BUG: df.__setitem__ casting ea to object when setting with scalar key and frame value

doc/source/whatsnew/v1.5.0.rst

@@ -968,6 +968,7 @@ Indexing
 - Bug in :meth:`Index.reindex` raising ``AssertionError`` when ``level`` was specified but no :class:`MultiIndex` was given; level is ignored now (:issue:`35132`)
 - Bug when setting a value too large for a :class:`Series` dtype failing to coerce to a common type (:issue:`26049`, :issue:`32878`)
 - Bug in :meth:`loc.__setitem__` treating ``range`` keys as positional instead of label-based (:issue:`45479`)
+- Bug in :meth:`DataFrame.__setitem__` casting extension array dtypes to object when setting with a scalar key and :class:`DataFrame` as value (:issue:`46896`)
 - Bug in :meth:`Series.__setitem__` when setting ``boolean`` dtype values containing ``NA`` incorrectly raising instead of casting to ``boolean`` dtype (:issue:`45462`)
 - Bug in :meth:`Series.loc` raising with boolean indexer containing ``NA`` when :class:`Index` did not match (:issue:`46551`)
 - Bug in :meth:`Series.__setitem__` where setting :attr:`NA` into a numeric-dtype :class:`Series` would incorrectly upcast to object-dtype rather than treating the value as ``np.nan`` (:issue:`44199`)

pandas/core/frame.py

@@ -4064,13 +4064,27 @@ def _set_item_frame_value(self, key, value: DataFrame) -> None:
             if isinstance(self.columns, MultiIndex) and isinstance(
                 loc, (slice, Series, np.ndarray, Index)
             ):
-                cols = maybe_droplevels(cols, key)
-                if len(cols) and not cols.equals(value.columns):
-                    value = value.reindex(cols, axis=1)
+                cols_droplevel = maybe_droplevels(cols, key)
+                if len(cols_droplevel) and not cols_droplevel.equals(value.columns):
+                    value = value.reindex(cols_droplevel, axis=1)
 
-        # now align rows
-        arraylike = _reindex_for_setitem(value, self.index)
-        self._set_item_mgr(key, arraylike)
+                for col, col_droplevel in zip(cols, cols_droplevel):
+                    self[col] = value[col_droplevel]
+                return
+
+            if is_scalar(cols):
+                self[cols] = value[value.columns[0]]
+                return
+
+            # now align rows
+            arraylike = _reindex_for_setitem(value, self.index)
+            self._set_item_mgr(key, arraylike)
+            return
+
+        if len(value.columns) != 1:
+            raise ValueError("Expected DataFrame with 1 column.")
+
+        self[key] = value[value.columns[0]]
 
     def _iset_item_mgr(
         self, loc: int | slice | np.ndarray, value, inplace: bool = False

pandas/tests/frame/indexing/test_coercion.py

@@ -164,7 +164,6 @@ def test_12499():
     tm.assert_frame_equal(df, expected)
 
 
-@pytest.mark.xfail(reason="Too many columns cast to float64")
 def test_20476():
     mi = MultiIndex.from_product([["A", "B"], ["a", "b", "c"]])
     df = DataFrame(-1, index=range(3), columns=mi)

pandas/tests/frame/indexing/test_setitem.py

@@ -91,8 +91,8 @@ def test_setitem_error_msmgs(self):
         # GH 4107, more descriptive error message
         df = DataFrame(np.random.randint(0, 2, (4, 4)), columns=["a", "b", "c", "d"])
 
-        msg = "incompatible index of inserted column with frame index"
-        with pytest.raises(TypeError, match=msg):
+        msg = "Expected DataFrame with 1 column."
+        with pytest.raises(ValueError, match=msg):
             df["gr"] = df.groupby(["b", "c"]).count()
 
     def test_setitem_benchmark(self):
@@ -741,6 +741,18 @@ def test_setitem_rhs_dataframe(self):
         df.isetitem(0, DataFrame({"a": [10, 11]}, index=[1, 2]))
         tm.assert_frame_equal(df, expected)
 
+    def test_setitem_frame_overwrite_with_ea_dtype(self, any_numeric_ea_dtype):
+        # GH#46896
+        df = DataFrame(columns=["a", "b"], data=[[1, 2], [3, 4]])
+        df["a"] = DataFrame({"a": [10, 11]}, dtype=any_numeric_ea_dtype)
+        expected = DataFrame(
+            {
+                "a": Series([10, 11], dtype=any_numeric_ea_dtype),
+                "b": [2, 4],
+            }
+        )
+        tm.assert_frame_equal(df, expected)
+
 
 class TestSetitemTZAwareValues:
     @pytest.fixture
@@ -903,6 +915,19 @@ def test_frame_setitem_rangeindex_into_new_col(self):
         expected = DataFrame({"a": ["b"], "b": [100]}, index=[1])
         tm.assert_frame_equal(result, expected)
 
+    def test_setitem_frame_keep_ea_dtype(self, any_numeric_ea_dtype):
+        # GH#46896
+        df = DataFrame(columns=["a", "b"], data=[[1, 2], [3, 4]])
+        df["c"] = DataFrame({"a": [10, 11]}, dtype=any_numeric_ea_dtype)
+        expected = DataFrame(
+            {
+                "a": [1, 3],
+                "b": [2, 4],
+                "c": Series([10, 11], dtype=any_numeric_ea_dtype),
+            }
+        )
+        tm.assert_frame_equal(df, expected)
+
 
 class TestDataFrameSetItemSlicing:
     def test_setitem_slice_position(self):