Non-Standard Geometries

library(tidyverse)
library(palmerpenguins)

Beyond the Basics of Data Visualization

Historically, there have been five plot types that are considered basic or foundational:

1. Histograms
2. Boxplots
3. Barcharts
4. Scatterplots
5. Line graphs¹ (or time series plots)

While you may have never been explicitly told these are the basic plot types, hopefully this list is not surprising to you. It’s likely the case that, similar to the normal distribution, you knew about and were using these before you even knew their names, possibly through Excel or something back before you had taken any statistics courses.

This week, we’re going to learn about plots / geometries outside of these basic plots!

Lollipop plot with geom_segment()

All too often, we see people (students, researchers, newspapers) create barplots for variables that are not counts (e.g., mean salary). A barplot is designed to display the frequency of each level of a categorical variable, so people get confused when they are used to plot other summary statistics.

The lollipop plot solves this problem! A lollipop plot is the combination of two geometries—geom_segment() and geom_point(). The lollipop stick is created with geom_segment() and each lollipop’s head is created with geom_point().

Let’s give this a look using the penguins data from the palmerpenguins package.

species_means <- penguins |> 
  group_by(species) |> 
  summarize(mean_flipper = mean(flipper_length_mm, na.rm = TRUE))

ggplot(data = species_means, 
       mapping = aes(x = mean_flipper, y = species)
       ) +
  geom_segment(aes(yend = species, 
                   x = 0, 
                   xend = mean_flipper), 
               color = "gray50") +
  geom_point(size = 4, color = "steelblue") +
  labs(x = "Mean Flipper Length (mm)", 
       y = "", 
       title = "Comparison of Flipper Lengths for Different Penguin Species"
       )

Figure 1: Lollipop plot of mean flipper length

Ridgeline plot with geom_density_ridge()

Boxplots are an easy way to assess how different the centers (medians) and spreads (IQR, range) are between groups. However, boxplots do not display the shape of a distribution. Specifically, boxplots hide distributions with multiple modes (peaks).

The original ridgeline plot—Joy Division, Unknown Pleasures

The ridgeline plot solves this problem! A ridgeline plot is comprised of vertically stacked density plots. Because the ridgelines are stacked vertically, the categorical variable needs to be mapped to the y-axis.

library(ggridges)

ggplot(data = penguins, 
       mapping = aes(x = flipper_length_mm, 
                     y = species, 
                     fill = species)) +
  geom_density_ridges(alpha = 0.5) +
  labs(x = "Flipper Length (mm)", 
       y = "", 
       title = "Comparison of Flipper Lengths for Different Penguin Species"
       ) +
  theme(legend.position = "none")

Figure 2: Ridgeline density plot of penguin flipper length

TipChanging the height of the ridges

If you’re not a fan of the ridges touching each other, then the scale argument was made just for you! If you set scale = 1 (inside geom_density_ridges()) then the top of each ridge will be the bottom of the next ridge.

ImportantAdding color

Adding color to a plot is a fun way to make it more engaging. Here, color ridges are much more exciting than gray ridges. However, when we incorporate a color (or fill) aesthetic, ggplot2 automatically creates a legend for us. That’s nice most of the time, but here the information from the legend (the values of species) is already encoded in y-axis. So, the legend has redundant information. We don’t want to have redundant information in our plots, which is why we removed the legend (theme(legend_position = “none")).

Area plot with geom_ribbon()

There are two types of area plots, one emphasizes the total area while the other emphasizes the area between two groups. The New York Times loves to use area plots, so we’ve pulled out two examples from the What’s going on in this graph collection:

Total Area

Notice how the area always starts at 0 and goes up to wherever the highest \(y\) value is for that group.

Area Between Groups

Notice how the area starts at the value of the lower group and goes up to the value of the higher group.

The total area plot can be made with geom_area() and the between area plot can be made with geom_ribbon(). Let’s explore each of these functions!

Total Area

Typically, area plots are variants of line plots. So, let’s use the gapminder data from the gapminder package to explore life expectancy over time. Since there are multiple countries for each continent (for each year), we will need to summarize these values before plotting²

library(gapminder)

continent_life_exp <- gapminder |> 
  group_by(year, continent) |> 
  summarize(mean_life = mean(lifeExp, na.rm = TRUE), 
            .groups = "drop")

head(continent_life_exp)

# A tibble: 6 × 3
   year continent mean_life
  <int> <fct>         <dbl>
1  1952 Africa         39.1
2  1952 Americas       53.3
3  1952 Asia           46.3
4  1952 Europe         64.4
5  1952 Oceania        69.3
6  1957 Africa         41.3

Okay, now that we’ve summarized our data, let’s try making an area plot!

ggplot(data = continent_life_exp, 
       mapping = aes(x = year, 
                       y = mean_life, 
                       color = continent)) +
  geom_line() +
  geom_area(mapping = aes(fill = continent), 
            position = "identity") +
  labs(x = "", 
       y = "", 
       title = "Mean Life Expectancy Over Time", 
       fill = "Continent", 
       color = "Continent") 

Figure 3: A first go at making an area plot

Huh, that doesn’t look great. It looks like we are only getting a plot for Oceania. We can try making the areas more transparent (with alpha) and see if we can uncover the other groups.

ggplot(data = continent_life_exp, 
       mapping = aes(x = year, 
                       y = mean_life, 
                       color = continent)) +
  geom_line() +
  geom_area(mapping = aes(fill = continent), 
            position = "identity", 
            alpha = 0.25) +
  labs(x = "", 
       y = "", 
       title = "Mean Life Expectancy Over Time") 

Figure 4: Incorporating transparency

Okay! We do have every area plot, it’s just hard to see them. Looking at the plot we can tell that the tallest line is Oceania and the shortest line is Africa. If we do some clever reordering of the levels of continent then we should be able to see every group.

TipReordering Factors Based on Another Variable

In this case, we want to reorder continent based on the mean life expectancy. We could reorder this by hand, but that sounds like a lot of typing and seems error prone.

Instead, we can use the fct_reorder() function from the forcats package to reorder continent based on mean_life. We need to be sure to specify .desc = TRUE since we want the order to be larges to smallest!

gapminder |> 
  group_by(year, continent) |> 
  summarize(mean_life = mean(lifeExp, na.rm = TRUE), 
            .groups = "drop") |> 
  mutate(continent = forcats::fct_reorder(continent,
                                          mean_life, 
                                          .desc = TRUE)
         ) |> 
  ggplot(mapping = aes(x = year, 
                       y = mean_life, 
                       color = continent)) +
  geom_line() +
  geom_area(mapping = aes(fill = continent), 
            position = "identity", 
            alpha = 0.6) +
  labs(x = "", 
       y = "", 
       title = "Mean Life Expectancy Over Time") 

Figure 5: A more polished area plot where every group is visible

Area Between Groups

Suppose instead of featuring every continent, we want our plot to display the profound differences in life expectancy between the largest and smallest groups. This is what a ribbon plot accomplishes!

If you look at the documentation for geom_ribbon() you will see that the function requires ymin and ymax aesthetics. These values tell ggplot() where the shading should start and end. Currently, the data we are plotting have one column for continent and one column for mean_life, but we need to have our data structured in such a way that the values for Africa are in one column (for ymin) and the values for Oceania are in a different column (for ymax). So, we are going to need to restructure our date.

ribbon_summaries <- gapminder |> 
  filter(continent %in% c("Oceania", "Africa")) |> 
  group_by(year, continent) |> 
  summarize(mean_life = mean(lifeExp, na.rm = TRUE), 
            .groups = "drop") |>
  pivot_wider(names_from = continent, 
              values_from = mean_life, 
              names_prefix = "mean_")

head(ribbon_summaries)

# A tibble: 6 × 3
   year mean_Africa mean_Oceania
  <int>       <dbl>        <dbl>
1  1952        39.1         69.3
2  1957        41.3         70.3
3  1962        43.3         71.1
4  1967        45.3         71.3
5  1972        47.5         71.9
6  1977        49.6         72.9

Now that I’ve pivoted the summary statistics, I have exactly the data I need! The mean_Africa column can be mapped to ymin and the mean_Oceania column can be mapped to ymax. Let’s give it a try!

gapminder |> 
  filter(continent %in% c("Oceania", "Africa")) |> 
  group_by(year, continent) |> 
  summarize(mean_life = mean(lifeExp, na.rm = TRUE), 
            .groups = "drop") |> 
  ggplot(mapping = aes(x = year, 
                       y = mean_life, 
                       color = continent)) +
  geom_line(linewidth = 2) +
  geom_ribbon(data = ribbon_summaries, 
              mapping = aes(x = year, 
                            ymin = mean_Africa, 
                            ymax = mean_Oceania
                            ), 
            position = "identity",
            inherit.aes = FALSE, 
            fill = "lightgray") +
  labs(x = "", 
       y = "", 
       title = "Profound Differences in Life Expectancy", 
       subtitle = "Comparing Continents with Highest and Lowest Life Expectancy") 

Figure 6: A ribbon plot highlighting differences in life expectancy between Africa and Oceania

Voila! We have a ribbon plot highlighting the differences in life expectancy between these two continents. The icing on the cake would be to reorder our legend so it goes in the same order as the plot or even remove the legend in favor of annotations, which we will learn in the next chapter.

Heatmap with geom_tile()

Heatmaps are a visualization which use color intensity to represent the magnitude of values in a dataset. With a heatmap, we can identify patterns quickly, spot trends and anomalies, understand relationships between variables, and compare large amounts of data efficiently.

A simple heatmap could help us spot trends in data collection. Based on the heatmap below, it appears that far more penguins were sampled on the Biscoe Island and the data collection in the first year (2007) was much smaller than the later years.

penguins |> 
  count(year, island) |> 
  ggplot(mapping = aes(x = year, y = island, fill = n)) +
  geom_tile() +
  labs(x = "", 
       y = "", 
       fill = "Penguins \nSampled", 
       title = "Number of Penguins Sampled Each Year", 
       subtitle = "Separated by Island of the Palmer Archipeligo")

Figure 7: Heatmap of sample sizes

Another common application of a heatmap is to visualize relationships (correlations) between variables. The first step is to get the pairwise correlations. We like the corrr package for this, since it is compatible with the tidyverse and it return a dataframe.

library(corrr)

penguins_cor <- penguins |>
  # Make year a category since we don't want to include it
  mutate(year = forcats::as_factor(year)) |> 
  select(where(is.numeric)) |> 
  correlate(method = "pearson", 
            use = "pairwise.complete.obs")

penguins_cor

# A tibble: 4 × 5
  term              bill_length_mm bill_depth_mm flipper_length_mm body_mass_g
  <chr>                      <dbl>         <dbl>             <dbl>       <dbl>
1 bill_length_mm            NA            -0.235             0.656       0.595
2 bill_depth_mm             -0.235        NA                -0.584      -0.472
3 flipper_length_mm          0.656        -0.584            NA           0.871
4 body_mass_g                0.595        -0.472             0.871      NA    

The only thing different about this output (as compared to the base R cor() function) is that the diagonal entries have correlations of NA instead of 1. Similar to making the area plots between groups, we should notice that these data are not currently in the layout ggplot() expects. Namely, the values of the second variable are spread across the columns. So, we will need to pivot our data before we plot.

penguins_cor <- penguins_cor |> 
  rename(term1 = term) |> 
  pivot_longer(cols = -term1, 
               names_to = "term2", 
               values_to = "correlation")

penguins_cor

# A tibble: 16 × 3
   term1             term2             correlation
   <chr>             <chr>                   <dbl>
 1 bill_length_mm    bill_length_mm         NA    
 2 bill_length_mm    bill_depth_mm          -0.235
 3 bill_length_mm    flipper_length_mm       0.656
 4 bill_length_mm    body_mass_g             0.595
 5 bill_depth_mm     bill_length_mm         -0.235
 6 bill_depth_mm     bill_depth_mm          NA    
 7 bill_depth_mm     flipper_length_mm      -0.584
 8 bill_depth_mm     body_mass_g            -0.472
 9 flipper_length_mm bill_length_mm          0.656
10 flipper_length_mm bill_depth_mm          -0.584
11 flipper_length_mm flipper_length_mm      NA    
12 flipper_length_mm body_mass_g             0.871
13 body_mass_g       bill_length_mm          0.595
14 body_mass_g       bill_depth_mm          -0.472
15 body_mass_g       flipper_length_mm       0.871
16 body_mass_g       body_mass_g            NA    

Okay, now that the data are in the correct orientation to plot, we want to think about the appearance of the plot. If we were to make a heatmap with these data, the x and y values would have the current labels of term1 and term2 (e.g., "bill_length_mm", "body_mass_g"). These don’t see like the labels we want for a nice looking visual!

Let’s first make a function to reformat the labels. The function should remove the units and the _ from the variable names, and convert the variable names to titles (i.e., Bill Length not bill length).

make_titles <- function(x){
  # expects a character or factor as an input
  stopifnot(is.character(x) | is.factor(x))
  
  # remove the units at the end of the name
  str_remove(x, pattern = "(g|mm)$") |> 
  # replace all _ with spaces
  str_replace_all(pattern = "_", replacement = " ") |> 
  # remove extra whitespace on left 
  str_trim(side = "both") |> 
  # convert the first letter of each word to upper case
  str_to_title()
}

Okay, now let’s use this function to make nice labels and get our heatmap!

penguins_cor |> 
  mutate(
    across(.cols = c(term1, term2), 
           .fns = ~ make_titles(.x)
           )
    ) |> 
  ggplot(mapping = aes(x = term1,
                       y = term2, 
                       fill = correlation)) +
  geom_tile() +
  labs(x = "", 
       y = "", 
       title = "Correlation Between Penguin Body Measurements", 
       fill = "Correlation"
       )

Figure 8: Heatmap of correlations, grey boxes represent perfect correlations (of 1)

Hexbin plot with geom_hex()

A scatterplot is our classic visualization to investigate the relationship between two quantitative variables. However, the usefulness of a scatterplot decreases as the size of the dataset grows. For example, the diamond dataset has 53940 observations on the size, cut, and price of various diamonds. When making a scatterplot of these observations, we end up with a plot that leaves something to be desired.

ggplot(data = diamonds, 
       mapping = aes(x = carat, 
                     y = price)
       ) +
  geom_point()

Figure 9: Overplotting in a scatterplot

You may have learned about using transparency (alpha) or point size (shape = ".") as methods to address this issue. Another option is to create a hexbin plot!

The hexbin R package contains binning and plotting functions for hexagonal bins. The geom_hex() function from this package is the tool we need to make a hexagonal bin plot.

library(hexbin)

ggplot(data = diamonds) +
  geom_hex(mapping = aes(x = carat, y = price)) +
  scale_y_continuous(
    labels = scales::label_currency(prefix = "$")
    ) +
  labs(x = "Carat of Diamond", 
       y = "", 
       fill = "Number of \nDiamonds",
       title = "Number of Diamonds Observed", 
       subtitle = "For each Price, Carat Combination")

Figure 10: Hexbin plot of frequency of observations for each carat and price combination

But wait, there’s more!

We’ve only begun to scratch the surface of “non-standard” geometries that could be used. The From Data to Viz website provides a more exhaustive list of the plethora of different data visualizations that could be made and the situations when they’d be most useful. Have you ever wondered when you would want to make a radar plot? Or maybe a parallel plot? This site has it all!

Check-inCheck In

1. Which of the following non-standard plots could be used to visualize categorical variables?

2. Which of the following non-standard plots could be used to visualize numerical variables?

(Some plots can go both places!)

• Bubble Plot

• Violin Plot

• Circular Barplot

• Tree Map

• Sankey Diagram

• Waffle Chart

• Stream Graph

• Radar Chart

• Bubble Plot

Footnotes

1. We’re of the opinion that line graphs are really just scatterplots with a little extra definition, but acknowledge the first four as core plot types.

2. As a reminder geom_line() requires there be one \(y\) value for every \(x\) value.