Restrict bootstrap warm-start to exact column-name matches and retry on failure

Author: remlapmot

pySEQTarget/SEQuential.py

@@ -279,7 +279,10 @@ def fit(self) -> None:
             models_list = _subgroup_fit(self, start_params=start)
             if not is_boot:
                 self._outcome_start_params = {
-                    val: {key: m.params.values for key, m in sg.items()}
+                    val: {
+                        key: (m.params.values, list(m.model.exog_names))
+                        for key, m in sg.items()
+                    }
                     for val, sg in zip(self._unique_subgroups, models_list)
                 }
             return models_list
@@ -310,7 +313,10 @@ def fit(self) -> None:
             )
 
         if not is_boot:
-            self._outcome_start_params = {k: m.params.values for k, m in models.items()}
+            self._outcome_start_params = {
+                k: (m.params.values, list(m.model.exog_names))
+                for k, m in models.items()
+            }
 
         if self.offload:
             offloaded_models = {}

pySEQTarget/analysis/_outcome_fit.py

@@ -104,10 +104,22 @@ def _outcome_fit(
 
     model = smf.glm(**glm_kwargs)
 
-    # Drop warm-start coefs if the design matrix column count doesn't match
-    # (e.g. categorical level dropped or added in a bootstrap resample).
-    if start_params is not None and len(start_params) != model.exog.shape[1]:
-        start_params = None
-
-    model_fit = model.fit(start_params=start_params)
+    # Drop warm-start coefs unless the design matrix columns match exactly
+    # by name — bootstrap resamples can shift categorical reference levels or
+    # column ordering, in which case the cached coefs are meaningless and
+    # IRLS can diverge into NaN/Inf and crash LAPACK.
+    if start_params is not None:
+        sp_values, sp_names = start_params
+        if list(model.exog_names) != list(sp_names):
+            start_params = None
+        else:
+            start_params = sp_values
+
+    try:
+        model_fit = model.fit(start_params=start_params)
+    except Exception:
+        if start_params is not None:
+            model_fit = model.fit()
+        else:
+            raise
     return model_fit