Work with missing data

  • 16 minutes

Most of the time, the datasets you want to use (or have to use) have missing values in them. How missing data is handled carries with it subtle trade-offs that can affect your final analysis and real-world outcomes.

pandas handles missing values in two ways. The first you've seen before, in previous sections: the keyword NaN, or Not a Number. This is actually a special value that is part of the IEEE floating-point specification. NaN is used only to indicate missing floating-point values.

For missing values that aren't floats, pandas uses the Python None object. Although it might seem confusing that you'll encounter two different kinds of values that say essentially the same thing, there are sound programmatic reasons for this design choice. In practice, this enables pandas to deliver a good compromise for the vast majority of cases. Notwithstanding this, both None and NaN carry restrictions, and you need to be mindful of these.

None: Non-float missing data

Because None comes from Python, you can't use it in NumPy and pandas arrays that aren't of data type object. Remember, NumPy arrays (and the data structures in pandas) can contain only a single type of data. This is what gives them their tremendous power for large-scale data and computational work, but it also limits their flexibility. Such arrays have to upcast to the “lowest common denominator,” the data type that will encompass everything in the array. When None is in the array, it means you are working with Python objects.

To see this in action, consider the following example array (note the dtype for it):

import numpy as np

example1 = np.array([2, None, 6, 8])
example1

This output is returned:

array([2, None, 6, 8], dtype=object)

The reality of upcast data types carries two side effects with it. First, operations will be carried out at the level of interpreted Python code rather than compiled NumPy code. Essentially, this means that any operations involving Series or DataFrames with None in them will be slower. Although you probably wouldn't notice this performance hit, it might become an issue for large datasets.

The second side effect stems from the first. Because None essentially drags Series or DataFrames back into the world of plain Python, using NumPy or pandas aggregations like sum() or min() on arrays that contain a None value will generally produce an error.

Try it:

example1.sum()

This output is returned:

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-108-ce9901ad18bd> in <module>
----> 1 example1.sum()

/opt/anaconda3/lib/python3.7/site-packages/numpy/core/_methods.py in _sum(a, axis, dtype, out, keepdims, initial)
     34 def _sum(a, axis=None, dtype=None, out=None, keepdims=False,
     35          initial=_NoValue):
---> 36     return umr_sum(a, axis, dtype, out, keepdims, initial)
     37 
     38 def _prod(a, axis=None, dtype=None, out=None, keepdims=False,

TypeError: unsupported operand type(s) for +: 'int' and 'NoneType'

Key takeaway

Addition (and other operations) between integers and None values is undefined, which can limit what you can do with datasets that contain them.

NaN: Missing float values

In contrast to None, NumPy (and therefore pandas) supports NaN for its fast, vectorized operations and ufuncs. The bad news is that any arithmetic that is performed on NaN always results in NaN.

For example:

np.nan + 1

This output is returned:

nan

Run this code in a cell:

np.nan * 0

This output is returned:

nan

The good news is that aggregations that are run on arrays that have NaN in them don't result in errors. But the results aren't uniformly useful.

Run this code in a cell:

example2 = np.array([2, np.nan, 6, 8]) 
example2.sum(), example2.min(), example2.max()

Here's the output:

(nan, nan, nan)

Try it yourself

What happens if you add np.nan and None together?


Hint (expand to reveal)

Remember: NaN is just for missing floating-point values. There's no NaN equivalent for integers, strings, or Booleans.



NaN and None: Null values in pandas

Even though NaN and None can behave somewhat differently, pandas is nevertheless built to handle them interchangeably.

To see what we mean, consider a Series of integers:

int_series = pd.Series([1, 2, 3], dtype=int)
int_series

This output is returned:

0    1
1    2
2    3
dtype: int64

Try it yourself

Now, set an element of int_series equal to None.

  • How does that element show up in the series?
  • What is the dtype of the series?

Hint (expand to reveal) 
int_series[1] = None
int_series

0    1.0
1    NaN
2    3.0
dtype: float64



In the process of upcasting data types to establish data homogeneity in Series and DataFrames, pandas will willingly switch missing values between None and NaN. Because of this design feature, it can be helpful to think of None and NaN as two different flavors of null in pandas. Indeed, some of the core methods you'll use to deal with missing values in pandas reflect this idea in their names:

  • isnull(): Generates a Boolean mask that indicates missing values
  • notnull(): Opposite of isnull()
  • dropna(): Returns a filtered version of the data
  • fillna(): Returns a copy of the data with missing values filled or imputed

These are important methods to master, so let's go over them each in some depth.

Detect null values

Both isnull() and notnull() are your primary methods for detecting null data. Both return Boolean masks over your data.

example3 = pd.Series([0, np.nan, '', None])

example3.isnull()

This output is returned:

0    False
1     True
2    False
3     True
dtype: bool

Look closely at the output. Does any of it surprise you? Although 0 is an arithmetic null, it's nevertheless a perfectly good integer, and pandas treats it as such. '' is a little more subtle. Although you used it earlier to represent an empty string value, it's nevertheless a string object and not a representation of null, as far as pandas is concerned.

Now, let's turn this around and use these methods in a manner more like you'll use them in practice. You can use Boolean masks directly as a Series or DataFrame index, which can be useful when trying to work with isolated missing (or present) values.

Try it yourself

Try running example3[example3.notnull()]. But before you do, what do you expect to see?


Hint (expand to reveal) 
example3[example3.notnull()]

0    0
2     
dtype: object



Key takeaway

Both the isnull() and notnull() methods produce similar results when you use them in DataFrames. They show the results and the index of those results, which helps you enormously as you wrestle with your data.

Drop null values

Beyond identifying missing values, pandas provides a convenient means to remove null values from Series and DataFrames. (Particularly on large datasets, it's often more advisable to simply remove missing [NA] values from your analysis than deal with them in other ways.) To see this in action, let's return to example3:

example3 = example3.dropna()
example3

Here's the output:

0    0
2     
dtype: object

Note that this should look like your output from example3[example3.notnull()]. The difference here is that, rather than just indexing on the masked values, dropna has removed those missing values from the Series example3.

Because DataFrames have two dimensions, they afford more options for dropping data.

Run this code in a cell:

example4 = pd.DataFrame([[1,      np.nan, 7], 
                         [2,      5,      8], 
                         [np.nan, 6,      9]])
example4

Here's the output:

|   | 0   | 1   | 2 | 
---------------------
| 0 | 1.0 | NaN | 7 |
| 1 | 2.0 | 5.0 | 8 |
| 2 | NaN | 6.0 | 9 |

Did you notice that pandas upcast two of the columns to floats to accommodate NaN?

You can't drop a single value from a DataFrame, You must drop full rows or columns. Depending on what you're doing, you might want to do one or the other, so pandas gives you options for both. In data science, columns generally represent variables and rows represent observations, so you are more likely to drop rows of data. The default setting for dropna() is to drop all rows that contain any null values:

example4.dropna()

This output is returned:

|   | 0   | 1   | 2 | 
---------------------
| 1 | 2.0 | 5.0 | 8 |

If necessary, you can drop NA values from columns. Use axis=1 to do so:

example4.dropna(axis='columns')

The output looks like this:

|   |  2 | 
----------
| 0 |  7 |
| 1 |  8 |
| 2 |  9 |

Notice that this can drop a lot of data that you might want to keep, particularly in smaller datasets. What if you want to drop only rows or columns that contain several null values, or even only all null values? You specify those settings in dropna with the how and thresh parameters.

By default, how=any. (If you want to check for yourself, or see what other parameters the method has, run example4.dropna? in a code cell.) You might alternatively specify how=all to drop only rows or columns that contain all null values. Let's expand our example DataFrame to see this in action.

Run this code in a cell:

example4[3] = np.nan
example4

This output is returned:

|   | 0   | 1   | 2 | 3   |
---------------------------
| 0 | 1.0 | NaN | 7 | NaN |
| 1 | 2.0 | 5.0 | 8 | NaN |
| 2 | NaN | 6.0 | 9 | NaN |

Try it yourself

How might you go about dropping just column 3? Remember that you need to supply both the axis parameter and the how parameter.


Hint (expand to reveal) 
example4.dropna(how = "all", axis="columns", inplace=True)

|   | 0   | 1   | 2 |
---------------------
| 0 | 1.0 | NaN | 7 |
| 1 | 2.0 | 5.0 | 8 |
| 2 | NaN | 6.0 | 9 |



The thresh parameter gives you more fine-grained control. You set the number of non-null values that a row or column needs to have in order to be kept.

Run this code in a cell:

example4.dropna(axis='rows', thresh=3)

This output is returned:

|   | 0   | 1   | 2 | 3   |
---------------------------
| 1 | 2.0 | 5.0 | 8 | NaN |

Here, the first and last row were dropped because they contain only two non-null values.

Fill null values

Depending on your dataset, sometimes it makes more sense to fill null values with valid ones rather than drop them. You might use isnull to do this in place, but that can be laborious, particularly if you have a lot of values to fill. Because this is such a common task in data science, pandas provides fillna. This returns a copy of the Series or DataFrame with the missing values replaced with one that you choose. Let's create another example Series to see how this works in practice.

Run this code in a cell:

example5 = pd.Series([1, np.nan, 2, None, 3], index=list('abcde'))
example5

The output looks like this:

a    1.0
b    NaN
c    2.0
d    NaN
e    3.0
dtype: float64

You can fill all the null entries with a single value, such as 0:

example5.fillna(0)

You get this:

a    1.0
b    0.0
c    2.0
d    0.0
e    3.0
dtype: float64

Try it yourself

What happens if you try to fill null values with a string, like ''?


Hint (expand to reveal) 
example5.fillna('')

a    1
b     
c    2
d     
e    3
dtype: object



You can forward-fill null values, which is to use the last valid value to fill a null:

example5.fillna(method='ffill')

The output looks like this:

a    1.0
b    1.0
c    2.0
d    2.0
e    3.0
dtype: float64

You can also back-fill to propagate the next valid value backward to fill a null:

example5.fillna(method='bfill')

This output is returned:

a    1.0
b    2.0
c    2.0
d    3.0
e    3.0
dtype: float64

As you might guess, this works the same with DataFrames, but you can also specify an axis along which to fill null values:

example4

The output looks like this:

|   | 0   | 1   | 2 | 3   |
---------------------------
| 0 | 1.0 | NaN | 7 | NaN |
| 1 | 2.0 | 5.0 | 8 | NaN |
| 2 | NaN | 6.0 | 9 | NaN |

To forward-fill null values in a DataFrame, run this code in a cell:

example4.fillna(method='ffill', axis=1)

The output looks like this:

|   | 0   | 1   | 2   | 3   |
-----------------------------
| 0 | 1.0 | 1.0 | 7.0 | 7.0 |
| 1 | 2.0 | 5.0 | 8.0 | 8.0 |
| 2 | NaN | 6.0 | 9.0 | 9.0 |

Notice that when a preceding value isn't available for forward-filling, the null value remains.

Try it yourself

What output does example4.fillna(method='bfill', axis=1) produce?


Hint (expand to reveal) 
|   | 0   | 1   | 2   | 3   |
-----------------------------
| 0 | 1.0 | 1.0 | 7.0 | NaN |
| 1 | 2.0 | 5.0 | 8.0 | NaN |
| 2 | 6.0 | 6.0 | 9.0 | NaN |



What about example4.fillna(method='ffill') and example4.fillna(method='bfill')?


Hint (expand to reveal) 
|   | 0   | 1   | 2 | 3   |
---------------------------
| 0 | 1.0 | NaN | 7 | NaN |
| 1 | 2.0 | 5.0 | 8 | NaN |
| 2 | 2.0 | 6.0 | 9 | NaN |

and

|   | 0   | 1   | 2 | 3   |
---------------------------
| 0 | 1.0 | 5.0 | 7 | NaN |
| 1 | 2.0 | 5.0 | 8 | NaN |
| 2 | NaN | 6.0 | 9 | NaN |



Can you think of a longer code snippet to write that can fill all the null values in example4?


Hint (expand to reveal) 
example4.fillna(method='ffill', axis="columns").fillna(method='ffill')

|   | 0   | 1   | 2   | 3   |
-----------------------------
| 0 | 1.0 | 5.0 | 7.0 | 7.0 |
| 1 | 2.0 | 5.0 | 8.0 | 8.0 |
| 2 | 2.0 | 6.0 | 9.0 | 9.0 |



You can be creative about how you use fillna. For example, let's look at example4 again. This time, you'll fill the missing values with the average of all the values in the DataFrame.

Run this code in a cell:

example4.fillna(example4.mean())

The output looks like this:

|   | 0   | 1   | 2 | 3   |
---------------------------
| 0 | 1.0 | 5.5 | 7 | NaN |
| 1 | 2.0 | 5.0 | 8 | NaN |
| 2 | 1.5 | 6.0 | 9 | NaN |

Notice that column 3 is still valueless. The default direction is to fill values row-wise.

Takeaway

There are multiple ways to deal with missing values in your datasets. The specific strategy you use (removing them, replacing them, or even how you replace them) should be dictated by the particulars of that data. You'll develop a better sense of how to deal with missing values the more you handle and interact with datasets.