Make the adjustment_model() priors= escape hatch reachable

The beta_{outcome} guard ran inside _inherit_outcome_priors, which is
called before the caller's priors= argument is merged, so the workaround
the error message advertised raised the very same error. Split the guard
into its own helper that also sees the user's overrides, so an explicit
priors={"beta_Y": ...} satisfies it and is used for the reduced equation.

Coefficient priors are still not inherited: beta is a vector over the
outcome's predictors, and a reduced equation's predictor set can differ
from the structural one. Dispersion inheritance is unchanged.

Fixes #463.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: juanitorduz

pathmc/_model.py

@@ -876,8 +876,11 @@ def adjustment_model(
         families : dict[str, str] | None
             Per-variable families for the reduced model only.
         priors : dict | None
-            Prior overrides for the reduced model, merged with inherited
-            outcome dispersion priors.
+            Prior overrides for the reduced equation, merged with the
+            outcome dispersion priors inherited from this model.
+            Coefficient priors are never inherited, since the reduced
+            predictor set can differ from the structural one — pass
+            ``beta_{outcome}`` here to set it on the reduced equation.
 
         Returns
         -------

pathmc/adjustment.py

@@ -227,20 +227,31 @@ def _validate_reduced_spec(
             )
 
 
-def _inherit_outcome_priors(
-    parent_priors: dict[str, Any],
+def _reject_uninheritable_beta_prior(
     construction_priors: dict[str, Any] | None,
+    user_priors: dict[str, Any] | None,
     outcome: str,
-) -> dict[str, Any]:
-    """Copy outcome dispersion priors from the parent; reject beta overrides."""
+) -> None:
+    """Raise when a parent beta prior would be silently replaced by defaults."""
     beta_key = f"beta_{outcome}"
-    if construction_priors and beta_key in construction_priors:
-        raise ValueError(
-            f"The structural model has a custom prior on '{beta_key}'. "
-            f"Adjustment models do not inherit coefficient priors. Pass "
-            f"priors= on adjustment_model() for the reduced equation."
-        )
+    if not construction_priors or beta_key not in construction_priors:
+        return
+    if user_priors is not None and beta_key in user_priors:
+        return
+    raise ValueError(
+        f"The structural model has a custom prior on '{beta_key}'. "
+        f"Adjustment models do not inherit coefficient priors because the "
+        f"reduced predictor set can differ. Pass "
+        f"priors={{'{beta_key}': ...}} to adjustment_model() to set it on "
+        f"the reduced equation."
+    )
+
 
+def _inherit_outcome_priors(
+    parent_priors: dict[str, Any],
+    outcome: str,
+) -> dict[str, Any]:
+    """Copy outcome dispersion priors from the parent."""
     inherited: dict[str, Any] = {}
     for suffix in _OUTCOME_PRIOR_SUFFIXES:
         key = f"{suffix}_{outcome}"
@@ -395,11 +406,12 @@ def from_path_model(
         construction_priors = (
             parent._construction.get("priors") if parent._construction else None
         )
-        inherited_priors = _inherit_outcome_priors(
-            parent._priors,
+        _reject_uninheritable_beta_prior(
             construction_priors,
+            priors,
             outcome_name,
         )
+        inherited_priors = _inherit_outcome_priors(parent._priors, outcome_name)
         reduced_defaults = default_priors(
             reduced_spec,
             families=outcome_families or None,

pathmc/skills/pathmc/SKILL.md

@@ -152,6 +152,10 @@ The DSL is lavaan-inspired:
    sets exist, pass `adjustment_set=` explicitly; pathmc does not pick
    among them. Pass `data=` when the parent structural model is
    data-free. Panel models are not supported on the adjustment path.
+   Outcome dispersion priors (`sigma_Y`, `nu_Y`, `alpha_disp_Y`) are
+   inherited from the parent, but coefficient priors are not, because
+   the reduced predictor set can differ: if the parent set a custom
+   `beta_Y`, pass `priors={"beta_Y": ...}` to `adjustment_model()`.
 10. **`predictions()` / `comparisons()` / `slopes()` share one API on
    `PathModel` and `AdjustmentModel`.** On the structural model they
    use truncated-factorization g-computation; on

tests/test_adjustment.py

@@ -250,6 +250,52 @@ def test_user_priors_on_adjustment_model(self, rng):
             == 3.0
         )
 
+    def test_parent_beta_override_satisfied_by_user_priors(self, rng):
+        df = _fork_df(rng)
+        model = pathmc.model(
+            "X ~ Z\nY ~ X + Z",
+            data=df,
+            priors={"beta_Y": Prior("Normal", mu=0, sigma=2.0)},
+        )
+        adjusted = model.adjustment_model(
+            "X -> Y",
+            priors={"beta_Y": Prior("Normal", mu=0, sigma=2.0)},
+        )
+        beta = adjusted.outcome_model._priors["beta_Y"].to_dict()
+        assert beta["dist"] == "Normal"
+        assert beta["kwargs"]["sigma"] == 2.0
+
+    def test_parent_beta_override_without_user_beta_still_raises(self, rng):
+        df = _fork_df(rng)
+        model = pathmc.model(
+            "X ~ Z\nY ~ X + Z",
+            data=df,
+            priors={"beta_Y": Prior("Normal", mu=0, sigma=2.0)},
+        )
+        with pytest.raises(ValueError, match="beta_Y"):
+            model.adjustment_model(
+                "X -> Y",
+                priors={"sigma_Y": Prior("HalfNormal", sigma=3.0)},
+            )
+
+    def test_user_beta_override_keeps_inherited_dispersion(self, rng):
+        df = _fork_df(rng)
+        model = pathmc.model(
+            "X ~ Z\nY ~ X + Z",
+            data=df,
+            priors={
+                "beta_Y": Prior("Normal", mu=0, sigma=2.0),
+                "sigma_Y": Prior("HalfNormal", sigma=4.0),
+            },
+        )
+        adjusted = model.adjustment_model(
+            "X -> Y",
+            priors={"beta_Y": Prior("Normal", mu=0, sigma=5.0)},
+        )
+        priors = adjusted.outcome_model._priors
+        assert priors["beta_Y"].to_dict()["kwargs"]["sigma"] == 5.0
+        assert priors["sigma_Y"].to_dict()["kwargs"]["sigma"] == 4.0
+
 
 class TestInnerModelTypes:
     def test_inner_is_pathmodel_with_pymc_model(self, rng):