9 Visualize with ggplot2
This chapter introduces data visualization with the R package ggplot2 (Wickham, Chang, et al., 2024).
Please note: As this chapter is still incomplete, please study Chapter 2: Visualizing data of the ds4psy book (Neth, 2023a) instead.
Preparation
Recommended readings for this chapter include:
Chapter 2: Visualizing data of the ds4psy book (Neth, 2023a)
Chapter 3: Data visualisation of the r4ds book (Wickham & Grolemund, 2017), or Chapter 1: Data visualization of its 2nd edition (Wickham, Çetinkaya-Rundel, et al., 2023)
Preflections
What are common elements of visualizations?
What is the relation between data and those (functional) elements?
What are aesthetic features of visualizations?
9.1 Introduction
The ggplot2 package (Wickham, Chang, et al., 2024) and the corresponding book ggplot2: Elegant graphics for data analysis (Wickham, 2016) provide an implementation of The Grammar of Graphics (Wilkinson, 2005), which develops a systematic way of thinking about — or a language and philosophy of — data visualization. The notion of a grammar is one that we are familiar with (e.g., when studying a foreign language), but its exact meaning remains difficult to define. Wilkinson (2005) notes that a grammar provides the rules that make languages expressive. Knowing the grammar of a language allows us to combine elementary concepts into well-formed sentences. Similarly, learning the grammar of graphics will allow us turning data into visualizations.
Learning how to use ggplot2 is — just like learning R — a journey, rather than a destination.
Hence, we should not be surprised if some concepts remain somewhat obscure for a while.
Fortunately, there is no need to understand all about ggplot()
to create awesome visualizations with it.
9.1.0.1 Terminology
Distinguish between ggplot2 and ggplot()
:
ggplot and ggplot2 denote R packages (in its version ggplot2 3.5.1), whereas
ggplot()
is the main function of those packages that generates a visualization.
Beyond this technical distinction, the grammar of graphics includes many new terms:
mapping data variables to visual dimensions or parameters (e.g., axes, groups)
distinguish a range of geoms (i.e., geometric objects, e.g., areas, bars, lines, points) that transform data via statistics (
stat
arguments orstat_*
functions)aesthetic features (e.g., colors, shapes, sizes) and descriptive elements (e.g., text captions, labels, legend, titles)
combining graphical elements into layers and viewing different facets of a visualization
9.1.1 Contents
This chapter provides only an introduction to the ggplot2 package. It covers a few basic types of visualizations and shows how they can be combined with aesthetic features (e.g., colors, labels, and themes) and functionality (e.g., combining or faceting plots) to create beautiful and powerful visualizations.
9.1.2 Data and tools
This chapter primarily uses the functions of the ggplot2 package:
but also some related packages:
In addition to using data from the datasets and ggplot2 packages, we use the penguins
dataset
from the palmerpenguins package (Horst et al., 2022):
# install.packages('palmerpenguins')
library(palmerpenguins)
![Meet the penguins of the Palmer Archipelago, Antarctica. (Artwork by @allison_horst.)](images/palmerpenguins/lter_penguins.png)
Figure 9.1: Meet the penguins of the Palmer Archipelago, Antarctica. (Artwork by @allison_horst.)
9.2 Essentials of ggplot2
Before we can draw our first plots, we will introduce a minimal template for ggplot()
commands and explain some related terminology.
9.2.1 Minimal template
Generally speaking, a plot takes some <DATA>
as input and creates a visualization by mapping data variables or values to (parts of) geometric objects.
A minimal template of a ggplot()
command can be reduced to the following structure:
# Minimal ggplot template:
ggplot(<DATA>) + # 1. specify data set to use
<GEOM_fun>(aes(<MAPPING>) # 2. specify geom + variable mapping(s)
The minimal template includes the following elements:
The
<DATA>
is a data frame or tibble that contains all data that is to be plotted and is shaped in suitable form (see below).
Its variable names are the levers by which the data values are being mapped to the plot.<GEOM_fun>
is a function that maps data to a geometric object (“geom”) according to an aesthetic mapping that is specified inaes(<MAPPING>)
. A mapping specifies a relation between two entities. Here, the mapping specifies the correspondence of variables to graphical elements, i.e., what goes where.-
A geom’s visual appearance is controlled by aesthetics (e.g., colors, shapes, sizes, …) and can be customized by keyword arguments (e.g.,
color
,fill
,shape
,size
…). There are two general ways and positions to do this:- within the aesthetic mapping (when varying visual features as a function of data properties), or
- by setting its arguments to specific values in
<arg_1 = val_1, ..., arg_n = val_n>
(when remaining constant).
Note that the functions that make up a ggplot()
expression (which are typically positioned on separate lines) are connected by the +
operator, rather than some other pipe operator.
9.2.2 Terminology
An obstacle to many technologies is that insiders tend to converse in special terms that appear to obscure rather than reveal insight. In this respect, ggplot2 is no exception. Fortunately, the number of needed terms is limited and the investment is worthwhile.
Two abstract notions that are relevant in the context of the ggplot2 package are geoms and mapping.
An implicit requirement of ggplot()
is that the to-be-plotted data must be in the right format (shape).
Geometric objects
Basic types of visualizations in ggplot2 involve geometric objects (so-called geoms), which are accessed via dedicated functions (<GEOM_fun>
).
When first encountering ggplot2, it makes sense to familiarize ourselves with basic geom functions that create basic types of visualizations.
Just like other R functions, geoms require specific input arguments to work.
As we get more experienced, we will realize that geoms can be combined to create more complex plots and can invoke particular computations (so-called stats).
Mapping data to visual elements
When creating visualizations, the main regularity that beginners tend to struggle with is to define the mapping between data and elements of the visualization. The notion of a mapping is a relational concept that essentially specifies what goes where. The what part typically refers to some part of the data (e.g., a variable), whereas the where part refers to some aspect or part of the visualization (e.g., an axis, geometric object, or aesthetic feature).
Data format
The <DATA>
provided to the data
argument of the ggplot()
function must be rectangular table (i.e., a data.frame
or tibble
).
Beyond this basic type, ggplot()
assumes that the data is formatted in a specific ways (in so-called “long” format, using factor variables to describe measurement values).
Essentially, this format ensures that some variables describe or provide handles on the values of others.
At this point, we do not need to worry about this and just work with existing sets of data that happen to be in the right shape.
(We will discuss corresponding data transformations in Chapter 14 on Tidying data.)
The data used in the subsequent examples is copied from the penguins
object of the palmerpenguins package (Horst et al., 2022).
We assign this data to an R object pg
and inspect it:
# Data:
pg <- palmerpenguins::penguins
# Inspect data:
dim(pg)
#> [1] 344 8
# Compact structure:
str(pg)
#> tibble [344 × 8] (S3: tbl_df/tbl/data.frame)
#> $ species : Factor w/ 3 levels "Adelie","Chinstrap",..: 1 1 1 1 1 1 1 1 1 1 ...
#> $ island : Factor w/ 3 levels "Biscoe","Dream",..: 3 3 3 3 3 3 3 3 3 3 ...
#> $ bill_length_mm : num [1:344] 39.1 39.5 40.3 NA 36.7 39.3 38.9 39.2 34.1 42 ...
#> $ bill_depth_mm : num [1:344] 18.7 17.4 18 NA 19.3 20.6 17.8 19.6 18.1 20.2 ...
#> $ flipper_length_mm: int [1:344] 181 186 195 NA 193 190 181 195 193 190 ...
#> $ body_mass_g : int [1:344] 3750 3800 3250 NA 3450 3650 3625 4675 3475 4250 ...
#> $ sex : Factor w/ 2 levels "female","male": 2 1 1 NA 1 2 1 2 NA NA ...
#> $ year : int [1:344] 2007 2007 2007 2007 2007 2007 2007 2007 2007 2007 ...
# Print some cases:
set.seed(100) # for reproducible randomness
s <- sample(1:nrow(pg), size = 10)
knitr::kable(pg[s, ], caption = "10 random cases (rows) of the `penguins` data.")
species | island | bill_length_mm | bill_depth_mm | flipper_length_mm | body_mass_g | sex | year |
---|---|---|---|---|---|---|---|
Gentoo | Biscoe | 45.2 | 15.8 | 215 | 5300 | male | 2008 |
Adelie | Biscoe | 45.6 | 20.3 | 191 | 4600 | male | 2009 |
Gentoo | Biscoe | 50.1 | 15.0 | 225 | 5000 | male | 2008 |
Adelie | Torgersen | NA | NA | NA | NA | NA | 2007 |
Chinstrap | Dream | 49.7 | 18.6 | 195 | 3600 | male | 2008 |
Chinstrap | Dream | 49.8 | 17.3 | 198 | 3675 | female | 2009 |
Adelie | Dream | 40.3 | 18.5 | 196 | 4350 | male | 2008 |
Adelie | Torgersen | 38.9 | 17.8 | 181 | 3625 | female | 2007 |
Gentoo | Biscoe | 47.3 | 15.3 | 222 | 5250 | male | 2007 |
Chinstrap | Dream | 43.2 | 16.6 | 187 | 2900 | female | 2007 |
The table shows the names of the 8 variables in our pg
data, which are rather self-explanatory.
For instance, the levels of the factor variables species
and island
can be used to group the other values (e.g., measurements of penguin physiology).
Note that each row of data refers to one observation of a penguin and the data contains some missing (NA
) values on some variables.
Do not worry if some of these terms remain unclear at this point. The following sections will provide plenty of examples that — hopefully — further explain and illustrate their meaning.
9.2.3 Plotting distributions
In Chapter 8, we used histograms and the hist()
function to visualize the distribution of variable values (see Section 8.2.1).
The corresponding geom function in ggplot2 is geom_histogram()
.
The data to be used is pg
and the only aesthetic mapping required for geom_histogram()
is to specify a continuous variable whose values should be mapped to the \(x\)-axis.
Let’s use the flipper_length_mm
variable for this purpose and create our first visualization with ggplot()
(Figure 9.2):
# Basic histogram:
ggplot(data = pg) +
geom_histogram(mapping = aes(x = flipper_length_mm))
#> `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.
#> Warning: Removed 2 rows containing non-finite outside the scale range
#> (`stat_bin()`).
![A histogram showing the distribution of flipper lengths of penguins (created by ggplot2).](i2ds_files/figure-html/ggplot-hist-01-1.png)
Figure 9.2: A histogram showing the distribution of flipper lengths of penguins (created by ggplot2).
Note that we succeeded in creating our first histogram in ggplot2.
This visualization is rather basic, but includes the bars of a histogram on a grey background with white grid lines (its signature theme_grey()
is based on Tufte, 2006) and two axes with appropriate labels.
As with the hist()
function (from the base R graphics package), the default behavior of the geom_histogram()
function is to categorize the values of the specified variable in discrete bins and display the counts of values per bin (as a bar chart).
Note also that evaluating our ggplot()
command printed a message and a warning.
Whereas the warning is due to our flipper_length_mm
variable containing 2 missing (NA
) values,
the message suggests that we could specify a numeric value to the bins
or to the binwidth
parameters to override the default setting of bins = 30
.
Just as we can use a natural language to say the same thing in different ways, the grammar of graphics allow for considerable flexibility in creating the same visualization. For instance, we can omit argument names of R functions, as long as the arguments (here data
and x
) are unambiguous and can move aesthetic mappings to the first line of the ggplot()
expression.
As a consequence, the following variants all create the same visualization:
# Basic histogram variants:
# A: explicit version:
ggplot(data = pg) +
geom_histogram(mapping = aes(x = flipper_length_mm), bins = 30) +
theme_grey()
# B: short version:
ggplot(pg) +
geom_histogram(aes(flipper_length_mm))
# C: moving aesthetic mapping to the 1st line:
ggplot(pg, aes(flipper_length_mm)) +
geom_histogram(bins = 30)
Adding colors, text labels, and themes
Before discovering more features of ggplot2, we should learn to improve its default visualizations.
The basic histogram of Figure 9.2 is informative, but can be embellished by adding colors, more informative text labels, and choosing a different theme.
Colors that do not vary by a data variable can be set as constants (i.e., outside the aes()
function) to color-related arguments of the current geom.
For the bars of geom_histogram()
, the color
argument refers to the border of the bars, whereas the bars themselves are colored by a fill
argument.
We can set these arguments to any of the 657 named R colors, available by evaluating colors()
.
(More complex color settings involving data variables and color scales will be introduced below.)
The best way to change default text labels is by using the labs()
function, which allows setting a range of labels by intuitive argument names.
A good visualization should usually have a descriptive title
and informative labels for its x
- and y
-axes.
The theme()
function of ggplot2 allows re-defining almost any aesthetic aspect of a plot.
Rather than specifying all of them manually, we can choose one of the theme_*()
functions that come with ggplot2.
An improved version of Figure 9.2 can be created as follows (Figure 9.3):
# Adding colors, labels and themes:
ggplot(pg) +
geom_histogram(aes(x = flipper_length_mm), binwidth = 2,
color = "grey20", fill = "deepskyblue") +
labs(title = "Distribution of penguin flipper lengths",
x = "Flipper length (in mm)", y = "Frequency") +
theme_bw()
![A better histogram showing the distribution of flipper lengths of penguins.](i2ds_files/figure-html/ggplot-hist-02-1.png)
Figure 9.3: A better histogram showing the distribution of flipper lengths of penguins.
The code for Figure 9.3 shows that ggplot()
commands can be viewed as a sequence of sub-commands, joined by the +
operator.
A neat feature of ggplot2 is that plots can be stored as R objects and then modified later.
For instance, the code of the previous chunk could be decomposed into two steps:
# Adding colors, labels and themes:
pg_1 <- ggplot(pg) +
geom_histogram(aes(x = flipper_length_mm), binwidth = 2,
color = "grey20", fill = "deepskyblue")
# pg_1 # basic plot with default settings
pg_2 <- pg_1 +
labs(title = "Distribution of penguin flipper lengths",
x = "Flipper length (in mm)", y = "Frequency") +
theme_bw()
pg_2 # annotated plot (with labels and modified theme)
![(ref:fig-ggplot-hist-2b)](i2ds_files/figure-html/ggplot-hist-02-steps-1.png)
Figure 9.4: (ref:fig-ggplot-hist-2b)
When storing a plot as an R object, evaluating the object prints the plot to the visualization area of RStudio.
Here, pg_1
provides the basic histogram (plus two color constants), and pg_2
adds text labels and changes the default plot theme.
Given the vast range of possible modifications, the best practice and strategy for working with ggplot2 is to first get the basic mechanics of the plot right (i.e., by adjusting geoms and variable mappings) before adding further bells and whistles for creating a more appealing visualization (e.g., by selecting aesthetics, text labels, or themes).
The modular structure of ggplot2 objects supports this strategy.
Grouping observations by mapping variables to aesthetics
Noting that “colors that do not vary by a data variable can be set as constants” (in the previous section) raises the question what other functions colors could serve.
A prominent function is to distinguish between different groups of observations.
This would require that a color element of our visualization is mapped to the levels of a categorical variable.
We can easily add this by moving a color argument into the aesthetic mapping function aes()
and assigning it to a categorical variable of our data.
For instance, the following code maps the factor variable species
to the fill
color of the histogram bars:
# Grouping by mapping aesthetics (fill color) to a data variable (species):
pg_3 <- ggplot(pg) +
geom_histogram(aes(x = flipper_length_mm, fill = species),
binwidth = 2, color = "grey20")
pg_3
![](i2ds_files/figure-html/ggplot-hist-03-1.png)
Note that moving fill = species
into the aes()
function had two effects:
First, the counts of observations (penguins) that were expressed in the bars are color-coded for the three different species of penguins.
(Note the difference to the constant color = "grey20"
setting, which lies outside the scope of the aes()
function.)
Additionally, a legend that describes the mapping of colors to species appeared to the right of the plotting area.
This is very useful default behavior, but we may want to adjust the aesthetic properties (e.g., the fill
color) to custom colors.
Adding color scales
Changing colors that are mapped to data variables typically requires specifying a color scale.
The range of color scale functions and corresponding palettes can be confusing and usually requires a lookup of the scale_color_*
function.
A popular option is to use one of the palettes of the RColorBrewer package (Neuwirth, 2022) that come pre-packaged with ggplot2.
The Brewer scales provide sequential, diverging and qualitative color palettes (see https://colorbrewer2.org for more information).
Looking up ?scale_colour_brewer
reveals that its qualitative scales are labeled as “Accent”, “Dark2”, “Paired”, “Pastel1”, “Pastel2”, “Set1”, “Set2”, and “Set3”.
As we aim to change the fill
colors, we can select the corresponding palettes by specifying scale_fill_brewer()
, e.g.,
# Grouping by aesthetics (and using a different color scale):
ggplot(pg) +
geom_histogram(aes(x = flipper_length_mm, fill = species), binwidth = 2, color = "grey20") +
scale_fill_brewer(palette = "Accent")
![](i2ds_files/figure-html/ggplot-colors-brewer-1.png)
When aiming to create a range of visualizations in a uniform style, it is advisable to define one or more palettes of custom colors. There are many R functions and packages supporting this task. We will use the unikn package, as it combines pleasing colors with useful color functions:
library(unikn) # for colors and color functions
# seecol(pal_unikn_pref) # view a (categorical) color palette
my_cols <- usecol(pal = c(Seeblau, Pinky, Seegruen), alpha = .67) # use 3 specific colors
my_cols <- usecol(pal = pal_unikn_pref, alpha = .67) # use an entire color scale
The usecol()
function allows defining a color palette (of a variable length n
) and add transparency (by setting the alpha
parameter to a value from 0 to 1).
Using color transparency is a primary way to prevent overplotting (see Chapters 8 on Visualize in R and Chapter 10 on Using colors for details and examples).
As we saved our plot as pg_3
above, we can add labels, apply our new custom color palette, and change the default theme as follows:
# Adding labels, color scale, and theme (to an existing plot):
pg_4 <- pg_3 +
labs(title = "Distribution of penguin flipper lengths (by species)",
x = "Flipper length (in mm)", y = "Frequency", fill = "Species:") +
scale_fill_manual(values = my_cols) +
theme_unikn()
pg_4
![](i2ds_files/figure-html/ggplot-hist-04-1.png)
Combining plots
Another good reason to save plots as R objects is to combine multiple plots later, e.g., by using the grid.arrange()
function of the gridExtra package (Auguie, 2017).
For instance, if we wanted to re-capitulate the journey from our first histogram (Figure 9.2 above) to our latest version of it we could re-create and save the former as a plot object pg_0
and combine it with our most recent plot object pg_4
:
# Re-create basic histogram (from above) and store as pg_0:
pg_0 <- ggplot(data = pg) +
geom_histogram(mapping = aes(x = flipper_length_mm))
library(gridExtra) # for combining plots
# Combine plots:
gridExtra::grid.arrange(pg_0, pg_4, nrow = 2)
![Combining two ggplot2 plots by using the grid.arrange() of the gridExtra package.](i2ds_files/figure-html/ggplot-hist-05-1.png)
Figure 9.5: Combining two ggplot2 plots by using the grid.arrange()
of the gridExtra package.
Practice
Here are some practice tasks for plotting distributions:
-
Playing with parameters: Re-create the basic histogram of Figure 9.2 and vary the
bins
orbinwidth
parameters.- What happens to the values on the \(y\)-axis when varying the parameters and why?
- What happens when we change the variable mapping from
x
toy
? - Which
binwidth
parameter corresponds to a value ofbins = 30
?
Penguin bill depths by island: Create a (series of) histogram(s) to show the distribution of bill depth for penguins by island.
-
Combining plots: Create a (series of) histogram(s) to show the distribution of bill length for penguins by island and combine it with the histogram for bill depth (created in 2.) into a single plot.
- Print both plots side-by-side (i.e., in two columns of a single row).
- If both plots show the same legend, remove one of them to only show one legend (on the right).
-
Alternative distributions: Study the documentation to
geom_histogram()
and explore its alternativesgeom_density()
andgeom_freqpoly()
.- Create a histogram, density plot, and frequency polygon to show the distribution of body mass (for the 3 species of penguins).
- What does the \(y\)-axis of a density plot show?
9.2.4 Plotting summaries
In addition to plotting distributions, a common type of visualization aims to show a summary of one or more variables. While there are many ways of doing this, we will focus on bar charts and box plots.
Bar charts
A bar chart seems simple, but is actually a quite complicated plot.
To realize this, we use a ggplot()
expression for our pg
data and geom_bar()
, mapping the factor variable species
to its \(x\)-axis:
![A basic bar chart (showing counts of cases).](i2ds_files/figure-html/ggplot-bar-01-1.png)
Figure 9.6: A basic bar chart (showing counts of cases).
Figure 9.6 illustrates that geom_bar()
does not simply plot given data values, but instead performs some computation.
In ggplot2, geoms that compute stuff are linked to so-called stat
(for statistics).
By default, geom_bar
groups observations into the categories specified by the variable levels mapped to x
and then counts the number of cases per category.
The following expression is a more explicit version of the previous code chunk (and would create the exact same plot as Figure 9.6):
![](i2ds_files/figure-html/ggplot-bar-01-expl-1.png)
The relation between geoms and stats
We have seen that geom_histogram()
categorized observations in our data into groups (bins) and counted their frequency (Figure 9.2).
Similarly, geom_bar()
automatically counted the observations in the levels of a variable mapped to x
(Figure 9.6).
This illustrates a hidden complexity in creating visualizations:
Many types of visualizations require computations or transformations of the input data.
If we provide raw data values to a ggplot2()
command, the geoms aim to guess which transformation we desire by linking geoms to stat
options (and corresponding functions).
The details of possible relations between geoms and stat
options are difficult to understand.
Rather than aiming to explain them here, we can only emphasize that geoms that compute values are linked to statistical functions that can also be invoked directly.
When asking for ggplot2 advice online, experts often provide nifty solutions that perform quite complicated data transformations in variable mappings.
Here are some examples that are — spoiler alert — likely to confuse you:
- We can instruct ggplot2 to count observations by mapping
..count..
to a variable:
![A bar chart counting the number of penguin observations by species.](i2ds_files/figure-html/ggplot-bar-02-1.png)
Figure 9.7: A bar chart counting the number of penguin observations by species.
- Instead of assigning
y
to..count..
, we can also ask for proportions (but then also need to specify thegroup
level):
# Compute proportions (in y and group mapping):
ggplot(pg) +
geom_bar(aes(x = species, y = ..prop.., group = 1))
![A bar chart computing proportions of penguin species.](i2ds_files/figure-html/ggplot-bar-03-1.png)
Figure 9.8: A bar chart computing proportions of penguin species.
- In case this cryptic code does not suffice to confuse you, we can even omit the
geom_
function altogether and directly ask for the summary of a given variable mapping (and specify thegeom
as an argument of thestat_summary()
function):
# Compute a bar chart of means (by using stat_summary):
ggplot(pg, aes(x = species, y = body_mass_g)) +
stat_summary(fun = mean, geom = "bar")
![A bar chart computing penguin’s mean body mass by species from data (without an explicit geom function).](i2ds_files/figure-html/ggplot-bar-04-1.png)
Figure 9.9: A bar chart computing penguin’s mean body mass by species from data (without an explicit geom function).
Do not worry if the last three examples remain rather confusing at this point!
They are shown here only to illustrate the intimate connection between data visualization and transformation.
Actually, computing values from data in visualization commands may be convenient and powerful, but is also error-prone and intransparent.
A better way of creating visualizations is to first compute all values that we are interested in (e.g., some measures of central tendency and variability) and then visualize these values.
Fortunately, novice users of ggplot2 only need to know that some geoms provide stat
options and choose an appropriate one (e.g., "count"
vs. "identity"
) if the default option fails.23
Better bar charts
Rather than relying on intransparent data transformations, a better way to create informative bar charts is to explicitly compute all values that we aim to visualize. This may require more effort, but also provides more control and is more transparent and reproducible.
As an example, we re-create the basic counts of Figure 9.6 and the mean chart of Figure 9.9 in a different way.
Interestingly, doing so will not require geom_bar()
, but rather geom_col()
.
- We first compute the counts of observations and means of the
body_mass_g
variable for thepg
data:
library(tidyverse)
# Compute summaries for body_mass_g of penguins per species:
tb <- pg %>%
group_by(species) %>%
summarise(n = n(),
mean_body_mass = mean(body_mass_g, na.rm = TRUE)
)
# Print tb:
knitr::kable(tb, caption = "A table of computed summary values.", digits = 2)
species | n | mean_body_mass |
---|---|---|
Adelie | 152 | 3700.66 |
Chinstrap | 68 | 3733.09 |
Gentoo | 124 | 5076.02 |
Note that we used a dplyr pipe to group, count and summarize cases (see Chapter 13 on Transforming data), but could have used base R functions to solve the same task.
We now can easily re-create Figure 9.6.
But as our summary table tb
already contains the desired counts (in a variable named n
), we no longer want geom_bar()
to count anything.
Thus, we could map the y
variable of geom_bar()
to n
and switch off the geom’s default counting behavior (by specifying stat = "identity"
):
![](i2ds_files/figure-html/ggplot-bar-06-1.png)
An easier way to achieve the same is to replace geom_bar()
by geom_col()
, as the latter uses stat = "identity"
by default:
![](i2ds_files/figure-html/ggplot-bar-07-1.png)
Why is using geom_col()
a better way of creating bar charts?
The main reason is that the computation of the values in tb
was entirely under our control and fully transparent.
A related benefit is that we can easily re-create Figure 9.9 by only changing the variable mapping of y
:
![](i2ds_files/figure-html/ggplot-bar-08-1.png)
This code is arguably much simpler than the one that created Figure 9.9.
Note also that the amount of data supplied to the corresponding ggplot()
functions is vastly different:
Whereas pg
is a raw data table of 344 rows and 8 columns, tb
is only a small summary table of 3 rows and 3 columns.
However, smaller is not always better.
If we ever wanted to visualize other aspects of the data, the data
argument provided as input to the ggplot()
function must provide corresponding variables.
To illustrate this point, we will further refine the aesthetic settings of our original bar chart (showing the counts of penguins by species in Figure 9.6) by the sex
variable.
As we have dropped this variable from tb
, we need to use the original pg
data for this purpose (or alternatively include sex
in tb
).
Adjusting aesthetics
As we have seen above, we can easily adjust colors, text labels, and themes to create prettier and more informative bar charts:
ggplot(pg) +
geom_bar(aes(x = species, fill = species), stat = "count") +
labs(title = "Number of penguins by species",
x = "Species", y = "Frequency", fill = "Species:") +
scale_fill_manual(values = my_cols) +
theme_unikn()
![](i2ds_files/figure-html/ggplot-bar-09-1.png)
Rather than mapping x
and fill
to the same variable, we could set fill
to a different variable of our pg
data.
For instance, let’s see what happens when we set fill
to sex
:
# Add sub-category (by mapping fill to sex):
ggplot(pg) +
geom_bar(aes(x = species, fill = sex)) +
scale_fill_manual(values = my_cols) +
theme_unikn()
![](i2ds_files/figure-html/ggplot-bar-10-1.png)
By mapping another variable to the fill
color of geom_bar()
we revealed additional information about our data.
Note also that the sub-categories of each bar are stacked. The reason is that the position
argument of geom_bar()
is set to "stack"
by default.
As the color mapping of the sex
variable may be somewhat un-intuitive, we also create an alternative color vector:
# Choose 3 colors (to be mapped to sex):
my_3_cols <- my_cols[c(2, 1, 8)]
# Show sub-category (as stacked bars):
ggplot(pg) +
geom_bar(aes(x = species, fill = sex), position = "stack") +
scale_fill_manual(values = my_3_cols) +
theme_unikn()
![](i2ds_files/figure-html/ggplot-bar-11-1.png)
An alternative position
setting is "dodge"
:
# Show sub-category (as dodged bars):
ggplot(pg) +
geom_bar(aes(x = species, fill = sex), position = "dodge") +
scale_fill_manual(values = my_3_cols) +
theme_unikn()
![](i2ds_files/figure-html/ggplot-bar-12-1.png)
However, if not all categories contain the same sub-categories (here sex
values), the width of the bars may need further adjustments.
Box plots
When aiming to visualize summary information of a continuous variable by the levels of some categorical variable, a good alternative is provided by a box plot.
ggplot(pg) +
geom_boxplot(aes(x = species, y = body_mass_g), fill = "gold")
![](i2ds_files/figure-html/ggplot-box-01-1.png)
Mapping fill color to an additional variable:
ggplot(pg) +
geom_boxplot(aes(x = species, y = body_mass_g, fill = island))
![](i2ds_files/figure-html/ggplot-box-02-1.png)
Overall, investing into manual data transformation and computations adds control and transparency to our visualizations and simplifies the code.
As an example, we have shown that bar charts showing means of some variable can be created by using geom_col()
rather than by using geom_bar()
.
However, when transforming data to be plotted we must make sure that the data
supplied as input to gglot()
contains all the variables and values that we want to visualize.
Better bar plots are often column plots: Pre-compute the values to display.
If we had pre-computed the counts, we could map them to y
and specify stat = "identity"
.
A good alternative to many bar charts — if they provide mean information — is provided by box plots.
Practice
Here are some practice tasks on plotting summaries in bar charts or box plots:
-
Understanding geoms:
Using the summary table
tb
from above, explain the result of the following command:
![](i2ds_files/figure-html/ggplot-bar-prac-01-1.png)
- Flipping coordinates: Evaluate the following expression and explain its result.
ggplot(pg) +
geom_bar(aes(x = species)) +
coord_flip()
- How can an identical plot be created without using `coord_flip()`?
-
Simple bar charts: Create a bar plot for the
pg
data showing the counts of penguins observed on each island.- by using
geom_bar()
- by using
geom_col()
- distinguish different penguin
species
as a sub-category
- by using
-
Misleading settings: Explain the output of the following command and find a better solution.
- Why is it misleading?
- How could it be fixed?
# Adding a factor variable:
ggplot(pg, aes(x = species, y = body_mass_g, fill = sex)) +
stat_summary(fun = mean, na.rm = TRUE, geom = "bar", position = "stack")
-
Create a boxplot that shows the mean flipper length of penguins on each of the 3 islands.
- Add
aes(fill = island))
togeom_boxplot()
and explain the result. - Change the
fill
aesthetic ofgeom_boxplot()
toaes(fill = species))
and explain the result.
- Add
ggplot(pg) +
geom_boxplot(aes(x = island, y = body_mass_g, fill = island))
# same as:
ggplot(pg) +
geom_boxplot(aes(x = island, y = body_mass_g, fill = island))
ggplot(pg) +
geom_boxplot(aes(x = island, y = body_mass_g, fill = species))
9.2.5 Plotting relations
Another common type of plot visualizes the relationship between two or more variables. Important types of plots that do this include scatterplots and visualizations of lines or trends. This section will introduce corresponding ggplot2 geoms.
Scatterplots
Scatterplots visualize the relation between two variables for a number of observations by corresponding points that are located in 2-dimensional space. Assuming two orthogonal axes (typically \(x\)- and \(y\)-axes), a primary variable is mapped to the \(x\)-axis, and a secondary variable is mapped to the \(y\)-axis of the plot. The points representing the individual observations then show the value of \(y\) as a function of \(x\).24
As an example of a simple scatterplot, we aim to solve the following task:
- Visualize the relationship between body mass and flipper length for (the 3 species of) penguins.
Solving this task in ggplot2 is simple and straightforward.
We provide our pg
data to ggplot()
and select the geometric object geom_point()
with the aesthetic mappings x = body_mass_g
and y = flipper_length_mm
:
ggplot(pg) +
geom_point(aes(x = body_mass_g, y = flipper_length_mm))
![](i2ds_files/figure-html/ggplot-points-00-1.png)
Overall, this basic scatterplot suggests a positive and possibly linear correlation between penguin’s body mass (mapped to the values on the \(x\)-axis) and their flipper length (mapped to the values of the \(y\)-axis). However, the example also illustrate a typical problem of scatterplots: When many points are clustered near each other or even at the same locations, they overlap or obscure each other — a phenomenon known as overplotting. There are many ways of preventing overplotting in ggplot2. In the context of scatterplots, a popular strategy against overplotting consists in using colors, color transparency, or grouping points into clusters by changing their aesthetic features.
The aesthetic features of points include colors, sizes, and symbol shapes.
As we have seen for other geoms, we can map either constant values or variables to aesthetic features of geom_point()
(Figure 9.10):
sp_01 <- ggplot(pg) +
geom_point(aes(x = body_mass_g, y = flipper_length_mm, # essential mappings
col = species, shape = species # aesthetic variables
), # vs.
alpha = .50, size = 2 # aesthetic constants
)
sp_01
![A scatterplot using geom_point() and an aesthetic grouping variable (species).](i2ds_files/figure-html/ggplot-points-01-1.png)
Figure 9.10: A scatterplot using geom_point()
and an aesthetic grouping variable (species
).
Note that Figure 9.10 mapped two aesthetic features (col
and shape
) to a variable (species
),
whereas two others (alpha
and size
) were mapped to constant values.
The effect of this difference is that the species
variable is used to group the geom’s visual elements (i.e., varying point color and shape by the different types of species), whereas their color transparency and size is set to constant values.
Finally, we can further improve our previous plot by choosing custom colors, text labels, and choosing another theme.
Since Figure 9.10 was saved as an R object (sp_01
), we can adjust the previous plot by adding labels, color scales, and theme functions:
sp_01 +
labs(title = "Penguin's flipper length by body mass (by species)",
x = "Body mass (in g)", y = "Flipper length (in mm)",
col = "Species:", shape = "Species:") +
scale_color_manual(values = my_cols) +
theme_bw()
![Adjusting our scatterplot’s text labels, color scale, and theme.](i2ds_files/figure-html/ggplot-points-02-1.png)
Figure 9.11: Adjusting our scatterplot’s text labels, color scale, and theme.
As before, tweaking aesthetics and adding text labels to the initial plot improved our visualization by making it both prettier and easier to interpret. (We will later see that faceting — i.e., splitting a plot into several sub-plots — is another way of preventing overplotting in ggplot.)
Practice
Here are some practice tasks on plotting relationships in scatterplots, lines or trends:
-
What is the relation between penguin’s bill length and bill depth?
- Create a scatterplot to visualize the relationship between both variables for the
pg
data. - Does this relationship vary for different species?
- Create a scatterplot to visualize the relationship between both variables for the
-
The following code builds on our previous scatterplot (Figure 9.10, saved above as
sp_01
), but maps the aesthetic featureshape
to the data variableisland
, rather than tospecies
.- Evaluate the code and the explain the resulting scatterplot.
- Criticize the plot’s trade-offs: What is good or bad about it?
- Try improving the plot so that the different types of
species
andisland
become more transparent.
ggplot(pg) +
geom_point(aes(x = body_mass_g, y = flipper_length_mm, # essential mappings
col = species, shape = island # aesthetic variables
), # vs.
alpha = .50, size = 2 # aesthetic constants
)
9.3 Advanced features of ggplot2
Using more advanced features of ggplot2 requires a more general template and some additional terminology:
- layers and viewing different
- facets of a visualization
The additional topics mentioned in this section are:
- Combining layers of geoms
- Splitting up plots into facets
- Using coordinate systems
- Saving plots
We will conclude by mentioning further ggplot2 extensions.
9.3.1 Generic template
A generic template for creating a visualization with additional bells and whistles has the following structure:
# Generic ggplot template:
ggplot(data = <DATA>) + # 1. specify data set to use
<GEOM_fun>(mapping = aes(<MAPPING>), # 2. specify geom + mappings
<arg_1 = val_1, ...) + # - optional arguments to geom
... # - additional geoms + mappings
<FACET_fun> + # - optional facet function
<LOOK_GOOD_fun> # - optional themes, colors, labels...
The generic template includes the following elements (beyond the <DATA>
and <GEOM_fun>
of the minimal template):
An optional
<FACET_fun>
uses one or more variable(s) to split a complex plot into multiple subplots.A sequence of optional
<LOOK_GOOD_fun>
adjust the visual features of plots (e.g., by adding titles and text labels, color scales, plot themes, or setting coordinate systems).
9.3.2 Layers of geoms
Geoms can be combined with each other.
When using multiple geoms (in layers): Specify common mappings globally, rather than locally.
9.3.3 Faceting
Explicit grouping by splitting plot into subplots: Faceting.
# Using a color-coded histogram (defined above):
pg_4
![](i2ds_files/figure-html/ggplot-faceting-1.png)
# Explicit grouping/splitting by 1 faceting variable:
pg_4 +
facet_grid(~species)
![](i2ds_files/figure-html/ggplot-faceting-2.png)
# Explicit grouping/splitting by a faceting grid:
pg_4 +
facet_grid(island~species)
![](i2ds_files/figure-html/ggplot-faceting-3.png)
9.4 Conclusion
As ggplot2 currently contains over 50 different geoms, the ones discussed in this chapter provide only an introductory glimpse of the available options.
The true power of ggplot2 results from its modular and extensible structure: It provides a set of tools that can be flexibly combined to create many different visualizations.
9.4.1 Summary
The R package ggplot2 provides a comprehensive toolbox for producing data visualizations. Unlike the collection of functions in base R graphics, ggplot2 uses a conceptual framework based on the grammar of graphics (Wilkinson, 2005). This allows us to construct a graph from composable elements, instead of being limited to a predefined set of charts.
Learning ggplot2 first involves getting a grasp on its terminology (e.g., aesthetic mappings, geoms, themes, layers, and facets) and its way of combining functions to create visualizations. A smart strategy when creating visualizations with ggplot2 for some data is to first select appropriate geoms and adjust variable mappings, before tuning aesthetics, labels, and themes.
9.4.2 Resources
Books and book chapters
The two main references on ggplot2 and its history are Wilkinson (2005) and Wickham (2016).
Introductory chapters on ggplot2 include:
- Chapter 2: Visualizing data of the ds4psy textbook (Neth, 2023a)
As the ggplot package is a key component and pre-cursor of the so-called tidyverse dialect or movement (Wickham et al., 2019), corresponding textbooks provide good introductions to ggplot2:
Chapter 3: Data visualization of the r4ds textbook (Wickham & Grolemund, 2017), or
Chapter 1: Data visualization of its 2nd edition (Wickham, Çetinkaya-Rundel, et al., 2023)
Online resources
One of the best starting points for learning ggplot2 is https://ggplot2.tidyverse.org/ and its vignettes:
Note also the helpful FAQ sections in the articles of https://ggplot2.tidyverse.org
Helpful insights into the relation between geoms and stats are provided by the following article:
- Demystifying stat layers in ggplot2 (by June Choe, 2020-09-26)
Further inspirations and tools for using ggplot2 include:
R Graphics Cookbook (Chang, 2012) provides hands-on advice on using ggplot2 and many useful recipes for data transformation.
Data Visualization with R (Kabacoff, 2018) relies heavily on the ggplot2 package, but also covers other approaches.
ggplot2 Extensions expand the range and scope of ggplot2 even further.
Cheatsheets
Here are some pointers to related Posit cheatsheets:
- Data visualization with ggplot2
![Data visualization with ggplot2 from Posit cheatsheets.](images/cheat/visualization.png)
Figure 9.12: Data visualization with ggplot2 from Posit cheatsheets.
The corresponding online reference provides an overview of key ggplot2 functionality.
9.4.3 Preview
We now learned to create visualizations in base R (in Chapter 8) and the ggplot2 package. Irrespective of the tools we use, colors are an important aesthetic for making more informative and pleasing visualizations. Chapter 10 on Using colors introduces the topic of color representation and show us how to find and manipulate color palettes.
9.5 Exercises
9.5.1 Histograms
In Chapter 8, we created our first histogram for a vector of numeric values x
as follows:
![](i2ds_files/figure-html/ggplot-hist-base-r-1.png)
- Re-create an analog histogram in ggplot2.
- What are the similarities and differences to the base R version?
- Add some aesthetics and labels to improve your histogram.
- Discuss relation between (and the use of
stats
in) histograms and bar charts.
Solution
In this example, the data x
consisted of a single vector.
However, as ggplot()
requires its data
to be in tabular form, we use data.frame()
to convert it into a data frame with one variable x
:
# Convert vector x into df:
df <- data.frame(x)
head(df)
#> x
#> 1 105.55735
#> 2 106.71259
#> 3 90.51432
#> 4 111.84809
#> 5 94.10383
#> 6 114.64747
Now we can fill in the minimal template and use the geom_histogram()
function for creating a histogram.
ggplot(data = df) +
geom_histogram(aes(x = x))
![](i2ds_files/figure-html/ggplot-hist-1.png)
It is interesting to study the commonalities and differences of the two basic histograms created.
Using ggplot()
seems a little harder than hist(x)
, but embeds creating a histogram in a visualization framework that is much more flexible and powerful than the hist()
function.
Studying the documentation of geom_histogram()
(and geom_bar()
, on which geom_histogram()
is based) reveals hidden complexity.
A difficulty is that both geoms note x
and y
as required aesthetics, but we succeeded by only providing x
.
The reason for this lies in the fact that they y
-values of our histogram were automatically computed by a default argument stat = "bin"
.
For continuous variables, this counts the number of values that are within a specific interval (a so-called bin).
Internally, the histogram above is actually computed as a bar chart:
![](i2ds_files/figure-html/ggplot-hist-as-bar-1.png)
Whereas some visualizations merely show existing data values, many visualizations first need to compute something (e.g., count the frequency of values within a specific interval). Specifying how to compute what is the purpose of the stat
function.
The relations between geoms and corresponding stat
functions is probably the most tricky part of ggplot2.
As long as we are using geoms with their default statistics, we do not need to worry about stat
.
But when visualizations go wrong with ggplot2, it is often due to a mismatch between a geom and a stat function.
Studying ?geom_histogram()
also reveals additional parameters that we can use to adjust our histogram.
Whereas hist()
used a breaks
parameter to adjust the number of categories, geom_histogram()
provides two corresponding parameters (called bins
and binwidth
). The following examples also specify some colors (by setting col
and fill
to named R colors):
ggplot(data = df) +
geom_histogram(aes(x = x),
bins = 10, col = "black", fill = "deepskyblue")
![](i2ds_files/figure-html/ggplot-hist-parameters-1.png)
ggplot(data = df) +
geom_histogram(aes(x = x),
binwidth = 10, col = "white", fill = "hotpink")
![](i2ds_files/figure-html/ggplot-hist-parameters-2.png)
Note that we specified all additional parameters (numeric values for bins
or binwidth
, and the color values of col
and fill
) outside of the aes()
function. As we will see, it is sometimes possible to use parameters inside of aes()
, but then they are used as variables, rather than as constants (i.e., fixed values).
9.5.2 Bar charts
Simple bar charts: Create a bar chart showing the average flipper length for each species of penguins.
Complex bar charts: Transform the
pg
data to compute means, SE values, and corresponding confidence intervals for theflipper_length_mm
variable, then usegeom_bar()
andgeom_errorbar()
to plot these means with confidence intervals.
- The following function allows computing the standard error (SE) of a variable:
# SE formula:
std_err <- function(x, na.rm = FALSE) {
# Handle NA values:
if (na.rm){
nr_na <- sum(is.na(x))
if (nr_na > 0){
x <- stats::na.omit(x)
message(paste0("Removed ", nr_na, " NA values."))
}
}
# Compute SE:
sqrt(stats::var(x, na.rm = na.rm)/length(x))
} # std_err().
# Check:
std_err(pg$body_mass_g)
#> [1] NA
std_err(pg$body_mass_g, na.rm = TRUE)
#> [1] 43.36473
std_err(pg$flipper_length_mm, na.rm = TRUE)
#> [1] 0.7603704
# Compute summaries for flipper_length_mm of penguins per species:
tb_2 <- pg %>%
group_by(species) %>%
summarise(n = n(),
mean_flipper_length = mean(flipper_length_mm, na.rm = TRUE),
se_flipper_length = std_err(flipper_length_mm, na.rm = TRUE),
mn_conf_min = mean_flipper_length - 1.96 * se_flipper_length,
mn_conf_max = mean_flipper_length + 1.96 * se_flipper_length
)
tb_2
#> # A tibble: 3 × 6
#> species n mean_flipper_length se_flipper_length mn_conf_min mn_conf_max
#> <fct> <int> <dbl> <dbl> <dbl> <dbl>
#> 1 Adelie 152 190. 0.532 189. 191.
#> 2 Chinstrap 68 196. 0.865 194. 198.
#> 3 Gentoo 124 217. 0.585 216. 218.
# Bar plots with error bars:
ggplot(tb_2, aes(x = species, y = mean_flipper_length, fill = species)) +
geom_bar(stat = "identity") +
geom_errorbar(aes(ymin = mn_conf_min, ymax = mn_conf_max), width = .50) +
scale_fill_manual(values = my_cols) +
theme_unikn()
![](i2ds_files/figure-html/ggplot-ex-bar-errorbars-1.png)
ggplot(tb_2, aes(x = species, y = mean_flipper_length, fill = species)) +
geom_col() +
geom_errorbar(aes(ymin = mn_conf_min, ymax = mn_conf_max), width = .50) +
scale_fill_manual(values = my_cols) +
theme_unikn()
![](i2ds_files/figure-html/ggplot-ex-bar-errorbars-2.png)
9.5.3 Scatterplots
In Chapter 8, we created a scatterplot for a vector of numeric values x
and y
as follows:
# Data:
x <- 11:43
y <- c(sample(5:15), sample(10:20), sample(15:25))
# Scatterplot (of points):
plot(x = x, y = y,
main = "A positive correlation")
![](i2ds_files/figure-html/ggplot-base-plot-1.png)
- Re-create an analog scatterplot in ggplot2.
- Create a scatterplot of fuel consumption on the highway (
hwy
) by engine displacement (displ
) for thempg
data (inggplot2::mpg
) in both base R and ggplot2.
Solution
Using ggplot2:
df <- data.frame(x, y)
ggplot(df) +
geom_point(aes(x = x, y = y))
![](i2ds_files/figure-html/ggplot-scatterplot-1.png)
A scatterplot from the mpg
data:
Note that we usually want to add labels and titles, as well as modify other aesthetic features of visualizations.
- In base R (and a transparent color
my_col
definition using the unikn package):
# Define some color (from unikn, with transparency):
my_col <- unikn::usecol(unikn::Bordeaux, alpha = 1/4)
# With aesthetics (see ?par):
plot(x = mpg$displ, y = mpg$hwy, type = "p",
col = my_col, pch = 16, cex = 1.5,
main = "A basic scatterplot",
xlab = "Displacement", ylab = "MPG on highway"
)
![](i2ds_files/figure-html/ggplot-scatterplot-mpg-base-1.png)
- Using ggplot2:
ggplot(mpg) +
geom_point(aes(x = displ, y = hwy), size = 2, col = my_col) +
labs(x = "Displacement", y = "MPG on highway",
title = "A basic scatterplot") +
theme_minimal()
![](i2ds_files/figure-html/ggplot-scatterplot-mpg-ggplot-1.png)
9.5.4 The rule of 72
In finance, the rule of 72 is a heuristic strategy for estimating the doubling time of an investment. Dividing the number (72) by the interest percentage per period (usually years) yields the approximate number of periods required for doubling the initial investment. (See Wikipedia for details: en | de.)
- Create a line graph that compares the true doubling time with the heuristic estimates for a range of (positive) interest rates.
Hints:
- Consider wrapping your code into a function.
- Consider creating an interactive application of your function (think Shiny).
9.5.5 Advanced ggplot expressions
The following ggplot()
expressions are copied from the documentation of the corresponding geoms.
Run the code, inspect the result, and then try to explain how they work:
- A facet of histograms:
ggplot(economics_long, aes(value)) +
facet_wrap(~variable, scales = 'free_x') +
geom_histogram(binwidth = function(x) 2 * IQR(x) / (length(x)^(1/3)))
9.5.6 More expercises
For more exercises on using ggplot2,
see the Exercises of Chapter 2 of the ds4psy textbook (Neth, 2023a).
Bonus: Re-create the base R plots of Section 8.5 (of Chapter 8) in ggplot2.
For more detailed explanations of the connection between geoms and stats, see the ggplot2 documentation or the online article Demystifying stat layers in ggplot2 (by June Choe, 2020-09-26).↩︎
Note the relation between the mathematical and the computational notion of a function: The scatterplot visualization shows a relation by mapping values on some dimension \(x\) to values on some dimension \(y\). If we view the \(x\)-values as inputs and the \(y\)-values as outputs, the underlying function is the relation that transforms the former into the latter.↩︎