BUG: inconsistent concat casting EA vs non-EA

doc/source/whatsnew/v1.3.0.rst

@@ -293,10 +293,11 @@ Groupby/resample/rolling
 
 Reshaping
 ^^^^^^^^^
-
 - Bug in :meth:`DataFrame.unstack` with missing levels led to incorrect index names (:issue:`37510`)
+- Bug in :func:`concat` incorrectly casting to ``object`` dtype in some cases when one or more of the operands is empty (:issue:`38843`)
 -
 
+
 Sparse
 ^^^^^^
 

pandas/core/dtypes/concat.py

@@ -128,7 +128,7 @@ def is_nonempty(x) -> bool:
     # marginal given that it would still require shape & dtype calculation and
     # np.concatenate which has them both implemented is compiled.
     non_empties = [x for x in to_concat if is_nonempty(x)]
-    if non_empties and axis == 0:
+    if non_empties:
         to_concat = non_empties
 
     typs = _get_dtype_kinds(to_concat)

pandas/tests/indexing/test_partial.py

@@ -170,11 +170,21 @@ def test_partial_setting_mixed_dtype(self):
         with pytest.raises(ValueError, match=msg):
             df.loc[0] = [1, 2, 3]
 
-        # TODO: #15657, these are left as object and not coerced
+    @pytest.mark.parametrize("dtype", [None, "int64", "Int64"])
+    def test_loc_setitem_expanding_empty(self, dtype):
         df = DataFrame(columns=["A", "B"])
-        df.loc[3] = [6, 7]
 
-        exp = DataFrame([[6, 7]], index=[3], columns=["A", "B"], dtype="object")
+        value = [6, 7]
+        if dtype == "int64":
+            value = np.array(value, dtype=dtype)
+        elif dtype == "Int64":
+            value = pd.array(value, dtype=dtype)
+
+        df.loc[3] = value
+
+        exp = DataFrame([[6, 7]], index=[3], columns=["A", "B"], dtype=dtype)
+        if dtype is not None:
+            exp = exp.astype(dtype)
         tm.assert_frame_equal(df, exp)
 
     def test_series_partial_set(self):

pandas/tests/reshape/concat/test_concat.py

@@ -474,11 +474,12 @@ def test_concat_will_upcast(dt, pdt):
         assert x.values.dtype == "float64"
 
 
-def test_concat_empty_and_non_empty_frame_regression():
+@pytest.mark.parametrize("dtype", ["int64", "Int64"])
+def test_concat_empty_and_non_empty_frame_regression(dtype):
     # GH 18178 regression test
-    df1 = DataFrame({"foo": [1]})
+    df1 = DataFrame({"foo": [1]}).astype(dtype)
     df2 = DataFrame({"foo": []})
-    expected = DataFrame({"foo": [1.0]})
+    expected = df1
     result = pd.concat([df1, df2])
     tm.assert_frame_equal(result, expected)
 

pandas/tests/reshape/concat/test_empty.py

@@ -202,12 +202,15 @@ def test_concat_empty_series_dtypes_sparse(self):
         expected = pd.SparseDtype("object")
         assert result.dtype == expected
 
-    def test_concat_empty_df_object_dtype(self):
+    @pytest.mark.parametrize("dtype", ["int64", "Int64"])
+    def test_concat_empty_df_object_dtype(self, dtype):
         # GH 9149
         df_1 = DataFrame({"Row": [0, 1, 1], "EmptyCol": np.nan, "NumberCol": [1, 2, 3]})
+        df_1["Row"] = df_1["Row"].astype(dtype)
         df_2 = DataFrame(columns=df_1.columns)
         result = pd.concat([df_1, df_2], axis=0)
-        expected = df_1.astype(object)
+        expected = df_1.copy()
+        expected["EmptyCol"] = expected["EmptyCol"].astype(object)  # TODO: why?
         tm.assert_frame_equal(result, expected)
 
     def test_concat_empty_dataframe_dtypes(self):