BUG: groupby with categorical drops empty groups when aggregating over a series

[aimboden]

• Series groupby excluding NaN groups with Categorical (DataFrame DOES include)
• sorting via a returned Interval-like-Index (string based)

Hello,

When grouping a DataFrame over more than one column including a categorical, the empty groups are kept in the aggregation result. A test for this behaviour was introduced in #8138.

However, when performing aggregation on only one column of the DataFrame, the empty groups are dropped. This seems inconsistent to me and I guess that it's an edge case that wasn't thought of at the time.

d = {'foo': [10, 8, 4, 1], 'bar': [10, 20, 30, 40],
     'baz': ['d', 'c', 'd', 'c']}
df = pd.DataFrame(d)
cat = pd.cut(df['foo'], np.linspace(0, 20, 5))
df['range'] = cat
groups = df.groupby(['range', 'baz'], as_index=True, sort=True)

# Expected result, fixed as part of #8138
fixed = groups.agg('mean')

# Inconsistent behaviour with series
inconsistent = groups['foo'].agg('mean')

# Expected result
expected = fixed['foo']

fixed

		bar	foo
range	baz
(0, 5]	c	1	40
	d	4	30
(10, 15]	c	NaN	NaN
	d	NaN	NaN
(15, 20]	c	NaN	NaN
	d	NaN	NaN
(5, 10]	c	8	20
	d	10	10

inconsistent

range	baz
(0, 5]	c	1
	d	4
(5, 10]	c	8
	d	10

expected

range	baz
(0, 5]	c	1
	d	4
(10, 15]	c	NaN
	d	NaN
(15, 20]	c	NaN
	d	NaN
(5, 10]	c	8
	d	10

Note the strange ordering of the categorical index. I would expect sorted = True to sort by categorical level and not by lexical order?

Also note that using as_index=False fails due to #8869

[jreback]

This needs an IntervalIndex, see #8707; these are just strings for now.

[aimboden]

I agree that the sorting part requires an IntervalIndex, but what about the dropping of empty groups? This was taken care of in the case of DataFrames, but still happens for Series.

[jreback]

hmm, should be consistent with DataFrame and (either both inlclude the nan groups as Frame does now, or both exclude - was some debate over this).