BUG: Fixed upcast dtype for datetime64 in merge

Closes pandas-dev#31208

Author: TomAugspurger

pandas/core/internals/concat.py

@@ -350,7 +350,7 @@ def _get_empty_dtype_and_na(join_units):
         dtype = upcast_classes["datetimetz"]
         return dtype[0], tslibs.NaT
     elif "datetime" in upcast_classes:
-        return np.dtype("M8[ns]"), tslibs.iNaT
+        return np.dtype("M8[ns]"), np.datetime64("NaT", "ns")
     elif "timedelta" in upcast_classes:
         return np.dtype("m8[ns]"), np.timedelta64("NaT", "ns")
     else:  # pragma

pandas/tests/reshape/merge/test_merge.py

@@ -2152,3 +2152,20 @@ def test_merge_multiindex_columns():
     expected["id"] = ""
 
     tm.assert_frame_equal(result, expected)
+
+
+def test_merge_datetime_upcast_dtype():
+    # https://github.com/pandas-dev/pandas/issues/31208
+    df1 = pd.DataFrame({"x": ["a", "b", "c"], "y": ["1", "2", "4"]})
+    df2 = pd.DataFrame(
+        {"y": ["1", "2", "3"], "z": pd.to_datetime(["2000", "2001", "2002"])}
+    )
+    result = pd.merge(df1, df2, how="left", on="y")
+    expected = pd.DataFrame(
+        {
+            "x": ["a", "b", "c"],
+            "y": ["1", "2", "4"],
+            "z": pd.to_datetime(["2000", "2001", "NaT"]),
+        }
+    )
+    tm.assert_frame_equal(result, expected)