BUG: Outer/right merge with EA dtypes cast to object

doc/source/whatsnew/v1.3.3.rst

@@ -17,7 +17,7 @@ Fixed regressions
 - Fixed regression in :class:`DataFrame` constructor failing to broadcast for defined :class:`Index` and len one list of :class:`Timestamp` (:issue:`42810`)
 - Performance regression in :meth:`core.window.ewm.ExponentialMovingWindow.mean` (:issue:`42333`)
 - Fixed regression in :meth:`.GroupBy.agg` incorrectly raising in some cases (:issue:`42390`)
--
+- Fixed regression in :meth:`merge` where columns with ``ExtensionDtype`` was cast to ``object`` in ``left`` and ``outer`` merge (:issue:`40073`)
 
 .. ---------------------------------------------------------------------------
 

pandas/core/reshape/merge.py

@@ -70,6 +70,7 @@
     Categorical,
     Index,
     MultiIndex,
+    Series,
 )
 from pandas.core import groupby
 import pandas.core.algorithms as algos
@@ -81,10 +82,7 @@
 from pandas.core.sorting import is_int64_overflow_possible
 
 if TYPE_CHECKING:
-    from pandas import (
-        DataFrame,
-        Series,
-    )
+    from pandas import DataFrame
     from pandas.core.arrays import DatetimeArray
 
 
@@ -904,17 +902,22 @@ def _maybe_add_join_keys(
                 # error: Item "bool" of "Union[Any, bool]" has no attribute "all"
                 if mask_left.all():  # type: ignore[union-attr]
                     key_col = Index(rvals)
+                    final_dtype = rvals.dtype
                 # error: Item "bool" of "Union[Any, bool]" has no attribute "all"
                 elif (
                     right_indexer is not None
                     and mask_right.all()  # type: ignore[union-attr]
                 ):
                     key_col = Index(lvals)
+                    final_dtype = lvals.dtype
                 else:
                     key_col = Index(lvals).where(~mask_left, rvals)
+                    final_dtype = lvals.dtype
 
                 if result._is_label_reference(name):
-                    result[name] = key_col
+                    result[name] = Series(
+                        key_col, dtype=final_dtype, index=result.index
+                    )
                 elif result._is_level_reference(name):
                     if isinstance(result.index, MultiIndex):
                         key_col.name = name

pandas/tests/reshape/merge/test_merge.py

@@ -354,8 +354,8 @@ def test_merge_join_key_dtype_cast(self):
         df = merge(df1, df2, how="outer")
 
         # GH13169
-        # this really should be bool
-        assert df["key"].dtype == "object"
+        # GH#40073
+        assert df["key"].dtype == "bool"
 
         df1 = DataFrame({"val": [1]})
         df2 = DataFrame({"val": [2]})
@@ -2487,3 +2487,42 @@ def test_mergeerror_on_left_index_mismatched_dtypes():
     df_2 = DataFrame(data=["X"], columns=["C"], index=[999])
     with pytest.raises(MergeError, match="Can only pass argument"):
         merge(df_1, df_2, on=["C"], left_index=True)
+
+
+@pytest.mark.parametrize(
+    "expected_data, how",
+    [
+        ([1, 2], "outer"),
+        ([], "inner"),
+        ([2], "right"),
+        ([1], "left"),
+    ],
+)
+@pytest.mark.parametrize(
+    "dtype", ["Float64", "Float32", "Int64", "Int32", "UInt64", "UInt32"]
+)
+def test_merge_EA_dtype(dtype, how, expected_data):
+    # GH#40073
+    d1 = DataFrame([(1,)], columns=["id"], dtype=dtype)
+    d2 = DataFrame([(2,)], columns=["id"], dtype=dtype)
+    result = merge(d1, d2, how=how)
+    expected = DataFrame(expected_data, columns=["id"], dtype=dtype)
+    tm.assert_frame_equal(result, expected)
+
+
+@pytest.mark.parametrize(
+    "expected_data, how",
+    [
+        (["a", "b"], "outer"),
+        ([], "inner"),
+        (["b"], "right"),
+        (["a"], "left"),
+    ],
+)
+def test_merge_string_dtype(how, expected_data):
+    # GH#40073
+    d1 = DataFrame([("a",)], columns=["id"], dtype="string")
+    d2 = DataFrame([("b",)], columns=["id"], dtype="string")
+    result = merge(d1, d2, how=how)
+    expected = DataFrame(expected_data, columns=["id"], dtype="string")
+    tm.assert_frame_equal(result, expected)