BUG: Groupby quantiles incorrect bins #33200 (#33644)

doc/source/whatsnew/v1.1.0.rst

@@ -826,6 +826,7 @@ Groupby/resample/rolling
 - Bug in :meth:`DataFrame.resample` where an ``AmbiguousTimeError`` would be raised when the resulting timezone aware :class:`DatetimeIndex` had a DST transition at midnight (:issue:`25758`)
 - Bug in :meth:`DataFrame.groupby` where a ``ValueError`` would be raised when grouping by a categorical column with read-only categories and ``sort=False`` (:issue:`33410`)
 - Bug in :meth:`GroupBy.first` and :meth:`GroupBy.last` where None is not preserved in object dtype (:issue:`32800`)
+- Bug in :meth:`GroupBy.quantile` causes the quantiles to be shifted when the ``by`` axis contains ``NaN`` (:issue:`33200`, :issue:`33569`)
 - Bug in :meth:`Rolling.min` and :meth:`Rolling.max`: Growing memory usage after multiple calls when using a fixed window (:issue:`30726`)
 - Bug in :meth:`Series.groupby` would raise ``ValueError`` when grouping by :class:`PeriodIndex` level (:issue:`34010`)
 - Bug in :meth:`GroupBy.agg`, :meth:`GroupBy.transform`, and :meth:`GroupBy.resample` where subclasses are not preserved (:issue:`28330`)

pandas/_libs/groupby.pyx

@@ -777,7 +777,13 @@ def group_quantile(ndarray[float64_t] out,
                 non_na_counts[lab] += 1
 
     # Get an index of values sorted by labels and then values
-    order = (values, labels)
+    if labels.any():
+        # Put '-1' (NaN) labels as the last group so it does not interfere
+        # with the calculations.
+        labels_for_lexsort = np.where(labels == -1, labels.max() + 1, labels)
+    else:
+        labels_for_lexsort = labels
+    order = (values, labels_for_lexsort)
     sort_arr = np.lexsort(order).astype(np.int64, copy=False)
 
     with nogil:

pandas/tests/groupby/test_quantile.py

@@ -181,15 +181,32 @@ def test_quantile_missing_group_values_no_segfaults():
         grp.quantile()
 
 
-def test_quantile_missing_group_values_correct_results():
-    # GH 28662
-    data = np.array([1.0, np.nan, 3.0, np.nan])
-    df = pd.DataFrame(dict(key=data, val=range(4)))
+@pytest.mark.parametrize(
+    "key, val, expected_key, expected_val",
+    [
+        ([1.0, np.nan, 3.0, np.nan], range(4), [1.0, 3.0], [0.0, 2.0]),
+        ([1.0, np.nan, 2.0, 2.0], range(4), [1.0, 2.0], [0.0, 2.5]),
+        (["a", "b", "b", np.nan], range(4), ["a", "b"], [0, 1.5]),
+        ([0], [42], [0], [42.0]),
+        ([], [], np.array([], dtype="float64"), np.array([], dtype="float64")),
+    ],
+)
+def test_quantile_missing_group_values_correct_results(
+    key, val, expected_key, expected_val
+):
+    # GH 28662, GH 33200, GH 33569
+    df = pd.DataFrame({"key": key, "val": val})
 
-    result = df.groupby("key").quantile()
     expected = pd.DataFrame(
-        [1.0, 3.0], index=pd.Index([1.0, 3.0], name="key"), columns=["val"]
+        expected_val, index=pd.Index(expected_key, name="key"), columns=["val"]
     )
+
+    grp = df.groupby("key")
+
+    result = grp.quantile(0.5)
+    tm.assert_frame_equal(result, expected)
+
+    result = grp.quantile()
     tm.assert_frame_equal(result, expected)