python - Prevent pandas from interpreting 'NA' as NaN in a string

Question

The pandas read_csv() method interprets 'NA' as nan (not a number) instead of a valid string.

In the simple case below note that the output in row 1, column 2 (zero based count) is 'nan' instead of 'NA'.

sample.tsv (tab delimited)

PDB CHAIN SP_PRIMARY RES_BEG RES_END PDB_BEG PDB_END SP_BEG SP_END
5d8b N P60490 1 146 1 146 1 146
5d8b NA P80377 1 126 1 126 1 126
5d8b O P60491 1 118 1 118 1 118

read_sample.py

import pandas as pd

df = pd.read_csv(
    'sample.tsv',
    sep='\t',
    encoding='utf-8',
)

for df_tuples in df.itertuples(index=True):
    print(df_tuples)

output

(0, u'5d8b', u'N', u'P60490', 1, 146, 1, 146, 1, 146)
(1, u'5d8b', nan, u'P80377', 1, 126, 1, 126, 1, 126)
(2, u'5d8b', u'O', u'P60491', 1, 118, 1, 118, 1, 118)

Additional Information

Re-writing the file with quotes for data in the 'CHAIN' column and then using the quotechar parameter quotechar='\'' has the same result. And passing a dictionary of types via the dtype parameter dtype=dict(valid_cols) does not change the result.

An old answer to Prevent pandas from automatically inferring type in read_csv suggests first using a numpy record array to parse the file, but given the ability to now specify column dtypes, this shouldn't be necessary.

Note that itertuples() is used to preserve dtypes as described in the iterrows documentation: "To preserve dtypes while iterating over the rows, it is better to use itertuples() which returns tuples of the values and which is generally faster as iterrows."

Example was tested on Python 2 and 3 with pandas version 0.16.2, 0.17.0, and 0.17.1.

Is there a way to capture a valid string 'NA' instead of it being converted to nan?

It's not that it's treating it as a number, but rather indicates that you have missing data. I don't know what the fix is in your case, but just thought you should know. — Jeff Mercado
@JeffMercado If you try Anton's example below, but remove keep_default_na=False, you'll see that NaN is actually <class 'float'>. — binarysubstrate
@binarysubstrate Maybe you can consider trimming your question a bit (eg only keeping the 'simple case'), this would make the question more focused and more useful for other readers (as it is a very good question!) — joris
@joris Good suggestion. I made an edit to simplify the question and reduce the length. — binarysubstrate

Anton Protopopov Anton Protopopov · Accepted Answer · 2015-11-27T07:30:05

You could use parameters keep_default_na and na_values to set all NA values by hand docs:

import pandas as pd
from io import StringIO

data = """
PDB CHAIN SP_PRIMARY RES_BEG RES_END PDB_BEG PDB_END SP_BEG SP_END
5d8b N P60490 1 146 1 146 1 146
5d8b NA P80377 _ 126 1 126 1 126
5d8b O P60491 1 118 1 118 1 118
"""

df = pd.read_csv(StringIO(data), sep=' ', keep_default_na=False, na_values=['_'])

In [130]: df
Out[130]:
    PDB CHAIN SP_PRIMARY  RES_BEG  RES_END  PDB_BEG  PDB_END  SP_BEG  SP_END
0  5d8b     N     P60490        1      146        1      146       1     146
1  5d8b    NA     P80377      NaN      126        1      126       1     126
2  5d8b     O     P60491        1      118        1      118       1     118

In [144]: df.CHAIN.apply(type)
Out[144]:
0    <class 'str'>
1    <class 'str'>
2    <class 'str'>
Name: CHAIN, dtype: object

EDIT

All default NA values from na-values (as of pandas 1.0.0):

The default NaN recognized values are ['-1.#IND', '1.#QNAN', '1.#IND', '-1.#QNAN', '#N/A N/A', '#N/A', 'N/A', 'n/a', 'NA', '', '#NA', 'NULL', 'null', 'NaN', '-NaN', 'nan', '-nan', ''].

python - Prevent pandas from interpreting 'NA' as NaN in a string

Additional Information

4 Answers