Indexing and Selecting in Pandas (part 1)

The topic of indexing and selecting data in pandas is core to using pandas, but it can be quite confusing. One reason for that is because over the years pandas has grown organically based on user requests so there are multiple way to select data out of a pandas DataFrame or Series. Reading through the documentation can be a real challenge, especially for beginners. For more advanced users, it is easy to learn a few techniques and just fall back on your favorite method to access data and not realize that there might be simpler, faster, or more reliable ways to both select and modify your data.

Since this can be a complicated and confusing topic, I’m choosing to break it into several smaller posts so that it doesn’t become overwhelming, working up from the basics to the more complex scenarios. The methods used for selecting and indexing are some of the most confusing methods to work with in pandas due to their different behavior with different argument types.

Indexing, starting with the basics

This will be the first of a series of posts covering indexing and selecting data. This post is just going to cover selecting data by index label or integer offsets. In the future, I’ll discuss slicing, boolean indexing, the query method, using isin, cross sections, and much more.

Our test data

As we get started, we want some data. But instead of making fake data to work with, let’s grab some real data online that is a little more interesting. I live near Chicago, so I looked at the Chicago Data Portal for some datasets. A smaller one I found was the Individual Landmarks dataset that gives an inventory of our famous Chicago landmarks. It has text, numeric, and date fields. I’m going to manipulate it a bit to make it useful for explaining more concepts.

These examples were executed using Python 3.8.6 and pandas 1.1.4.

>>> import pandas as pd
>>> import numpy as np
>>>
>>> # you should be able to grab this dataset as an unauthenticated user, but you can be rate limited
>>> df = pd.read_json("https://data.cityofchicago.org/resource/tdab-kixi.json")
>>> df.head(3)
                             landmark_name     id                    address  ... :@computed_region_awaf_s7ux  date_built               architect
0  Vassar Swiss Underwear Company Building  L-265  2543 - 2545 W Diversey Av  ...                        24.0         NaN                     NaN
1                     Mathilde Eliel House  L- 89            4122 S Ellis Av  ...                         1.0        1886        Adler & Sullivan
2                       Manhattan Building  L-139          431 S Dearborn St  ...                        48.0        1891  William LeBaron Jenney

[3 rows x 15 columns]
>>> df.dtypes
landmark_name                   object
id                              object
address                         object
landmark_designation_date       object
latitude                       float64
longitude                      float64
location                        object
:@computed_region_rpca_8um6    float64
:@computed_region_vrxf_vc4k    float64
:@computed_region_6mkv_f3dw    float64
:@computed_region_bdys_3d7i    float64
:@computed_region_43wa_7qmu    float64
:@computed_region_awaf_s7ux    float64
date_built                      object
architect                       object
dtype: object
>>> # let's transform that date from object to a datetime
>>> df['landmark_designation_date'] = pd.to_datetime(df['landmark_designation_date'])
>>> # also trimming down the columns
>>> df = df[['landmark_name', 'id', 'address', 'landmark_designation_date',
...          'latitude', 'longitude', 'location', 'date_built', 'architect']]
>>> df.columns
Index(['landmark_name', 'id', 'address', 'landmark_designation_date',
       'latitude', 'longitude', 'location', 'date_built', 'architect'],
      dtype='object')

Axes

The two main data structures in pandas both have at least one axis. A Series has one axis, the index. A DataFrame has two axes, the index and the columns. It’s useful to note here that in all the DataFrame functions that can be applied to either rows or columns, an axis of 0 refers to the index, an axis of 1 refers to the columns.

We can inspect these in our sample DataFrame. We’ll pick the landmark_name column as a sample Series to demonstrate the basics for a Series. You can see the column (which is a Series) and the entire DataFrame share the same index.

>>> s = df['landmark_name']
>>> print("Series index:", s.index)
Series index: RangeIndex(start=0, stop=317, step=1)
>>> print("DataFrame index:", df.index)
DataFrame index: RangeIndex(start=0, stop=317, step=1)

Index

In pandas, an Index (or a subclass) allows for the data structures that use it to support lookups (or selection), data alignment (think of time-series data especially, where all the observations needs to be aligned with their observation time), and reindexing (changing the underlying index to have different values, but keeping the data aligned). There are a number of types of indices, but for now, we’ll just look at the simple RangeIndex that our current DataFrame is using, which will have integer values.

Basic selecting with []

We’re going to start with the basic form of selecting, using the [] operator, which in Python maps to a class’s __getitem__ function (if you’re familiar with objects in Python, if not, don’t worry about that for now). Depending on whether the pandas object is a Series or a DataFrame and the arguments you pass into this function, you will get very different results. Let’s start with the basics, invoking with a single argument.

Series

With a Series, the call will return a single scalar value that matches the value at that label in the index. If you pass in a value for a label that doesn’t exist, you will get a KeyError raised. Also, if you pass in an integer and your index has that value, it will return it. But if you don’t have an integer value in your index, it will return the value by position. This is convenient, but can be confusing.

Now I’m going to give this DataFrame (and Series) a new index because the RangeIndex could make much of the following examples very confusing. It’s important for us to differentiate between accessing elements by label and by position. If our index labels are integers, you will not be able to see the difference! Since this dataset already has a unique id column, we’ll use that instead.

>>> df.index = df['id'].str.replace(' ', '') # remove spaces for this example
>>> s = df['landmark_name']
>>> df.index
Index(['L-265', 'L-89', 'L-139', 'L-12', 'L-88', 'L-318', 'L-85', 'L-149',
       'L-286', 'L-71',
       ...
       'L-241', 'L-133', 'L-169', 'L-277', 'L-164', 'L-310', 'L-103', 'L-236',
       'L-65', 'L-224'],
      dtype='object', name='id', length=317)

Now that our index doesn’t contain int values, when we call it with ints they will be evaluated as positional arguments. If you index had int values, they would be found first rather than posititional values. Confusing, isn’t it? This is one reason why you want to read on and see why there are better ways to do this.

>>> print("The value for L-265:", s['L-265'])
The value for L-265: Vassar Swiss Underwear Company Building
>>> print("The first value:", s[0])
The first value: Vassar Swiss Underwear Company Building
>>> print("The value for L-139:", s['L-139'])
The value for L-139: Manhattan Building
>>> print("The third value:", s[2])
The third value: Manhattan Building
>>> try:
...     s['L-900']
... except KeyError as ke:
...     print("Exception: ", ke)
...
Exception:  'L-900'

While I rarely use it, there is a get method available, which will return None if the argument is not in the index instead of raising a KeyError.

>>> print("The first value:", s.get(0))
The first value: Vassar Swiss Underwear Company Building
>>> print("Is there a value at 'L-900'?: ", s.get('L-900'))
Is there a value at 'L-900'?:  None

DataFrame

Now with a DataFrame, calls to [] are used for selecting from the column index, not the row index. This can be confusing since it’s different from a Series when passing in integer values. Instead, we pass in column names.

>>> try:
...     print("First element in a Series:", s[0])
...     print("First row in a DataFrame?:", df[0])
... except KeyError as ke:
...     print("Nope, that's not how you select rows in a DataFrame")
...
First element in a Series: Vassar Swiss Underwear Company Building
Nope, that's not how you select rows in a DataFrame
>>> df['landmark_name']
id
L-265              Vassar Swiss Underwear Company Building
L-89                                  Mathilde Eliel House
L-139                                   Manhattan Building
L-12     Machinery Hall at Illinois Institute of Techno...
L-88                                Melissa Ann Elam House
                               ...
L-310     (Former) Schlitz BreweryTied-House@1944 N.Oakley
L-103                                           Getty Tomb
L-236                         Engine Company 129, Truck 50
L-65                                  James Charnley House
L-224                         Beeson House and Coach House
Name: landmark_name, Length: 317, dtype: object

We can also select a list of columns, in any order (even repeated).

>>> df[['landmark_name', 'address', 'landmark_name']]

Attribute access for columns on a DataFrame or values in a Series.

Another way to select a DataFrame column or an element in a Series is using the attribute operator, .. Pandas will automagically create accessors for all DataFrame columns or for all labels in a Series, provided their name translates to valid Python. This can make life easier for you, especially in an environment with tab-completion like IPython or a Jupyter notebook, but in general, it’s best not to use these attributes in production code. Why not?

  • Your column names may collide with method names on a DataFrame itself, and in that case you will be accessing something you weren’t intending to.
  • Column names often may not be valid Python identifiers since they may contain spaces or start with (or just be) numbers, so you have to use the longer form anyway.
  • Using [] with quoted strings makes your code very clear for others (and future self) to read. 
  • Assigning to a non-existing attribute won’t create a new column, it will just create an attribute on the object. (We’ll talk about modifying data in subsequent posts, you can read this article for details on adding columns to a DataFrame)

So hopefully this list reinforces why it’s just a bad habit to rely on attribute access. In our data example, we can use attribute access for some of our column names, but because the primary index doesn’t make valid Python identifiers, we can’t use it on our Series (L-265 is not a valid Python identifier, for example). But it does work for some situations and may save you a few keystrokes when doing exploratory analysis.

Indexing in pandas can be so confusing

There are so many ways to do the same thing! What is the difference between .loc, .iloc, .ix, and []?  You can read the official documentation but there's so much of it and it seems so confusing. You can ask a question on Stack Overflow, but you're just as likely to get too many different and confusing answers as no answer at all. And existing answers don't fit your scenario.

You just need to get started with the basics.

What if you could quickly learn the basics of indexing and selecting data in pandas with clear examples and instructions on why and when you should use each one? What if the examples were all consistent, used realistic data, and included extra relevant background information?

Master the basics of pandas indexing with my free ebook. You'll learn what you need to get comfortable with pandas indexing. Covered topics include:

  • what an index is and why it is needed
  • how to select data in both a Series and DataFrame.
  • the difference between .loc, .iloc, .ix, and [] and when (and if) you should use them.
  • slicing, and how pandas slicing compares to regular Python slicing
  • boolean indexing
  • selecting via callable
  • how to use where and mask.
  • how to use query, and how it can help performance
  • time series indexing

Because it's highly focused, you'll learn the basics of indexing and be able to fall back on this knowledge time and again as you use other features in pandas.

Just give me your email and you'll get the free 57 page e-book, along with helpful articles about Python, pandas, and related technologies once or twice a month. Unsubscribe at any time.

Invalid email address
>>> df.landmark_name

Selecting with .loc

We also have the .loc attribute which is intended for selection and indexing by label, similar to [] on a Series. However, using .loc and .iloc will make it more clear in your code what your intentions are, and they behave differently. Note that .loc will raise KeyError when an element doesn’t exist at that label.

Also note that this is not a method, so we don’t call it, but rather we use the indexing operator ([]) on it. 

Series

Let’s start with a Series. First, note that we don’t select by location, we can select by index label.

>>> try:
...     print("Selecting by location?", s.loc[0])
... except KeyError as ke:
...     print("Nope, not with .loc: ", ke)
Nope, not with .loc:  0
>>> print("Yes, do it by label: ", s.loc['L-139'])
Yes, do it by label:  Manhattan Building

We can also pass in a list of labels, and they will be returned as a Series.

>>> s.loc[['L-12', 'L-265', 'L-224']]
id
L-12     Machinery Hall at Illinois Institute of Techno...
L-265              Vassar Swiss Underwear Company Building
L-224                         Beeson House and Coach House
Name: landmark_name, dtype: object

Note that this can be a list with a single element, but passing in a list returns a Series (with one element), not a scalar.

>>> s.loc[['L-12']]
id
L-12    Machinery Hall at Illinois Institute of Techno...
Name: landmark_name, dtype: object

DataFrame

On a DataFrame, a single argument to .loc will return a Series for the row matching the label. I’ll select one of my favorite Chicago buildings, the Rookery. If you ever are in Chicago, check out the lobby of this beautiful building, remodeled by Frank Lloyd Wright, and note that one of the columns has its marble cladding missing so you can see the original post underneath. Wright left it open so the original design could still be seen.

>>> df.loc['L-156']
landmark_name                                                 Rookery Building
id                                                                       L-156
address                                                       209 S LaSalle St
landmark_designation_date                            1972-07-05 07:00:00+00:00
latitude                                                               41.8791
longitude                                                             -87.6318
location                     {'latitude': '41.8790761299', 'longitude': '-8...
date_built                                                             1885-88
architect                                                       Burnham & Root
Name: L-156, dtype: object

If passed a list of labels, .loc will return a DataFrame of the matching rows. This is just selecting rows by index, but selecting multiple rows. Note that all of the elements in the list have to be in the index or KeyError is raised.

>>> df.loc[['L-12', 'L-11', 'L-13']]
                                          landmark_name     id  ...  date_built                architect
id                                                              ...
L-12  Machinery Hall at Illinois Institute of Techno...  L- 12  ...        1901  Patton, Fisher & Miller
L-11                        South Shore Cultural Center  L- 11  ...  1906,09,16           Marshall & Fox
L-13  Main Building at Illinois Institute of Technology  L- 13  ...     1891-93          Patton & Fisher

[3 rows x 9 columns]

But, DataFrame‘s .loc can take multiple arguments. The first argument is an indexer for rows, the second argument is an indexer for the columns. So if we wanted a row and the landmark_name column, we can get back a scalar.

>>> df.loc['L-12','landmark_name']
'Machinery Hall at Illinois Institute of Technology'

Now I’m going to move on from .loc for now, but will circle back to talk about some more advanced ways of selecting data with it.

Selecting with .iloc.

.iloc is a separate method for use with purely integer based indexing, starting from 0. It’s very similar to the behavior of .loc, as we’ll see, but raises IndexError when the indexer is out of bounds, or the wrong type.

Series

For a Series.iloc returns a scalar, just like .loc.

'Vassar Swiss Underwear Company Building'

Relative indexing is allowed, and if you try to access a non-existent element, IndexError is raised.

>>> s.iloc[-1]
'Beeson House and Coach House'
>>> try:
...     s.iloc[9999]
... except IndexError as ix:
...     print("You ran over the end of your Series: ", ix)
You ran over the end of your Series:  single positional indexer is out-of-bounds

Passing in a list returns a Series. Again, all elements in the list need to be in the index, or IndexError is raised.

>>> s.iloc[[0,1,2]]
id
L-265    Vassar Swiss Underwear Company Building
L-89                        Mathilde Eliel House
L-139                         Manhattan Building
Name: landmark_name, dtype: object

I’m going to stop here for this post, and I know this seems quite unsatisfying to only pull data out by index, but understanding how to use the index will make future types of selections much easier to understand. For now, make sure you remember that .loc selects items by label, .iloc by integer position, and [] is sort of a hybrid.

To learn more about indexing, take a look on my articles on slicing as well as boolean indexing which will give us much more power to get more out of the three methods that were covered in this post: [].loc, and .iloc.

5 thoughts on “Indexing and Selecting in Pandas (part 1)”

  1. I found the text highly informative. I suggest correcting the information on using .iloc for Series. The statement (above the example s.iloc[[0,1,2]]) “Passing in a list returns a DataFrame.” should be “Passing in a list returns a Series.”

  2. Thanks for this. I’m an absolute hack, but I’ve used Pandas for a few years. And these distinctions were never clear to me before. This is very well explained.

    1. Thank you. I’m glad it helped clear things up. I hope you’ll read the followup posts as well, these methods are loaded with a lot of capability, but can also be totally confusing, as I’m realizing as I dig in more.

  3. This distills the essence very nicely, thanks for the post. As a next topic, looking at performance implications of various ways of access data from existing padas object might be of interest – when the copies are made or not is not quite transparent.

Have anything to say about this topic?