Supporting Statistical Analysis for Research

## 4.7 Coding missing values

### 4.7.1 Data concepts - Conditionally created variables

In section 4.5 we examined how a test of a variable creates a results in a conditional variable. These condition variables can be used to create a new variable from two other variables, a true variable and false variable. The result variable for an observation will take the value of the true variable when the condition variable for that observation is true. Otherwise the observations will take the value from the false variable for that observation. This is visually displayed in figure 4.3.

As an example, let's consider using a conditionally created variable to replace values of "?" with na in a variable named x. This is visually displayed in figure 4.4. The condition variable would be the result of a testing x for equality with "?". The true value, what to do when x is equal to "?", would be a variable containing na for all observations. The false value, what to do when x is not equal to "?", would be x. The resulting variable could be saved in the data frame as x, overwriting the original variable x, saved with a new name, or used for something else.

### 4.7.2 Programming skills - Identifying missing data

The first place to check for information about missing data is the documentation for the data. If the data has a code book, it will often indicate values used to identify missing data.

Another approach is to visually scan the data in the data editor or a text editor. This is useful if the data set is smallish and can be open in an excel like format. This method can be difficult if the data set is large or can not be opened in a way that displays columns.

Making exploratory plots of the data is another good method to identify missing value indicators. Unusual and unexpected values become visible in plots.

### 4.7.3 Examples - R

These examples use the airAccs.csv data set.

1. We begin by using the same code as in the prior sections of this chapter to load the tidyverse, import the csv file, and rename the variables.

Note, the lubridate package is also loaded for use in these examples.

library(tidyverse)
library(lubridate)
airAccs_path <- file.path("..", "datasets", "airAccs.csv")
air_accidents_in <- read_csv(airAccs_path, col_types = cols())
Warning: Missing column names filled in: 'X1' [1]
air_accidents_in <-
air_accidents_in %>%
rename(
obs_num = 1,
date = Date,
plane_type = planeType,
aboard = Aboard,
ground = Ground
)

air_accidents <- air_accidents_in

glimpse(air_accidents)
Observations: 5,666
Variables: 8
$obs_num <dbl> 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, ...$ date       <date> 1908-09-17, 1912-07-12, 1913-08-06, 1913-09-09, 19...
$location <chr> "Fort Myer, Virginia", "Atlantic City, New Jersey",...$ operator   <chr> "Military - U.S. Army", "Military - U.S. Navy", "Pr...
$plane_type <chr> "Wright Flyer III", "Dirigible", "Curtiss seaplane"...$ dead       <dbl> 1, 5, 1, 14, 30, 21, 19, 20, 22, 19, 27, 20, 20, 23...
$aboard <dbl> 2, 5, 1, 20, 30, 41, 19, 20, 22, 19, 28, 20, 20, 23...$ ground     <dbl> 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ...
2. Identifying values that need to be coded as missing.

The description file for the airAccs.csv data set, airAccs.txt, does not provide any information on missing values.

A visual scan can be done using the data browser in RStudio. Take a look at the air_accidents data frame and see if there appears to be any missing data.

From the visual scan of the data you can see that there are ? values in the operator and plane_type columns. (A few of the rows are displayed below.)

air_accidents %>%
filter(operator == "?" | plane_type == "?") %>%
select("location", "operator", "plane_type")
# A tibble: 40 x 3
location                     operator             plane_type
<chr>                        <chr>                <chr>
1 Near Yarmouth, England       ?                    Zepplin LZ-95 (air sh~
2 Pao Ting Fou, China          ?                    ?
3 Fuhlsbuttel, Germany         ?                    LVG C VI
4 Venice, Italy                ?                    de Havilland DH-9
5 Cabrerolles, France          Grands Express Aeri~ ?
6 Toul, France                 CIDNA                ?
7 New York, New York           ?                    Sikorsky S-25
8 Rio de Janeiro, Brazil       ?                    ?
9 Rio de Janeiro, Brazil       ?                    Junkers G24
# ... with 30 more rows

Note: the entries with ? could be changed to NA using the na parameter of read_csv(). This is not done here to demonstrate techniques that operate on a single column.

There are also values in the plane_type column that include ?s with other text. These partial ? entries will be left as they are, since there may be important information in the non-? part of the entries. The entries with only ? will be changed to NA.

3. Coding values as NA.

The mutate() method can be used to change already existing columns or create new columns. The parameter name is the column name to be changed or added. The parameter value is what the column name is set to. We will use mutate() to replace the ? values.

The if_else() function selects between two possible values for each row based on the condition value for the row.

A common way if_else() is used to to set one of the two values to be the variable (i.e. the existing value of the variable.) Then the variable remains the same except when a particular condition is met. This is how we will use the if_else() function in this example.

The tidyverse conditional functions/methods do more type checking than base R functions. The tidyverse defines a set of NA_*_ values for character, real, and integer. The NA value used here is NA_character_.

air_accidents <-
air_accidents %>%
mutate(
operator = if_else(operator == "?", NA_character_, operator)
)

air_accidents %>%
filter(operator == "?" | plane_type == "?") %>%
select("location", "operator", "plane_type")
# A tibble: 25 x 3
location                       operator                       plane_type
<chr>                          <chr>                          <chr>
1 Pao Ting Fou, China            <NA>                           ?
2 Cabrerolles, France            Grands Express Aeriens         ?
3 Toul, France                   CIDNA                          ?
4 Rio de Janeiro, Brazil         <NA>                           ?
6 San Barbra, Honduras           <NA>                           ?
7 Miami, Florida                 Pan American Airways           ?
8 Poona, India                   Military - Indian Air Force    ?
9 Off Hampton Roads, Virginia    Military - US Navy             ?
10 Seljord, Norway                Military - U.S. Army Air Corps ?
# ... with 15 more rows

The ? values have been changed to NA's.

4. Other approaches to coding values as NA.

The recode() function will replace all occurrences of a specific value, ? in our example, with a different value.

The na_if() function is used to change all occurrences of a specific value, ? in our example, to NA. The na_if() function is a special case of the recode() function.

# restore original values of the variables
air_accidents <-
air_accidents %>%
mutate(
operator = pull(air_accidents_in, operator),
plane_type = pull(air_accidents_in, plane_type)
)

air_accidents <-
air_accidents %>%
mutate(
operator = na_if(operator, "?"),
plane_type = recode(plane_type, "?" = NA_character_)
)

air_accidents %>%
filter(operator == "?" | plane_type == "?") %>%
select("location", "operator", "plane_type")
# A tibble: 0 x 3
# ... with 3 variables: location <chr>, operator <chr>, plane_type <chr>

The mutate(), if_else(), and recode() functions are used in a large number of wrangling tasks. Becoming familiar with them will serve you well.

### 4.7.4 Examples - Python

These examples use the airAccs.csv data set.

1. We begin by using the same code as in the prior sections of this chapter to load the packages, import the csv file, and rename the variables.

from pathlib import Path
import pandas as pd
import numpy as np
airAccs_path = Path('..') / 'datasets' / 'airAccs.csv'
air_accidents_in = (
air_accidents_in
.rename(
columns={
air_accidents_in.columns[0]: 'obs_num',
'Date': 'date',
'planeType': 'plane_type',
'Aboard': 'aboard',
'Ground': 'ground'}))

air_accidents = air_accidents_in.copy(deep=True)

print(air_accidents.dtypes)
obs_num         int64
date           object
location       object
operator       object
plane_type     object
aboard        float64
ground        float64
dtype: object
2. Identifying values that need to be coded as missing.

The description file for the airAccs.csv data set, airAccs.txt, does not provide any information on missing values.

A visual scan of this data set can be done using Excel. Take a look at the air_accidents.csv file and see if there appears to be any missing data.

From the visual scan of the data you can see that there are ? values in the operator and plane_type columns. (A few of the rows are displayed below.)

(air_accidents
.query('operator == "?" | plane_type == "?"')
.copy()
.loc[:, ["location", "operator", "plane_type"]]
.pipe(print))
                           location  ...                plane_type
16           Near Yarmouth, England  ...  Zepplin LZ-95 (air ship)
63              Pao Ting Fou, China  ...                         ?
66             Fuhlsbuttel, Germany  ...                  LVG C VI
70                    Venice, Italy  ...         de Havilland DH-9
89              Cabrerolles, France  ...                         ?
100                    Toul, France  ...                         ?
110              New York, New York  ...             Sikorsky S-25
143          Rio de Janeiro, Brazil  ...                         ?
169          Rio de Janeiro, Brazil  ...               Junkers G24

[10 rows x 3 columns]

Note: the entries with ? could be changed to NaN using the na_values parameter of read_csv(). This is not done here to demonstrate techniques that operate on a single column.

There are also values in the plane_type column that include ?s with other text. These partial ? entries will be left as they are, since there may be important information in the non-? part of the entries. The entries with only ? will be changed to NA.

3. Coding values as NaN

The assign() method can be used to change already existing columns or create new columns. The parameter name is the column name to be changed or added. The parameter value is what the column name is set to. We will use assign() to replace the ? values.

A comprehension can be used to conditionally build a variable. The syntax for the list comprehension that we will use is

[<True_value> if <condition> else <False_value> for x in <variable> ]

The comprehension loops through all the rows of the variable testing each value against the condition given. If condition is True, the True_value is used for that row, otherwise the False_value is used. The x is used to represent the value of the variable at a row. Other variable names besides x can be used.

A common use of a comprehension is to set one of the two values to be the variable being tested. Then the variable remains the same except when a particular condition is met. This is how it will be used in this example.

air_accidents = (
air_accidents
.assign(
operator=[np.NaN if x == '?' else x
for x in air_accidents['operator']],
plane_type=[np.NaN if x == '?' else x
for x in air_accidents['plane_type']]))

(air_accidents
.query('operator == "?" | plane_type == "?"')
.copy()
.loc[:, ["location", "operator", "plane_type"]]
.pipe(print))
Empty DataFrame
Columns: [location, operator, plane_type]
Index: []
4. Another approach to coding values as NaN

The pandas replace() function uses a dictionary to specify what specific value is to be change to and which variables to apply the replacements to.

A dictionary is specified as matched pairs of object inside of the curly brackets. The : operator identifies the name (on the left side) and its paired value (on the right side.) The names are called keys because they are used to look up a value. Commas separate the entries in the dictionary. For example,

{'a': 1, 'b': 2}

associates 1 with the key a and 2 with with the key b.

The replace() function nests dictionaries inside of a dictionary. For example, the dictionary that replaces ? in the operator variable would be,

{'operator': {'?': np.NaN}}.

This associates {'?': np.NaN} with the key (a variable name) operator. The {'?': np.NaN} dictionary tells replace() to change ? to NaN. If other values were to be changed for the operator value, the {'?': np.NaN} dictionary would include those with each separated by a comma. If changes to another variable were to be made, the {'operator': {'?': np.NaN}} dictionary would include those names and their dictionaries, each separated by a comma as well. For example,

{'operator': {'?': np.NaN, '': np.NaN}, 'location': {'?': np.NaN}}

would change both ? and the empty string to NaN in the operator variable and ? to NaN in the location variable.

# restore original values of the variables
air_accidents = (
air_accidents
.assign(
operator=air_accidents_in['operator'],
plane_type=air_accidents_in['plane_type']))
# new code
air_accidents = (
air_accidents
.replace({
'operator': {'?': np.NaN},
'plane_type': {'?': np.NaN}}))

(air_accidents
.query('operator == "?" | plane_type == "?"')
.copy()
.loc[:, ["location", "operator", "plane_type"]]
.pipe(print))
Empty DataFrame
Columns: [location, operator, plane_type]
Index: []
5. Conditionally create a variable when neither condition is constant.

The prior examples in this section have worked because one of the two test conditions has been a constant value, np.NaN. In later sections you will need to choose between two variable values based on a test. Techniques that allow you to do so will be demonstrated here. These examples will replace the ? in the operator with np.NaN, but could be used to replace it with a variable that has different values for different observations.

The first method will use the zip() function inside of a comprehension. The zip() function creates sets of observations, called tuples, that a comprehension can iterate through.

# restore original values of the variables
air_accidents= (
air_accidents
.assign(
na_var = np.NaN,
operator=air_accidents_in['operator']))
# new code
air_accidents = (
air_accidents
.assign(
operator=[na if operator == '?' else operator
for operator,  na in
zip(air_accidents['operator'],air_accidents['na_var'])]))

(air_accidents
.query('operator == "?"')
.copy()
.loc[:, ["location", "operator", "plane_type"]]
.pipe(print))
Empty DataFrame
Columns: [location, operator, plane_type]
Index: []

The next method uses np.where() to select between two possible values. The parameters are the condition, true value, and false value. Note, np.where() is different than the pd.where() method.

# restore original values of the variables
air_accidents= (
air_accidents
.assign(
na_var = np.NaN,
operator=air_accidents_in['operator']))
# new code
air_accidents = (
air_accidents
.assign(
operator=np.where(
air_accidents['operator'] == '?',
air_accidents['na_var'],
air_accidents['operator'])))

(air_accidents
.query('operator == "?"')
.copy()
.loc[:, ["location", "operator", "plane_type"]]
.pipe(print))
Empty DataFrame
Columns: [location, operator, plane_type]
Index: []

The pd.mask(), pd.where() and np.select()are other methods and functions that are useful when creating variables based on a condition.

6. Using subsetting to code values as NaN

Subsetting is done by using the loc[] method to subset conditionally on the operator variable having the value ?. All of the subsetted values are then changed to NaN.

# restore original values of the operator variable
air_accidents['operator'] = air_accidents_in['operator']

air_accidents.loc[air_accidents['operator']=='?', 'operator'] = np.NaN

(air_accidents
.query('operator == "?" | plane_type == "?"')
.copy()
.loc[:, ["location", "operator", "plane_type"]]
.pipe(print))
Empty DataFrame
Columns: [location, operator, plane_type]
Index: []

This does explicitly what replace() does. Using replace() would be the preferred approach for what we have done here, because this subsetting approach does not allow for method chaining.

### 4.7.5 Exercises

These exercises use the PSID.csv data set that was imported in the prior section.

1. Import the PSID.csv data set.

2. Code NAs for the kids variable.

3. Display observations that contain missing values in the Kids variable.