Fix bug in combine first with string dtype and NA only in one level of index (pandas-dev#37568)

Author: phofl

doc/source/whatsnew/v1.2.0.rst

@@ -531,7 +531,7 @@ Reshaping
 - Bug in :meth:`Series.transform` would give incorrect results or raise when the argument ``func`` was dictionary (:issue:`35811`)
 - Bug in :meth:`DataFrame.pivot` did not preserve :class:`MultiIndex` level names for columns when rows and columns both multiindexed (:issue:`36360`)
 - Bug in :func:`join` returned a non deterministic level-order for the resulting :class:`MultiIndex` (:issue:`36910`)
--
+- Bug in :meth:`DataFrame.combine_first()` caused wrong alignment with dtype ``string`` and one level of ``MultiIndex`` containing only ``NA`` (:issue:`37591`)
 
 Sparse
 ^^^^^^

pandas/core/arrays/categorical.py

@@ -322,7 +322,7 @@ def __init__(
             # sanitize_array coerces np.nan to a string under certain versions
             # of numpy
             values = maybe_infer_to_datetimelike(values, convert_dates=True)
-            if not isinstance(values, np.ndarray):
+            if not isinstance(values, (np.ndarray, ExtensionArray)):
                 values = com.convert_to_list_like(values)
 
                 # By convention, empty lists result in object dtype:

pandas/tests/frame/methods/test_combine_first.py

@@ -353,3 +353,15 @@ def test_combine_first_with_asymmetric_other(self, val):
         exp = DataFrame({"isBool": [True], "isNum": [val]})
 
         tm.assert_frame_equal(res, exp)
+
+    def test_combine_first_string_dtype_only_na(self):
+        # GH: 37519
+        df = DataFrame({"a": ["962", "85"], "b": [pd.NA] * 2}, dtype="string")
+        df2 = DataFrame({"a": ["85"], "b": [pd.NA]}, dtype="string")
+        df.set_index(["a", "b"], inplace=True)
+        df2.set_index(["a", "b"], inplace=True)
+        result = df.combine_first(df2)
+        expected = DataFrame(
+            {"a": ["962", "85"], "b": [pd.NA] * 2}, dtype="string"
+        ).set_index(["a", "b"])
+        tm.assert_frame_equal(result, expected)