BUG: Categorical.from_codes casts nullable int categories to object

[arw2019]

For int64 the construction yields:

In [18]: import numpy as np
    ...: import pandas as pd
    ...: import pandas._testing as tm
    ...: 
    ...: integer_dtype="int64"
    ...: 
    ...: cats = pd.array(range(5), dtype=integer_dtype)
    ...: codes = np.random.randint(5, size=3)
    ...: categorical_dtype = pd.CategoricalDtype(cats)
    ...: arr = pd.Categorical.from_codes(codes, dtype=categorical_dtype)
    ...: arr
Out[18]: 
[0, 2, 0]
Categories (5, int64): [0, 1, 2, 3, 4]

whereas for nullable int categories:

In [19]: import numpy as np
    ...: import pandas as pd
    ...: import pandas._testing as tm
    ...: 
    ...: integer_dtype="Int64"
    ...: 
    ...: cats = pd.array(range(5), dtype=integer_dtype)
    ...: codes = np.random.randint(5, size=3)
    ...: categorical_dtype = pd.CategoricalDtype(cats)
    ...: arr = pd.Categorical.from_codes(codes, dtype=categorical_dtype)
    ...: arr
Out[19]: 
[3, 3, 4]
Categories (5, object): [0, 1, 2, 3, 4]

I expect instead in the nullable int case:

Out[19]: 
[3, 3, 4]
Categories (5, Int64): [0, 1, 2, 3, 4]

[anders-wind]

I can confirm that this happens on version 1.2.1.

[jorisvandenbossche]

In general we still need to add support for "nullable" categorical (however it may look like).

Very practical reason currently is that we don't yet support storing EAs in the index, and Categorical uses Index as the categories. Now, personally I would say that for the nullable dtypes, we could also look into storing the categories as a plain array instead of Index.