Part 1 is here.

Steps to a ggplot2 plot

4. Geoms

Now we can plot iris!

Compare the ggplot2 code to the base R code we saw earlier:

ggplot2

ggplot(iris) + 
  # This aes() call sets up the PLOT LEVEL aesthetics
  aes(x = Sepal.Length, y = Sepal.Width, color = Species) + 
  geom_point()

base R

plot(iris$Sepal.Length, iris$Sepal.Width, col = iris$Species)
legend(7, 4.3, 
       unique(iris$Species),
       col=1:length(iris$Species), 
       pch = 1)

Now the advantages of ggplot2 and the grammar-of-graphics approach start to get clear.

Same data, different geoms

Geoms have different aesthetic requirements, and not every geom works with every dataset (visually or computationally):

# Note now the aes() call is INSIDE the ggplot() call--makes it obvious that these
# are the initial, PLOT LEVEL aesthetics
iris_plot <- ggplot(iris_long, aes(x = Sepal, 
                                   y = Petal, 
                                   color = Species))

geom_point

iris_plot + geom_point()

geom_line

iris_plot + geom_line()

geom_bin2d

iris_plot + geom_bin2d()

geom_violin

iris_plot + geom_violin()
## Warning: position_dodge requires non-overlapping x intervals

geom_dotplot

iris_plot + geom_dotplot()
## `stat_bindot()` using `bins = 30`. Pick better value with `binwidth`.
## Error: geom_dotplot requires the following missing aesthetics: y

Using multiple geoms

You can have as many geoms in a plot as you want. Later geoms are drawn in front of earlier ones:

iris_plot + geom_density_2d() + geom_point()

Geoms and aesthetics

Because of the layering principle we discussed above, later steps and in particular geoms override earlier ones as the plot is built up.

In particular, if you define aesthetics in geoms they override any earlier definitions but only within that geom:

# Note that, as before, aesthetics are mapped to variables inside an aes() call
# If we want to map an aesthetic to a constant value we do so OUTSIDE aes()
# E.g., + geom_point(color = "red")
iris_plot + geom_density_2d() + geom_point(aes(color = dimension))

This doesn’t make much sense, but this might:

iris_plot + 
  geom_density_2d(aes(linetype = dimension)) + 
  geom_point(aes(shape = dimension))

What aesthetics are operative in this plot?

  • At the plot level, x = Sepal, y = Petal, and color = Species
  • The plot aesthetics hold for geom_point but it also uses shape = dimension
  • The plot aesthetics hold for geom_density2d but it also uses linetype = dimension

Geoms that compute things

Some geoms don’t take both x and y aesthetics; rather, they take just one and compute the other, or they transform one of the aesthetics by some computation. How this happens is beyond our scope here (come back next week!) but let’s look at a couple of examples.

geom_histogram

ggplot(iris_long) +
  aes(x = Sepal) +
  geom_histogram()
## `stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

geom_boxplot

ggplot(iris_long) +
  aes(x = Species, y = Sepal) +
  geom_boxplot()

geom_boxplot (2)

ggplot(iris_long) +
  aes(x = Species, y = Sepal) +
  geom_boxplot() +
  geom_jitter(alpha = 0.5, aes(color = dimension))

Notice how the final boxplot-and-point plot makes it clear that the boxplot-only plot is highly deceptive: there are two groups of data here (two dimensions that were measured), and you probably wouldn’t want to present them combined together.

Remember one of the 10 data visualization principles above: visualize your data, not just summaries!

Data-less geoms

Some geoms are handy for annotation or to help with interpretation:

iris_plot + 
  geom_density_2d(aes(linetype = dimension)) + 
  geom_point(aes(shape = dimension)) +
  # vertical line
  geom_vline(xintercept = 5, size = 3) +
  # horizontal line
  geom_hline(yintercept = 6, color = "purple", linetype = 2) +
  # a-b line
  geom_abline(size = 10, alpha = 0.25, slope = 0.3, intercept = 1)

There’s also a useful annotate() function; see its documentation.

What geom should I use?

The question here is really: what are you trying to communicate?

(See notes under Data visualization section above.)

If you want to… …try this geom
show the relationship between two variables geom_point geom_jitter
show values over time or a series geom_bar geom_line
show data distribution (1D) geom_histogram geom_density
show data distribution (against another variable) geom_boxplot geom_violin
show data distribution (2D) geom_hex geom_bin2d
analyze trends geom_smooth geom_line

Smoothers and models

We frequently would like to fit, or show, trend lines with point data. The easiest way to do this is with geom_smooth():

iris_plot + geom_point() + geom_smooth()
## `geom_smooth()` using method = 'loess' and formula 'y ~ x'

What’s happened here?

  • By default, geom_smooth uses a loess smoother, which is just a form of local regression
  • This is noted in the diagnostic message printed by geom_smooth
  • geom_smooth inherited the plot aesthetics (x, y, color), so it fit separately to each color (group of data)

More likely, we would like a simple linear regression—i.e., trend lines. For this we just need to override the default and tell is to use R’s lm function:

iris_plot + geom_point() + geom_smooth(method = "lm")

We might also want to fit an overall trend line, to the pooled data (i.e., without groups). You might be tempted in this case to override the color aesthetic to a constant value. This works, but it’s not ideal. Better:

iris_plot + geom_point() + 
  geom_smooth(method = "lm") +   # per-group trend line
  geom_smooth(method = "lm",  
              color = "black", 
              group = 1,         # pooled; intent is clear
              linetype = 2)

Behind the scenes:

  • Our color aesthetic controls the group aesthetic in this plot
  • group is ultimately what determines how the data are split up for computation and plotting
  • Setting group = 1 makes it crystal clear what we want: a single group for the second geom_smooth.

Finally, we can use any smoothing function we want, including custom ones.

ggplot(iris_long, aes(Sepal, Petal)) + 
  geom_point() +
  geom_smooth(method = "lm") +
  geom_smooth(method = "lm", formula = y ~ poly(x, 2), color = "red")

3. Labels

Labeling your plots well is important.

As we’ve seen, ggplot2 takes its default axis and aesthetic labels from the columns you specify. These are easy to change using the xlab and ylab functions; the ggtitle function is available as well:

iris_plot + geom_point() + 
  geom_smooth() +
  xlab("What is a sepal again?") +
  ylab("Petal (cm)") +
  ggtitle("This is starting to look like a real plot!")

2. Themes

ggplot2’s theme system is powerful and sometimes confusing. Like the rest of the ggplot system, it uses the idea of inheritance: you can apply themes, or aspects of them, to entire plots, sub-elements, or small details, and changes cascade down.

The simplest step is to apply a theme to an entire plot:

Theme examples

theme_gray

iris_plot + geom_point() + theme_gray()

theme_dark

iris_plot + geom_point() + theme_dark()

theme_minimal

iris_plot + geom_point() + theme_minimal()

theme_cowplot

Many more themes are available in other R packages and online repositories:

library(cowplot)
iris_plot + geom_point() + theme_cowplot()

theme_economist

Many more themes are available in other R packages and online repositories:

library(ggthemes)
iris_plot + geom_point() + theme_economist()

The theming system is also how we change specific aspects of plots.

Theme inheritance

Let’s say we want to apply some formatting to text in our plot. But, which text?

All text

iris_plot + geom_point() + ggtitle("Iris") +
  theme(text = element_text(size = 20, color = "red"))

Axis text

iris_plot + geom_point() + ggtitle("Iris") +
  theme(axis.text = element_text(size = 20, color = "red"))

x axis text

iris_plot + geom_point() + ggtitle("Iris") +
  theme(axis.text.x = element_text(size = 20, color = "red"))

Bottom x axis text

iris_plot + geom_point() + ggtitle("Iris") +
  theme(axis.text.x.bottom = element_text(size = 20, color = "red"))

All titles

iris_plot + geom_point() + ggtitle("Iris") +
  theme(title = element_text(size = 20, color = "red"))

Axis titles

iris_plot + geom_point() + ggtitle("Iris") +
  theme(axis.title = element_text(size = 20, color = "red"))

We can do similar things with any other aspect of the plot: grid lines, legends, backgrounds, etc. From the help page:

Theme elements inherit properties from other theme elements hierarchically. For example, axis.title.x.bottom inherits from axis.title.x which inherits from axis.title, which in turn inherits from text. All text elements inherit directly or indirectly from text; all lines inherit from line, and all rectangular objects inherit from rect. This means that you can modify the appearance of multiple elements by setting a single high-level component.

Interestingly (I just realized this in preparing these slides) this seems to be wrong or at least incomplete? I don’t understand why axis.text isn’t inheriting from text (which is what the documentation says). 🤷

1. Facets

Facets are multi-panel plots that show different subsets of a data frame. They work best on discrete variables and can help clear up a busy plot.

  • facet_grid is a rigid m x n matrix and best used with multiple variables
  • facet_wrap is a long ribbon of panels that can be wrapped into any number of columns using ncol and best used with a single variable

Faceting examples

Which of these is the best visualization? Why?

None

iris_plot + geom_point()

by Species

iris_plot + geom_point() + facet_wrap(~Species)

by dimension

iris_plot + geom_point() + facet_wrap(~dimension)

by both

iris_plot + geom_point() + facet_grid(Species ~ dimension)

At this point, we’ve pretty much arrived at our final plot!

Accessibility

At a minimum, be aware that some of the default colors in ggplot2 have equal luminance and so can be difficult to distinguish, particularly for colorblind viewers.

We can change colors and palettes, however; for example, the Viridis color scales included with ggplot2 are designed to be perceptually uniform in both color and black-and-white. They are also designed to be perceived by viewers with common forms of color blindness.

Use viridsd() with discrete data, and viridisc() with continuous data.

The RColorBrewer package is another good option worth looking into.

Default

tx_housing  # down-sampled from the "txhousing" dataset

scale_colour_viridis_d

tx_housing + scale_color_viridis_d()

Using and saving plots

Note that ggplot() objects are just like any other R object, and can be printed (displaying the plot), passed to functions, saved to disk, etc.

The primary way to save a plot image is the ggsave() function:

p <- ggplot(cars, aes(speed, dist)) + geom_point()
print(p)
ggsave("cars_plot.pdf")

# You can specify file types and dimensions:
q <- ggplot(iris, aes(Sepal.Length)) + geom_histogram()
print(q)
ggsave("iris_plot.png", width = 8, height = 5)

# By default ggsave assumes you mean "save the last plot generated"
# but you can save arbitrary objects:
ggsave("cars_plot.jpg", plot = p + theme_bw())

A complicated example

This figure is from a recent paper we published with many co-authors in GCB:

This seems—and is—complex, but you now have the tools to see what’s happening in the code that generates the figure.

# Whittaker biome plot
library(plotbiomes)
p_inset <- whittaker_base_plot() +
  geom_point(data = cosore_points, 
             aes(x = mat_cosore, y = map_cosore / 10),
             color = "black", shape = 4) +
  coord_cartesian(ylim = c(0, 500)) +
  theme(axis.title = element_blank(),
        axis.text = element_text(size = 8),
        legend.text = element_text(size = 7),
        legend.key.size = unit(0.4, "lines"),
        legend.position = c(0.35, 0.75),
        legend.title = element_blank(),
        panel.grid.major = element_blank(),
        panel.grid.minor = element_blank(),
        panel.background = element_rect(fill = "white"),
        panel.border = element_rect(colour = "black", 
                                    fill = NA, size = 0.5))

# SP's main climate space plot
p <- ggplot() +
  geom_hex(data = map_mat_global,
           aes(x = mat, y = map / 10), 
           bins = 100, na.rm = TRUE) +
  scale_fill_viridis_c(name = "Grid cells", begin = 0.85, end = 0) +
  geom_point(data = cosore_points, 
             aes(x = mat_cosore, y = map_cosore / 10),
             color = "black", shape = 4, size = 1.5, na.rm = TRUE) +
  theme_minimal() +
  coord_cartesian(ylim = c(0, 500)) +
  labs(x = "MAT (°C)", y = "MAP (cm)")

# Inset the first inside the second
library(cowplot, quietly = TRUE)
ggdraw() +
  draw_plot(p) +
  draw_plot(p_inset, x = 0.1, y = 0.52, width = 0.4, height = 0.45)

Fancier things

This workshop has covered a lot of ground, and necessarily skipped over and/or fudged some concepts for clarity and time.

Important things we haven’t talked about include:

  • Statistical transformations
  • Scales and coordinate systems
  • Plotting dates, maps, animations
  • Swapping in new data to an existing plot
  • Combining different plots in a larger figure

Some of these are built into ggplot2, but there’s also a whole ecosystem of extension packages that people have written.

Resources

This is a great walkthrough of the evolution of a complex and beautiful visualization:

This presentation was written in R Markdown.

  • The presentation in html format is here
  • The repository is here
  • The Intermediate ggplot workshop slides are here