BUG : ValueError in case on NaN value in groupby columns (#24850)

Author: cgangwar11

doc/source/whatsnew/v0.24.0.rst

@@ -1786,6 +1786,7 @@ Groupby/Resample/Rolling
 - Bug in :meth:`DataFrame.groupby` did not respect the ``observed`` argument when selecting a column and instead always used ``observed=False`` (:issue:`23970`)
 - Bug in :func:`pandas.core.groupby.SeriesGroupBy.pct_change` or :func:`pandas.core.groupby.DataFrameGroupBy.pct_change` would previously work across groups when calculating the percent change, where it now correctly works per group (:issue:`21200`, :issue:`21235`).
 - Bug preventing hash table creation with very large number (2^32) of rows (:issue:`22805`)
+- Bug in groupby when grouping on categorical causes ``ValueError`` and incorrect grouping if ``observed=True`` and ``nan`` is present in categorical column (:issue:`24740`, :issue:`21151`).
 
 Reshaping
 ^^^^^^^^^

pandas/core/groupby/grouper.py

@@ -299,6 +299,7 @@ def __init__(self, index, grouper=None, obj=None, name=None, level=None,
                 self._labels = self.grouper.codes
                 if observed:
                     codes = algorithms.unique1d(self.grouper.codes)
+                    codes = codes[codes != -1]
                 else:
                     codes = np.arange(len(categories))
 

pandas/tests/groupby/test_categorical.py

@@ -420,6 +420,39 @@ def test_observed_groups(observed):
     tm.assert_dict_equal(result, expected)
 
 
+def test_observed_groups_with_nan(observed):
+    # GH 24740
+    df = pd.DataFrame({'cat': pd.Categorical(['a', np.nan, 'a'],
+                       categories=['a', 'b', 'd']),
+                       'vals': [1, 2, 3]})
+    g = df.groupby('cat', observed=observed)
+    result = g.groups
+    if observed:
+        expected = {'a': Index([0, 2], dtype='int64')}
+    else:
+        expected = {'a': Index([0, 2], dtype='int64'),
+                    'b': Index([], dtype='int64'),
+                    'd': Index([], dtype='int64')}
+    tm.assert_dict_equal(result, expected)
+
+
+def test_dataframe_categorical_with_nan(observed):
+    # GH 21151
+    s1 = pd.Categorical([np.nan, 'a', np.nan, 'a'],
+                        categories=['a', 'b', 'c'])
+    s2 = pd.Series([1, 2, 3, 4])
+    df = pd.DataFrame({'s1': s1, 's2': s2})
+    result = df.groupby('s1', observed=observed).first().reset_index()
+    if observed:
+        expected = DataFrame({'s1': pd.Categorical(['a'],
+                              categories=['a', 'b', 'c']), 's2': [2]})
+    else:
+        expected = DataFrame({'s1': pd.Categorical(['a', 'b', 'c'],
+                              categories=['a', 'b', 'c']),
+                              's2': [2, np.nan, np.nan]})
+    tm.assert_frame_equal(result, expected)
+
+
 def test_datetime():
     # GH9049: ensure backward compatibility
     levels = pd.date_range('2014-01-01', periods=4)