fix bug where np.bincount default arg minlength must be None for np<1.13

Author: tp

doc/source/whatsnew/v0.23.4.txt

@@ -31,6 +31,7 @@ Bug Fixes
 **Groupby/Resample/Rolling**
 
 - Bug where calling :func:`DataFrameGroupBy.agg` with a list of functions including ``ohlc`` as the non-initial element would raise a ``ValueError`` (:issue:`21716`)
+- Bug in :func:`pandas.core.groupby.SeriesGroupBy.count` when using numpy < 1.13 and ngroups=0 (:issue:`21956`).
 -
 
 **Conversion**

pandas/core/groupby/generic.py

@@ -46,6 +46,7 @@
 from pandas.core.index import Index, MultiIndex, CategoricalIndex
 from pandas.core.arrays.categorical import Categorical
 from pandas.core.internals import BlockManager, make_block
+from pandas.compat.numpy import _np_version_under1p13
 
 from pandas.plotting._core import boxplot_frame_groupby
 
@@ -1207,7 +1208,10 @@ def count(self):
 
         mask = (ids != -1) & ~isna(val)
         ids = ensure_platform_int(ids)
-        out = np.bincount(ids[mask], minlength=ngroups or 0)
+        minlength = ngroups or 0
+        if _np_version_under1p13 and minlength == 0:
+            minlength = None
+        out = np.bincount(ids[mask], minlength=minlength)
 
         return Series(out,
                       index=self.grouper.result_index,

pandas/tests/groupby/test_counting.py

@@ -212,3 +212,13 @@ def test_count_with_datetimelike(self, datetimelike):
         expected = DataFrame({'y': [2, 1]}, index=['a', 'b'])
         expected.index.name = "x"
         assert_frame_equal(expected, res)
+
+    def test_count_with_only_nans_in_first_group(self):
+        # GH21956
+        df = DataFrame({'A': [np.nan, np.nan], 'B': ['a', 'b'], 'C': [1, 2]})
+        result = df.groupby(['A', 'B']).C.count()
+        mi = MultiIndex(levels=[[], ['a', 'b']],
+                        labels=[[], []],
+                        names=['A', 'B'])
+        expected = Series([], index=mi, dtype=np.int64, name='C')
+        assert_series_equal(result, expected, check_index_type=False)