if_else shouldn't coerce nan to strings

[machow]

Edit: Removed pasted code snipper, replaced with reproducible example.

Problem: if_else can end up converting np.nan to a string: "nan".

Cause: if_else tries to downcast using numpy at the very end

Example

import numpy as np
from siuba import if_else

if_else(
    pd.Series([True, True, False, False]),
    np.nan,
    pd.Series([123, 456, "Number", "Percent"]),
)

# Result: array(['nan', 'nan', 'Number', 'Percent'], dtype='<U32')

The root cause in is the method used to downcast at the end. Here is the problem shown just with numpy...

np.array([np.nan, np.nan, "Number", "Percent"])

# Result: array(['nan', 'nan', 'Number', 'Percent'], dtype='<U32')

[machow]

I think the solution to this problem is to replace the very last line...

    return np.array(list(result))

with

    return pd.Series(result)

There are two reasons...

1. It fixes the issue with converting np.nan to a string
2. Since this is if_else dispatching on a pd.Series, it makes sense for it to return a Series

Alternatively, we could engage whatever mechanism pandas uses for converting the final array, since we don't really need all the Series machinery (e.g. an index). However, I suspect as pandas supports String arrays, etc.., through Series, it will become harder to avoid them.

[machow]

Instead of fix 2 (returning a Series), I opted to just keep the np array dtype "O". The reasoning is..

• inferring a type to convert to is one of pandas slowest operations
• as a result, anything that infers a type won't scale when run N_group times

(this is why pandas goes to great lengths to wait to convert types, until the very end of a set of operations)