Data Wrangling with tidyr

Overview

Teaching: 25 min
Exercises: 15 min

Questions

• How can I reformat a dataframe to meet my needs?

Objectives

• Describe the concept of a wide and a long table format and for which purpose those formats are useful.

• Describe the roles of variable names and their associated values when a table is reshaped.

• Reshape a dataframe from long to wide format and back with the pivot_wider and pivot_longer commands from the tidyr package.

• Export a dataframe to a csv file.

dplyr pairs nicely with tidyr which enables you to swiftly convert between different data formats (long vs. wide) for plotting and analysis. To learn more about tidyr after the workshop, you may want to check out this handy data tidying with tidyr cheatsheet.

To make sure everyone will use the same dataset for this lesson, we’ll read again the SAFI dataset that we downloaded earlier.

## load the tidyverse
library(tidyverse)
library(here)

interviews <- read_csv(here("data", "SAFI_clean.csv"), na = "NULL")

## inspect the data
interviews

## preview the data
# view(interviews)

Reshaping with pivot_wider() and pivot_longer()

There are essentially three rules that define a “tidy” dataset:

1. Each variable has its own column
2. Each observation has its own row
3. Each value must have its own cell

This graphic visually represents the three rules that define a “tidy” dataset:

R for Data Science, Wickham H and Grolemund G (https://r4ds.had.co.nz/index.html) © Wickham, Grolemund 2017 This image is licenced under Attribution-NonCommercial-NoDerivs 3.0 United States (CC-BY-NC-ND 3.0 US)

In this section we will explore how these rules are linked to the different data formats researchers are often interested in: “wide” and “long”. This tutorial will help you efficiently transform your data shape regardless of original format. First we will explore qualities of the interviews data and how they relate to these different types of data formats.

Long and wide data formats

In the interviews data, each row contains the values of variables associated with each record collected (each interview in the villages), where it is stated that the key_ID was “added to provide a unique Id for each observation” and the instance_ID “does this as well but it is not as convenient to use.”

However, with some inspection, we notice that there are more than one row in the dataset with the same key_ID (as seen below). However, the instanceIDs associated with these duplicate key_IDs are not the same. Thus, we should think of instanceID as the unique identifier for observations!

interviews %>%
  select(key_ID, village, interview_date, instanceID)

# A tibble: 131 × 4
   key_ID village  interview_date      instanceID                               
    <dbl> <chr>    <dttm>              <chr>                                    
 1      1 God      2016-11-17 00:00:00 uuid:ec241f2c-0609-46ed-b5e8-fe575f6cefef
 2      1 God      2016-11-17 00:00:00 uuid:099de9c9-3e5e-427b-8452-26250e840d6e
 3      3 God      2016-11-17 00:00:00 uuid:193d7daf-9582-409b-bf09-027dd36f9007
 4      4 God      2016-11-17 00:00:00 uuid:148d1105-778a-4755-aa71-281eadd4a973
 5      5 God      2016-11-17 00:00:00 uuid:2c867811-9696-4966-9866-f35c3e97d02d
 6      6 God      2016-11-17 00:00:00 uuid:daa56c91-c8e3-44c3-a663-af6a49a2ca70
 7      7 God      2016-11-17 00:00:00 uuid:ae20a58d-56f4-43d7-bafa-e7963d850844
 8      8 Chirodzo 2016-11-16 00:00:00 uuid:d6cee930-7be1-4fd9-88c0-82a08f90fb5a
 9      9 Chirodzo 2016-11-16 00:00:00 uuid:846103d2-b1db-4055-b502-9cd510bb7b37
10     10 Chirodzo 2016-12-16 00:00:00 uuid:8f4e49bc-da81-4356-ae34-e0d794a23721
# … with 121 more rows

As seen in the code below, for each interview date in each village no instanceIDs are the same. Thus, this format is what is called a “long” data format, where each observation occupies only one row in the dataframe.

interviews %>%
  filter(village == "Chirodzo") %>%
  select(key_ID, village, interview_date, instanceID) %>%
  sample_n(size = 10)

# A tibble: 10 × 4
   key_ID village  interview_date      instanceID                               
    <dbl> <chr>    <dttm>              <chr>                                    
 1     48 Chirodzo 2016-11-16 00:00:00 uuid:e180899c-7614-49eb-a97c-40ed013a38a2
 2     44 Chirodzo 2016-11-17 00:00:00 uuid:f9fadf44-d040-4fca-86c1-2835f79c4952
 3     52 Chirodzo 2016-11-16 00:00:00 uuid:6db55cb4-a853-4000-9555-757b7fae2bcf
 4     62 Chirodzo 2016-11-16 00:00:00 uuid:c6597ecc-cc2a-4c35-a6dc-e62c71b345d6
 5     61 Chirodzo 2016-11-16 00:00:00 uuid:2401cf50-8859-44d9-bd14-1bf9128766f2
 6    192 Chirodzo 2017-06-03 00:00:00 uuid:f94409a6-e461-4e4c-a6fb-0072d3d58b00
 7     68 Chirodzo 2016-11-16 00:00:00 uuid:ef04b3eb-b47d-412e-9b09-4f5e08fc66f9
 8    199 Chirodzo 2017-06-04 00:00:00 uuid:ffc83162-ff24-4a87-8709-eff17abc0b3b
 9      9 Chirodzo 2016-11-16 00:00:00 uuid:846103d2-b1db-4055-b502-9cd510bb7b37
10     35 Chirodzo 2016-11-17 00:00:00 uuid:ff7496e7-984a-47d3-a8a1-13618b5683ce

We notice that the layout or format of the interviews data is in a format that adheres to rules 1-3, where

• each column is a variable
• each row is an observation
• each value has its own cell

This is called a “long” data format. But, we notice that each column represents a different variable. In the “longest” data format there would only be three columns, one for the id variable, one for the observed variable, and one for the observed value (of that variable). This data format is quite unsightly and difficult to work with, so you will rarely see it in use.

Alternatively, in a “wide” data format we see modifications to rule 1, where each column no longer represents a single variable. Instead, columns can represent different levels/values of a variable. For instance, in some data you encounter the researchers may have chosen for every survey date to be a different column.

These may sound like dramatically different data layouts, but there are some tools that make transitions between these layouts much simpler than you might think! The gif below shows how these two formats relate to each other, and gives you an idea of how we can use R to shift from one format to the other.

Long and wide dataframe layouts mainly affect readability. You may find that visually you may prefer the “wide” format, since you can see more of the data on the screen. However, all of the R functions we have used thus far expect for your data to be in a “long” data format. This is because the long format is more machine readable and is closer to the formatting of databases.

Questions which warrant different data formats

In interviews, each row contains the values of variables associated with each record (the unit), values such as the village of the respondent, the number of household members, or the type of wall their house had. This format allows for us to make comparisons across individual surveys, but what if we wanted to look at differences in households grouped by different types of housing construction materials?

To facilitate this comparison we would need to create a new table where each row (the unit) was comprised of values of variables associated with housing material (e.g. the respondent_wall_type). In practical terms this means the values of the wall construction materials in respondent_wall_type (e.g. muddaub, burntbricks, cement, sunbricks) would become the names of column variables and the cells would contain values of TRUE or FALSE, for whether that house had a wall made of that material.

Once we we’ve created this new table, we can explore the relationship within and between villages. The key point here is that we are still following a tidy data structure, but we have reshaped the data according to the observations of interest.

Alternatively, if the interview dates were spread across multiple columns, and we were interested in visualizing, within each village, how irrigation conflicts have changed over time. This would require for the interview date to be included in a single column rather than spread across multiple columns. Thus, we would need to transform the column names into values of a variable.

We can do both these of transformations with two tidyr functions, pivot_wider() and pivot_longer().

Pivoting wider

pivot_wider() takes three principal arguments:

1. the data
2. the names_from column variable whose values will become new column names.
3. the values_from column variable whose values will fill the new column variables.

Further arguments include values_fill which, if set, fills in missing values with the value provided.

Let’s use pivot_wider() to transform interviews to create new columns for each type of wall construction material. We will make use of the pipe operator as have done before. Because both the names_from and values_from parameters must come from column values, we will create a dummy column (we’ll name it wall_type_logical) to hold the value TRUE, which we will then place into the appropriate column that corresponds to the wall construction material for that respondent. When using mutate() if you give a single value, it will be used for all observations in the dataset.

For each row in our newly pivoted table, only one of the newly created wall type columns will have a value of TRUE, since each house can only be made of one wall type. The default value that pivot_wider uses to fill the other wall types is NA.

If instead of the default value being NA, we wanted these values to be FALSE, we can insert a default value into the values_fill argument. By including values_fill = list(wall_type_logical = FALSE) inside pivot_wider(), we can fill the remainder of the wall type columns for that row with the value FALSE.

interviews_wide <- interviews %>%
    mutate(wall_type_logical = TRUE) %>%
    pivot_wider(names_from = respondent_wall_type,
                values_from = wall_type_logical,
                values_fill = list(wall_type_logical = FALSE))

View the interviews_wide dataframe and notice that there is no longer a column titled respondent_wall_type. This is because there is a default parameter in pivot_wider() that drops the original column. The values that were in that column have now become columns named muddaub, burntbricks, sunbricks, and cement. You can use dim(interviews) and dim(interviews_wide) to see how the number of columns has changed between the two datasets.

Pivoting longer

The opposing situation could occur if we had been provided with data in the form of interviews_wide, where the building materials are column names, but we wish to treat them as values of a respondent_wall_type variable instead.

In this situation we are gathering these columns turning them into a pair of new variables. One variable includes the column names as values, and the other variable contains the values in each cell previously associated with the column names. We will do this in two steps to make this process a bit clearer.

pivot_longer() takes four principal arguments:

1. the data
2. cols are the names of the columns we use to fill the a new values variable (or to drop).
3. the names_to column variable we wish to create from the cols provided.
4. the values_to column variable we wish to create and fill with values associated with the cols provided.

To recreate our original dataframe, we will use the following:

1. the data - interviews_wide
2. a list of cols (columns) that are to be reshaped; these can be specified using a : if the columns to be reshaped are in one area of the dataframe, or with a vector (c()) command if the columns are spread throughout the dataframe.
3. the names_to column will be a character string of the name the column these columns will be collapsed into (“respondent_wall_type”).
4. the values_to column will be a character string of the name of the column the values of the collapsed columns will be inserted into (“wall_type_logical”). This column will be populated with values of TRUE or FALSE.

interviews_long <- interviews_wide %>%
  pivot_longer(cols = c("muddaub", "cement", "sunbricks", "burntbricks"),
               names_to = "respondent_wall_type",
               values_to = "wall_type_logical")

This creates a dataframe with 524 rows (4 rows per interview respondent). The four rows for each respondent differ only in the value of the “respondent_wall_type” and “wall_type_logical” columns. View the data to see what this looks like.

Only one row for each interview respondent is informative–we know that if the house walls are made of “sunbrick” they aren’t made of any other the other materials. Therefore, it would make sense to filter our dataset to only keep values where wall_type_logical is TRUE. Because wall_type_logical is already either TRUE or FALSE, when passing the column name to filter(), it will automatically already only keep rows where this column has the value TRUE. We can then remove the wall_type_logical column.

We do all of these steps together in the next chunk of code:

interviews_long <- interviews_wide %>%
    pivot_longer(cols = c(burntbricks, cement, muddaub, sunbricks),
                 names_to = "respondent_wall_type",
                 values_to = "wall_type_logical") %>%
    filter(wall_type_logical) %>%
    select(-wall_type_logical)

View both interviews_long and interviews_wide and compare their structure.

Applying pivot_wider() to clean our data

Now that we’ve learned about pivot_longer() and pivot_wider() we’re going to put these functions to use to fix a problem with the way that our data is structured. In the spreadsheets lesson, we learned that it’s best practice to have only a single piece of information in each cell of your spreadsheet. In this dataset, we have several columns which contain multiple pieces of information. For example, the items_owned column contains information about whether our respondents owned a fridge, a television, etc. To make this data easier to analyze, we will split this column and create a new column for each item. Each cell in that column will either be TRUE or FALSE and will indicate whether that interview respondent owned that item (similar to what we did previously with wall_type).

interviews_items_owned <- interviews %>%
  separate_rows(items_owned, sep = ";") %>%
  replace_na(list(items_owned = "no_listed_items")) %>%
  mutate(items_owned_logical = TRUE) %>%
    pivot_wider(names_from = items_owned,
                values_from = items_owned_logical,
                values_fill = list(items_owned_logical = FALSE))

nrow(interviews_items_owned)

[1] 131

There are a couple of new concepts in this code chunk, so let’s walk through it line by line. First we create a new object (interviews_items_owned) based on the interviews dataframe.

interviews_items_owned <- interviews %>%

Then we use the new function separate_rows() from the tidyr package to separate the values of items_owned based on the presence of semi-colons (;). The values of this variable were multiple items separated by semi-colons, so this action creates a row for each item listed in a household’s possession. Thus, we end up with a long format version of the dataset, with multiple rows for each respondent. For example, if a respondent has a television and a solar panel, that respondent will now have two rows, one with “television” and the other with “solar panel” in the items_owned column.

separate_rows(items_owned, sep = ";") %>%

You may notice that one of the columns is called ´NA´. This is because some of the respondents did not own any of the items that was in the interviewer’s list. We can use the replace_na() function to change these NA values to something more meaningful. The replace_na() function expects for you to give it a list() of columns that you would like to replace the NA values in, and the value that you would like to replace the NAs. This ends up looking like this:

replace_na(list(items_owned = "no_listed_items")) %>%

Next, we create a new variable named items_owned_logical, which has one value (TRUE) for every row. This makes sense, since each item in every row was owned by that household. We are constructing this variable so that when spread the items_owned across multiple columns, we can fill the values of those columns with logical values describing whether the household did (TRUE) or didn’t (FALSE) own that particular item.

mutate(items_owned_logical = TRUE) %>%

Lastly, we use pivot_wider() to switch from long format to wide format. This creates a new column for each of the unique values in the items_owned column, and fills those columns with the values of items_owned_logical. We also declare that for items that are missing, we want to fill those cells with the value of FALSE instead of NA.

pivot_wider(names_from = items_owned,
            values_from = items_owned_logical,
            values_fill = list(items_owned_logical = FALSE))

View the interviews_items_owned dataframe. It should have 131 rows (the same number of rows you had originally), but extra columns for each item. How many columns were added?

This format of the data allows us to do interesting things, like make a table showing the number of respondents in each village who owned a particular item:

interviews_items_owned %>%
  filter(bicycle) %>%
  group_by(village) %>%
  count(bicycle)

# A tibble: 3 × 3
# Groups:   village [3]
  village  bicycle     n
  <chr>    <lgl>   <int>
1 Chirodzo TRUE       17
2 God      TRUE       23
3 Ruaca    TRUE       20

Or below we calculate the average number of items from the list owned by respondents in each village. This code uses the rowSums() function to count the number of TRUE values in the bicycle to car columns for each row, hence its name. Note that we replaced NA values with the value no_listed_items, so we must exclude this value in the aggregation. We then group the data by villages and calculate the mean number of items, so each average is grouped by village.

interviews_items_owned %>%
    mutate(number_items = rowSums(select(., bicycle:car))) %>%
    group_by(village) %>%
    summarize(mean_items = mean(number_items))

# A tibble: 3 × 2
  village  mean_items
  <chr>         <dbl>
1 Chirodzo       4.62
2 God            4.07
3 Ruaca          5.63

Exercise

1. Create a new dataframe (named interviews_months_lack_food) that has one column for each month and records TRUE or FALSE for whether each interview respondent was lacking food in that month.

Solution

interviews_months_lack_food <- interviews %>%
  separate_rows(months_lack_food, sep = ";") %>%
  mutate(months_lack_food_logical  = TRUE) %>%
  pivot_wider(names_from = months_lack_food,
              values_from = months_lack_food_logical,
              values_fill = list(months_lack_food_logical = FALSE))

1. How many months (on average) were respondents without food if they did belong to an irrigation association? What about if they didn’t?

Solution

interviews_months_lack_food %>%
  mutate(number_months = rowSums(select(., Jan:May))) %>%
  group_by(memb_assoc) %>%
  summarize(mean_months = mean(number_months))

# A tibble: 3 × 2
  memb_assoc mean_months
  <chr>            <dbl>
1 no                2   
2 yes               2.30
3 <NA>              2.82

Exporting data

Now that you have learned how to use dplyr and tidyr to wrangle your raw data, you may want to export these new data sets to share them with your collaborators or for archival purposes.

Similar to the read_csv() function used for reading CSV files into R, there is a write_csv() function that generates CSV files from dataframes.

Before using write_csv(), we are going to create a new folder, data_output, in our working directory that will store this generated dataset. We don’t want to write generated datasets in the same directory as our raw data. It’s good practice to keep them separate. The data folder should only contain the raw, unaltered data, and should be left alone to make sure we don’t delete or modify it. In contrast, our script will generate the contents of the data_output directory, so even if the files it contains are deleted, we can always re-generate them.

In preparation for our next lesson on plotting, we are going to create a version of the dataset where each of the columns includes only one data value. To do this, we will use pivot_wider to expand the months_lack_food and items_owned columns. We will also create a couple of summary columns.

interviews_plotting <- interviews %>%
  ## pivot wider by items_owned
  separate_rows(items_owned, sep = ";") %>%
  ## if there were no items listed, changing NA to no_listed_items
  replace_na(list(items_owned = "no_listed_items")) %>%
  mutate(items_owned_logical = TRUE) %>%
  pivot_wider(names_from = items_owned,
              values_from = items_owned_logical,
              values_fill = list(items_owned_logical = FALSE)) %>%
  ## pivot wider by months_lack_food
  separate_rows(months_lack_food, sep = ";") %>%
  mutate(months_lack_food_logical = TRUE) %>%
  pivot_wider(names_from = months_lack_food,
              values_from = months_lack_food_logical,
              values_fill = list(months_lack_food_logical = FALSE)) %>%
  ## add some summary columns
  mutate(number_months_lack_food = rowSums(select(., Jan:May))) %>%
  mutate(number_items = rowSums(select(., bicycle:car)))

Now we can save this dataframe to our data_output directory.

write_csv (interviews_plotting, file = "data_output/interviews_plotting.csv")

Key Points

• Use the tidyr package to change the layout of dataframes.

• Use pivot_wider() to go from long to wide format.

• Use pivot_longer() to go from wide to long format.