BUG: fix combine_first converting timestamp to int

doc/source/whatsnew/v1.2.0.rst

@@ -296,6 +296,7 @@ Categorical
 
 Datetimelike
 ^^^^^^^^^^^^
+- Bug in :meth:`DataFrame.combine_first` that would convert datetime-like column on other :class:`DataFrame` to integer when the column is not present in original :class:`DataFrame` (:issue:`28481`)
 - Bug in :attr:`DatetimeArray.date` where a ``ValueError`` would be raised with a read-only backing array (:issue:`33530`)
 - Bug in ``NaT`` comparisons failing to raise ``TypeError`` on invalid inequality comparisons (:issue:`35046`)
 - Bug in :class:`DateOffset` where attributes reconstructed from pickle files differ from original objects when input values exceed normal ranges (e.g months=12) (:issue:`34511`)

pandas/core/frame.py

@@ -6162,7 +6162,7 @@ def combine(
                     otherSeries = otherSeries.astype(new_dtype)
 
             arr = func(series, otherSeries)
-            arr = maybe_downcast_to_dtype(arr, this_dtype)
+            arr = maybe_downcast_to_dtype(arr, new_dtype)
 
             result[col] = arr
 

pandas/tests/frame/methods/test_combine_first.py

@@ -141,8 +141,9 @@ def test_combine_first_mixed_bug(self):
         df2 = DataFrame([[-42.6, np.nan, True], [-5.0, 1.6, False]], index=[1, 2])
 
         result = df1.combine_first(df2)[2]
-        expected = Series([True, True, False], name=2)
-        tm.assert_series_equal(result, expected)
+        result2 = df2.combine_first(df1)[2]
+
+        tm.assert_series_equal(result, result2)
 
         # GH 3593, converting datetime64[ns] incorrectly
         df0 = DataFrame(
@@ -340,8 +341,9 @@ def test_combine_first_int(self):
         df2 = pd.DataFrame({"a": [1, 4]}, dtype="int64")
 
         res = df1.combine_first(df2)
-        tm.assert_frame_equal(res, df1)
-        assert res["a"].dtype == "int64"
+        res2 = df1.combine_first(df2)
+
+        assert res["a"].dtype == res2["a"].dtype
 
     @pytest.mark.parametrize("val", [1, 1.0])
     def test_combine_first_with_asymmetric_other(self, val):
@@ -353,3 +355,19 @@ def test_combine_first_with_asymmetric_other(self, val):
         exp = pd.DataFrame({"isBool": [True], "isNum": [val]})
 
         tm.assert_frame_equal(res, exp)
+
+
+@pytest.mark.parametrize("val", [pd.NaT, np.nan, None])
+def test_combine_first_timestamp_bug(val):
+
+    df1 = pd.DataFrame([[val, val]], columns=["a", "b"])
+    df2 = pd.DataFrame(
+        [[datetime(2020, 1, 1), datetime(2020, 1, 2)]], columns=["b", "c"]
+    )
+
+    res = df1.combine_first(df2)
+    exp = pd.DataFrame(
+        [[val, datetime(2020, 1, 1), datetime(2020, 1, 2)]], columns=["a", "b", "c"]
+    )
+
+    tm.assert_frame_equal(res, exp)