BUG: DataFrame.stack losing EA dtypes with multi-columns and mixed dtypes

doc/source/whatsnew/v2.1.0.rst

@@ -191,7 +191,7 @@ Groupby/resample/rolling
 
 Reshaping
 ^^^^^^^^^
--
+- Bug in :meth:`DataFrame.stack` losing extension dtypes when columns is a :class:`MultiIndex` and frame contains mixed dtypes (:issue:`45740`)
 -
 
 Sparse

pandas/core/reshape/reshape.py

@@ -15,7 +15,10 @@
 from pandas.util._decorators import cache_readonly
 from pandas.util._exceptions import find_stack_level
 
-from pandas.core.dtypes.cast import maybe_promote
+from pandas.core.dtypes.cast import (
+    find_common_type,
+    maybe_promote,
+)
 from pandas.core.dtypes.common import (
     ensure_platform_int,
     is_1d_only_ea_dtype,
@@ -745,25 +748,19 @@ def _convert_level_number(level_num: int, columns: Index):
             chunk.columns = level_vals_nan.take(chunk.columns.codes[-1])
             value_slice = chunk.reindex(columns=level_vals_used).values
         else:
-            if frame._is_homogeneous_type and is_extension_array_dtype(
-                frame.dtypes.iloc[0]
-            ):
+            subset = this[this.columns[loc]]
+            dtype = find_common_type(subset.dtypes.tolist())
+            if is_extension_array_dtype(dtype):
                 # TODO(EA2D): won't need special case, can go through .values
                 #  paths below (might change to ._values)
-                dtype = this[this.columns[loc]].dtypes.iloc[0]
-                subset = this[this.columns[loc]]
-
                 value_slice = dtype.construct_array_type()._concat_same_type(
-                    [x._values for _, x in subset.items()]
+                    [x._values.astype(dtype, copy=False) for _, x in subset.items()]
                 )
                 N, K = subset.shape
                 idx = np.arange(N * K).reshape(K, N).T.ravel()
                 value_slice = value_slice.take(idx)
-
-            elif frame._is_mixed_type:
-                value_slice = this[this.columns[loc]].values
             else:
-                value_slice = this.values[:, loc]
+                value_slice = subset.values
 
         if value_slice.ndim > 1:
             # i.e. not extension

pandas/tests/frame/test_stack_unstack.py

@@ -1146,6 +1146,27 @@ def test_stack_multi_columns_non_unique_index(self, index, columns):
         expected_codes = np.asarray(new_index.codes)
         tm.assert_numpy_array_equal(stacked_codes, expected_codes)
 
+    @pytest.mark.parametrize(
+        "vals1, vals2, dtype1, dtype2, expected_dtype",
+        [
+            ([1, 2], [3.0, 4.0], "Int64", "Float64", "Float64"),
+            ([1, 2], ["foo", "bar"], "Int64", "string", "object"),
+        ],
+    )
+    def test_stack_multi_columns_mixed_extension_types(
+        self, vals1, vals2, dtype1, dtype2, expected_dtype
+    ):
+        # GH45740
+        df = DataFrame(
+            {
+                ("A", 1): Series(vals1, dtype=dtype1),
+                ("A", 2): Series(vals2, dtype=dtype2),
+            }
+        )
+        result = df.stack()
+        expected = df.astype(object).stack().astype(expected_dtype)
+        tm.assert_frame_equal(result, expected)
+
     @pytest.mark.parametrize("level", [0, 1])
     def test_unstack_mixed_extension_types(self, level):
         index = MultiIndex.from_tuples([("A", 0), ("A", 1), ("B", 1)], names=["a", "b"])
@@ -2181,9 +2202,11 @@ def test_stack_nullable_dtype(self):
         df[df.columns[0]] = df[df.columns[0]].astype(pd.Float64Dtype())
         result = df.stack("station")
 
-        # TODO(EA2D): we get object dtype because DataFrame.values can't
-        #  be an EA
-        expected = df.astype(object).stack("station")
+        expected = (
+            df.astype(object)
+            .stack("station")
+            .astype({"r": pd.Float64Dtype(), "t_mean": pd.Int64Dtype()})
+        )
         tm.assert_frame_equal(result, expected)
 
     def test_unstack_mixed_level_names(self):