DEPR: object inference in to_stata (#56536)

* DEPR: object inference in to_stata

* Whatsnew

* Fix broken test

* alphabetize

Author: jbrockmendel

doc/source/whatsnew/v3.0.0.rst

@@ -456,6 +456,7 @@ Other Deprecations
 - Deprecated allowing non-keyword arguments in :meth:`Series.to_string` except ``buf``. (:issue:`57280`)
 - Deprecated behavior of :meth:`.DataFrameGroupBy.groups` and :meth:`.SeriesGroupBy.groups`, in a future version ``groups`` by one element list will return tuple instead of scalar. (:issue:`58858`)
 - Deprecated behavior of :meth:`Series.dt.to_pytimedelta`, in a future version this will return a :class:`Series` containing python ``datetime.timedelta`` objects instead of an ``ndarray`` of timedelta; this matches the behavior of other :meth:`Series.dt` properties. (:issue:`57463`)
+- Deprecated converting object-dtype columns of ``datetime.datetime`` objects to datetime64 when writing to stata (:issue:`56536`)
 - Deprecated lowercase strings ``d``, ``b`` and ``c`` denoting frequencies in :class:`Day`, :class:`BusinessDay` and :class:`CustomBusinessDay` in favour of ``D``, ``B`` and ``C`` (:issue:`58998`)
 - Deprecated lowercase strings ``w``, ``w-mon``, ``w-tue``, etc. denoting frequencies in :class:`Week` in favour of ``W``, ``W-MON``, ``W-TUE``, etc. (:issue:`58998`)
 - Deprecated parameter ``method`` in :meth:`DataFrame.reindex_like` / :meth:`Series.reindex_like` (:issue:`58667`)

pandas/io/stata.py

@@ -393,14 +393,22 @@ def parse_dates_safe(
                 d["days"] = np.asarray(diff).astype("m8[D]").view("int64")
 
         elif infer_dtype(dates, skipna=False) == "datetime":
+            warnings.warn(
+                # GH#56536
+                "Converting object-dtype columns of datetimes to datetime64 when "
+                "writing to stata is deprecated. Call "
+                "`df=df.infer_objects(copy=False)` before writing to stata instead.",
+                FutureWarning,
+                stacklevel=find_stack_level(),
+            )
             if delta:
                 delta = dates._values - stata_epoch
 
                 def f(x: timedelta) -> float:
-                    return US_PER_DAY * x.days + 1000000 * x.seconds + x.microseconds
+                    return US_PER_DAY * x.days + 1_000_000 * x.seconds + x.microseconds
 
                 v = np.vectorize(f)
-                d["delta"] = v(delta)
+                d["delta"] = v(delta) // 1_000  # convert back to ms
             if year:
                 year_month = dates.apply(lambda x: 100 * x.year + x.month)
                 d["year"] = year_month._values // 100

pandas/tests/io/test_stata.py

@@ -1030,7 +1030,13 @@ def test_big_dates(self, datapath, temp_file):
         # {c : c[-2:] for c in columns}
         path = temp_file
         expected.index.name = "index"
-        expected.to_stata(path, convert_dates=date_conversion)
+        msg = (
+            "Converting object-dtype columns of datetimes to datetime64 "
+            "when writing to stata is deprecated"
+        )
+        exp_object = expected.astype(object)
+        with tm.assert_produces_warning(FutureWarning, match=msg):
+            exp_object.to_stata(path, convert_dates=date_conversion)
         written_and_read_again = self.read_dta(path)
 
         tm.assert_frame_equal(