library(tidyverse)
us = read_csv("https://github.com/nytimes/covid-19-data/raw/master/rolling-averages/us-states.csv")
pop = read_csv("us_census_population_totals_est_2020-2021.csv")
us = us %>%
left_join(pop %>% rename(population = `2021`, state = State) %>% select(state, population)) %>%
mutate(cases_per_100k = signif(cases/population*1e5, digits = 3)) %>%
mutate(deaths_per_100k = signif(deaths/population*1e5, digits = 3)) %>%
mutate(day = factor(weekdays(date), levels = c("Sunday", "Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday"))) %>%
mutate(month = factor(months(date), levels = c("January", "February", "March", "April", "May", "June", "July", "August", "September", "October", "November", "December"))) %>%
filter(!is.na(population), cases >= 0, deaths >= 0)Finessing plots in ggplot2
Outline for today
- Reminders
- Principles of displaying data
- Modifying plot elements
- Themes
Principles of displaying data
While there is a lot of art in designing figures to display data, there is also some science. Researchers interested in designing effective figures have found some helpful rules of thumb that take advantage of simple intuitions as well as empirical results from psychology and neuroscience.
More data. Less ink.
In his book, “The Visual Display of Quantitative Information”1, Edward Tufte states that
Data-ink is the non-erasable core of the graphic, the non-redundant ink arranged in response to variation in the numbers represented
and emphasizes that the “redundant data-ink” should be minimized. In other words, use as few visual elements as necessary to display your data. For example, a bar chart simply shows the relative magnitude of different factors, and thus needs only bars lined up next to one another for visual comparison. Yet, many bar charts come with “chart junk”, which are elements that are unnecessary for displaying the data. The example below shows how removing “chart junk” can make a bar chart much simpler, easier to read, and even more attractive (at least in terms of elegance and simplicity).
This rule can and should be applied to tables as well. The example below shows how removing the chart junk from the table can make it visually much simpler and easier to read without losing the ability to easily distinguish rows or compare across columns. The example also displays some useful rules for tables, such as removing unnecessary horizontal lines, aligning text and numbers correctly, and using row spacing to help distinguish rows.
Visual properties of graphical elements
Nobel prize-winning work in neuroscience by David Hubel and Torsten Wiesel (among others) showed that the visual cortex is designed to recognize certain basic visual features, such as orientation and contrast that distinguishes edges. These basic features are then assembled into more complex visual objects in other brain regions. Knowing that some visual features may be more “basic” than others with respect to how they are processed by the brain means that you can leverage those features to make graphics easier to read.
An example of how some visual elements are more basic than others comes from the work by William Cleveland and Robert McGill on the speed and accuracy that people have in distinguishing specific graphical elements 2. The table below shows these elements and their rank from most to least accurately distinguishable.
| Rank | Graphical element |
|---|---|
| 1 | Positions on a common scale |
| 2 | Positions on the same but nonaligned scales |
| 3 | Lengths |
| 4 | Angles, slopes |
| 5 | Area |
| 6 | Volume, color saturation |
| 7 | Color hue |
The figure below gives you a sense of what each of these elements are.
As an example, a pie chart, which uses angles to indicate the relative size of a category, can be harder to read than a bar chart, which uses positions on a common scale. In other words, never use a pie chart!
Gestalt principles
Gestalt (“shape” in German) principles come from German psychologists in the early 20th century who tried to come up with the rules for perception. These rules are built on common sense intuitions and can be useful in composing figures, particularly with respect to grouping related parts of a figure. The general rule is that objects that look alike, are close to one another, connected by lines or enclosed together belong together somehow.
- Similarity. Objects with similar color, shape, or orientation are grouped together.
- Proximity. Objects close to each other are grouped together.
- Connection. Objects linked to each other are grouped together.
- Enclosure. Objects enclosed together are grouped together.
(Bad) Examples
In “The Visual Display of Quantitative Information”, Tufte said of this graphic
This may well be the worst graphic ever to find its way into print”
Here is another example to get one’s blood boiling from (ironically) a 2012 report on “World Happiness”.
Modifying plot elements
With all the basic tools of ggplot2, you can already implement many of the visual design principles described above. The remaining changes you might need to make include altering the labels, annotations, coordinate system or scaling, color scaling, or plot size.
Labels
You have already seen how to add simple labels to simple plots, but now you will add labels to ggplot2 plots. Adding labels in ggplot2 is accomplished with the labs() function. For example, if you load the COVID-19 data,
you can plot and then add a title easily
cases_plot_ky_ny_fl = us %>%
filter(state %in% c("Kentucky", "New York", "Florida")) %>%
ggplot(aes(x=date, y=cases_per_100k, color=state))
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases")`geom_smooth()` using method = 'loess' and formula 'y ~ x'
You can also add a subtitle, which is additional detail below the title, and a caption, which should describe the data in the plot.
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases",
subtitle = "Data aggregated by the New York Times",
caption = "Though Kentucky was often keeping lower cases numbers than Florida,
the combination of the Delta variant and a lack mask requirements may have led it
achieve similar case rates as Florida during late summer 2021.")Labels can also be added to the axes and the legend.
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases",
x = "Date", y = "Cases per 100k persons", color = "State")`geom_smooth()` using method = 'loess' and formula 'y ~ x'
Mathematical symbols can be added by using expression() instead of the quotation characters ““. The quote function also works in simple cases. Check ?plotmath for options. For example,
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = expression("This is an integral" ~ integral(f(x)*dx, a, b) ~ "that doesn't mean anything"),
x = expression(x[y]^z),
y = expression(frac(y,x) == frac(alpha, beta)))`geom_smooth()` using method = 'loess' and formula 'y ~ x'
Note above that we glue together math expressions and normal text by putting the normal text in a string and gluing it to the expression with a ~.
Annotations
Adding annotations to plots can be very important and people often do this in programs such as Adobe Illustrator. However, taking the plot to another program makes generating the figure much more complicated and breaks the “reproducible science” method using RMarkdown where any change in the data should easily be converted into updated figures and documents.
One way to add text to a plot is with geom_text, which is like geom_point, but has a label option. For example, you can label the days where each state reached it largest number of cases per 100k. The code below first groups the data by state, since you want to use a label for each state. Then, it filters the rows to include only the ones that rank first when sorted into descending order based on casses_per_100k. Finally, it uses this data table for the geom_text.
maxcases_ky_ny_fl = us %>%
filter(state %in% c("Kentucky", "New York", "Florida")) %>%
group_by(state) %>% filter(row_number(desc(cases_per_100k)) == 1)
# note: row_number here is what is doing the ordering and `desc` tells it to order greatest to least
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases", x = "Date", y = "Cases per 100k persons", color = "State") +
geom_text(aes(label = c("Florida max", "New York max", "Kentucky max")), data = maxcases_ky_ny_fl, show.legend = FALSE)`geom_smooth()` using method = 'loess' and formula 'y ~ x'
You can see in the above that you actually need to create a new data table for the text annotations that has the right names for the x and y variables. This is because ggplot2 only understands how to plot dataframes, not other things. This is both the source of its power and limitations. Thus, to put labels on plots with dates on the x-axis, we have to create a data frame with dates as the location we want the text. Likewise, we have to give the y-value as cases per 100k.
library(lubridate) # for the `ymd` function
label = tibble(date = ymd("2021-07-01"), cases_per_100k = 250, label = "This is a label in the middle of the plot")
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases", x = "Date", y = "Cases per 100k persons", color = "State") +
geom_text(aes(label = label), data = label, vjust = "bottom", hjust = "center", color = "black", show.legend = FALSE)
The text has a “justification” in reference to the (x,y) location of the point you specify. You can set the vertical (vjust) and horizontal (hjust) justication above using the options below. By setting “bottom” and “center” as above, the coordinate is at the bottom and in the center of the text, which means the text is centered above the point.
Coordinate systems
Coordinate systems in ggplot2 can be complex, but usually you will only want to flip the x and y axes with coord_flip() as in previous examples.
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases", x = "Date", y = "Cases per 100k persons", color = "State") +
coord_flip()`geom_smooth()` using method = 'loess' and formula 'y ~ x'
There is a coordinate system for “polar” coordinates that effectively produces a pie chart. Since pie charts are bad (see above), avoid this unless your data really are in polar coordinates.
You can use a coordinate transform to put the x, y, or both axes on a log scale. The function to accomplish this is coord_trans() where function names are given for the x and y arguments (e.g, log10).
library(gapminder)
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) +
geom_smooth(method = "lm") +
coord_trans(x = "log10", y = "log10")`geom_smooth()` using formula 'y ~ x'
Above, you can notice that the straight line (since “lm” was used to plot the line) is curved, which indicates that the line was fit on the untransformed data (i.e., a straight line plotted on a log-log plot is curved). Below, you will see how to change the scales to a log scale so that the line is fit on the transformed data.
Scales
Scales control how the data maps to aesthetics, which includes whether the data is on an arithmetic or log scale, how data maps to colors, and how the scale values themselves are displayed (i.e., the tick marks). By default, ggplot2 takes the scatter plot below
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) 
and adds
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) +
scale_x_continuous() +
scale_y_continuous() +
scale_colour_discrete()
You can alter properties of these scales including where tick marks are, the labels of those marks, etc. Modifying the x-tick spacing and getting ride of the y-labels looks like this:
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) +
scale_x_continuous(breaks = seq(10000, 100000, by = 10000)) +
scale_y_continuous(labels = NULL)
Changing the scales to log values can be done with the scale_x_log10 and scale_y_log10 functions.
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) +
scale_x_log10() +
scale_y_log10() +
geom_smooth(method = "lm")`geom_smooth()` using formula 'y ~ x'
Above, you can see that the fit “lm” line is straight, which means it was applied to the transformed data. Looking into the coord_trans docs, we find an explanation for this:
The difference between transforming the scales and transforming the coordinate system is that scale transformation occurs BEFORE statistics, and coordinate transformation afterwards.
Finally, you can change the color scale for the discrete variables plotted. One common alternative set of color scales are the “ColorBrewer” (http://colorbrewer2.org/) scales that are designed to work well with color blind folks and can be loaded with library(RColorBrewer).
library(RColorBrewer)
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) +
scale_x_log10() + scale_y_log10() +
scale_colour_brewer(palette = "Dark2")
You can also set the color scale manually, which is nice for making Kentucky blue and Florida orange.
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases", x = "Date", y = "Cases per 100k persons", color = "State") +
scale_colour_manual(values = c(Florida = "orange", Kentucky = "blue", `New York` = "red"))`geom_smooth()` using method = 'loess' and formula 'y ~ x'
Zooming
You can “zoom” by either taking a subset of the data and plotting that or by changing the x and y limits in the coordinate system. The latter option is better for really “zooming” into a region whereas the former is better when you care only about that subset. To do the latter,
ggplot(gapminder, aes(x = gdpPercap, y = lifeExp)) +
geom_point(aes(color = continent)) +
coord_cartesian(xlim = c(1000, 2000), ylim = c(50, 70))
Themes
More generally, you can modify non-data elements of the plot with a theme. There are eight themes included with ggplot2:
ggplot2 themesApplying them just requires adding the specific function:
cases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases", x = "Date", y = "Cases per 100k persons", color = "State") +
scale_colour_manual(values = c(Florida = "orange", Kentucky = "blue", `New York` = "red")) +
theme_bw()`geom_smooth()` using method = 'loess' and formula 'y ~ x'
Hadley Wickham has some text defending the default theme with the gray background. I won’t detail his reasons since I think that theme is frankly ugly and the gray background is an example of “chart junk” we just discussed.
Claus O. Wilke theme (cowplot)
Claus O. Wilke, an evolutionary biologist at UT Austin, has put together a theme that he describes as
a publication-ready theme for ggplot2, one that requires a minimum amount of fiddling with sizes of axis labels, plot backgrounds, etc.
Once you load the package, you can use the theme.
library(cowplot)
Attaching package: 'cowplot'The following object is masked from 'package:lubridate':
stampcases_plot_ky_ny_fl + geom_point(alpha = 0.3) +
geom_smooth(lwd = 0.5, span = 0.1) +
labs(title = "COVID-19 Cases", x = "Date", y = "Cases per 100k persons", color = "State") +
scale_colour_manual(values = c(Florida = "orange", Kentucky = "blue", `New York` = "red")) +
theme_cowplot(font_size = 12)`geom_smooth()` using method = 'loess' and formula 'y ~ x'
The theme is meant to work well with saving figures (coming in another class session), adding annotations (cowplot does not require creating a data table), and placing subplots in arbitrary arrangements in the plot. For more information, check out https://cran.r-project.org/web/packages/cowplot/vignettes/introduction.html. Wilke also has a book on data visualization https://clauswilke.com/dataviz/ that might be of interest.
Lab 
Problems
- Create a plot of per capita deaths using the
usCOVID-19 dataset (use code block from above) using three states- Use color to denote each state in the plot
- Use the
labsfunction to add x and y labels and a title. - Add
geom_textannotation for the days with the highest death counts for each state - Bonus 2 points: add the day of the week in the
geom_textannotation (e.g., “New York - Sunday”) without typing it in manually
- Create a plot from any of the datasets we have used previously that includes the following
- Use color to represent the value of some variable in the data
- Descriptive labels for the axes and title
- Appropriate tick mark breaks and labels (only if defaults are bad)
- Non-ggplot2 default theme (pick your favorite)
- Bonus 1 points: change color of the x and y tick labels to blue
- Find a bad figure in a scientific paper in your field
- Save the figure and include it as a .jpg or .png with your .Rmd and load the figure into your .Rmd file as an image.
- Describe what is wrong with the figure using graphics principles discussed today.
- Describe how you would fix the figure.
- Bonus 5 points: load in the data and actually fixing the figure!