3.1 ggplot2: The Grammar of Graphics

1. Why we plot

Numbers do not tell the whole story. Anscombe’s quartet is four data sets with the same means, variances, and correlation. The scatterplots are not the same: one is a line, one is a curve, one is a line with an outlier, one is a vertical stack with an outlier. The Datasaurus dozen is the same idea with a T-Rex in it.

Plot when you explore and when you explain. Do not add decoration that does not carry information.

R has three plotting systems:

• Base R: start a blank plot, then add pieces (plot(), lines(), rug())
• lattice: one large function call
• ggplot2: layers that follow a grammar of graphics

This course uses ggplot2. It comes with library(tidyverse). You do not load it again.

data("mpg")

mpg is a tibble of cars: engine size (displ), city and highway mpg (cty, hwy), drive (drv), class, and more. Type ?mpg.

A Base R scatterplot with a loess curve and rugs takes several unconnected calls. The ggplot2 version is one chain:

ggplot(mpg, mapping = aes(x = displ, y = hwy)) +
  geom_point() +
  geom_smooth(se = FALSE, method = loess) +
  geom_rug() +
  xlab("Displacement") +
  ylab("Highway")

qplot() is a short ggplot2 helper for a quick look. We will write ggplot() so the layers stay visible.

2. The grammar, in four pieces

The grammar of graphics treats a plot as layers. You can get most of the work done with four of them:

• data: a data frame
• aesthetic mappings: which column is x, y, color, shape, size, alpha, fill, … Created with aes()
• geoms: the marks, geom_point(), geom_histogram(), geom_boxplot(), …
• facets: a panel per level of a categorical variable, facet_wrap() or facet_grid()

There are also stats (binning for a histogram), positions (jitter, dodge, stack), and coordinate systems. We will change those when we need them.

ggplot() makes the canvas. It needs data (a data frame) and, usually, mapping = aes(...).

ggplot(data = mpg, mapping = aes(x = displ, y = hwy))

That draws axes and an empty panel. There is no geom yet.

Add a layer with + at the end of the line, then the next function on the next line. A + at the start of a line will error.

ggplot(data = mpg, mapping = aes(x = displ, y = hwy)) +
  geom_point()

You can pipe the data in. The first argument of ggplot() is then filled, so you only write the mapping:

mpg |>
  ggplot(mapping = aes(x = displ, y = cty)) +
  geom_point()

Larger displ goes with smaller cty: a negative relationship.

3. Which variable is x

You choose the axes from the question. The response usually goes on y. The explanatory variable goes on x.

“Plot A vs B” and “plot A as a function of B” both put B on x and A on y. Swapping the axes does not change the association. It does change what a reader thinks is doing the explaining.

ggplot(data = mpg, mapping = aes(x = cty, y = hwy)) +
  geom_point()

cty and hwy move together: a positive relationship.

4. A third variable: categorical

A categorical variable has a short list of groups. In mpg, drv is f (front), r (rear), or 4.

Map it to color (usually clearer) or shape (harder once you have many levels):

ggplot(data = mpg, mapping = aes(x = displ, y = hwy, color = drv)) +
  geom_point()

ggplot(data = mpg, mapping = aes(x = displ, y = hwy, shape = drv)) +
  geom_point()

Fifty states as fifty colors is a mess. Recode into fewer groups, or facet (note 3.3).

size and alpha have a natural order (small to large, faint to dark). A categorical variable usually does not. Mapping drv to size or alpha is legal and usually a bad idea.

5. A third variable: quantitative

A third number can go on color (a spectrum), alpha (transparency), or size (a bubble chart; worst when points overlap).

ggplot(data = mpg, mapping = aes(x = displ, y = hwy, color = cty)) +
  geom_point()

ggplot(data = mpg, mapping = aes(x = displ, y = hwy, color = cty)) +
  geom_point() +
  viridis::scale_color_viridis()

ggplot(data = mpg, mapping = aes(x = displ, y = hwy, alpha = cty)) +
  geom_point()

ggplot(data = mpg, mapping = aes(x = displ, y = hwy, size = cty)) +
  geom_point()

shape has a short list of symbols. Mapping a continuous variable to shape errors.

Stacking color, size, alpha, and shape at once is possible and usually unreadable. If you cannot say the sentence the plot is making, drop a mapping.

6. Practice

1. Load iris (data(iris); head(iris)). Which columns are quantitative? Which is categorical?

2. Scatterplot petal length vs petal width, colored by species. What do you see?

3. Map a categorical column in mpg to size or alpha. Why is that a poor default?