Table Of Contents

Search

Enter search terms or a module, class or function name.

Indexing and Selecting Data

The axis labeling information in pandas objects serves many purposes:

  • Identifies data (i.e. provides metadata) using known indicators, important for analysis, visualization, and interactive console display
  • Enables automatic and explicit data alignment
  • Allows intuitive getting and setting of subsets of the data set

In this section, we will focus on the final point: namely, how to slice, dice, and generally get and set subsets of pandas objects. The primary focus will be on Series and DataFrame as they have received more development attention in this area. Expect more work to be invested in higher-dimensional data structures (including Panel) in the future, especially in label-based advanced indexing.

Note

The Python and NumPy indexing operators [] and attribute operator . provide quick and easy access to pandas data structures across a wide range of use cases. This makes interactive work intuitive, as there’s little new to learn if you already know how to deal with Python dictionaries and NumPy arrays. However, since the type of the data to be accessed isn’t known in advance, directly using standard operators has some optimization limits. For production code, we recommended that you take advantage of the optimized pandas data access methods exposed in this chapter.

Warning

Whether a copy or a reference is returned for a setting operation, may depend on the context. This is sometimes called chained assignment and should be avoided. See Returning a View versus Copy

Warning

In 0.15.0 Index has internally been refactored to no longer subclass ndarray but instead subclass PandasObject, similarly to the rest of the pandas objects. This should be a transparent change with only very limited API implications (See the Internal Refactoring)

Warning

Indexing on an integer-based Index with floats has been clarified in 0.18.0, for a summary of the changes, see here.

See the MultiIndex / Advanced Indexing for MultiIndex and more advanced indexing documentation.

See the cookbook for some advanced strategies

Different Choices for Indexing

New in version 0.11.0.

Object selection has had a number of user-requested additions in order to support more explicit location based indexing. pandas now supports three types of multi-axis indexing.

  • .loc is primarily label based, but may also be used with a boolean array. .loc will raise KeyError when the items are not found. Allowed inputs are:

    • A single label, e.g. 5 or 'a', (note that 5 is interpreted as a label of the index. This use is not an integer position along the index)
    • A list or array of labels ['a', 'b', 'c']
    • A slice object with labels 'a':'f', (note that contrary to usual python slices, both the start and the stop are included!)
    • A boolean array

    See more at Selection by Label

  • .iloc is primarily integer position based (from 0 to length-1 of the axis), but may also be used with a boolean array. .iloc will raise IndexError if a requested indexer is out-of-bounds, except slice indexers which allow out-of-bounds indexing. (this conforms with python/numpy slice semantics). Allowed inputs are:

    • An integer e.g. 5
    • A list or array of integers [4, 3, 0]
    • A slice object with ints 1:7
    • A boolean array

    See more at Selection by Position

  • .ix supports mixed integer and label based access. It is primarily label based, but will fall back to integer positional access unless the corresponding axis is of integer type. .ix is the most general and will support any of the inputs in .loc and .iloc. .ix also supports floating point label schemes. .ix is exceptionally useful when dealing with mixed positional and label based hierarchical indexes.

    However, when an axis is integer based, ONLY label based access and not positional access is supported. Thus, in such cases, it’s usually better to be explicit and use .iloc or .loc.

    See more at Advanced Indexing and Advanced Hierarchical.

Getting values from an object with multi-axes selection uses the following notation (using .loc as an example, but applies to .iloc and .ix as well). Any of the axes accessors may be the null slice :. Axes left out of the specification are assumed to be :. (e.g. p.loc['a'] is equiv to p.loc['a', :, :])

Object Type Indexers
Series s.loc[indexer]
DataFrame df.loc[row_indexer,column_indexer]
Panel p.loc[item_indexer,major_indexer,minor_indexer]

Basics

As mentioned when introducing the data structures in the last section, the primary function of indexing with [] (a.k.a. __getitem__ for those familiar with implementing class behavior in Python) is selecting out lower-dimensional slices. Thus,

Object Type Selection Return Value Type
Series series[label] scalar value
DataFrame frame[colname] Series corresponding to colname
Panel panel[itemname] DataFrame corresponding to the itemname

Here we construct a simple time series data set to use for illustrating the indexing functionality:

In [1]: dates = pd.date_range('1/1/2000', periods=8)

In [2]: df = pd.DataFrame(np.random.randn(8, 4), index=dates, columns=['A', 'B', 'C', 'D'])

In [3]: df
Out[3]: 
                   A         B         C         D
2000-01-01  0.469112 -0.282863 -1.509059 -1.135632
2000-01-02  1.212112 -0.173215  0.119209 -1.044236
2000-01-03 -0.861849 -2.104569 -0.494929  1.071804
2000-01-04  0.721555 -0.706771 -1.039575  0.271860
2000-01-05 -0.424972  0.567020  0.276232 -1.087401
2000-01-06 -0.673690  0.113648 -1.478427  0.524988
2000-01-07  0.404705  0.577046 -1.715002 -1.039268
2000-01-08 -0.370647 -1.157892 -1.344312  0.844885

In [4]: panel = pd.Panel({'one' : df, 'two' : df - df.mean()})

In [5]: panel
Out[5]: 
<class 'pandas.core.panel.Panel'>
Dimensions: 2 (items) x 8 (major_axis) x 4 (minor_axis)
Items axis: one to two
Major_axis axis: 2000-01-01 00:00:00 to 2000-01-08 00:00:00
Minor_axis axis: A to D

Note

None of the indexing functionality is time series specific unless specifically stated.

Thus, as per above, we have the most basic indexing using []:

In [6]: s = df['A']

In [7]: s[dates[5]]
Out[7]: -0.67368970808837059

In [8]: panel['two']
Out[8]: 
                   A         B         C         D
2000-01-01  0.409571  0.113086 -0.610826 -0.936507
2000-01-02  1.152571  0.222735  1.017442 -0.845111
2000-01-03 -0.921390 -1.708620  0.403304  1.270929
2000-01-04  0.662014 -0.310822 -0.141342  0.470985
2000-01-05 -0.484513  0.962970  1.174465 -0.888276
2000-01-06 -0.733231  0.509598 -0.580194  0.724113
2000-01-07  0.345164  0.972995 -0.816769 -0.840143
2000-01-08 -0.430188 -0.761943 -0.446079  1.044010

You can pass a list of columns to [] to select columns in that order. If a column is not contained in the DataFrame, an exception will be raised. Multiple columns can also be set in this manner:

In [9]: df
Out[9]: 
                   A         B         C         D
2000-01-01  0.469112 -0.282863 -1.509059 -1.135632
2000-01-02  1.212112 -0.173215  0.119209 -1.044236
2000-01-03 -0.861849 -2.104569 -0.494929  1.071804
2000-01-04  0.721555 -0.706771 -1.039575  0.271860
2000-01-05 -0.424972  0.567020  0.276232 -1.087401
2000-01-06 -0.673690  0.113648 -1.478427  0.524988
2000-01-07  0.404705  0.577046 -1.715002 -1.039268
2000-01-08 -0.370647 -1.157892 -1.344312  0.844885

In [10]: df[['B', 'A']] = df[['A', 'B']]

In [11]: df
Out[11]: 
                   A         B         C         D
2000-01-01 -0.282863  0.469112 -1.509059 -1.135632
2000-01-02 -0.173215  1.212112  0.119209 -1.044236
2000-01-03 -2.104569 -0.861849 -0.494929  1.071804
2000-01-04 -0.706771  0.721555 -1.039575  0.271860
2000-01-05  0.567020 -0.424972  0.276232 -1.087401
2000-01-06  0.113648 -0.673690 -1.478427  0.524988
2000-01-07  0.577046  0.404705 -1.715002 -1.039268
2000-01-08 -1.157892 -0.370647 -1.344312  0.844885

You may find this useful for applying a transform (in-place) to a subset of the columns.

Attribute Access

You may access an index on a Series, column on a DataFrame, and an item on a Panel directly as an attribute:

In [12]: sa = pd.Series([1,2,3],index=list('abc'))

In [13]: dfa = df.copy()
In [14]: sa.b
Out[14]: 2

In [15]: dfa.A
Out[15]: 
2000-01-01   -0.282863
2000-01-02   -0.173215
2000-01-03   -2.104569
2000-01-04   -0.706771
2000-01-05    0.567020
2000-01-06    0.113648
2000-01-07    0.577046
2000-01-08   -1.157892
Freq: D, Name: A, dtype: float64

In [16]: panel.one
Out[16]: 
                   A         B         C         D
2000-01-01  0.469112 -0.282863 -1.509059 -1.135632
2000-01-02  1.212112 -0.173215  0.119209 -1.044236
2000-01-03 -0.861849 -2.104569 -0.494929  1.071804
2000-01-04  0.721555 -0.706771 -1.039575  0.271860
2000-01-05 -0.424972  0.567020  0.276232 -1.087401
2000-01-06 -0.673690  0.113648 -1.478427  0.524988
2000-01-07  0.404705  0.577046 -1.715002 -1.039268
2000-01-08 -0.370647 -1.157892 -1.344312  0.844885

You can use attribute access to modify an existing element of a Series or column of a DataFrame, but be careful; if you try to use attribute access to create a new column, it fails silently, creating a new attribute rather than a new column.

In [17]: sa.a = 5

In [18]: sa
Out[18]: 
a    5
b    2
c    3
dtype: int64

In [19]: dfa.A = list(range(len(dfa.index)))  # ok if A already exists

In [20]: dfa
Out[20]: 
            A         B         C         D
2000-01-01  0  0.469112 -1.509059 -1.135632
2000-01-02  1  1.212112  0.119209 -1.044236
2000-01-03  2 -0.861849 -0.494929  1.071804
2000-01-04  3  0.721555 -1.039575  0.271860
2000-01-05  4 -0.424972  0.276232 -1.087401
2000-01-06  5 -0.673690 -1.478427  0.524988
2000-01-07  6  0.404705 -1.715002 -1.039268
2000-01-08  7 -0.370647 -1.344312  0.844885

In [21]: dfa['A'] = list(range(len(dfa.index)))  # use this form to create a new column

In [22]: dfa
Out[22]: 
            A         B         C         D
2000-01-01  0  0.469112 -1.509059 -1.135632
2000-01-02  1  1.212112  0.119209 -1.044236
2000-01-03  2 -0.861849 -0.494929  1.071804
2000-01-04  3  0.721555 -1.039575  0.271860
2000-01-05  4 -0.424972  0.276232 -1.087401
2000-01-06  5 -0.673690 -1.478427  0.524988
2000-01-07  6  0.404705 -1.715002 -1.039268
2000-01-08  7 -0.370647 -1.344312  0.844885

Warning

  • You can use this access only if the index element is a valid python identifier, e.g. s.1 is not allowed. See here for an explanation of valid identifiers.
  • The attribute will not be available if it conflicts with an existing method name, e.g. s.min is not allowed.
  • Similarly, the attribute will not be available if it conflicts with any of the following list: index, major_axis, minor_axis, items, labels.
  • In any of these cases, standard indexing will still work, e.g. s['1'], s['min'], and s['index'] will access the corresponding element or column.
  • The Series/Panel accesses are available starting in 0.13.0.

If you are using the IPython environment, you may also use tab-completion to see these accessible attributes.

You can also assign a dict to a row of a DataFrame:

In [23]: x = pd.DataFrame({'x': [1, 2, 3], 'y': [3, 4, 5]})

In [24]: x.iloc[1] = dict(x=9, y=99)

In [25]: x
Out[25]: 
   x   y
0  1   3
1  9  99
2  3   5

Slicing ranges

The most robust and consistent way of slicing ranges along arbitrary axes is described in the Selection by Position section detailing the .iloc method. For now, we explain the semantics of slicing using the [] operator.

With Series, the syntax works exactly as with an ndarray, returning a slice of the values and the corresponding labels:

In [26]: s[:5]
Out[26]: 
2000-01-01   -0.282863
2000-01-02   -0.173215
2000-01-03   -2.104569
2000-01-04   -0.706771
2000-01-05    0.567020
Freq: D, Name: A, dtype: float64

In [27]: s[::2]
Out[27]: 
2000-01-01   -0.282863
2000-01-03   -2.104569
2000-01-05    0.567020
2000-01-07    0.577046
Freq: 2D, Name: A, dtype: float64

In [28]: s[::-1]
Out[28]: 
2000-01-08   -1.157892
2000-01-07    0.577046
2000-01-06    0.113648
2000-01-05    0.567020
2000-01-04   -0.706771
2000-01-03   -2.104569
2000-01-02   -0.173215
2000-01-01   -0.282863
Freq: -1D, Name: A, dtype: float64

Note that setting works as well:

In [29]: s2 = s.copy()

In [30]: s2[:5] = 0

In [31]: s2
Out[31]: 
2000-01-01    0.000000
2000-01-02    0.000000
2000-01-03    0.000000
2000-01-04    0.000000
2000-01-05    0.000000
2000-01-06    0.113648
2000-01-07    0.577046
2000-01-08   -1.157892
Freq: D, Name: A, dtype: float64

With DataFrame, slicing inside of [] slices the rows. This is provided largely as a convenience since it is such a common operation.

In [32]: df[:3]
Out[32]: 
                   A         B         C         D
2000-01-01 -0.282863  0.469112 -1.509059 -1.135632
2000-01-02 -0.173215  1.212112  0.119209 -1.044236
2000-01-03 -2.104569 -0.861849 -0.494929  1.071804

In [33]: df[::-1]
Out[33]: 
                   A         B         C         D
2000-01-08 -1.157892 -0.370647 -1.344312  0.844885
2000-01-07  0.577046  0.404705 -1.715002 -1.039268
2000-01-06  0.113648 -0.673690 -1.478427  0.524988
2000-01-05  0.567020 -0.424972  0.276232 -1.087401
2000-01-04 -0.706771  0.721555 -1.039575  0.271860
2000-01-03 -2.104569 -0.861849 -0.494929  1.071804
2000-01-02 -0.173215  1.212112  0.119209 -1.044236
2000-01-01 -0.282863  0.469112 -1.509059 -1.135632

Selection By Label

Warning

Whether a copy or a reference is returned for a setting operation, may depend on the context. This is sometimes called chained assignment and should be avoided. See Returning a View versus Copy

Warning

.loc is strict when you present slicers that are not compatible (or convertible) with the index type. For example using integers in a DatetimeIndex. These will raise a TypeError.
In [34]: dfl = pd.DataFrame(np.random.randn(5,4), columns=list('ABCD'), index=pd.date_range('20130101',periods=5))

In [35]: dfl
Out[35]: 
                   A         B         C         D
2013-01-01  1.075770 -0.109050  1.643563 -1.469388
2013-01-02  0.357021 -0.674600 -1.776904 -0.968914
2013-01-03 -1.294524  0.413738  0.276662 -0.472035
2013-01-04 -0.013960 -0.362543 -0.006154 -0.923061
2013-01-05  0.895717  0.805244 -1.206412  2.565646
In [4]: dfl.loc[2:3]
TypeError: cannot do slice indexing on <class 'pandas.tseries.index.DatetimeIndex'> with these indexers [2] of <type 'int'>

String likes in slicing can be convertible to the type of the index and lead to natural slicing.

In [36]: dfl.loc['20130102':'20130104']
Out[36]: 
                   A         B         C         D
2013-01-02  0.357021 -0.674600 -1.776904 -0.968914
2013-01-03 -1.294524  0.413738  0.276662 -0.472035
2013-01-04 -0.013960 -0.362543 -0.006154 -0.923061

pandas provides a suite of methods in order to have purely label based indexing. This is a strict inclusion based protocol. At least 1 of the labels for which you ask, must be in the index or a KeyError will be raised! When slicing, the start bound is included, AND the stop bound is included. Integers are valid labels, but they refer to the label and not the position.

The .loc attribute is the primary access method. The following are valid inputs:

  • A single label, e.g. 5 or 'a', (note that 5 is interpreted as a label of the index. This use is not an integer position along the index)
  • A list or array of labels ['a', 'b', 'c']
  • A slice object with labels 'a':'f' (note that contrary to usual python slices, both the start and the stop are included!)
  • A boolean array
In [37]: s1 = pd.Series(np.random.randn(6),index=list('abcdef'))

In [38]: s1
Out[38]: 
a    1.431256
b    1.340309
c   -1.170299
d   -0.226169
e    0.410835
f    0.813850
dtype: float64

In [39]: s1.loc['c':]
Out[39]: 
c   -1.170299
d   -0.226169
e    0.410835
f    0.813850
dtype: float64

In [40]: s1.loc['b']
Out[40]: 1.3403088497993827

Note that setting works as well:

In [41]: s1.loc['c':] = 0

In [42]: s1
Out[42]: 
a    1.431256
b    1.340309
c    0.000000
d    0.000000
e    0.000000
f    0.000000
dtype: float64

With a DataFrame

In [43]: df1 = pd.DataFrame(np.random.randn(6,4),
   ....:                    index=list('abcdef'),
   ....:                    columns=list('ABCD'))
   ....: 

In [44]: df1
Out[44]: 
          A         B         C         D
a  0.132003 -0.827317 -0.076467 -1.187678
b  1.130127 -1.436737 -1.413681  1.607920
c  1.024180  0.569605  0.875906 -2.211372
d  0.974466 -2.006747 -0.410001 -0.078638
e  0.545952 -1.219217 -1.226825  0.769804
f -1.281247 -0.727707 -0.121306 -0.097883

In [45]: df1.loc[['a','b','d'],:]
Out[45]: 
          A         B         C         D
a  0.132003 -0.827317 -0.076467 -1.187678
b  1.130127 -1.436737 -1.413681  1.607920
d  0.974466 -2.006747 -0.410001 -0.078638

Accessing via label slices

In [46]: df1.loc['d':,'A':'C']
Out[46]: 
          A         B         C
d  0.974466 -2.006747 -0.410001
e  0.545952 -1.219217 -1.226825
f -1.281247 -0.727707 -0.121306

For getting a cross section using a label (equiv to df.xs('a'))

In [47]: df1.loc['a']
Out[47]: 
A    0.132003
B   -0.827317
C   -0.076467
D   -1.187678
Name: a, dtype: float64

For getting values with a boolean array

In [48]: df1.loc['a']>0
Out[48]: 
A     True
B    False
C    False
D    False
Name: a, dtype: bool

In [49]: df1.loc[:,df1.loc['a']>0]
Out[49]: 
          A
a  0.132003
b  1.130127
c  1.024180
d  0.974466
e  0.545952
f -1.281247

For getting a value explicitly (equiv to deprecated df.get_value('a','A'))

# this is also equivalent to ``df1.at['a','A']``
In [50]: df1.loc['a','A']
Out[50]: 0.13200317033032932

Selection By Position

Warning

Whether a copy or a reference is returned for a setting operation, may depend on the context. This is sometimes called chained assignment and should be avoided. See Returning a View versus Copy

pandas provides a suite of methods in order to get purely integer based indexing. The semantics follow closely python and numpy slicing. These are 0-based indexing. When slicing, the start bounds is included, while the upper bound is excluded. Trying to use a non-integer, even a valid label will raise a IndexError.

The .iloc attribute is the primary access method. The following are valid inputs:

  • An integer e.g. 5
  • A list or array of integers [4, 3, 0]
  • A slice object with ints 1:7
  • A boolean array
In [51]: s1 = pd.Series(np.random.randn(5), index=list(range(0,10,2)))

In [52]: s1
Out[52]: 
0    0.695775
2    0.341734
4    0.959726
6   -1.110336
8   -0.619976
dtype: float64

In [53]: s1.iloc[:3]
Out[53]: 
0    0.695775
2    0.341734
4    0.959726
dtype: float64

In [54]: s1.iloc[3]
Out[54]: -1.1103361028911669

Note that setting works as well:

In [55]: s1.iloc[:3] = 0

In [56]: s1
Out[56]: 
0    0.000000
2    0.000000
4    0.000000
6   -1.110336
8   -0.619976
dtype: float64

With a DataFrame

In [57]: df1 = pd.DataFrame(np.random.randn(6,4),
   ....:                    index=list(range(0,12,2)),
   ....:                    columns=list(range(0,8,2)))
   ....: 

In [58]: df1
Out[58]: 
           0         2         4         6
0   0.149748 -0.732339  0.687738  0.176444
2   0.403310 -0.154951  0.301624 -2.179861
4  -1.369849 -0.954208  1.462696 -1.743161
6  -0.826591 -0.345352  1.314232  0.690579
8   0.995761  2.396780  0.014871  3.357427
10 -0.317441 -1.236269  0.896171 -0.487602

Select via integer slicing

In [59]: df1.iloc[:3]
Out[59]: 
          0         2         4         6
0  0.149748 -0.732339  0.687738  0.176444
2  0.403310 -0.154951  0.301624 -2.179861
4 -1.369849 -0.954208  1.462696 -1.743161

In [60]: df1.iloc[1:5,2:4]
Out[60]: 
          4         6
2  0.301624 -2.179861
4  1.462696 -1.743161
6  1.314232  0.690579
8  0.014871  3.357427

Select via integer list

In [61]: df1.iloc[[1,3,5],[1,3]]
Out[61]: 
           2         6
2  -0.154951 -2.179861
6  -0.345352  0.690579
10 -1.236269 -0.487602
In [62]: df1.iloc[1:3,:]
Out[62]: 
          0         2         4         6
2  0.403310 -0.154951  0.301624 -2.179861
4 -1.369849 -0.954208  1.462696 -1.743161
In [63]: df1.iloc[:,1:3]
Out[63]: 
           2         4
0  -0.732339  0.687738
2  -0.154951  0.301624
4  -0.954208  1.462696
6  -0.345352  1.314232
8   2.396780  0.014871
10 -1.236269  0.896171
# this is also equivalent to ``df1.iat[1,1]``
In [64]: df1.iloc[1,1]
Out[64]: -0.15495077442490321

For getting a cross section using an integer position (equiv to df.xs(1))

In [65]: df1.iloc[1]
Out[65]: 
0    0.403310
2   -0.154951
4    0.301624
6   -2.179861
Name: 2, dtype: float64

Out of range slice indexes are handled gracefully just as in Python/Numpy.

# these are allowed in python/numpy.
# Only works in Pandas starting from v0.14.0.
In [66]: x = list('abcdef')

In [67]: x
Out[67]: ['a', 'b', 'c', 'd', 'e', 'f']

In [68]: x[4:10]
Out[68]: ['e', 'f']

In [69]: x[8:10]
Out[69]: []

In [70]: s = pd.Series(x)

In [71]: s
Out[71]: 
0    a
1    b
2    c
3    d
4    e
5    f
dtype: object

In [72]: s.iloc[4:10]
Out[72]: 
4    e
5    f
dtype: object

In [73]: s.iloc[8:10]
Out[73]: Series([], dtype: object)

Note

Prior to v0.14.0, iloc would not accept out of bounds indexers for slices, e.g. a value that exceeds the length of the object being indexed.

Note that this could result in an empty axis (e.g. an empty DataFrame being returned)

In [74]: dfl = pd.DataFrame(np.random.randn(5,2), columns=list('AB'))

In [75]: dfl
Out[75]: 
          A         B
0 -0.082240 -2.182937
1  0.380396  0.084844
2  0.432390  1.519970
3 -0.493662  0.600178
4  0.274230  0.132885

In [76]: dfl.iloc[:,2:3]
Out[76]: 
Empty DataFrame
Columns: []
Index: [0, 1, 2, 3, 4]

In [77]: dfl.iloc[:,1:3]
Out[77]: 
          B
0 -2.182937
1  0.084844
2  1.519970
3  0.600178
4  0.132885

In [78]: dfl.iloc[4:6]
Out[78]: 
         A         B
4  0.27423  0.132885

A single indexer that is out of bounds will raise an IndexError. A list of indexers where any element is out of bounds will raise an IndexError

dfl.iloc[[4,5,6]]
IndexError: positional indexers are out-of-bounds

dfl.iloc[:,4]
IndexError: single positional indexer is out-of-bounds

Selecting Random Samples

A random selection of rows or columns from a Series, DataFrame, or Panel with the sample() method. The method will sample rows by default, and accepts a specific number of rows/columns to return, or a fraction of rows.

In [79]: s = pd.Series([0,1,2,3,4,5])

# When no arguments are passed, returns 1 row.
In [80]: s.sample()
Out[80]: 
1    1
dtype: int64

# One may specify either a number of rows:
In [81]: s.sample(n=3)
Out[81]: 
2    2
5    5
4    4
dtype: int64

# Or a fraction of the rows:
In [82]: s.sample(frac=0.5)
Out[82]: 
5    5
0    0
2    2
dtype: int64

By default, sample will return each row at most once, but one can also sample with replacement using the replace option:

In [83]: s = pd.Series([0,1,2,3,4,5])

 # Without replacement (default):
In [84]: s.sample(n=6, replace=False)
Out[84]: 
5    5
4    4
2    2
0    0
1    1
3    3
dtype: int64

 # With replacement:
In [85]: s.sample(n=6, replace=True)
Out[85]: 
1    1
2    2
5    5
5    5
5    5
3    3
dtype: int64

By default, each row has an equal probability of being selected, but if you want rows to have different probabilities, you can pass the sample function sampling weights as weights. These weights can be a list, a numpy array, or a Series, but they must be of the same length as the object you are sampling. Missing values will be treated as a weight of zero, and inf values are not allowed. If weights do not sum to 1, they will be re-normalized by dividing all weights by the sum of the weights. For example:

In [86]: s = pd.Series([0,1,2,3,4,5])

In [87]: example_weights = [0, 0, 0.2, 0.2, 0.2, 0.4]

In [88]: s.sample(n=3, weights=example_weights)
Out[88]: 
2    2
5    5
4    4
dtype: int64

# Weights will be re-normalized automatically
In [89]: example_weights2 = [0.5, 0, 0, 0, 0, 0]

In [90]: s.sample(n=1, weights=example_weights2)
Out[90]: 
0    0
dtype: int64

When applied to a DataFrame, you can use a column of the DataFrame as sampling weights (provided you are sampling rows and not columns) by simply passing the name of the column as a string.

In [91]: df2 = pd.DataFrame({'col1':[9,8,7,6], 'weight_column':[0.5, 0.4, 0.1, 0]})

In [92]: df2.sample(n = 3, weights = 'weight_column')
Out[92]: 
   col1  weight_column
0     9            0.5
2     7            0.1
1     8            0.4

sample also allows users to sample columns instead of rows using the axis argument.

In [93]: df3 = pd.DataFrame({'col1':[1,2,3], 'col2':[2,3,4]})

In [94]: df3.sample(n=1, axis=1)
Out[94]: 
   col2
0     2
1     3
2     4

Finally, one can also set a seed for sample‘s random number generator using the random_state argument, which will accept either an integer (as a seed) or a numpy RandomState object.

In [95]: df4 = pd.DataFrame({'col1':[1,2,3], 'col2':[2,3,4]})

# With a given seed, the sample will always draw the same rows.
In [96]: df4.sample(n=2, random_state=2)
Out[96]: 
   col1  col2
2     3     4
1     2     3

In [97]: df4.sample(n=2, random_state=2)
Out[97]: 
   col1  col2
2     3     4
1     2     3

Setting With Enlargement

New in version 0.13.

The .loc/.ix/[] operations can perform enlargement when setting a non-existant key for that axis.

In the Series case this is effectively an appending operation

In [98]: se = pd.Series([1,2,3])

In [99]: se
Out[99]: 
0    1
1    2
2    3
dtype: int64

In [100]: se[5] = 5.

In [101]: se
Out[101]: 
0    1.0
1    2.0
2    3.0
5    5.0
dtype: float64

A DataFrame can be enlarged on either axis via .loc

In [102]: dfi = pd.DataFrame(np.arange(6).reshape(3,2),
   .....:                 columns=['A','B'])
   .....: 

In [103]: dfi
Out[103]: 
   A  B
0  0  1
1  2  3
2  4  5

In [104]: dfi.loc[:,'C'] = dfi.loc[:,'A']

In [105]: dfi
Out[105]: 
   A  B  C
0  0  1  0
1  2  3  2
2  4  5  4

This is like an append operation on the DataFrame.

In [106]: dfi.loc[3] = 5

In [107]: dfi
Out[107]: 
   A  B  C
0  0  1  0
1  2  3  2
2  4  5  4
3  5  5  5

Fast scalar value getting and setting

Since indexing with [] must handle a lot of cases (single-label access, slicing, boolean indexing, etc.), it has a bit of overhead in order to figure out what you’re asking for. If you only want to access a scalar value, the fastest way is to use the at and iat methods, which are implemented on all of the data structures.

Similarly to loc, at provides label based scalar lookups, while, iat provides integer based lookups analogously to iloc

In [108]: s.iat[5]
Out[108]: 5

In [109]: df.at[dates[5], 'A']
Out[109]: 0.1136484096888855

In [110]: df.iat[3, 0]
Out[110]: -0.70677113363008448

You can also set using these same indexers.

In [111]: df.at[dates[5], 'E'] = 7

In [112]: df.iat[3, 0] = 7

at may enlarge the object in-place as above if the indexer is missing.

In [113]: df.at[dates[-1]+1, 0] = 7

In [114]: df
Out[114]: 
                   A         B         C         D    E    0
2000-01-01 -0.282863  0.469112 -1.509059 -1.135632  NaN  NaN
2000-01-02 -0.173215  1.212112  0.119209 -1.044236  NaN  NaN
2000-01-03 -2.104569 -0.861849 -0.494929  1.071804  NaN  NaN
2000-01-04  7.000000  0.721555 -1.039575  0.271860  NaN  NaN
2000-01-05  0.567020 -0.424972  0.276232 -1.087401  NaN  NaN
2000-01-06  0.113648 -0.673690 -1.478427  0.524988  7.0  NaN
2000-01-07  0.577046  0.404705 -1.715002 -1.039268  NaN  NaN
2000-01-08 -1.157892 -0.370647 -1.344312  0.844885  NaN  NaN
2000-01-09       NaN       NaN       NaN       NaN  NaN  7.0

Boolean indexing

Another common operation is the use of boolean vectors to filter the data. The operators are: | for or, & for and, and ~ for not. These must be grouped by using parentheses.

Using a boolean vector to index a Series works exactly as in a numpy ndarray:

In [115]: s = pd.Series(range(-3, 4))

In [116]: s
Out[116]: 
0   -3
1   -2
2   -1
3    0
4    1
5    2
6    3
dtype: int64

In [117]: s[s > 0]
Out[117]: 
4    1
5    2
6    3
dtype: int64

In [118]: s[(s < -1) | (s > 0.5)]
Out[118]: 
0   -3
1   -2
4    1
5    2
6    3
dtype: int64

In [119]: s[~(s < 0)]
Out[119]: 
3    0
4    1
5    2
6    3
dtype: int64

You may select rows from a DataFrame using a boolean vector the same length as the DataFrame’s index (for example, something derived from one of the columns of the DataFrame):

In [120]: df[df['A'] > 0]
Out[120]: 
                   A         B         C         D    E   0
2000-01-04  7.000000  0.721555 -1.039575  0.271860  NaN NaN
2000-01-05  0.567020 -0.424972  0.276232 -1.087401  NaN NaN
2000-01-06  0.113648 -0.673690 -1.478427  0.524988  7.0 NaN
2000-01-07  0.577046  0.404705 -1.715002 -1.039268  NaN NaN

List comprehensions and map method of Series can also be used to produce more complex criteria:

In [121]: df2 = pd.DataFrame({'a' : ['one', 'one', 'two', 'three', 'two', 'one', 'six'],
   .....:                     'b' : ['x', 'y', 'y', 'x', 'y', 'x', 'x'],
   .....:                     'c' : np.random.randn(7)})
   .....: 

# only want 'two' or 'three'
In [122]: criterion = df2['a'].map(lambda x: x.startswith('t'))

In [123]: df2[criterion]
Out[123]: 
       a  b         c
2    two  y  1.450520
3  three  x  0.206053
4    two  y -0.251905

# equivalent but slower
In [124]: df2[[x.startswith('t') for x in df2['a']]]
Out[124]: 
       a  b         c
2    two  y  1.450520
3  three  x  0.206053
4    two  y -0.251905

# Multiple criteria
In [125]: df2[criterion & (df2['b'] == 'x')]
Out[125]: 
       a  b         c
3  three  x  0.206053

Note, with the choice methods Selection by Label, Selection by Position, and Advanced Indexing you may select along more than one axis using boolean vectors combined with other indexing expressions.

In [126]: df2.loc[criterion & (df2['b'] == 'x'),'b':'c']
Out[126]: 
   b         c
3  x  0.206053

Indexing with isin

Consider the isin method of Series, which returns a boolean vector that is true wherever the Series elements exist in the passed list. This allows you to select rows where one or more columns have values you want:

In [127]: s = pd.Series(np.arange(5), index=np.arange(5)[::-1], dtype='int64')

In [128]: s
Out[128]: 
4    0
3    1
2    2
1    3
0    4
dtype: int64

In [129]: s.isin([2, 4, 6])
Out[129]: 
4    False
3    False
2     True
1    False
0     True
dtype: bool

In [130]: s[s.isin([2, 4, 6])]
Out[130]: 
2    2
0    4
dtype: int64

The same method is available for Index objects and is useful for the cases when you don’t know which of the sought labels are in fact present:

In [131]: s[s.index.isin([2, 4, 6])]
Out[131]: 
4    0
2    2
dtype: int64

# compare it to the following
In [132]: s[[2, 4, 6]]
Out[132]: 
2    2.0
4    0.0
6    NaN
dtype: float64

In addition to that, MultiIndex allows selecting a separate level to use in the membership check:

In [133]: s_mi = pd.Series(np.arange(6),
   .....:                  index=pd.MultiIndex.from_product([[0, 1], ['a', 'b', 'c']]))
   .....: 

In [134]: s_mi
Out[134]: 
0  a    0
   b    1
   c    2
1  a    3
   b    4
   c    5
dtype: int64

In [135]: s_mi.iloc[s_mi.index.isin([(1, 'a'), (2, 'b'), (0, 'c')])]
Out[135]: 
0  c    2
1  a    3
dtype: int64

In [136]: s_mi.iloc[s_mi.index.isin(['a', 'c', 'e'], level=1)]
Out[136]: 
0  a    0
   c    2
1  a    3
   c    5
dtype: int64

DataFrame also has an isin method. When calling isin, pass a set of values as either an array or dict. If values is an array, isin returns a DataFrame of booleans that is the same shape as the original DataFrame, with True wherever the element is in the sequence of values.

In [137]: df = pd.DataFrame({'vals': [1, 2, 3, 4], 'ids': ['a', 'b', 'f', 'n'],
   .....:                    'ids2': ['a', 'n', 'c', 'n']})
   .....: 

In [138]: values = ['a', 'b', 1, 3]

In [139]: df.isin(values)
Out[139]: 
     ids   ids2   vals
0   True   True   True
1   True  False  False
2  False  False   True
3  False  False  False

Oftentimes you’ll want to match certain values with certain columns. Just make values a dict where the key is the column, and the value is a list of items you want to check for.

In [140]: values = {'ids': ['a', 'b'], 'vals': [1, 3]}

In [141]: df.isin(values)
Out[141]: 
     ids   ids2   vals
0   True  False   True
1   True  False  False
2  False  False   True
3  False  False  False

Combine DataFrame’s isin with the any() and all() methods to quickly select subsets of your data that meet a given criteria. To select a row where each column meets its own criterion:

In [142]: values = {'ids': ['a', 'b'], 'ids2': ['a', 'c'], 'vals': [1, 3]}

In [143]: row_mask = df.isin(values).all(1)

In [144]: df[row_mask]
Out[144]: 
  ids ids2  vals
0   a    a     1

The where() Method and Masking

Selecting values from a Series with a boolean vector generally returns a subset of the data. To guarantee that selection output has the same shape as the original data, you can use the where method in Series and DataFrame.

To return only the selected rows

In [145]: s[s > 0]
Out[145]: 
3    1
2    2
1    3
0    4
dtype: int64

To return a Series of the same shape as the original

In [146]: s.where(s > 0)
Out[146]: 
4    NaN
3    1.0
2    2.0
1    3.0
0    4.0
dtype: float64

Selecting values from a DataFrame with a boolean criterion now also preserves input data shape. where is used under the hood as the implementation. Equivalent is df.where(df < 0)

In [147]: df[df < 0]
Out[147]: 
                   A         B         C         D
2000-01-01       NaN       NaN -0.863838       NaN
2000-01-02 -1.048089 -0.025747 -0.988387       NaN
2000-01-03       NaN       NaN       NaN -0.055758
2000-01-04       NaN -0.489682       NaN -0.034571
2000-01-05 -2.484478 -0.281461       NaN       NaN
2000-01-06       NaN -0.977349       NaN -0.064034
2000-01-07 -1.282782       NaN -1.071357       NaN
2000-01-08       NaN       NaN       NaN -0.744471

In addition, where takes an optional other argument for replacement of values where the condition is False, in the returned copy.

In [148]: df.where(df < 0, -df)
Out[148]: 
                   A         B         C         D
2000-01-01 -1.266143 -0.299368 -0.863838 -0.408204
2000-01-02 -1.048089 -0.025747 -0.988387 -0.094055
2000-01-03 -1.262731 -1.289997 -0.082423 -0.055758
2000-01-04 -0.536580 -0.489682 -0.369374 -0.034571
2000-01-05 -2.484478 -0.281461 -0.030711 -0.109121
2000-01-06 -1.126203 -0.977349 -1.474071 -0.064034
2000-01-07 -1.282782 -0.781836 -1.071357 -0.441153
2000-01-08 -2.353925 -0.583787 -0.221471 -0.744471

You may wish to set values based on some boolean criteria. This can be done intuitively like so:

In [149]: s2 = s.copy()

In [150]: s2[s2 < 0] = 0

In [151]: s2
Out[151]: 
4    0
3    1
2    2
1    3
0    4
dtype: int64

In [152]: df2 = df.copy()

In [153]: df2[df2 < 0] = 0

In [154]: df2
Out[154]: 
                   A         B         C         D
2000-01-01  1.266143  0.299368  0.000000  0.408204
2000-01-02  0.000000  0.000000  0.000000  0.094055
2000-01-03  1.262731  1.289997  0.082423  0.000000
2000-01-04  0.536580  0.000000  0.369374  0.000000
2000-01-05  0.000000  0.000000  0.030711  0.109121
2000-01-06  1.126203  0.000000  1.474071  0.000000
2000-01-07  0.000000  0.781836  0.000000  0.441153
2000-01-08  2.353925  0.583787  0.221471  0.000000

By default, where returns a modified copy of the data. There is an optional parameter inplace so that the original data can be modified without creating a copy:

In [155]: df_orig = df.copy()

In [156]: df_orig.where(df > 0, -df, inplace=True);

In [157]: df_orig
Out[157]: 
                   A         B         C         D
2000-01-01  1.266143  0.299368  0.863838  0.408204
2000-01-02  1.048089  0.025747  0.988387  0.094055
2000-01-03  1.262731  1.289997  0.082423  0.055758
2000-01-04  0.536580  0.489682  0.369374  0.034571
2000-01-05  2.484478  0.281461  0.030711  0.109121
2000-01-06  1.126203  0.977349  1.474071  0.064034
2000-01-07  1.282782  0.781836  1.071357  0.441153
2000-01-08  2.353925  0.583787  0.221471  0.744471

alignment

Furthermore, where aligns the input boolean condition (ndarray or DataFrame), such that partial selection with setting is possible. This is analogous to partial setting via .ix (but on the contents rather than the axis labels)

In [158]: df2 = df.copy()

In [159]: df2[ df2[1:4] > 0 ] = 3

In [160]: df2
Out[160]: 
                   A         B         C         D
2000-01-01  1.266143  0.299368 -0.863838  0.408204
2000-01-02 -1.048089 -0.025747 -0.988387  3.000000
2000-01-03  3.000000  3.000000  3.000000 -0.055758
2000-01-04  3.000000 -0.489682  3.000000 -0.034571
2000-01-05 -2.484478 -0.281461  0.030711  0.109121
2000-01-06  1.126203 -0.977349  1.474071 -0.064034
2000-01-07 -1.282782  0.781836 -1.071357  0.441153
2000-01-08  2.353925  0.583787  0.221471 -0.744471

New in version 0.13.

Where can also accept axis and level parameters to align the input when performing the where.

In [161]: df2 = df.copy()

In [162]: df2.where(df2>0,df2['A'],axis='index')
Out[162]: 
                   A         B         C         D
2000-01-01  1.266143  0.299368  1.266143  0.408204
2000-01-02 -1.048089 -1.048089 -1.048089  0.094055
2000-01-03  1.262731  1.289997  0.082423  1.262731
2000-01-04  0.536580  0.536580  0.369374  0.536580
2000-01-05 -2.484478 -2.484478  0.030711  0.109121
2000-01-06  1.126203  1.126203  1.474071  1.126203
2000-01-07 -1.282782  0.781836 -1.282782  0.441153
2000-01-08  2.353925  0.583787  0.221471  2.353925

This is equivalent (but faster than) the following.

In [163]: df2 = df.copy()

In [164]: df.apply(lambda x, y: x.where(x>0,y), y=df['A'])
Out[164]: 
                   A         B         C         D
2000-01-01  1.266143  0.299368  1.266143  0.408204
2000-01-02 -1.048089 -1.048089 -1.048089  0.094055
2000-01-03  1.262731  1.289997  0.082423  1.262731
2000-01-04  0.536580  0.536580  0.369374  0.536580
2000-01-05 -2.484478 -2.484478  0.030711  0.109121
2000-01-06  1.126203  1.126203  1.474071  1.126203
2000-01-07 -1.282782  0.781836 -1.282782  0.441153
2000-01-08  2.353925  0.583787  0.221471  2.353925

mask

mask is the inverse boolean operation of where.

In [165]: s.mask(s >= 0)
Out[165]: 
4   NaN
3   NaN
2   NaN
1   NaN
0   NaN
dtype: float64

In [166]: df.mask(df >= 0)
Out[166]: 
                   A         B         C         D
2000-01-01       NaN       NaN -0.863838       NaN
2000-01-02 -1.048089 -0.025747 -0.988387       NaN
2000-01-03       NaN       NaN       NaN -0.055758
2000-01-04       NaN -0.489682       NaN -0.034571
2000-01-05 -2.484478 -0.281461       NaN       NaN
2000-01-06       NaN -0.977349       NaN -0.064034
2000-01-07 -1.282782       NaN -1.071357       NaN
2000-01-08       NaN       NaN       NaN -0.744471

The query() Method (Experimental)

New in version 0.13.

DataFrame objects have a query() method that allows selection using an expression.

You can get the value of the frame where column b has values between the values of columns a and c. For example:

In [167]: n = 10

In [168]: df = pd.DataFrame(np.random.rand(n, 3), columns=list('abc'))

In [169]: df
Out[169]: 
          a         b         c
0  0.687704  0.582314  0.281645
1  0.250846  0.610021  0.420121
2  0.624328  0.401816  0.932146
3  0.011763  0.022921  0.244186
4  0.590198  0.325680  0.890392
5  0.598892  0.296424  0.007312
6  0.634625  0.803069  0.123872
7  0.924168  0.325076  0.303746
8  0.116822  0.364564  0.454607
9  0.986142  0.751953  0.561512

# pure python
In [170]: df[(df.a < df.b) & (df.b < df.c)]
Out[170]: 
          a         b         c
3  0.011763  0.022921  0.244186
8  0.116822  0.364564  0.454607

# query
In [171]: df.query('(a < b) & (b < c)')
Out[171]: 
          a         b         c
3  0.011763  0.022921  0.244186
8  0.116822  0.364564  0.454607

Do the same thing but fall back on a named index if there is no column with the name a.

In [172]: df = pd.DataFrame(np.random.randint(n / 2, size=(n, 2)), columns=list('bc'))

In [173]: df.index.name = 'a'

In [174]: df
Out[174]: 
   b  c
a      
0  0  4
1  2  2
2  1  4
3  3  3
4  0  4
5  1  1
6  0  3
7  2  4
8  3  1
9  3  2

In [175]: df.query('a < b and b < c')
Out[175]: 
Empty DataFrame
Columns: [b, c]
Index: []

If instead you don’t want to or cannot name your index, you can use the name index in your query expression:

In [176]: df = pd.DataFrame(np.random.randint(n, size=(n, 2)), columns=list('bc'))

In [177]: df
Out[177]: 
   b  c
0  9  8
1  0  5
2  4  7
3  2  0
4  5  1
5  4  0
6  2  0
7  6  7
8  3  4
9  9  7

In [178]: df.query('index < b < c')
Out[178]: 
   b  c
2  4  7

Note

If the name of your index overlaps with a column name, the column name is given precedence. For example,

In [179]: df = pd.DataFrame({'a': np.random.randint(5, size=5)})

In [180]: df.index.name = 'a'

In [181]: df.query('a > 2') # uses the column 'a', not the index
Out[181]: 
   a
a   
2  4
3  4
4  3

You can still use the index in a query expression by using the special identifier ‘index’:

In [182]: df.query('index > 2')
Out[182]: 
   a
a   
3  4
4  3

If for some reason you have a column named index, then you can refer to the index as ilevel_0 as well, but at this point you should consider renaming your columns to something less ambiguous.

MultiIndex query() Syntax

You can also use the levels of a DataFrame with a MultiIndex as if they were columns in the frame:

In [183]: n = 10

In [184]: colors = np.random.choice(['red', 'green'], size=n)

In [185]: foods = np.random.choice(['eggs', 'ham'], size=n)

In [186]: colors
Out[186]: 
array(['green', 'green', 'red', 'red', 'green', 'green', 'red', 'green',
       'red', 'green'], 
      dtype='|S5')

In [187]: foods
Out[187]: 
array(['eggs', 'eggs', 'ham', 'ham', 'eggs', 'ham', 'eggs', 'ham', 'eggs',
       'eggs'], 
      dtype='|S4')

In [188]: index = pd.MultiIndex.from_arrays([colors, foods], names=['color', 'food'])

In [189]: df = pd.DataFrame(np.random.randn(n, 2), index=index)

In [190]: df
Out[190]: 
                   0         1
color food                    
green eggs  1.555563 -0.823761
      eggs  0.535420 -1.032853
red   ham   1.469725  1.304124
      ham   1.449735  0.203109
green eggs -1.032011  0.969818
      ham  -0.962723  1.382083
red   eggs -0.938794  0.669142
green ham  -0.433567 -0.273610
red   eggs  0.680433 -0.308450
green eggs -0.276099 -1.821168

In [191]: df.query('color == "red"')
Out[191]: 
                   0         1
color food                    
red   ham   1.469725  1.304124
      ham   1.449735  0.203109
      eggs -0.938794  0.669142
      eggs  0.680433 -0.308450

If the levels of the MultiIndex are unnamed, you can refer to them using special names:

In [192]: df.index.names = [None, None]

In [193]: df
Out[193]: 
                   0         1
green eggs  1.555563 -0.823761
      eggs  0.535420 -1.032853
red   ham   1.469725  1.304124
      ham   1.449735  0.203109
green eggs -1.032011  0.969818
      ham  -0.962723  1.382083
red   eggs -0.938794  0.669142
green ham  -0.433567 -0.273610
red   eggs  0.680433 -0.308450
green eggs -0.276099 -1.821168

In [194]: df.query('ilevel_0 == "red"')
Out[194]: 
                 0         1
red ham   1.469725  1.304124
    ham   1.449735  0.203109
    eggs -0.938794  0.669142
    eggs  0.680433 -0.308450

The convention is ilevel_0, which means “index level 0” for the 0th level of the index.

query() Use Cases

A use case for query() is when you have a collection of DataFrame objects that have a subset of column names (or index levels/names) in common. You can pass the same query to both frames without having to specify which frame you’re interested in querying

In [195]: df = pd.DataFrame(np.random.rand(n, 3), columns=list('abc'))

In [196]: df
Out[196]: 
          a         b         c
0  0.449167  0.447421  0.557792
1  0.427877  0.690714  0.035221
2  0.543944  0.507917  0.468025
3  0.973561  0.727371  0.700486
4  0.515391  0.337711  0.107630
5  0.697879  0.541870  0.043587
6  0.074629  0.717019  0.661852
7  0.524683  0.446253  0.066302
8  0.744757  0.361155  0.205344
9  0.948737  0.555495  0.357893

In [197]: df2 = pd.DataFrame(np.random.rand(n + 2, 3), columns=df.columns)

In [198]: df2
Out[198]: 
           a         b         c
0   0.164291  0.785148  0.847700
1   0.953696  0.111044  0.252125
2   0.033633  0.892920  0.037910
3   0.662580  0.219559  0.243745
4   0.885036  0.790218  0.224283
5   0.736107  0.139168  0.302827
6   0.657803  0.713897  0.611185
7   0.136624  0.984960  0.195246
8   0.123436  0.627712  0.618673
9   0.371660  0.047902  0.480088
10  0.062993  0.185760  0.568018
11  0.483467  0.445289  0.309040

In [199]: expr = '0.0 <= a <= c <= 0.5'

In [200]: map(lambda frame: frame.query(expr), [df, df2])
Out[200]: 
[Empty DataFrame
 Columns: [a, b, c]
 Index: [],           a         b         c
 2  0.033633  0.892920  0.037910
 7  0.136624  0.984960  0.195246
 9  0.371660  0.047902  0.480088]

query() Python versus pandas Syntax Comparison

Full numpy-like syntax

In [201]: df = pd.DataFrame(np.random.randint(n, size=(n, 3)), columns=list('abc'))

In [202]: df
Out[202]: 
   a  b  c
0  8  8  1
1  1  1  9
2  9  9  8
3  7  0  1
4  8  6  6
5  8  2  8
6  3  2  2
7  9  4  8
8  7  2  7
9  9  4  8

In [203]: df.query('(a < b) & (b < c)')
Out[203]: 
Empty DataFrame
Columns: [a, b, c]
Index: []

In [204]: df[(df.a < df.b) & (df.b < df.c)]
Out[204]: 
Empty DataFrame
Columns: [a, b, c]
Index: []

Slightly nicer by removing the parentheses (by binding making comparison operators bind tighter than &/|)

In [205]: df.query('a < b & b < c')
Out[205]: 
Empty DataFrame
Columns: [a, b, c]
Index: []

Use English instead of symbols

In [206]: df.query('a < b and b < c')
Out[206]: 
Empty DataFrame
Columns: [a, b, c]
Index: []

Pretty close to how you might write it on paper

In [207]: df.query('a < b < c')
Out[207]: 
Empty DataFrame
Columns: [a, b, c]
Index: []

The in and not in operators

query() also supports special use of Python’s in and not in comparison operators, providing a succinct syntax for calling the isin method of a Series or DataFrame.

# get all rows where columns "a" and "b" have overlapping values
In [208]: df = pd.DataFrame({'a': list('aabbccddeeff'), 'b': list('aaaabbbbcccc'),
   .....:                    'c': np.random.randint(5, size=12),
   .....:                    'd': np.random.randint(9, size=12)})
   .....: 

In [209]: df
Out[209]: 
    a  b  c  d
0   a  a  1  7
1   a  a  0  1
2   b  a  2  0
3   b  a  1  7
4   c  b  0  8
5   c  b  1  1
6   d  b  1  0
7   d  b  1  7
8   e  c  4  2
9   e  c  2  7
10  f  c  2  2
11  f  c  4  6

In [210]: df.query('a in b')
Out[210]: 
   a  b  c  d
0  a  a  1  7
1  a  a  0  1
2  b  a  2  0
3  b  a  1  7
4  c  b  0  8
5  c  b  1  1

# How you'd do it in pure Python
In [211]: df[df.a.isin(df.b)]
Out[211]: 
   a  b  c  d
0  a  a  1  7
1  a  a  0  1
2  b  a  2  0
3  b  a  1  7
4  c  b  0  8
5  c  b  1  1

In [212]: df.query('a not in b')
Out[212]: 
    a  b  c  d
6   d  b  1  0
7   d  b  1  7
8   e  c  4  2
9   e  c  2  7
10  f  c  2  2
11  f  c  4  6

# pure Python
In [213]: df[~df.a.isin(df.b)]
Out[213]: 
    a  b  c  d
6   d  b  1  0
7   d  b  1  7
8   e  c  4  2
9   e  c  2  7
10  f  c  2  2
11  f  c  4  6

You can combine this with other expressions for very succinct queries:

# rows where cols a and b have overlapping values and col c's values are less than col d's
In [214]: df.query('a in b and c < d')
Out[214]: 
   a  b  c  d
0  a  a  1  7
1  a  a  0  1
3  b  a  1  7
4  c  b  0  8

# pure Python
In [215]: df[df.b.isin(df.a) & (df.c < df.d)]
Out[215]: 
    a  b  c  d
0   a  a  1  7
1   a  a  0  1
3   b  a  1  7
4   c  b  0  8
7   d  b  1  7
9   e  c  2  7
11  f  c  4  6

Note

Note that in and not in are evaluated in Python, since numexpr has no equivalent of this operation. However, only the in/not in expression itself is evaluated in vanilla Python. For example, in the expression

df.query('a in b + c + d')

(b + c + d) is evaluated by numexpr and then the in operation is evaluated in plain Python. In general, any operations that can be evaluated using numexpr will be.

Special use of the == operator with list objects

Comparing a list of values to a column using ==/!= works similarly to in/not in

In [216]: df.query('b == ["a", "b", "c"]')
Out[216]: 
    a  b  c  d
0   a  a  1  7
1   a  a  0  1
2   b  a  2  0
3   b  a  1  7
4   c  b  0  8
5   c  b  1  1
6   d  b  1  0
7   d  b  1  7
8   e  c  4  2
9   e  c  2  7
10  f  c  2  2
11  f  c  4  6

# pure Python
In [217]: df[df.b.isin(["a", "b", "c"])]
Out[217]: 
    a  b  c  d
0   a  a  1  7
1   a  a  0  1
2   b  a  2  0
3   b  a  1  7
4   c  b  0  8
5   c  b  1  1
6   d  b  1  0
7   d  b  1  7
8   e  c  4  2
9   e  c  2  7
10  f  c  2  2
11  f  c  4  6

In [218]: df.query('c == [1, 2]')
Out[218]: 
    a  b  c  d
0   a  a  1  7
2   b  a  2  0
3   b  a  1  7
5   c  b  1  1
6   d  b  1  0
7   d  b  1  7
9   e  c  2  7
10  f  c  2  2

In [219]: df.query('c != [1, 2]')
Out[219]: 
    a  b  c  d
1   a  a  0  1
4   c  b  0  8
8   e  c  4  2
11  f  c  4  6

# using in/not in
In [220]: df.query('[1, 2] in c')
Out[220]: 
    a  b  c  d
0   a  a  1  7
2   b  a  2  0
3   b  a  1  7
5   c  b  1  1
6   d  b  1  0
7   d  b  1  7
9   e  c  2  7
10  f  c  2  2

In [221]: df.query('[1, 2] not in c')
Out[221]: 
    a  b  c  d
1   a  a  0  1
4   c  b  0  8
8   e  c  4  2
11  f  c  4  6

# pure Python
In [222]: df[df.c.isin([1, 2])]
Out[222]: 
    a  b  c  d
0   a  a  1  7
2   b  a  2  0
3   b  a  1  7
5   c  b  1  1
6   d  b  1  0
7   d  b  1  7
9   e  c  2  7
10  f  c  2  2

Boolean Operators

You can negate boolean expressions with the word not or the ~ operator.

In [223]: df = pd.DataFrame(np.random.rand(n, 3), columns=list('abc'))

In [224]: df['bools'] = np.random.rand(len(df)) > 0.5

In [225]: df.query('~bools')
Out[225]: 
          a         b         c  bools
1  0.023729  0.458852  0.284632  False
2  0.004236  0.442347  0.767020  False
3  0.890592  0.568747  0.332692  False
5  0.843864  0.902394  0.538418  False
6  0.428277  0.870859  0.380827  False
7  0.321819  0.559198  0.382567  False
8  0.119041  0.715950  0.448156  False
9  0.606506  0.300726  0.666418  False

In [226]: df.query('not bools')
Out[226]: 
          a         b         c  bools
1  0.023729  0.458852  0.284632  False
2  0.004236  0.442347  0.767020  False
3  0.890592  0.568747  0.332692  False
5  0.843864  0.902394  0.538418  False
6  0.428277  0.870859  0.380827  False
7  0.321819  0.559198  0.382567  False
8  0.119041  0.715950  0.448156  False
9  0.606506  0.300726  0.666418  False

In [227]: df.query('not bools') == df[~df.bools]
Out[227]: 
      a     b     c bools
1  True  True  True  True
2  True  True  True  True
3  True  True  True  True
5  True  True  True  True
6  True  True  True  True
7  True  True  True  True
8  True  True  True  True
9  True  True  True  True

Of course, expressions can be arbitrarily complex too

# short query syntax
In [228]: shorter = df.query('a < b < c and (not bools) or bools > 2')

# equivalent in pure Python
In [229]: longer = df[(df.a < df.b) & (df.b < df.c) & (~df.bools) | (df.bools > 2)]

In [230]: shorter
Out[230]: 
          a         b        c  bools
2  0.004236  0.442347  0.76702  False

In [231]: longer
Out[231]: 
          a         b        c  bools
2  0.004236  0.442347  0.76702  False

In [232]: shorter == longer
Out[232]: 
      a     b     c bools
2  True  True  True  True

Performance of query()

DataFrame.query() using numexpr is slightly faster than Python for large frames

_images/query-perf.png

Note

You will only see the performance benefits of using the numexpr engine with DataFrame.query() if your frame has more than approximately 200,000 rows

_images/query-perf-small.png

This plot was created using a DataFrame with 3 columns each containing floating point values generated using numpy.random.randn().

Duplicate Data

If you want to identify and remove duplicate rows in a DataFrame, there are two methods that will help: duplicated and drop_duplicates. Each takes as an argument the columns to use to identify duplicated rows.

  • duplicated returns a boolean vector whose length is the number of rows, and which indicates whether a row is duplicated.
  • drop_duplicates removes duplicate rows.

By default, the first observed row of a duplicate set is considered unique, but each method has a keep parameter to specify targets to be kept.

  • keep='first' (default): mark / drop duplicates except for the first occurrence.
  • keep='last': mark / drop duplicates except for the last occurrence.
  • keep=False: mark / drop all duplicates.
In [233]: df2 = pd.DataFrame({'a': ['one', 'one', 'two', 'two', 'two', 'three', 'four'],
   .....:                     'b': ['x', 'y', 'x', 'y', 'x', 'x', 'x'],
   .....:                     'c': np.random.randn(7)})
   .....: 

In [234]: df2
Out[234]: 
       a  b         c
0    one  x -0.599770
1    one  y -0.120145
2    two  x -1.403265
3    two  y  0.162169
4    two  x  0.232777
5  three  x  1.048589
6   four  x  0.070084

In [235]: df2.duplicated('a')
Out[235]: 
0    False
1     True
2    False
3     True
4     True
5    False
6    False
dtype: bool

In [236]: df2.duplicated('a', keep='last')
Out[236]: 
0     True
1    False
2     True
3     True
4    False
5    False
6    False
dtype: bool

In [237]: df2.duplicated('a', keep=False)
Out[237]: 
0     True
1     True
2     True
3     True
4     True
5    False
6    False
dtype: bool

In [238]: df2.drop_duplicates('a')
Out[238]: 
       a  b         c
0    one  x -0.599770
2    two  x -1.403265
5  three  x  1.048589
6   four  x  0.070084

In [239]: df2.drop_duplicates('a', keep='last')
Out[239]: 
       a  b         c
1    one  y -0.120145
4    two  x  0.232777
5  three  x  1.048589
6   four  x  0.070084

In [240]: df2.drop_duplicates('a', keep=False)
Out[240]: 
       a  b         c
5  three  x  1.048589
6   four  x  0.070084

Also, you can pass a list of columns to identify duplications.

In [241]: df2.duplicated(['a', 'b'])
Out[241]: 
0    False
1    False
2    False
3    False
4     True
5    False
6    False
dtype: bool

In [242]: df2.drop_duplicates(['a', 'b'])
Out[242]: 
       a  b         c
0    one  x -0.599770
1    one  y -0.120145
2    two  x -1.403265
3    two  y  0.162169
5  three  x  1.048589
6   four  x  0.070084

To drop duplicates by index value, use Index.duplicated then perform slicing. Same options are available in keep parameter.

In [243]: df3 = pd.DataFrame({'a': np.arange(6),
   .....:                     'b': np.random.randn(6)},
   .....:                    index=['a', 'a', 'b', 'c', 'b', 'a'])
   .....: 

In [244]: df3
Out[244]: 
   a         b
a  0 -0.774247
a  1 -0.436052
b  2  0.274093
c  3  1.386100
b  4  0.705452
a  5  0.515768

In [245]: df3.index.duplicated()
Out[245]: array([False,  True, False, False,  True,  True], dtype=bool)

In [246]: df3[~df3.index.duplicated()]
Out[246]: 
   a         b
a  0 -0.774247
b  2  0.274093
c  3  1.386100

In [247]: df3[~df3.index.duplicated(keep='last')]
Out[247]: 
   a         b
c  3  1.386100
b  4  0.705452
a  5  0.515768

In [248]: df3[~df3.index.duplicated(keep=False)]
Out[248]: 
   a       b
c  3  1.3861

Dictionary-like get() method

Each of Series, DataFrame, and Panel have a get method which can return a default value.

In [249]: s = pd.Series([1,2,3], index=['a','b','c'])

In [250]: s.get('a')               # equivalent to s['a']
Out[250]: 1

In [251]: s.get('x', default=-1)
Out[251]: -1

The select() Method

Another way to extract slices from an object is with the select method of Series, DataFrame, and Panel. This method should be used only when there is no more direct way. select takes a function which operates on labels along axis and returns a boolean. For instance:

In [252]: df.select(lambda x: x == 'A', axis=1)
Out[252]: 
                   A
2000-01-01 -0.226151
2000-01-02  1.380589
2000-01-03  1.185125
2000-01-04 -0.124870
2000-01-05 -0.926415
2000-01-06 -0.447053
2000-01-07  0.183787
2000-01-08 -1.544483

The lookup() Method

Sometimes you want to extract a set of values given a sequence of row labels and column labels, and the lookup method allows for this and returns a numpy array. For instance,

In [253]: dflookup = pd.DataFrame(np.random.rand(20,4), columns = ['A','B','C','D'])

In [254]: dflookup.lookup(list(range(0,10,2)), ['B','C','A','B','D'])
Out[254]: array([ 0.9242,  0.3338,  0.7562,  0.6023,  0.7212])

Index objects

The pandas Index class and its subclasses can be viewed as implementing an ordered multiset. Duplicates are allowed. However, if you try to convert an Index object with duplicate entries into a set, an exception will be raised.

Index also provides the infrastructure necessary for lookups, data alignment, and reindexing. The easiest way to create an Index directly is to pass a list or other sequence to Index:

In [255]: index = pd.Index(['e', 'd', 'a', 'b'])

In [256]: index
Out[256]: Index([u'e', u'd', u'a', u'b'], dtype='object')

In [257]: 'd' in index
Out[257]: True

You can also pass a name to be stored in the index:

In [258]: index = pd.Index(['e', 'd', 'a', 'b'], name='something')

In [259]: index.name
Out[259]: 'something'

The name, if set, will be shown in the console display:

In [260]: index = pd.Index(list(range(5)), name='rows')

In [261]: columns = pd.Index(['A', 'B', 'C'], name='cols')

In [262]: df = pd.DataFrame(np.random.randn(5, 3), index=index, columns=columns)

In [263]: df
Out[263]: 
cols         A         B         C
rows                              
0     0.352642  1.996218 -1.293063
1    -0.393590  0.034083  0.293336
2    -0.448824  0.224249  0.709838
3    -0.331820 -0.548158 -1.313391
4    -1.549151  0.635278 -1.103599

In [264]: df['A']
Out[264]: 
rows
0    0.352642
1   -0.393590
2   -0.448824
3   -0.331820
4   -1.549151
Name: A, dtype: float64

Setting metadata

New in version 0.13.0.

Indexes are “mostly immutable”, but it is possible to set and change their metadata, like the index name (or, for MultiIndex, levels and labels).

You can use the rename, set_names, set_levels, and set_labels to set these attributes directly. They default to returning a copy; however, you can specify inplace=True to have the data change in place.

See Advanced Indexing for usage of MultiIndexes.

In [265]: ind = pd.Index([1, 2, 3])

In [266]: ind.rename("apple")
Out[266]: Int64Index([1, 2, 3], dtype='int64', name=u'apple')

In [267]: ind
Out[267]: Int64Index([1, 2, 3], dtype='int64')

In [268]: ind.set_names(["apple"], inplace=True)

In [269]: ind.name = "bob"

In [270]: ind
Out[270]: Int64Index([1, 2, 3], dtype='int64', name=u'bob')

New in version 0.15.0.

set_names, set_levels, and set_labels also take an optional level` argument

In [271]: index = pd.MultiIndex.from_product([range(3), ['one', 'two']], names=['first', 'second'])

In [272]: index
Out[272]: 
MultiIndex(levels=[[0, 1, 2], [u'one', u'two']],
           labels=[[0, 0, 1, 1, 2, 2], [0, 1, 0, 1, 0, 1]],
           names=[u'first', u'second'])

In [273]: index.levels[1]
Out[273]: Index([u'one', u'two'], dtype='object', name=u'second')

In [274]: index.set_levels(["a", "b"], level=1)
Out[274]: 
MultiIndex(levels=[[0, 1, 2], [u'a', u'b']],
           labels=[[0, 0, 1, 1, 2, 2], [0, 1, 0, 1, 0, 1]],
           names=[u'first', u'second'])

Set operations on Index objects

Warning

In 0.15.0. the set operations + and - were deprecated in order to provide these for numeric type operations on certain index types. + can be replace by .union() or |, and - by .difference().

The two main operations are union (|), intersection (&) These can be directly called as instance methods or used via overloaded operators. Difference is provided via the .difference() method.

In [275]: a = pd.Index(['c', 'b', 'a'])

In [276]: b = pd.Index(['c', 'e', 'd'])

In [277]: a | b
Out[277]: Index([u'a', u'b', u'c', u'd', u'e'], dtype='object')

In [278]: a & b
Out[278]: Index([u'c'], dtype='object')

In [279]: a.difference(b)
Out[279]: Index([u'a', u'b'], dtype='object')

Also available is the sym_diff (^) operation, which returns elements that appear in either idx1 or idx2 but not both. This is equivalent to the Index created by idx1.difference(idx2).union(idx2.difference(idx1)), with duplicates dropped.

In [280]: idx1 = pd.Index([1, 2, 3, 4])

In [281]: idx2 = pd.Index([2, 3, 4, 5])

In [282]: idx1.sym_diff(idx2)
Out[282]: Int64Index([1, 5], dtype='int64')

In [283]: idx1 ^ idx2
Out[283]: Int64Index([1, 5], dtype='int64')

Missing values

New in version 0.17.1.

Important

Even though Index can hold missing values (NaN), it should be avoided if you do not want any unexpected results. For example, some operations exclude missing values implicitly.

Index.fillna fills missing values with specified scalar value.

In [284]: idx1 = pd.Index([1, np.nan, 3, 4])

In [285]: idx1
Out[285]: Float64Index([1.0, nan, 3.0, 4.0], dtype='float64')

In [286]: idx1.fillna(2)
Out[286]: Float64Index([1.0, 2.0, 3.0, 4.0], dtype='float64')

In [287]: idx2 = pd.DatetimeIndex([pd.Timestamp('2011-01-01'), pd.NaT, pd.Timestamp('2011-01-03')])

In [288]: idx2
Out[288]: DatetimeIndex(['2011-01-01', 'NaT', '2011-01-03'], dtype='datetime64[ns]', freq=None)

In [289]: idx2.fillna(pd.Timestamp('2011-01-02'))
Out[289]: DatetimeIndex(['2011-01-01', '2011-01-02', '2011-01-03'], dtype='datetime64[ns]', freq=None)

Set / Reset Index

Occasionally you will load or create a data set into a DataFrame and want to add an index after you’ve already done so. There are a couple of different ways.

Set an index

DataFrame has a set_index method which takes a column name (for a regular Index) or a list of column names (for a MultiIndex), to create a new, indexed DataFrame:

In [290]: data
Out[290]: 
     a    b  c    d
0  bar  one  z  1.0
1  bar  two  y  2.0
2  foo  one  x  3.0
3  foo  two  w  4.0

In [291]: indexed1 = data.set_index('c')

In [292]: indexed1
Out[292]: 
     a    b    d
c               
z  bar  one  1.0
y  bar  two  2.0
x  foo  one  3.0
w  foo  two  4.0

In [293]: indexed2 = data.set_index(['a', 'b'])

In [294]: indexed2
Out[294]: 
         c    d
a   b          
bar one  z  1.0
    two  y  2.0
foo one  x  3.0
    two  w  4.0

The append keyword option allow you to keep the existing index and append the given columns to a MultiIndex:

In [295]: frame = data.set_index('c', drop=False)

In [296]: frame = frame.set_index(['a', 'b'], append=True)

In [297]: frame
Out[297]: 
           c    d
c a   b          
z bar one  z  1.0
y bar two  y  2.0
x foo one  x  3.0
w foo two  w  4.0

Other options in set_index allow you not drop the index columns or to add the index in-place (without creating a new object):

In [298]: data.set_index('c', drop=False)
Out[298]: 
     a    b  c    d
c                  
z  bar  one  z  1.0
y  bar  two  y  2.0
x  foo  one  x  3.0
w  foo  two  w  4.0

In [299]: data.set_index(['a', 'b'], inplace=True)

In [300]: data
Out[300]: 
         c    d
a   b          
bar one  z  1.0
    two  y  2.0
foo one  x  3.0
    two  w  4.0

Reset the index

As a convenience, there is a new function on DataFrame called reset_index which transfers the index values into the DataFrame’s columns and sets a simple integer index. This is the inverse operation to set_index

In [301]: data
Out[301]: 
         c    d
a   b          
bar one  z  1.0
    two  y  2.0
foo one  x  3.0
    two  w  4.0

In [302]: data.reset_index()
Out[302]: 
     a    b  c    d
0  bar  one  z  1.0
1  bar  two  y  2.0
2  foo  one  x  3.0
3  foo  two  w  4.0

The output is more similar to a SQL table or a record array. The names for the columns derived from the index are the ones stored in the names attribute.

You can use the level keyword to remove only a portion of the index:

In [303]: frame
Out[303]: 
           c    d
c a   b          
z bar one  z  1.0
y bar two  y  2.0
x foo one  x  3.0
w foo two  w  4.0

In [304]: frame.reset_index(level=1)
Out[304]: 
         a  c    d
c b               
z one  bar  z  1.0
y two  bar  y  2.0
x one  foo  x  3.0
w two  foo  w  4.0

reset_index takes an optional parameter drop which if true simply discards the index, instead of putting index values in the DataFrame’s columns.

Note

The reset_index method used to be called delevel which is now deprecated.

Adding an ad hoc index

If you create an index yourself, you can just assign it to the index field:

data.index = index

Returning a view versus a copy

When setting values in a pandas object, care must be taken to avoid what is called chained indexing. Here is an example.

In [305]: dfmi = pd.DataFrame([list('abcd'),
   .....:                      list('efgh'),
   .....:                      list('ijkl'),
   .....:                      list('mnop')],
   .....:                     columns=pd.MultiIndex.from_product([['one','two'],
   .....:                                                         ['first','second']]))
   .....: 

In [306]: dfmi
Out[306]: 
    one          two       
  first second first second
0     a      b     c      d
1     e      f     g      h
2     i      j     k      l
3     m      n     o      p

Compare these two access methods:

In [307]: dfmi['one']['second']
Out[307]: 
0    b
1    f
2    j
3    n
Name: second, dtype: object
In [308]: dfmi.loc[:,('one','second')]
Out[308]: 
0    b
1    f
2    j
3    n
Name: (one, second), dtype: object

These both yield the same results, so which should you use? It is instructive to understand the order of operations on these and why method 2 (.loc) is much preferred over method 1 (chained [])

dfmi['one'] selects the first level of the columns and returns a DataFrame that is singly-indexed. Then another python operation dfmi_with_one['second'] selects the series indexed by 'second' happens. This is indicated by the variable dfmi_with_one because pandas sees these operations as separate events. e.g. separate calls to __getitem__, so it has to treat them as linear operations, they happen one after another.

Contrast this to df.loc[:,('one','second')] which passes a nested tuple of (slice(None),('one','second')) to a single call to __getitem__. This allows pandas to deal with this as a single entity. Furthermore this order of operations can be significantly faster, and allows one to index both axes if so desired.

Why does assignment fail when using chained indexing?

The problem in the previous section is just a performance issue. What’s up with the SettingWithCopy warning? We don’t usually throw warnings around when you do something that might cost a few extra milliseconds!

But it turns out that assigning to the product of chained indexing has inherently unpredictable results. To see this, think about how the Python interpreter executes this code:

dfmi.loc[:,('one','second')] = value
# becomes
dfmi.loc.__setitem__((slice(None), ('one', 'second')), value)

But this code is handled differently:

dfmi['one']['second'] = value
# becomes
dfmi.__getitem__('one').__setitem__('second', value)

See that __getitem__ in there? Outside of simple cases, it’s very hard to predict whether it will return a view or a copy (it depends on the memory layout of the array, about which pandas makes no guarantees), and therefore whether the __setitem__ will modify dfmi or a temporary object that gets thrown out immediately afterward. That’s what SettingWithCopy is warning you about!

Note

You may be wondering whether we should be concerned about the loc property in the first example. But dfmi.loc is guaranteed to be dfmi itself with modified indexing behavior, so dfmi.loc.__getitem__ / dfmi.loc.__setitem__ operate on dfmi directly. Of course, dfmi.loc.__getitem__(idx) may be a view or a copy of dfmi.

Sometimes a SettingWithCopy warning will arise at times when there’s no obvious chained indexing going on. These are the bugs that SettingWithCopy is designed to catch! Pandas is probably trying to warn you that you’ve done this:

def do_something(df):
   foo = df[['bar', 'baz']]  # Is foo a view? A copy? Nobody knows!
   # ... many lines here ...
   foo['quux'] = value       # We don't know whether this will modify df or not!
   return foo

Yikes!

Evaluation order matters

Furthermore, in chained expressions, the order may determine whether a copy is returned or not. If an expression will set values on a copy of a slice, then a SettingWithCopy exception will be raised (this raise/warn behavior is new starting in 0.13.0)

You can control the action of a chained assignment via the option mode.chained_assignment, which can take the values ['raise','warn',None], where showing a warning is the default.

In [309]: dfb = pd.DataFrame({'a' : ['one', 'one', 'two',
   .....:                            'three', 'two', 'one', 'six'],
   .....:                     'c' : np.arange(7)})
   .....: 

# This will show the SettingWithCopyWarning
# but the frame values will be set
In [310]: dfb['c'][dfb.a.str.startswith('o')] = 42

This however is operating on a copy and will not work.

>>> pd.set_option('mode.chained_assignment','warn')
>>> dfb[dfb.a.str.startswith('o')]['c'] = 42
Traceback (most recent call last)
     ...
SettingWithCopyWarning:
     A value is trying to be set on a copy of a slice from a DataFrame.
     Try using .loc[row_index,col_indexer] = value instead

A chained assignment can also crop up in setting in a mixed dtype frame.

Note

These setting rules apply to all of .loc/.iloc/.ix

This is the correct access method

In [311]: dfc = pd.DataFrame({'A':['aaa','bbb','ccc'],'B':[1,2,3]})

In [312]: dfc.loc[0,'A'] = 11

In [313]: dfc
Out[313]: 
     A  B
0   11  1
1  bbb  2
2  ccc  3

This can work at times, but is not guaranteed, and so should be avoided

In [314]: dfc = dfc.copy()

In [315]: dfc['A'][0] = 111

In [316]: dfc
Out[316]: 
     A  B
0  111  1
1  bbb  2
2  ccc  3

This will not work at all, and so should be avoided

>>> pd.set_option('mode.chained_assignment','raise')
>>> dfc.loc[0]['A'] = 1111
Traceback (most recent call last)
     ...
SettingWithCopyException:
     A value is trying to be set on a copy of a slice from a DataFrame.
     Try using .loc[row_index,col_indexer] = value instead

Warning

The chained assignment warnings / exceptions are aiming to inform the user of a possibly invalid assignment. There may be false positives; situations where a chained assignment is inadvertently reported.