pandas.eval with a boolean series with missing data

Question

Description

I am using pandas.eval on a boolean series with missing data.

To do this I use an indexer to mark non-null values and .loc to only apply .eval on the rows with non-missing data.

Applying the logical not operator using the expression ~bool or not(bool) returns -1 or -2.

I understand that this is because my boolean series is casted as object type because of the missing values, but I am wondering :

• Why the -1 and -2 output ?
• What would be the proper way to use .eval on a boolean series with missing data ?

Example

Here is a reproducible example using pandas 0.20.3.

df = pd.DataFrame({'bool': [True, False, None]})
    bool
0   True
1  False
2   None

indexer = ~pd.isnull(df['bool'])
0     True
1     True
2    False
Name: bool, dtype: bool

df.loc[indexer].eval('~bool')
0    -2
1    -1
Name: bool, dtype: object

Answer 1

For eval, ~ maps to op.invert as seen in the source code here.

_unary_ops_syms = '+', '-', '~', 'not'
_unary_ops_funcs = op.pos, op.neg, op.invert, op.invert
_unary_ops_dict = dict(zip(_unary_ops_syms, _unary_ops_funcs))

Thus when your Series is of good old object type, what you're seeing here is

>>> ~True
-2
>>> ~False
-1

# or with your Series
>>> ~pd.Series(True, dtype='object')
0    -2
dtype: object

Where you want

>>> ~pd.Series(True)
0    False
dtype: bool

The outputs ~True -> -2 and ~False -> -1 are because bool is a subclass of int in Python, and -2, -1 are the bitwise complements of 1 and 0 respectively.

---

The obvious solution is to either convert the Series to bool type beforehand with astype(bool) in an extra setp, or if for some reason you cannot do so before the eval,

>>> df.loc[indexer].eval('~bool.astype("bool")')
0    False
1     True
Name: bool, dtype: bool