``ValueError`` raised if
+        ``rank_deficient_action="error"``). Based on Peduzzi et al.
+        (1996). Only applies to IPW and DR estimation methods.
     pscore_fallback : str, default="error"
         Action when propensity score estimation fails:
         - "error": Raise the exception (default)

docs/methodology/REGISTRY.md

@@ -412,7 +412,9 @@ The multiplier bootstrap uses random weights w_i with E[w]=0 and Var(w)=1:
   - Fallback: Controlled by `pscore_fallback` parameter (default `"error"`).
     If IRLS fails entirely (LinAlgError/ValueError) and `pscore_fallback="error"`,
     the error is raised. If `pscore_fallback="unconditional"`, falls back to
-    unconditional propensity score (all covariates dropped) with warning.
+    unconditional propensity score with warning. For IPW, this effectively
+    drops all covariates. For DR, the propensity model is unconditional but
+    the outcome-regression component still uses covariates.
   - **Note:** `pscore_fallback` default changed from unconditional to error.
     Set `pscore_fallback="unconditional"` for legacy behavior.
   - **Note:** When `pscore_fallback="unconditional"` triggers, the propensity-

tests/test_staggered.py

@@ -3511,7 +3511,7 @@ def test_dr_fallback_warning(self):
                 )
 
             fallback_warns = [
-                x for x in w if "covariates dropped" in str(x.message)
+                x for x in w if "unconditional propensity" in str(x.message)
             ]
             assert len(fallback_warns) > 0, "Expected fallback warning in DR path"
             assert results.overall_att is not None
@@ -3708,7 +3708,7 @@ def test_cs_pscore_fallback_unconditional_opt_in(self):
                     covariates=["x1"],
                 )
             fallback_warns = [
-                x for x in w if "covariates dropped" in str(x.message)
+                x for x in w if "unconditional propensity" in str(x.message)
             ]
             assert len(fallback_warns) > 0
             assert results.overall_att is not None