diff --git a/doc/source/whatsnew/v1.3.0.rst b/doc/source/whatsnew/v1.3.0.rst
index fb8eecdaa275e..326597763d1bc 100644
--- a/doc/source/whatsnew/v1.3.0.rst
+++ b/doc/source/whatsnew/v1.3.0.rst
@@ -585,7 +585,7 @@ Reshaping
 - Bug in :meth:`DataFrame.append` returning incorrect dtypes with combinations of ``datetime64`` and ``timedelta64`` dtypes (:issue:`39574`)
 - Bug in :meth:`DataFrame.pivot_table` returning a ``MultiIndex`` for a single value when operating on and empty ``DataFrame`` (:issue:`13483`)
 - Allow :class:`Index` to be passed to the :func:`numpy.all` function (:issue:`40180`)
--
+- Bug in :meth:`DataFrame.stack` not preserving ``CategoricalDtype`` in a ``MultiIndex`` (:issue:`36991`)
 
 Sparse
 ^^^^^^
diff --git a/pandas/core/reshape/reshape.py b/pandas/core/reshape/reshape.py
index 271bb2ca8dd75..ff6ba3f8f4164 100644
--- a/pandas/core/reshape/reshape.py
+++ b/pandas/core/reshape/reshape.py
@@ -600,6 +600,33 @@ def stack_multiple(frame, level, dropna=True):
     return result
 
 
+def _stack_multi_column_index(columns: MultiIndex) -> MultiIndex:
+    """Creates a MultiIndex from the first N-1 levels of this MultiIndex."""
+    if len(columns.levels) <= 2:
+        return columns.levels[0]._rename(name=columns.names[0])
+
+    levs = [
+        [lev[c] if c >= 0 else None for c in codes]
+        for lev, codes in zip(columns.levels[:-1], columns.codes[:-1])
+    ]
+
+    # Remove duplicate tuples in the MultiIndex.
+    tuples = zip(*levs)
+    unique_tuples = (key for key, _ in itertools.groupby(tuples))
+    new_levs = zip(*unique_tuples)
+
+    # The dtype of each level must be explicitly set to avoid inferring the wrong type.
+    # See GH-36991.
+    return MultiIndex.from_arrays(
+        [
+            # Not all indices can accept None values.
+            Index(new_lev, dtype=lev.dtype) if None not in new_lev else new_lev
+            for new_lev, lev in zip(new_levs, columns.levels)
+        ],
+        names=columns.names[:-1],
+    )
+
+
 def _stack_multi_columns(frame, level_num=-1, dropna=True):
     def _convert_level_number(level_num, columns):
         """
@@ -634,20 +661,7 @@ def _convert_level_number(level_num, columns):
         level_to_sort = _convert_level_number(0, this.columns)
         this = this.sort_index(level=level_to_sort, axis=1)
 
-    # tuple list excluding level for grouping columns
-    if len(frame.columns.levels) > 2:
-        levs = []
-        for lev, level_codes in zip(this.columns.levels[:-1], this.columns.codes[:-1]):
-            if -1 in level_codes:
-                lev = np.append(lev, None)
-            levs.append(np.take(lev, level_codes))
-        tuples = list(zip(*levs))
-        unique_groups = [key for key, _ in itertools.groupby(tuples)]
-        new_names = this.columns.names[:-1]
-        new_columns = MultiIndex.from_tuples(unique_groups, names=new_names)
-    else:
-        new_columns = this.columns.levels[0]._rename(name=this.columns.names[0])
-        unique_groups = new_columns
+    new_columns = _stack_multi_column_index(this.columns)
 
     # time to ravel the values
     new_data = {}
@@ -658,7 +672,7 @@ def _convert_level_number(level_num, columns):
     level_vals_used = np.take(level_vals_nan, level_codes)
     levsize = len(level_codes)
     drop_cols = []
-    for key in unique_groups:
+    for key in new_columns:
         try:
             loc = this.columns.get_loc(key)
         except KeyError:
diff --git a/pandas/tests/frame/test_stack_unstack.py b/pandas/tests/frame/test_stack_unstack.py
index fd23ea3a7621c..4082f21254e52 100644
--- a/pandas/tests/frame/test_stack_unstack.py
+++ b/pandas/tests/frame/test_stack_unstack.py
@@ -1065,6 +1065,27 @@ def test_stack_preserve_categorical_dtype(self, ordered, labels):
 
         tm.assert_series_equal(result, expected)
 
+    @pytest.mark.parametrize("ordered", [False, True])
+    @pytest.mark.parametrize(
+        "labels,data",
+        [
+            (list("xyz"), [10, 11, 12, 13, 14, 15]),
+            (list("zyx"), [14, 15, 12, 13, 10, 11]),
+        ],
+    )
+    def test_stack_multi_preserve_categorical_dtype(self, ordered, labels, data):
+        # GH-36991
+        cidx = pd.CategoricalIndex(labels, categories=sorted(labels), ordered=ordered)
+        cidx2 = pd.CategoricalIndex(["u", "v"], ordered=ordered)
+        midx = MultiIndex.from_product([cidx, cidx2])
+        df = DataFrame([sorted(data)], columns=midx)
+        result = df.stack([0, 1])
+
+        s_cidx = pd.CategoricalIndex(sorted(labels), ordered=ordered)
+        expected = Series(data, index=MultiIndex.from_product([[0], s_cidx, cidx2]))
+
+        tm.assert_series_equal(result, expected)
+
     def test_stack_preserve_categorical_dtype_values(self):
         # GH-23077
         cat = pd.Categorical(["a", "a", "b", "c"])