BUG: Wrong dtype when resetting a multiindex with missing values. (#1…

doc/source/whatsnew/v0.25.0.rst

@@ -1068,6 +1068,8 @@ MultiIndex
 ^^^^^^^^^^
 
 - Bug in which incorrect exception raised by :class:`Timedelta` when testing the membership of :class:`MultiIndex` (:issue:`24570`)
+- Bug in :meth:`DataFrame.reset_index` where dtype was sometimes not preserved for :class:`MultiIndex` that is empty or with missing values (:issue:`19602`)
+- Bug in which :meth:`DataFrame.reset_index` did not work for :class:`MultiIndex` with :class:`CategoricalIndex` levels with missing values (:issue:`24206`)
 -
 
 I/O

pandas/core/frame.py

@@ -83,7 +83,6 @@
 from pandas.core import algorithms, common as com, nanops, ops
 from pandas.core.accessor import CachedAccessor
 from pandas.core.arrays import Categorical, ExtensionArray
-from pandas.core.arrays.datetimelike import DatetimeLikeArrayMixin as DatetimeLikeArray
 from pandas.core.arrays.sparse import SparseFrameAccessor
 from pandas.core.generic import NDFrame, _shared_docs
 from pandas.core.index import (
@@ -4610,33 +4609,19 @@ def _maybe_casted_values(index, labels=None):
 
             # if we have the labels, extract the values with a mask
             if labels is not None:
-                mask = labels == -1
-
-                # we can have situations where the whole mask is -1,
-                # meaning there is nothing found in labels, so make all nan's
-                if mask.all():
-                    values = np.empty(len(mask))
-                    values.fill(np.nan)
+                if isinstance(values, np.ndarray):
+                    mask = labels == -1
+                    # we can have situations where the whole mask is -1,
+                    # meaning there is nothing found in labels, so make all nan's
+                    if mask.all():
+                        values = np.empty(len(mask), dtype=values.dtype)
+                        values.fill(np.nan)
+                    else:
+                        values = values.take(labels)
+                        if mask.any():
+                            values, _ = maybe_upcast_putmask(values, mask, np.nan)
                 else:
-                    values = values.take(labels)
-
-                    # TODO(https://github.com/pandas-dev/pandas/issues/24206)
-                    # Push this into maybe_upcast_putmask?
-                    # We can't pass EAs there right now. Looks a bit
-                    # complicated.
-                    # So we unbox the ndarray_values, op, re-box.
-                    values_type = type(values)
-                    values_dtype = values.dtype
-
-                    if issubclass(values_type, DatetimeLikeArray):
-                        values = values._data
-
-                    if mask.any():
-                        values, changed = maybe_upcast_putmask(values, mask, np.nan)
-
-                    if issubclass(values_type, DatetimeLikeArray):
-                        values = values_type(values, dtype=values_dtype)
-
+                    values = values.take(labels, allow_fill=True)
             return values
 
         new_index = ibase.default_index(len(new_obj))
@@ -4680,7 +4665,7 @@ def _maybe_casted_values(index, labels=None):
                     missing = self.columns.nlevels - len(name_lst)
                     name_lst += [col_fill] * missing
                     name = tuple(name_lst)
-                # to ndarray and maybe infer different dtype
+                # to array-like and maybe infer different dtype
                 level_values = _maybe_casted_values(lev, lab)
                 new_obj.insert(0, name, level_values)
 

pandas/tests/frame/test_alter_axes.py

@@ -12,6 +12,7 @@
 
 from pandas import (
     Categorical,
+    CategoricalIndex,
     DataFrame,
     DatetimeIndex,
     Index,
@@ -1186,6 +1187,45 @@ def test_reset_index_multiindex_nan(self):
         rs = df.set_index(["A", "B"]).reset_index()
         tm.assert_frame_equal(rs, df)
 
+    def test_reset_index_multiindex_datetime_all_nan(self):
+        # GH 19602
+        df = DataFrame({0: DatetimeIndex([]), 1: []}, columns=[0, 1])
+        rs = df.set_index([0, 1]).reset_index()
+        tm.assert_frame_equal(rs, df)
+
+        idx = MultiIndex(
+            levels=[DatetimeIndex([]), DatetimeIndex(["2015-01-01 11:00:00"])],
+            codes=[[-1, -1], [0, -1]],
+            names=[0, 1],
+        )
+        df = DataFrame(index=idx).reset_index()
+
+        xp = DataFrame(
+            {
+                0: DatetimeIndex([np.nan, np.nan]),
+                1: DatetimeIndex(["2015-01-01 11:00:00", np.nan]),
+            },
+            columns=[0, 1],
+        )
+        tm.assert_frame_equal(df, xp)
+
+    def test_reset_index_multiindex_categorical_with_nan(self):
+        # GH 24206
+        idx = MultiIndex(
+            [CategoricalIndex(["A", "B"]), CategoricalIndex(["a", "b"])],
+            [[0, 0, 1, 1], [0, 1, 0, -1]],
+        )
+        df = DataFrame({"col": range(len(idx))}, index=idx).reset_index()
+        xp = DataFrame(
+            {
+                "level_0": CategoricalIndex(["A", "A", "B", "B"]),
+                "level_1": CategoricalIndex(["a", "b", "a", np.nan]),
+                "col": [0, 1, 2, 3],
+            },
+            columns=["level_0", "level_1", "col"],
+        )
+        tm.assert_frame_equal(df, xp)
+
     def test_reset_index_with_datetimeindex_cols(self):
         # GH5818
         #