In j-effendy/dataviz:
knitr::opts_chunk$set(echo = FALSE, message = FALSE, warning = FALSE)
library(ggplot2)
library(tidyverse)
library(ebviz)
library(colorspace)
library(patchwork)
library(pals)
\newpage
Abstract
The work involved in the productions of visualizations in an edited book
is often laborious. With many authors involved, it is difficult to produce
an edited book that has a cohesive design. This project aims to reduce
workflow required in producing visualizations that are readable, cohesive, and
accessible in a Multilevel Regression and Poststratification (MRP) edited book.
The project contributes to a ggplot theme and accessible colour schemes. The
theme is developed based on Tufte's principle of data graphical excellence.
In addition, the project also develops two colour schemes for each type of
colour palettes: qualitative, sequential, and diverging palettes. These theme
and colour palettes are bundled together in an R package called ebviz.
An R package is used for this project as authors can easily install the package
and use its functions for their research.
Introduction
This project relates to an edited book about Multilevel Regression and
Poststratification (MRP). As is the case with other books, there will be
numerous visualizations that will be used within the book. These plots will need to be
readable, cohesive, and accessible. There is a group of authors (around one to
five authors) who work together on a chapter. The editors then work to combine
all the chapters to finalise the book. With many authors contributing to the
book, there is a need to balance authors' autonomy so that the final version of
the book have a consistent look/design. One mean to achieve this is by producing
visualizations that meet a set of criteria.
A style guide can be used to create a cohesive edited book. Implementing a
style guide, however, is time consuming and expensive. The editors/designers
need to spend a lot of effort and time to update or reproduce figures. Another
option that can be taken is by creating an R package that contains theme and
colour scale functions that work with ggplot2 [@ggplot2]. This will allow users
(in this case, authors) to add a theme and a colour scale function to their
plots. This is easy to use, and it will cause minimal disruptions to authors'
workflow. More importantly, this will allow the book to have visualizations that are
readable, cohesive, and accessible.
There are several aspects that need to be considered to create a cohesive edited
book. Firstly, a ggplot theme needs to be developed in a way that allows readers
to extract information from the visualizations as much as possible. In addition
to theme, various colour scales that are accessible need to be developed for
different data types (e.g. categorical and numerical information). The theme
and colour scales will be wrapped in an R package called ebviz. The R package
contains the following items:
- Function for the theme and colour scales.
- Documentations.
- Testing.
In the next section, the project will discuss how a theme that allows viewers
to extract information as much as possible from a graph was developed for
the project. Next, the project will explore the theory behind colour scales in
data visualizations, and describe the methods of choosing colour palettes for
the project. In the fourth section, the report will explain how the theme
and colour schemes will be wrapped in an R package. To put it briefly, the
report contributes themes, colour schemes, and a package called ebviz.
\newpage
Themes
The first aspect that needs to be considered is how a good visualization looks
like, and whether there are general rule that can be applied across different
types of plots. However, before looking at principles of good data visualizations,
it is better to understand what constitute a visualization. Azzam, Evergreen, and
Germuth [-@azzam2013data] stated that data visualization is defined by three
criteria:
- A process that is based quantitative or qualitative data
- The result of the process is a graph that is representative of the raw data
- The graph is readable and support exploration, examination, and communications
of data.
Without careful consideration, data visualization has the
potential to cause confusion resulting in misunderstanding or error in decision
making [@evergreen2011eval]. In addition, inefficient visual decoding may result in
users to not detect important aspect of the data [@cleveland1993model].
Unfortunately, inefficient graphs are not uncommon occurrence, especially with
increasing number of options in data visualization over the years. Tufte
[-@tufte2001visual] found that prevalent errors and data misrepresentation were
found in mass media. Another study found that a total of 1,365 cases of unnecessary
graphical elements were found out of 20,080 visualizations from a set of articles
published between 1996 and 2016 [@friedman2021data].
With widespread error in data visualizations, guidelines for better graphs
for quantitative and qualitative data were established. According to Tufte
[-@tufte2001visual], visualizations need to bring viewer's attention to the
substance of the data as opposed to graphic design, methodology or something
else. In addition, graphs should prevent distortion of the data and should
encourage viewers to compare different parts of data. Another suggestion from
Tufte is to avoid the usage of chartjunk. Any decorations that do not provide
new information to readers are considered to be chartjunk. This can easily cause
viewers to be confused or distracted from the data. Some examples of chartjunk
are moiré effect (the illusion of vibration caused by excessive patterns),
heavy grid lines, noisy background, and a fake perspective (i.e. 3-D charts).
Cleveland and McGill [-@cleveland1984graphical] conducted a study of effective
data visualization based on graphical perception to investigate humans' ability
to visually decode information encoded on graphs. In the study, people were
given elementary perceptual tasks to decode information from graphs. The tasks
are then ordered based on the accuracy of the performance. The order of the
elementary tasks from most to least accurate is:
- Position along a common scale
- Positions along nonaligned scales
- Length, direction, angle
- Area
- Volume, curvature
- Shading, colour saturation
This shows that there are general rules that can be applied across different
types of plots, and this can be computationally coded once for efficiency.
The next section will discuss how this can be implemented through the use of
ggplot2 [@ggplot2] theme.
Roles of Themes in Data Visualization
Themes in ggplot [@ggplot2] package are non-data components of data visualizations.
Themes allow user to customize plot the background, axis line, ticks, legend
position, grid, etc. As a default, ggplot [@ggplot2] uses theme_grey for plots.
However, ggplot [@ggplot2] provides other complete themes that users can use, such as
theme_bw() and theme_minimal().
Theme is especially important for a research book where there are many authors
contributing to the book. Without a set theme, it is likely that each author
will apply their own visualization styles for their research. Not only does
setting a theme for a research book allow authors to avoid spending their time
to customize their plots' appearance, but it also allows research book to
have consistency in their visualizations.
As stated in previous section, customizing plot themes is imperative in data
visualizations as non-data components impacts readers' ability to comprehend the
graph. The next section will discuss how general principles of graphical excellence
from Tufte can be applied to themes.
Criteria for good themes
There are five principles of data graphics that are proposed by Tufte
[-@tufte2001visual]. This project focuses on two of the principles that
specifically relate to a theme. In his book, Tufte introduced an important
concept: data-ink. Data-ink is defined as a non-redundant ink; it is ink used
on a graph that represent the underlying data.
$$
\text{Data-ink ratio} = \frac{\text{data-ink}}{\text{total ink used to print the graphic}}
$$
Good graphical representations can be achieved by reducing non-data ink as this
a way to maximize data-ink ratio [@tufte2001visual]. Graphical inks that
provide no new information about the underlying data are not beneficial to
readers and sometimes result in graphical clutters. However, maximizing the
data-ink ratio needs to be done cautiously as components of the graph that
provide useful information can be removed by mistake.
Design choices on data visualizations need to allow information to be presented in an
organised fashion. One option is to make the colour of non-data components (axis, labels,
grid lines, borders) to be grey, and the data component to be brightly
coloured, making the data to be the focus of attention [@stone2006choosing].
This was found to assist readers to better understand the roles and relationship between
different elements of the visualizations.
Potential Themes
g1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) +
geom_point(size=3) +
labs(title="Scatter Plot",
subtitle = "theme_grey()",
caption = "Data: mtcars")
g1
The default ggplot2 theme is theme_grey(), with grey background and white
grid lines. These colour combinations reduces the contrast between the background
and grid lines and gives the data plots more value-added appearance
[@carr1994using]. Moreover, grey background allows the plot to have a similar
colour to the rest of the texts. As a result, the graphics blend well with
the flow of a text [@WickhamHadleygEGf].
While having a grey background and white grid
lines have their benefits, some steps can be made to improve the theme. As
shown in figure \@ref(fig:scatter-grey), having a grey panel background
on top of a white plot background increases non-data components.
In addition, having both major and minor grid lines distracts readers from the
real data, and rarely needed as data visualizations are not meant to provide
precise quantitative value [@few2005grid].
A ggplot theme is developed for this project that is not only based on
Tufte principles of data graphics, but also to create a theme that is suitable
for a wide range of plots. There are two potential themes that were developed for
ebviz: theme_seagull() and theme_wombat().
Seagull Theme
w1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) +
geom_point(size=3) +
labs(title="Scatter Plot",
subtitle = "theme_wombat()",
caption = "Data: mtcars") +
theme_wombat()
w2 <- ggplot(mtcars, aes(factor(carb),fill=factor(carb))) +
geom_bar() +
scale_y_continuous(expand = c(0,0)) +
scale_x_discrete(expand = c(0,0)) +
labs(title="Bar Plot",
subtitle = "theme_wombat()",
caption = "Data: mtcars") +
theme_wombat()
w3 <- ggplot(economics_long, aes(date, value01, colour = variable)) +
geom_line() +
scale_y_continuous(expand = c(0,0)) +
scale_x_date(expand = c(0,0)) +
labs(title = "Line Plot",
subtitle = "theme_wombat()",
caption = "Data: economics_long") +
theme_wombat()
s1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) +
geom_point(size=3) +
labs(title="Scatter Plot",
subtitle = "theme_seagull()",
caption = "Data: mtcars") +
theme_seagull()
s2 <- ggplot(mtcars, aes(factor(carb),fill=factor(carb))) +
geom_bar() +
scale_y_continuous(expand = c(0,0)) +
scale_x_discrete(expand = c(0,0)) +
labs(title="Bar Plot",
subtitle = "theme_seagull()",
caption = "Data: mtcars") +
theme_seagull()
s3 <- ggplot(economics_long, aes(date, value01, colour = variable)) +
geom_line() +
scale_y_continuous(expand = c(0,0)) +
scale_x_date(expand = c(0,0)) +
labs(title = "Line Plot",
subtitle = "theme_seagull()",
caption = "Data: economics_long") +
theme_seagull()
bw1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) +
geom_point(size=3) +
labs(title="Scatter Plot",
subtitle = "theme_bw()",
caption = "Data: mtcars") +
theme_bw()
bw2 <- ggplot(mtcars, aes(factor(carb),fill=factor(carb))) +
geom_bar() +
labs(title="Bar Plot",
subtitle = "theme_bw()",
caption = "Data: mtcars") +
theme_bw()
bw3 <- ggplot(economics_long, aes(date, value01, colour = variable)) +
geom_line() +
labs(title = "Line Plot",
subtitle = "theme_bw()",
caption = "Data: economics_long") +
theme_bw()
(s1 + w1 + bw1) / (s2 + w2 + bw2) / (s3 + w3 + bw3)
Plot examples with theme_seagull() can be seen in figure \@ref(fig:theme-comparison).
This ggplot2 theme is built based on theme_minimal(). The idea behind this
theme is to reduce non-data components as much as possible by maximizing white
space around the plot. The %+replace% operator is used to change the elements of
the plots. Axis line on the y axis is removed, and axis line on the x axis is
thicker. The major vertical grid lines and minor grid lines on both axis are removed.
The grid lines, as shown in figure \@ref(fig:theme-comparison), are visually
muted so that it does not distract the data points. Cleveland
[-@cleveland1993visualizing] found that grid lines enhance readers' perceptions
in making visual comparison and accuracy of information extraction.
Wombat Theme
Another potential theme for this project is called theme_wombat().
This theme is developed based on ggplot2 basic theme theme_bw().
The %+replace% operator is also used to develop theme_wombat(). It can be seen
in figure \@ref(fig:theme-comparison) that minor grid lines on both axis are
removed for this theme. Similar to theme_seagull(), theme_wombat() has
white background and light grey grid lines.
Themes Comparison
This section is going to compare the potential themes theme_seagull() and
theme_wombat() with ggplot2 [@ggplot2] built in theme, theme_bw().
The theme_bw is chosen for this comparison as it is one of the built-in themes
that minimizes non-data components while still preserving light coloured grid lines.
Side by side comparison of scatter plots, bar charts, and line charts between
the three themes can be found in figure \@ref(fig:theme-comparison). One of
the differences between ebviz potential themes and theme_bw() is the size
of the text. The axis title, plot title, subtitle, and caption with
theme_seagull() and theme_wombat() are easier to read in comparison to
theme_bw(). Secondly, the legend position is located on the bottom of the
plot as opposed to the right of the plot. The reason to the change in legend
position is to maximize the data-ink ratio of the plots. Putting the legend
on the right of the plot takes up the space that can be used to display the
underlying data. Another way of maximizing data-ink ratio is by reducing
the margin around the plot. As shown in the bar plots and line plots in figure \@ref(fig:theme-comparison), there are no extra space between the bars/lines
with the axis lines in the plots with theme_seagull() and theme_wombat().
One of the objectives of theme_seagull() is to have a theme that increases
data-ink ratio as much as possible. In comparison to other two themes,
theme_seagull() has the least amount of grid lines and borders around the
plot, making it to have the highest data-ink ratio. While there is no vertical
grid lines, major horizontal grid lines are preserved as they aid readers to
approximate and compare values between data points.
The theme_wombat() is developed to be the middle ground between theme_seagull()
and theme_bw(). In comparison to the ggplot2 built-in theme, this theme
is less cluttered as it does not have minor grid lines on both axis. Unlike
theme_seagull(), however, this theme retains major vertical grid lines.
The Chosen Theme
As stated above, theme_seagull() has the least clutter between the two
potential themes.However, theme_wombat() is likely to assist readers
to perform comparison more accurately than theme_seagull()
while still maintaining a high level of data-ink ratio. Using the scatter plot
example in figure \@ref(fig:theme-comparison), the major vertical grid lines
from theme_wombat() aids readers' ability to estimate the value of data points along the
x axis. This shows that theme_wombat() is a good compromise of minimizing
clutters as well as allowing readers to extract information from the graph as
much as possible.
Choosing Colour Palletes
In addition to themes, colour also plays an important role in data
visualizations. Proper use of colour in a visualization enhances the quality
of information being presented [@zeileis2009escaping]. One of the most important
use of colour in data visualizations is to differentiate one element from one
another [@stone2006choosing]. However, a poor choice of colour palette can impede
viewers' ability to effectively extract information from the visualization.
With a simple visual feature, some objects can be easily recognised without
requiring much attention. The target object will "pop-out"
in field of vision regardless of how many distractors surrounding it
[@treisman1985preattentive]. The use of colour in data visualization can facilitate
this "pop-out" as it makes some elements to be easier to be found.
set.seed(20211028)
x <- rnorm(20)
y <- rnorm(20)
color <- c("red","red","blue","red","blue","red","red","red","red","red","red",
"blue","red","red","red","red","red","red","red","red")
pop_out <- data.frame(x,y,color)
p1 <- pop_out %>% ggplot(aes(x = x,
y = y,
color = color)) +
geom_point() +
theme_wombat() +
theme(legend.position = "none",
axis.title.x = element_blank(),
axis.title.y = element_blank(),
axis.text = element_blank(),
axis.ticks = element_blank()) +
ggtitle("Color")
p2 <- pop_out %>% ggplot(aes(x = x,
y = y,
shape = color)) +
geom_point() +
theme_wombat() +
theme(legend.position = "none",
axis.title.x = element_blank(),
axis.title.y = element_blank(),
axis.text = element_blank(),
axis.ticks = element_blank()) +
ggtitle("Shape")
p1 + p2
Figure \@ref(fig:pop-out) is an example, which is adapted from Healy's [-@healy2018data]
book about data visualization, that demonstrates how easy it is for colour to be
recognised in a visualization. The figure contains three "odd" observations that
are different than the rest. To represent this, the first panel uses colour to differentiate
the different kinds of observations. The three observations are coloured red, and
the rest are blue. On the other hand, the second panel uses shape to represent
this feature. Circle is used to represent the "odd" observations, and triangle is
used for the rest. While readers have no problem in identifying the three
observations in the second panel, it is much easier to find these observations
on the first panel. This shows that finding the observations of interest in the
first panel requires significantly lower level of attentions compared to the
second panel.
In addition to assisting viewers in differentiating one attribute to another,
colours can also be used to evoke viewers' emotions. Emotion is important
in visualizations for story telling, communicative intent, and engagement
[@bartram2017affective]. It was also found that emotions can significantly
improve many cognitive processes such as attention, learning, decision making,
and problem-solving [@harrison2012exploring].
This project will focus to create colour schemes that assist users to easily
differentiate one data attribute from another. To achieve this, it is important
to choose these schemes using a colour space that works well with human perception
of colour. The next section will discuss the different options of colour space,
and which one will be used for the project.
Colour Spaces
To effectively use colours in visualizations, it is imperative to understand
how human colour perception works. It was theorized that human colour vision evolved
in three distinct stages [@ihaka2003colour]:
- Perception of light and dark (monochrome)
- Perception of yellow/blue contrast
- Perception of green/red contrast
As a result of these three colour axes, colours are naturally described as
locations in a three dimensional space. While these colour axis provide
physiological description, human perception of colours do not corresponds to
these axes. Human perception of colour corresponds to the polar coordinates
of the colour coordinate (yellow/blue and green/red), and a dark/light axis
[@zeileis2006choosing].
Unfortunately, not all software packages use colour space was developed based on
human perception. For example, the most common colour map used by software
packages, the Red-Green-Blue (RGB) [@stauffer2015somewhere], was primarily
developed to correspond with colour generation on computer screens rather than
to correspond with human perception of colour [@zeileis2009escaping].
Zeileis, Hornik, and Murrell further stated that it is practically impossible
for a human to control the perceptual properties of a colour within this colour
model as there is no single dimension that relates to hue or brightness of a
colour
One popular alternative to RGB system is Hue-Saturation-Value (HSV). This model
is a simple transformation of RGB system and aims to capture the dimensions of
human visual perceptions [@smith1978color]. Unfortunately, the result of mapping
HSV's three dimensions to three dimensions of human perceptions is very poor
[@zeileis2006choosing]. Moreover, it is fairly difficult to find a set of colours
that are in harmony in HSV colour space [@zeileis2006choosing]. Furthermore, HSV encourages
the use of colours that are highly saturated. While these colours are good to
attract viewers' attention, it is difficult to look at for a long time
[@zeileis2009escaping].
HCL, which is short for Hue-Chroma-Luminance, was developed to overcome many
problems encountered by other colour spaces. This colour space is developed
based on the three colour attributes that corresponds to human perceptions:
hue, chroma, and luminance [@sarifuddin2005new]. It was found that the three
dimensions from this colour space matches human visual perception well
[@zeileis2019colorspace]. In general, HCL colour space can be used to generate
three types of colour scales: qualitative, sequential, and diverging.
Qualitative Palettes
This colour palette is designed for coding categorical information where the
categories do not have particular ordering [@zeileis2019colorspace]. In other
words, this colour palette will give the same perceptual weight to each category
so that there will be no perception of one group to be larger or more important
than any other one [@zeileis2006choosing]. This colour palette is commonly
applied to bar plots or box plots.
Sequential Palettes
Sequential pallets are used for coding ordered/numerical information that span
in a certain interval [@zeileis2006choosing]. The low values of the interval
are considered to be uninteresting, whereas high values are considered to be
interesting [@zeileis2019colorspace]. Low values are usually depicted by lighter
colours, while darker colours are used for high values [@harrower2003colorbrewer].
This palette is usually used on heatmaps.
Diverging Palletes
Similar to sequential palette, diverging pallets are designed for coding ordered/
numerical information that ranges in an interval. The difference between
diverging and sequential palletes is that diverging pallets include a neutral
value [@zeileis2006choosing]. This palette should be used when an important data
class or break points need to be highlighted [@harrower2003colorbrewer]. Harrower
and Brewer further stated that this critical data class or break points is
highlighted by a change in in hue and luminance, and should represent a critical
value of the data (e.g. mean, median, or zero). For instance, the average GDP
per-capita in a choropleth map can be emphasized (i.e. set as neutral value) so
that the values above and below the average can be shown in different hues.
Colour palette of choice
In addition to plot themes, this project also contributes two colour schemes for
each of the three colour palettes type mentioned in the previous section.
The pals packages used to build the colour palettes
and assess them. The chosen colour palettes need to be assessed to ensure
that they achieve their intended purpose. For example, the qualitative colour palette
needs to avoid implicit ordering as this palette is intended for representing categorical
variables. In addition, we need to ensure that people with colour vision
impairments are able to distinguish the different colours in the chosen palettes.
The pal.safe() function from pals package allows us to see how colour palettes
are seen by deutan, protan, and tritan colour-blind people.
The methodology of choosing colour palettes that are not only accessible but also
enhance the quality of visualizations will be discussed in this section. Firstly, the
report will explore common strategies used in choosing qualitative, sequential,
and diverging palettes. Then, examples of palettes made from these strategies will
be assessed. Lastly, the report will show how colour schemes for
qualitative, sequential and diverging palettes are chosen for the project.
quokka_palette <- c("#8C01E6", "#93B1EB", "#AB313D", "#D6CA24")
quoll_palette <- c('#8D5698', '#EFD36F', '#247A3E', '#17B0A8')
seal_palette <- colorspace::sequential_hcl(7,
h = c(0, -140),
c = c(80,40),
l = c(40, 75),
power = c(1,1))
snake_palette <- colorspace::sequential_hcl(6,
h = c(250, 90),
c = c(40, 55),
l = c(33,98),
power = c(0.5,1),
rev = TRUE)
dingo_palette <- colorspace::diverge_hcl(7,
h = c(260, 0),
c = 80,
l = c(30, 90),
power = 1.5)
dolphin_palette <- colorspace::diverging_hcl(n = 7,
h = c(265, 45),
c = 80,
l = c(70, 95),
power = c(0.5, 2),
rev = TRUE)
# colorblind filter
pal.safe(quokka_palette, main = "Quokka Palette")
pal.safe(quoll_palette, main = "Quoll Palette")
pal.safe(seal_palette, main = "Seal Palette")
pal.safe(snake_palette, main = "Snake Palette")
pal.safe(dingo_palette, main = "Dingo Palette")
pal.safe(dolphin_palette, main = "Dolphin Palette")
pal.safe(qualitative_hcl(4, palette = "harmonic"), main = "Harmonic Palette")
Chosen Qualitative Palettes
One common strategy to choose this type of palette is by keeping chroma and luminance
relatively constant, and hue varied; making all colours in the palette to balance
towards the same grey [@zeileis2009escaping]. While this strategy is relatively
simple, keeping chroma and luminance fixed may not be the best approach to
choose colour palette that are accessible.
Panel (g) in figure \@ref(fig:specplot) shows an example of a qualitative colour palette from
colorspace package called harmonic. As shown in the figure, chroma
and luminance remain constant, while the hue varies across the four colours.
It is fair to say that this colour set works relatively well to show categorical
information as it does not give implied ordering across different categories.
This method, however, does not have the same result for people with visual
impairment. It can be seen in figure \@ref(fig:color-blind) that people
with colour-blindness will have difficulty in distinguishing some colours from
harmonic palette. For instance, there are three colours that look similar as seen
by people with tritan colour-blindness. Moreover, this colour palette results in
implied ordering of different data group for people with visual impairment.
specplot(quokka_palette, main = "Quokka Palette")
specplot(quoll_palette, main = "Quoll Palette")
specplot(seal_palette, main = "Seal Palette")
specplot(snake_palette, main = "Snake Palette")
specplot(dingo_palette, main = "Dingo Palette")
specplot(dolphin_palette, main = "Dolphin Palette")
specplot(qualitative_hcl(4, palette = "harmonic"), main = "Harmonic Palette")
The two qualitative colour palettes developed for the project are called
Quokka and Quoll palettes. Panel (a) and (b) in figure \@ref(fig:specplot) shows
the HCL colour spectrum plot for these palettes. In
contrast with the Harmonic Palette, Quokka and Quoll palettes have differing
hues, chroma and luminance across all colours. As a result, each colour in
these palettes can easily be distinguished by people with deutan, protan, and
tritan colour-blindness (see figure \@ref(fig:color-blind)). Furthermore,
natural ordering implied by Quokka and Quoll is less prominent than Harmonic
palette.
Chosen Sequential Palettes
The strategy to choose a sequential palette is different to that of qualitative
palette. One easy way to create a sequential palette is by choosing two values
for hue (h1 and h2), chroma (c1 and c2), luminance (l1 and l2), and power
transformations (p1 and p2) for chroma and luminance [@zeileis2006choosing].
These values of hues, chroma, luminance, and power transformation are used
as inputs to sequential_hcl function to produce a set of sequential palette.
The power transformation allows the chroma and luminance to increase from low
values to high values in a nonlinear fashion.
The two sequential palettes created for this project are Seal and
Snake palettes. As shown in panel (c) and (d) in figure \@ref(fig:specplot),
these palettes use multi-hue schemes as opposed to single-hue schemes. Multi-hue
palettes were chosen for the project as it provides better contrast between different
values compared to a single-hue sequential palette [@harrower2003colorbrewer].
Moreover, multi-hue sequential palettes avoid the use of extreme colours, as well
as allowing viewers to better discriminate the middle colours [@zeileis2019colorspace].
In addition, figure \@ref(fig:color-blind) shows that people with deutan, protan,
and tritan colour blindness are able to distinguish the colours in the palette.
This means that the colour palettes will allow readers with colour vision impairment
to see implied order and how the data ranges from low to high values.
Chosen Diverging Palettes
The approach in choosing diverging palettes is similar to choosing sequential
palettes. Diverging palettes can be built by choosing two values for
hues (h1 and h2), chroma (c1 and c2), luminance (l1 and l2), and power
transformation (p1 and p2) for chroma and luminance [@zeileis2006choosing].
These values will be taken as inputs to function diverge_hcl() to create a diverging
palette. It will automatically choose the neutral colour for the palette.
Three conditions need to be met by a diverging palette to avoid colour
confusions. Colour confusions are caused by naming confusions, colour vision
impairment, and simultaneous contrast confusion [@brewer1996guidelines].
Naming confusion happens when people use two basic colour names for the same
colour (e.g. a colour was called red when presented once, then pink at
another time). On the other hand, simultaneous contrast
refers to a phenomena where colours are affected by their opposite (e.g.
a colour will appear lighter when surrounded by a dark colour, and a colour will
appear warmer when surrounded by a cool colour) [@brewer1992review]. This can be
prevented by avoiding selecting pairs of colours that are opposite in the colour
wheel for diverging palettes [@brewer1996guidelines].
Dingo and Dolphin palettes are diverging palettes developed for the project.
These palettes are shown in panel (e) and (f) in figure \@ref(fig:specplot).
Red/blue pairs and orange/blue hue pairs are chosen for Dingo and Dolphin
palettes respectively as these pairs create no naming confusions and
simultaneous contrast confusions [@brewer1996guidelines]. Moreover, these
palettes create no colour confusion due to colour vision impairment as colours in
Dingo and Dolphin palettes can be easily distinguished by people with colour-blindness
(shown in figure \@ref(fig:color-blind)). Another reason for choosing these pairs
is because red/blue pair is often used in the topic of the edited book, MRP,
to show electoral result (e.g. Democratic and Republican party). Therefore,
it is important to have a red/blue scale and a non-red/blue scale for functionality
in the book.
R Package
One of the objectives of this project is to create a solution for authors and
editors of the research book to create visualizations that are readable,
cohesive, and accessible. The solution needs to be easy to use so that it
does not disrupt authors' workflow. To achieve this, R code necessary
to create the chosen ggplot2 [@ggplot2] theme and colour schemes will be bundled in an R
package called ebviz. R packages are extensions to R that can comprise of
codes, data, documentation, and tests [@wickham2015r].
One of the reasons for creating an R package for this project is because of the
ease of use. Codes, documentation, and tests in an R package are built in a
standardized format so that any R user can install it [@wickham2015r]. In
addition, it is fairly simple to create an R package. Packages such as roxygen2
[@roxygen2], devtools [@devtools], and testthat [@testthat] provide great
assistance in documenting and testing the package. Another benefit of using an
R package is that any updates to the theme and colour schemes are not time
consuming for the users. Any updates to themes and colour schemes simply need to
be applied to the code in the R package. There is no need to update the code
for each plots in the book; it just needs to be recompiled. This shows that
using an R package will allow us to achieve one of the aims of the project:
to provide a solution that causes minimal disruption to authors' work.
For users to be able to easily access an R package, it needs to be stored in a
repository. For this project, the ebviz package will be stored in GitHub.
GitHub is a popular choice for a repository as it has the integration with git,
a version control software. Version control records changes applied to the
package to a file. This allows user to recall specific versions of the package
at later time [@loeliger2012version]. An alternate option to GitHub is
Comprehensive R Archive Network (CRAN). CRAN, however, is not suitable for this
package as ebviz has a fairly small scope and is intended for a small group
of people (i.e. authors of the MRP research book). CRAN is usually a repository
used for more complex R packages intended for large group of people.
In the next section, the report will discuss how the ebviz package was
created. It will discuss each component of the package, and what tools
were used in developing the package.
ebviz Components
One of the most important components of a package is the R code. One firm
rule for an R package is that its R functions need to be stored in R scripts
[@wickham2015r]. Following this rule, the ebviz has its R code in R
scripts, which are stored in R directory. The chosen colour palettes can be
accessed from the scale_colour_ebviz and scale_fill_ebviz functions which
are stored in colour.R. These functions take the name of the palette as an
input (e.g. quokka). The ggplot theme chosen for the project can be used by
function theme_wombat() which is stored in an R script called theme.R.
In addition to R codes, documentation is also a crucial aspect of a well-made
package [@wickham2015r]. In the absence of documentation, users will not have a guide
that assists them to use the package properly. Proper documentation also assists
the authors of the package for future use. This assists authors to remember the
purpose of the functions. One type of documentation is an object documentation;
this can be accessed by the ? operator or help(). Information such as function
description, usage, argument(s), and examples can be found in the documentation.
The roxygen2 [@roxygen2], devtools [@devtools], and usethis [@usethis]
packages are used in creating object documentation.
Another crucial aspect to package development is testing. While testing add
extra work to package development, it ensures that the package works
as intended. It also ensures that the package works for edge cases - the less
common uses that a user might use. In addition, testing ensures error and warning
messages to appear when required (and not appear when they are not supposed to).
In other words, testing provide protection and quality assurance for users.
We use the packages testthat [@testthat] and usethis [@usethis] to write
the tests for the ebviz package. One example of a test that was used for
this package is one that ensures that users use valid input when using
scale_colour_ebviz and scale_fill_ebviz.
Conclusions and Limitations
One of the challenges in producing a research book is creating a book that has
a consistent look or design. Typically, there is a group of authors that work on
every chapter of an edited book. These authors will have their own unique style in
producing their visualizations for the book. While this is not a bad thing,
visualizations style needs to be standardized to keep the look of the book
consistent. Not only do these visualizations need to have a consistent design,
but they also need to be readable and accessible.
The aim of the project is to provide a solution that can ensure that
visualizations in the research book meet the criteria mentioned in the previous
paragraph in an efficient manner. To achieve this objective, a ggplot theme was
developed for the project. The chosen theme was developed based on Tufte's
principles of graphical excellence. In addition, several colour scales for various
data types were also created. The colour palettes were developed using a colour
space that was based on human perception. Furthermore, to ensure that color is
accessible, we visualized how these colours as they would be seen by people with
colour visual impairment. These theme and colour scales were developed to ensure
that the viewers to extract information as much as possible from the
visualizations. The theme and colour schemes are bundled in an R package for
efficiency.
One limitation in the package is related to colour palettes, specifically
the chosen qualitative palettes. As mentioned in the previous section,
this palette is specifically used to show un-ordered categorical information.
It is therefore imperative to avoid implying that one group is larger or more
important than other groups. However, removing implied ordering while ensuring
accessibility proved to be difficult. Figure \@ref(fig:color-blind) shows that
while the Quokka and Quoll palettes are better than Harmonic palette in
avoiding implied ordering for people with colour visual impairment, there are still
some colours that look paired. In addition, Quokka and Quoll only have four
colours. It is difficult to produce a qualitative palette that have more than
four colours while still ensuring that people with colour impairment are able to
distinguish the colours from one another.
One area of future work for this project is to work on writing additional
testings. For instance, a test can be developed to ensure that a warning message will appear
when the theme function is applied to a non-ggplot object. Another test that
can be applied to the package is to ensure an error to pop-up when applying
a diverging palette to a categorical variable. Another approach that can be
done in the future is by having the users (i.e. the authors of the MRP books)
to test the functionality of the package for a period of time. This allows
the users to provide feedback, and this can be used to improve the package.
These testings are beneficial to protect the quality of the package.
In summary, this project contributes an R package containing a ggplot theme and
six colour palettes developed to assist MRP book authors in creating visualizations that
are readable, cohesive, and accessible.
\newpage
Reference
j-effendy/dataviz documentation built on Dec. 20, 2021, 8:04 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
knitr::opts_chunk$set(echo = FALSE, message = FALSE, warning = FALSE) library(ggplot2) library(tidyverse) library(ebviz) library(colorspace) library(patchwork) library(pals)
\newpage
Abstract
The work involved in the productions of visualizations in an edited book
is often laborious. With many authors involved, it is difficult to produce
an edited book that has a cohesive design. This project aims to reduce
workflow required in producing visualizations that are readable, cohesive, and
accessible in a Multilevel Regression and Poststratification (MRP) edited book.
The project contributes to a ggplot theme and accessible colour schemes. The
theme is developed based on Tufte's principle of data graphical excellence.
In addition, the project also develops two colour schemes for each type of
colour palettes: qualitative, sequential, and diverging palettes. These theme
and colour palettes are bundled together in an R package called ebviz.
An R package is used for this project as authors can easily install the package
and use its functions for their research.
Introduction
This project relates to an edited book about Multilevel Regression and Poststratification (MRP). As is the case with other books, there will be numerous visualizations that will be used within the book. These plots will need to be readable, cohesive, and accessible. There is a group of authors (around one to five authors) who work together on a chapter. The editors then work to combine all the chapters to finalise the book. With many authors contributing to the book, there is a need to balance authors' autonomy so that the final version of the book have a consistent look/design. One mean to achieve this is by producing visualizations that meet a set of criteria.
A style guide can be used to create a cohesive edited book. Implementing a
style guide, however, is time consuming and expensive. The editors/designers
need to spend a lot of effort and time to update or reproduce figures. Another
option that can be taken is by creating an R package that contains theme and
colour scale functions that work with ggplot2 [@ggplot2]. This will allow users
(in this case, authors) to add a theme and a colour scale function to their
plots. This is easy to use, and it will cause minimal disruptions to authors'
workflow. More importantly, this will allow the book to have visualizations that are
readable, cohesive, and accessible.
There are several aspects that need to be considered to create a cohesive edited
book. Firstly, a ggplot theme needs to be developed in a way that allows readers
to extract information from the visualizations as much as possible. In addition
to theme, various colour scales that are accessible need to be developed for
different data types (e.g. categorical and numerical information). The theme
and colour scales will be wrapped in an R package called ebviz. The R package
contains the following items:
- Function for the theme and colour scales.
- Documentations.
- Testing.
In the next section, the project will discuss how a theme that allows viewers
to extract information as much as possible from a graph was developed for
the project. Next, the project will explore the theory behind colour scales in
data visualizations, and describe the methods of choosing colour palettes for
the project. In the fourth section, the report will explain how the theme
and colour schemes will be wrapped in an R package. To put it briefly, the
report contributes themes, colour schemes, and a package called ebviz.
\newpage
Themes
The first aspect that needs to be considered is how a good visualization looks like, and whether there are general rule that can be applied across different types of plots. However, before looking at principles of good data visualizations, it is better to understand what constitute a visualization. Azzam, Evergreen, and Germuth [-@azzam2013data] stated that data visualization is defined by three criteria:
- A process that is based quantitative or qualitative data
- The result of the process is a graph that is representative of the raw data
- The graph is readable and support exploration, examination, and communications of data.
Without careful consideration, data visualization has the potential to cause confusion resulting in misunderstanding or error in decision making [@evergreen2011eval]. In addition, inefficient visual decoding may result in users to not detect important aspect of the data [@cleveland1993model]. Unfortunately, inefficient graphs are not uncommon occurrence, especially with increasing number of options in data visualization over the years. Tufte [-@tufte2001visual] found that prevalent errors and data misrepresentation were found in mass media. Another study found that a total of 1,365 cases of unnecessary graphical elements were found out of 20,080 visualizations from a set of articles published between 1996 and 2016 [@friedman2021data].
With widespread error in data visualizations, guidelines for better graphs for quantitative and qualitative data were established. According to Tufte [-@tufte2001visual], visualizations need to bring viewer's attention to the substance of the data as opposed to graphic design, methodology or something else. In addition, graphs should prevent distortion of the data and should encourage viewers to compare different parts of data. Another suggestion from Tufte is to avoid the usage of chartjunk. Any decorations that do not provide new information to readers are considered to be chartjunk. This can easily cause viewers to be confused or distracted from the data. Some examples of chartjunk are moiré effect (the illusion of vibration caused by excessive patterns), heavy grid lines, noisy background, and a fake perspective (i.e. 3-D charts).
Cleveland and McGill [-@cleveland1984graphical] conducted a study of effective data visualization based on graphical perception to investigate humans' ability to visually decode information encoded on graphs. In the study, people were given elementary perceptual tasks to decode information from graphs. The tasks are then ordered based on the accuracy of the performance. The order of the elementary tasks from most to least accurate is:
- Position along a common scale
- Positions along nonaligned scales
- Length, direction, angle
- Area
- Volume, curvature
- Shading, colour saturation
This shows that there are general rules that can be applied across different
types of plots, and this can be computationally coded once for efficiency.
The next section will discuss how this can be implemented through the use of
ggplot2 [@ggplot2] theme.
Roles of Themes in Data Visualization
Themes in ggplot [@ggplot2] package are non-data components of data visualizations.
Themes allow user to customize plot the background, axis line, ticks, legend
position, grid, etc. As a default, ggplot [@ggplot2] uses theme_grey for plots.
However, ggplot [@ggplot2] provides other complete themes that users can use, such as
theme_bw() and theme_minimal().
Theme is especially important for a research book where there are many authors contributing to the book. Without a set theme, it is likely that each author will apply their own visualization styles for their research. Not only does setting a theme for a research book allow authors to avoid spending their time to customize their plots' appearance, but it also allows research book to have consistency in their visualizations.
As stated in previous section, customizing plot themes is imperative in data visualizations as non-data components impacts readers' ability to comprehend the graph. The next section will discuss how general principles of graphical excellence from Tufte can be applied to themes.
Criteria for good themes
There are five principles of data graphics that are proposed by Tufte [-@tufte2001visual]. This project focuses on two of the principles that specifically relate to a theme. In his book, Tufte introduced an important concept: data-ink. Data-ink is defined as a non-redundant ink; it is ink used on a graph that represent the underlying data.
$$ \text{Data-ink ratio} = \frac{\text{data-ink}}{\text{total ink used to print the graphic}} $$
Good graphical representations can be achieved by reducing non-data ink as this a way to maximize data-ink ratio [@tufte2001visual]. Graphical inks that provide no new information about the underlying data are not beneficial to readers and sometimes result in graphical clutters. However, maximizing the data-ink ratio needs to be done cautiously as components of the graph that provide useful information can be removed by mistake.
Design choices on data visualizations need to allow information to be presented in an organised fashion. One option is to make the colour of non-data components (axis, labels, grid lines, borders) to be grey, and the data component to be brightly coloured, making the data to be the focus of attention [@stone2006choosing]. This was found to assist readers to better understand the roles and relationship between different elements of the visualizations.
Potential Themes
g1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) + geom_point(size=3) + labs(title="Scatter Plot", subtitle = "theme_grey()", caption = "Data: mtcars") g1
The default ggplot2 theme is theme_grey(), with grey background and white
grid lines. These colour combinations reduces the contrast between the background
and grid lines and gives the data plots more value-added appearance
[@carr1994using]. Moreover, grey background allows the plot to have a similar
colour to the rest of the texts. As a result, the graphics blend well with
the flow of a text [@WickhamHadleygEGf].
While having a grey background and white grid lines have their benefits, some steps can be made to improve the theme. As shown in figure \@ref(fig:scatter-grey), having a grey panel background on top of a white plot background increases non-data components. In addition, having both major and minor grid lines distracts readers from the real data, and rarely needed as data visualizations are not meant to provide precise quantitative value [@few2005grid].
A ggplot theme is developed for this project that is not only based on
Tufte principles of data graphics, but also to create a theme that is suitable
for a wide range of plots. There are two potential themes that were developed for
ebviz: theme_seagull() and theme_wombat().
Seagull Theme
w1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) + geom_point(size=3) + labs(title="Scatter Plot", subtitle = "theme_wombat()", caption = "Data: mtcars") + theme_wombat() w2 <- ggplot(mtcars, aes(factor(carb),fill=factor(carb))) + geom_bar() + scale_y_continuous(expand = c(0,0)) + scale_x_discrete(expand = c(0,0)) + labs(title="Bar Plot", subtitle = "theme_wombat()", caption = "Data: mtcars") + theme_wombat() w3 <- ggplot(economics_long, aes(date, value01, colour = variable)) + geom_line() + scale_y_continuous(expand = c(0,0)) + scale_x_date(expand = c(0,0)) + labs(title = "Line Plot", subtitle = "theme_wombat()", caption = "Data: economics_long") + theme_wombat() s1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) + geom_point(size=3) + labs(title="Scatter Plot", subtitle = "theme_seagull()", caption = "Data: mtcars") + theme_seagull() s2 <- ggplot(mtcars, aes(factor(carb),fill=factor(carb))) + geom_bar() + scale_y_continuous(expand = c(0,0)) + scale_x_discrete(expand = c(0,0)) + labs(title="Bar Plot", subtitle = "theme_seagull()", caption = "Data: mtcars") + theme_seagull() s3 <- ggplot(economics_long, aes(date, value01, colour = variable)) + geom_line() + scale_y_continuous(expand = c(0,0)) + scale_x_date(expand = c(0,0)) + labs(title = "Line Plot", subtitle = "theme_seagull()", caption = "Data: economics_long") + theme_seagull() bw1 <- ggplot(mtcars, aes(mpg,disp,color=factor(carb))) + geom_point(size=3) + labs(title="Scatter Plot", subtitle = "theme_bw()", caption = "Data: mtcars") + theme_bw() bw2 <- ggplot(mtcars, aes(factor(carb),fill=factor(carb))) + geom_bar() + labs(title="Bar Plot", subtitle = "theme_bw()", caption = "Data: mtcars") + theme_bw() bw3 <- ggplot(economics_long, aes(date, value01, colour = variable)) + geom_line() + labs(title = "Line Plot", subtitle = "theme_bw()", caption = "Data: economics_long") + theme_bw() (s1 + w1 + bw1) / (s2 + w2 + bw2) / (s3 + w3 + bw3)
Plot examples with theme_seagull() can be seen in figure \@ref(fig:theme-comparison).
This ggplot2 theme is built based on theme_minimal(). The idea behind this
theme is to reduce non-data components as much as possible by maximizing white
space around the plot. The %+replace% operator is used to change the elements of
the plots. Axis line on the y axis is removed, and axis line on the x axis is
thicker. The major vertical grid lines and minor grid lines on both axis are removed.
The grid lines, as shown in figure \@ref(fig:theme-comparison), are visually
muted so that it does not distract the data points. Cleveland
[-@cleveland1993visualizing] found that grid lines enhance readers' perceptions
in making visual comparison and accuracy of information extraction.
Wombat Theme
Another potential theme for this project is called theme_wombat().
This theme is developed based on ggplot2 basic theme theme_bw().
The %+replace% operator is also used to develop theme_wombat(). It can be seen
in figure \@ref(fig:theme-comparison) that minor grid lines on both axis are
removed for this theme. Similar to theme_seagull(), theme_wombat() has
white background and light grey grid lines.
Themes Comparison
This section is going to compare the potential themes theme_seagull() and
theme_wombat() with ggplot2 [@ggplot2] built in theme, theme_bw().
The theme_bw is chosen for this comparison as it is one of the built-in themes
that minimizes non-data components while still preserving light coloured grid lines.
Side by side comparison of scatter plots, bar charts, and line charts between
the three themes can be found in figure \@ref(fig:theme-comparison). One of
the differences between ebviz potential themes and theme_bw() is the size
of the text. The axis title, plot title, subtitle, and caption with
theme_seagull() and theme_wombat() are easier to read in comparison to
theme_bw(). Secondly, the legend position is located on the bottom of the
plot as opposed to the right of the plot. The reason to the change in legend
position is to maximize the data-ink ratio of the plots. Putting the legend
on the right of the plot takes up the space that can be used to display the
underlying data. Another way of maximizing data-ink ratio is by reducing
the margin around the plot. As shown in the bar plots and line plots in figure \@ref(fig:theme-comparison), there are no extra space between the bars/lines
with the axis lines in the plots with theme_seagull() and theme_wombat().
One of the objectives of theme_seagull() is to have a theme that increases
data-ink ratio as much as possible. In comparison to other two themes,
theme_seagull() has the least amount of grid lines and borders around the
plot, making it to have the highest data-ink ratio. While there is no vertical
grid lines, major horizontal grid lines are preserved as they aid readers to
approximate and compare values between data points.
The theme_wombat() is developed to be the middle ground between theme_seagull()
and theme_bw(). In comparison to the ggplot2 built-in theme, this theme
is less cluttered as it does not have minor grid lines on both axis. Unlike
theme_seagull(), however, this theme retains major vertical grid lines.
The Chosen Theme
As stated above, theme_seagull() has the least clutter between the two
potential themes.However, theme_wombat() is likely to assist readers
to perform comparison more accurately than theme_seagull()
while still maintaining a high level of data-ink ratio. Using the scatter plot
example in figure \@ref(fig:theme-comparison), the major vertical grid lines
from theme_wombat() aids readers' ability to estimate the value of data points along the
x axis. This shows that theme_wombat() is a good compromise of minimizing
clutters as well as allowing readers to extract information from the graph as
much as possible.
Choosing Colour Palletes
In addition to themes, colour also plays an important role in data visualizations. Proper use of colour in a visualization enhances the quality of information being presented [@zeileis2009escaping]. One of the most important use of colour in data visualizations is to differentiate one element from one another [@stone2006choosing]. However, a poor choice of colour palette can impede viewers' ability to effectively extract information from the visualization.
With a simple visual feature, some objects can be easily recognised without requiring much attention. The target object will "pop-out" in field of vision regardless of how many distractors surrounding it [@treisman1985preattentive]. The use of colour in data visualization can facilitate this "pop-out" as it makes some elements to be easier to be found.
set.seed(20211028) x <- rnorm(20) y <- rnorm(20) color <- c("red","red","blue","red","blue","red","red","red","red","red","red", "blue","red","red","red","red","red","red","red","red") pop_out <- data.frame(x,y,color) p1 <- pop_out %>% ggplot(aes(x = x, y = y, color = color)) + geom_point() + theme_wombat() + theme(legend.position = "none", axis.title.x = element_blank(), axis.title.y = element_blank(), axis.text = element_blank(), axis.ticks = element_blank()) + ggtitle("Color") p2 <- pop_out %>% ggplot(aes(x = x, y = y, shape = color)) + geom_point() + theme_wombat() + theme(legend.position = "none", axis.title.x = element_blank(), axis.title.y = element_blank(), axis.text = element_blank(), axis.ticks = element_blank()) + ggtitle("Shape") p1 + p2
Figure \@ref(fig:pop-out) is an example, which is adapted from Healy's [-@healy2018data] book about data visualization, that demonstrates how easy it is for colour to be recognised in a visualization. The figure contains three "odd" observations that are different than the rest. To represent this, the first panel uses colour to differentiate the different kinds of observations. The three observations are coloured red, and the rest are blue. On the other hand, the second panel uses shape to represent this feature. Circle is used to represent the "odd" observations, and triangle is used for the rest. While readers have no problem in identifying the three observations in the second panel, it is much easier to find these observations on the first panel. This shows that finding the observations of interest in the first panel requires significantly lower level of attentions compared to the second panel.
In addition to assisting viewers in differentiating one attribute to another, colours can also be used to evoke viewers' emotions. Emotion is important in visualizations for story telling, communicative intent, and engagement [@bartram2017affective]. It was also found that emotions can significantly improve many cognitive processes such as attention, learning, decision making, and problem-solving [@harrison2012exploring].
This project will focus to create colour schemes that assist users to easily differentiate one data attribute from another. To achieve this, it is important to choose these schemes using a colour space that works well with human perception of colour. The next section will discuss the different options of colour space, and which one will be used for the project.
Colour Spaces
To effectively use colours in visualizations, it is imperative to understand how human colour perception works. It was theorized that human colour vision evolved in three distinct stages [@ihaka2003colour]:
- Perception of light and dark (monochrome)
- Perception of yellow/blue contrast
- Perception of green/red contrast
As a result of these three colour axes, colours are naturally described as locations in a three dimensional space. While these colour axis provide physiological description, human perception of colours do not corresponds to these axes. Human perception of colour corresponds to the polar coordinates of the colour coordinate (yellow/blue and green/red), and a dark/light axis [@zeileis2006choosing].
Unfortunately, not all software packages use colour space was developed based on human perception. For example, the most common colour map used by software packages, the Red-Green-Blue (RGB) [@stauffer2015somewhere], was primarily developed to correspond with colour generation on computer screens rather than to correspond with human perception of colour [@zeileis2009escaping]. Zeileis, Hornik, and Murrell further stated that it is practically impossible for a human to control the perceptual properties of a colour within this colour model as there is no single dimension that relates to hue or brightness of a colour
One popular alternative to RGB system is Hue-Saturation-Value (HSV). This model is a simple transformation of RGB system and aims to capture the dimensions of human visual perceptions [@smith1978color]. Unfortunately, the result of mapping HSV's three dimensions to three dimensions of human perceptions is very poor [@zeileis2006choosing]. Moreover, it is fairly difficult to find a set of colours that are in harmony in HSV colour space [@zeileis2006choosing]. Furthermore, HSV encourages the use of colours that are highly saturated. While these colours are good to attract viewers' attention, it is difficult to look at for a long time [@zeileis2009escaping].
HCL, which is short for Hue-Chroma-Luminance, was developed to overcome many problems encountered by other colour spaces. This colour space is developed based on the three colour attributes that corresponds to human perceptions: hue, chroma, and luminance [@sarifuddin2005new]. It was found that the three dimensions from this colour space matches human visual perception well [@zeileis2019colorspace]. In general, HCL colour space can be used to generate three types of colour scales: qualitative, sequential, and diverging.
Qualitative Palettes
This colour palette is designed for coding categorical information where the categories do not have particular ordering [@zeileis2019colorspace]. In other words, this colour palette will give the same perceptual weight to each category so that there will be no perception of one group to be larger or more important than any other one [@zeileis2006choosing]. This colour palette is commonly applied to bar plots or box plots.
Sequential Palettes
Sequential pallets are used for coding ordered/numerical information that span in a certain interval [@zeileis2006choosing]. The low values of the interval are considered to be uninteresting, whereas high values are considered to be interesting [@zeileis2019colorspace]. Low values are usually depicted by lighter colours, while darker colours are used for high values [@harrower2003colorbrewer]. This palette is usually used on heatmaps.
Diverging Palletes
Similar to sequential palette, diverging pallets are designed for coding ordered/ numerical information that ranges in an interval. The difference between diverging and sequential palletes is that diverging pallets include a neutral value [@zeileis2006choosing]. This palette should be used when an important data class or break points need to be highlighted [@harrower2003colorbrewer]. Harrower and Brewer further stated that this critical data class or break points is highlighted by a change in in hue and luminance, and should represent a critical value of the data (e.g. mean, median, or zero). For instance, the average GDP per-capita in a choropleth map can be emphasized (i.e. set as neutral value) so that the values above and below the average can be shown in different hues.
Colour palette of choice
In addition to plot themes, this project also contributes two colour schemes for
each of the three colour palettes type mentioned in the previous section.
The pals packages used to build the colour palettes
and assess them. The chosen colour palettes need to be assessed to ensure
that they achieve their intended purpose. For example, the qualitative colour palette
needs to avoid implicit ordering as this palette is intended for representing categorical
variables. In addition, we need to ensure that people with colour vision
impairments are able to distinguish the different colours in the chosen palettes.
The pal.safe() function from pals package allows us to see how colour palettes
are seen by deutan, protan, and tritan colour-blind people.
The methodology of choosing colour palettes that are not only accessible but also enhance the quality of visualizations will be discussed in this section. Firstly, the report will explore common strategies used in choosing qualitative, sequential, and diverging palettes. Then, examples of palettes made from these strategies will be assessed. Lastly, the report will show how colour schemes for qualitative, sequential and diverging palettes are chosen for the project.
quokka_palette <- c("#8C01E6", "#93B1EB", "#AB313D", "#D6CA24") quoll_palette <- c('#8D5698', '#EFD36F', '#247A3E', '#17B0A8') seal_palette <- colorspace::sequential_hcl(7, h = c(0, -140), c = c(80,40), l = c(40, 75), power = c(1,1)) snake_palette <- colorspace::sequential_hcl(6, h = c(250, 90), c = c(40, 55), l = c(33,98), power = c(0.5,1), rev = TRUE) dingo_palette <- colorspace::diverge_hcl(7, h = c(260, 0), c = 80, l = c(30, 90), power = 1.5) dolphin_palette <- colorspace::diverging_hcl(n = 7, h = c(265, 45), c = 80, l = c(70, 95), power = c(0.5, 2), rev = TRUE) # colorblind filter pal.safe(quokka_palette, main = "Quokka Palette") pal.safe(quoll_palette, main = "Quoll Palette") pal.safe(seal_palette, main = "Seal Palette") pal.safe(snake_palette, main = "Snake Palette") pal.safe(dingo_palette, main = "Dingo Palette") pal.safe(dolphin_palette, main = "Dolphin Palette") pal.safe(qualitative_hcl(4, palette = "harmonic"), main = "Harmonic Palette")
Chosen Qualitative Palettes
One common strategy to choose this type of palette is by keeping chroma and luminance relatively constant, and hue varied; making all colours in the palette to balance towards the same grey [@zeileis2009escaping]. While this strategy is relatively simple, keeping chroma and luminance fixed may not be the best approach to choose colour palette that are accessible.
Panel (g) in figure \@ref(fig:specplot) shows an example of a qualitative colour palette from
colorspace package called harmonic. As shown in the figure, chroma
and luminance remain constant, while the hue varies across the four colours.
It is fair to say that this colour set works relatively well to show categorical
information as it does not give implied ordering across different categories.
This method, however, does not have the same result for people with visual
impairment. It can be seen in figure \@ref(fig:color-blind) that people
with colour-blindness will have difficulty in distinguishing some colours from
harmonic palette. For instance, there are three colours that look similar as seen
by people with tritan colour-blindness. Moreover, this colour palette results in
implied ordering of different data group for people with visual impairment.
specplot(quokka_palette, main = "Quokka Palette") specplot(quoll_palette, main = "Quoll Palette") specplot(seal_palette, main = "Seal Palette") specplot(snake_palette, main = "Snake Palette") specplot(dingo_palette, main = "Dingo Palette") specplot(dolphin_palette, main = "Dolphin Palette") specplot(qualitative_hcl(4, palette = "harmonic"), main = "Harmonic Palette")
The two qualitative colour palettes developed for the project are called Quokka and Quoll palettes. Panel (a) and (b) in figure \@ref(fig:specplot) shows the HCL colour spectrum plot for these palettes. In contrast with the Harmonic Palette, Quokka and Quoll palettes have differing hues, chroma and luminance across all colours. As a result, each colour in these palettes can easily be distinguished by people with deutan, protan, and tritan colour-blindness (see figure \@ref(fig:color-blind)). Furthermore, natural ordering implied by Quokka and Quoll is less prominent than Harmonic palette.
Chosen Sequential Palettes
The strategy to choose a sequential palette is different to that of qualitative
palette. One easy way to create a sequential palette is by choosing two values
for hue (h1 and h2), chroma (c1 and c2), luminance (l1 and l2), and power
transformations (p1 and p2) for chroma and luminance [@zeileis2006choosing].
These values of hues, chroma, luminance, and power transformation are used
as inputs to sequential_hcl function to produce a set of sequential palette.
The power transformation allows the chroma and luminance to increase from low
values to high values in a nonlinear fashion.
The two sequential palettes created for this project are Seal and Snake palettes. As shown in panel (c) and (d) in figure \@ref(fig:specplot), these palettes use multi-hue schemes as opposed to single-hue schemes. Multi-hue palettes were chosen for the project as it provides better contrast between different values compared to a single-hue sequential palette [@harrower2003colorbrewer]. Moreover, multi-hue sequential palettes avoid the use of extreme colours, as well as allowing viewers to better discriminate the middle colours [@zeileis2019colorspace]. In addition, figure \@ref(fig:color-blind) shows that people with deutan, protan, and tritan colour blindness are able to distinguish the colours in the palette. This means that the colour palettes will allow readers with colour vision impairment to see implied order and how the data ranges from low to high values.
Chosen Diverging Palettes
The approach in choosing diverging palettes is similar to choosing sequential
palettes. Diverging palettes can be built by choosing two values for
hues (h1 and h2), chroma (c1 and c2), luminance (l1 and l2), and power
transformation (p1 and p2) for chroma and luminance [@zeileis2006choosing].
These values will be taken as inputs to function diverge_hcl() to create a diverging
palette. It will automatically choose the neutral colour for the palette.
Three conditions need to be met by a diverging palette to avoid colour confusions. Colour confusions are caused by naming confusions, colour vision impairment, and simultaneous contrast confusion [@brewer1996guidelines]. Naming confusion happens when people use two basic colour names for the same colour (e.g. a colour was called red when presented once, then pink at another time). On the other hand, simultaneous contrast refers to a phenomena where colours are affected by their opposite (e.g. a colour will appear lighter when surrounded by a dark colour, and a colour will appear warmer when surrounded by a cool colour) [@brewer1992review]. This can be prevented by avoiding selecting pairs of colours that are opposite in the colour wheel for diverging palettes [@brewer1996guidelines].
Dingo and Dolphin palettes are diverging palettes developed for the project. These palettes are shown in panel (e) and (f) in figure \@ref(fig:specplot). Red/blue pairs and orange/blue hue pairs are chosen for Dingo and Dolphin palettes respectively as these pairs create no naming confusions and simultaneous contrast confusions [@brewer1996guidelines]. Moreover, these palettes create no colour confusion due to colour vision impairment as colours in Dingo and Dolphin palettes can be easily distinguished by people with colour-blindness (shown in figure \@ref(fig:color-blind)). Another reason for choosing these pairs is because red/blue pair is often used in the topic of the edited book, MRP, to show electoral result (e.g. Democratic and Republican party). Therefore, it is important to have a red/blue scale and a non-red/blue scale for functionality in the book.
R Package
One of the objectives of this project is to create a solution for authors and
editors of the research book to create visualizations that are readable,
cohesive, and accessible. The solution needs to be easy to use so that it
does not disrupt authors' workflow. To achieve this, R code necessary
to create the chosen ggplot2 [@ggplot2] theme and colour schemes will be bundled in an R
package called ebviz. R packages are extensions to R that can comprise of
codes, data, documentation, and tests [@wickham2015r].
One of the reasons for creating an R package for this project is because of the
ease of use. Codes, documentation, and tests in an R package are built in a
standardized format so that any R user can install it [@wickham2015r]. In
addition, it is fairly simple to create an R package. Packages such as roxygen2
[@roxygen2], devtools [@devtools], and testthat [@testthat] provide great
assistance in documenting and testing the package. Another benefit of using an
R package is that any updates to the theme and colour schemes are not time
consuming for the users. Any updates to themes and colour schemes simply need to
be applied to the code in the R package. There is no need to update the code
for each plots in the book; it just needs to be recompiled. This shows that
using an R package will allow us to achieve one of the aims of the project:
to provide a solution that causes minimal disruption to authors' work.
For users to be able to easily access an R package, it needs to be stored in a
repository. For this project, the ebviz package will be stored in GitHub.
GitHub is a popular choice for a repository as it has the integration with git,
a version control software. Version control records changes applied to the
package to a file. This allows user to recall specific versions of the package
at later time [@loeliger2012version]. An alternate option to GitHub is
Comprehensive R Archive Network (CRAN). CRAN, however, is not suitable for this
package as ebviz has a fairly small scope and is intended for a small group
of people (i.e. authors of the MRP research book). CRAN is usually a repository
used for more complex R packages intended for large group of people.
In the next section, the report will discuss how the ebviz package was
created. It will discuss each component of the package, and what tools
were used in developing the package.
ebviz Components
One of the most important components of a package is the R code. One firm
rule for an R package is that its R functions need to be stored in R scripts
[@wickham2015r]. Following this rule, the ebviz has its R code in R
scripts, which are stored in R directory. The chosen colour palettes can be
accessed from the scale_colour_ebviz and scale_fill_ebviz functions which
are stored in colour.R. These functions take the name of the palette as an
input (e.g. quokka). The ggplot theme chosen for the project can be used by
function theme_wombat() which is stored in an R script called theme.R.
In addition to R codes, documentation is also a crucial aspect of a well-made
package [@wickham2015r]. In the absence of documentation, users will not have a guide
that assists them to use the package properly. Proper documentation also assists
the authors of the package for future use. This assists authors to remember the
purpose of the functions. One type of documentation is an object documentation;
this can be accessed by the ? operator or help(). Information such as function
description, usage, argument(s), and examples can be found in the documentation.
The roxygen2 [@roxygen2], devtools [@devtools], and usethis [@usethis]
packages are used in creating object documentation.
Another crucial aspect to package development is testing. While testing add
extra work to package development, it ensures that the package works
as intended. It also ensures that the package works for edge cases - the less
common uses that a user might use. In addition, testing ensures error and warning
messages to appear when required (and not appear when they are not supposed to).
In other words, testing provide protection and quality assurance for users.
We use the packages testthat [@testthat] and usethis [@usethis] to write
the tests for the ebviz package. One example of a test that was used for
this package is one that ensures that users use valid input when using
scale_colour_ebviz and scale_fill_ebviz.
Conclusions and Limitations
One of the challenges in producing a research book is creating a book that has a consistent look or design. Typically, there is a group of authors that work on every chapter of an edited book. These authors will have their own unique style in producing their visualizations for the book. While this is not a bad thing, visualizations style needs to be standardized to keep the look of the book consistent. Not only do these visualizations need to have a consistent design, but they also need to be readable and accessible.
The aim of the project is to provide a solution that can ensure that visualizations in the research book meet the criteria mentioned in the previous paragraph in an efficient manner. To achieve this objective, a ggplot theme was developed for the project. The chosen theme was developed based on Tufte's principles of graphical excellence. In addition, several colour scales for various data types were also created. The colour palettes were developed using a colour space that was based on human perception. Furthermore, to ensure that color is accessible, we visualized how these colours as they would be seen by people with colour visual impairment. These theme and colour scales were developed to ensure that the viewers to extract information as much as possible from the visualizations. The theme and colour schemes are bundled in an R package for efficiency.
One limitation in the package is related to colour palettes, specifically the chosen qualitative palettes. As mentioned in the previous section, this palette is specifically used to show un-ordered categorical information. It is therefore imperative to avoid implying that one group is larger or more important than other groups. However, removing implied ordering while ensuring accessibility proved to be difficult. Figure \@ref(fig:color-blind) shows that while the Quokka and Quoll palettes are better than Harmonic palette in avoiding implied ordering for people with colour visual impairment, there are still some colours that look paired. In addition, Quokka and Quoll only have four colours. It is difficult to produce a qualitative palette that have more than four colours while still ensuring that people with colour impairment are able to distinguish the colours from one another.
One area of future work for this project is to work on writing additional testings. For instance, a test can be developed to ensure that a warning message will appear when the theme function is applied to a non-ggplot object. Another test that can be applied to the package is to ensure an error to pop-up when applying a diverging palette to a categorical variable. Another approach that can be done in the future is by having the users (i.e. the authors of the MRP books) to test the functionality of the package for a period of time. This allows the users to provide feedback, and this can be used to improve the package. These testings are beneficial to protect the quality of the package.
In summary, this project contributes an R package containing a ggplot theme and six colour palettes developed to assist MRP book authors in creating visualizations that are readable, cohesive, and accessible.
\newpage
Reference
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.