BFF—Become a fan of facets

About

“Small multiples” is a design principle for data visualizations advocated by Edward Tufte. I won’t describe the principle in detail, and even less present the work of a master in the field.
Suffice to say that there is no serious discussion about the display of quantitative information without a reference to E. Tufte—a.k.a. “The Leonardo da Vinci of data” (NYTimes) and “The Galileo of graphics.” (Bloomberg). So, disregard his principles at your own peril!
This page of the Pew Research Center has illustrations of the principle. Figure 1 shows an example published by Statistics Portugal with the last census data—okay, full disclosure: I did it for them.

Figure 1: Evolution of average age in Portuguese municipalities, 2011–2021, per NUTS-II. Highlighted in each NUTS-II are the 2021 min and max, the average (black), and the highest increase (orange); all decreases in light blue—e.g., Aljezur. Source: INE, Census.

Idea

“Small multiples” refers to the process of dividing a plot into multiple smaller subplots based on one or more categorical variables. They are typically dense in data. But the division of the data into subsets makes it easier to compare patterns and trends across different groups.
Thus they are particularly useful when exploring relationships between variables while considering the influence of categorical factors.
In {ggplot2}, the process is called facetting. It is achieved using the facet_wrap() or facet_grid() functions, depending on whether you want a one-dimensional or two-dimensional layout, respectively

Case-in-point

There are innumerous applications of the principle. Most of them are probably even better cases than the case presented here.
I suggest that bar plots can advantageously be presented in facets.
Figure 2 shows how many goals each striker scored against each of the main opponents. The case in point is to offer an alternative to Figure 2, using facets [facet_wrap()].

Figure 2: Bar plot to improve on. Number of goals each striker scored against each of the main opponents.

Packages

Install the following packages [install.packages()] if they are not present in your machine.

library(tibble)
library(ggplot2)
library(ggthemes)

0

The data used for the illustration is created by hand [tibble()].

1

This is a package with cool themes for {ggplot}. I’m using it below only for an aesthetic suggestion, so no real need.

Code

The data

The plots below use a random data set created as follows.

set.seed(42)
teams <- c("Leeds", "Brighton", "Luton", "Wolves", "Everton")
teams <- factor(teams,
                levels = teams)
                 
df <- tibble(
  striker = sample(LETTERS[1:5],
                   26,
                   replace = TRUE),
  opponent = sample(teams,
                    26,
                    replace = TRUE,
                    prob = c(0.1, 0.1, 0.1, 0.35, 0.35)
                    ))

1

The functions below use a pseudo-random number generator. To guarantee that the example is reproducible, i.e., that the “random” numbers will remain the same, we must set the seed of the random number generator. Of course, a different seed will result in different random numbers. I usually choose the number 42 as the seed because 42 is the answer to the ultimate question of life, the universe, and everything.

2

This is a character vector that I use multiple times below. So it is good to create it before.

3

The ‘opponent’ is a categorical variable [factor()] with the given levels.

4

This creates a tibble by hand.

5

For the variable ‘striker’, choose randomly [sample()] 26 of the first 5 capital letters [LETTERS[1:5]]. Of course, some must be repeated.

6

Same for variable ‘opponent’, but this time, each team has a different probability of being selected [prob()].

The “bad” plot

I first recreate Figure 2.

p1 <- df |>
  ggplot(aes(x = opponent, fill = striker )) +
  geom_bar(position = position_dodge())

p1

1

The object p1: pipe the above df into the ggplot() function.

2

Two mappings: x axis and the color of fill.

3

The plot is a bar plot. The argument position_dodge() puts the bars next to one another, instead of stacking them.

Figure 3: Again, the bar plot to improve on.

Bars with same width

An obvious problem with Figure 3 is the variation of the bars’ widths. Here, the problem is surprisingly with the package: Excel does not make it that bad.

p2 <- df |>
  ggplot(aes(x = opponent, fill = striker )) +
  geom_bar(position = position_dodge(preserve = "single"))

p2

1

The preserve = "single" makes the trick.

Figure 4: Same plot as Figure 3, but with bars with same width.

This solution is not quite satisfying, yet. Do you see why?

Enters facet_()

Notice in Figure 4 that the bars, despite having the same width, do not appear in a fixed position—i.e., the strikers do not always appear in the same order because the plot drops strikers who did not score against an opponent. One may think that the colors are “there for that”, for distinguishing the strikers. But I still find it confusing.
Enters the facet() function. This function will allow us to create the same plot for each level of the categorical variable, as in Figure 5.

p3 <- df |>
  ggplot(aes(x = striker, fill = striker )) +
  geom_bar(position = position_dodge() ) +
  facet_wrap(facet = vars(opponent),
             nrow = 1)

p3

1

Here is the magic. facet_wrap() is used because I only want to subset the data in one categorical variable [vars()].

2

I force the plot to be in one row, slightly compressing each of the facets. Otherwise, ggplot() would probably divide the facets into two lines, which I find less appropriate for comparisons.

Figure 5: Same plot as Figure 3, but divided for each level of the variable ‘opponent’, i.e., with a facet for each of them.

Remove redundancy

With facets, I find that the color mapping becomes redundant since we already identify the striker on the x-axis.

p4 <- df |>
  ggplot(aes(x = striker)) +
  geom_bar(position = position_dodge(),
           alpha = 0.5) +
  facet_wrap(facet = vars(opponent),
             nrow = 1) 

p4

1

No fill mapping anymore.

2

The alpha argument sets the transparency of the color. The default, 1, is too dark for my taste.

Figure 6: Same plot as Figure 5, but without the redundant color fill mapping.

Aesthetics

Want to see how the plot looks like under different theme options? Here is an example. In particular, appreciate how easy it is to change so much a plot with a single line of code. Careful, do not be driven away by the simplicity and start using multiple themes in the same document!

p5 <- p4 +
  theme_fivethirtyeight()

p5  

1

The “crazy” part. You can use the very plot above and add [+] new code to it.

2

This is just one of the numerous themes offered by the package {ggthemes}.

Figure 7: Same plot as above, but under a fully predefined theme.

Advanced R

As you know, every part of the plot can be customized. I dot encourage to pursue the elusive quest for the most “beautiful” plot. But sometimes some changes are required for your publication. These are generally obtained through options of the theme [theme()]. I offer a few examples here specifically related to the facets. Notice that the plot actually becomes less beautiful.

p6 <- p4 +
  theme(axis.title = element_blank(),
        strip.text.x = element_text(size = 12,
                                    face = "bold"),
        strip.background = element_rect(fill = "white",
                                        color = "black"),
        panel.spacing.x = unit(1, "lines"))

p6 

0

Crazy again. I just use the same plot as above and add [+] code.

1

Remove axis titles [element_blank()]. Not specific to facets.

2

The strip is the “title” on the top of the facet. I change the text [element_text()] to a bigger size and to bold face.

3

The strip background has a fill—inside—and a color—border—for which the color can be changed.

4

This is a useful option. It changes the space [panel.spacing.x] between the facets on the horizontal axis.

Figure 8: Same plot as above, but other theme options.