1  Anatomy of a ggplot

This chapter is about the underlying features of ggplot2.

You can get really far with ggplot2 by copying an example, changing the variable names, and a bit of use of an LLM.

One of the challenges with this, is that you might not have internalised the mental model of how ggplot plots are built. The interface to ggplot2 helps you think about your data, build plots quickly for exploration, but also present finely made graphics.

In this chapter we will take plots, tease them apart, and put them back together.

The greatest value of a picture is when it forces us to notice what we never expected to see.

— John Tukey

Overview

Duration 45 minutes

Questions

• What is this plot telling me, and what is it made of?
• What are the pieces of a ggplot?
• How do I build a plot up?
• How do I map a variable onto colour, shape, or size?
• Why does colour inside aes() do something different to colour outside it?
• How do I add a second layer?

What you need this session

• A session of RStudio open
• Something to draw on, and something to draw with
• The following packages installed

library(countdown)
library(ggplot2)
library(naniar)
library(dplyr)
library(tidyr)

Motivation

From https://www.research.autodesk.com/publications/same-stats-different-graphs

1.1 The data

We will be looking at some data from the naniar R package: oceanbuoys. The helpfile tells us:

?oceanbuoys

Real-time data from moored ocean buoys for improved detection, understanding and prediction of El Ni’o and La Ni’a. The data is collected by the Tropical Atmosphere Ocean project (https://www.pmel.noaa.gov/gtmba/pmel-theme/pacific-ocean-tao).

Along with a list of the variables.

The data look like the following

year	latitude	longitude	sea_temp_c	air_temp_c	humidity	wind_ew	wind_ns
1997	0	-110	27.59	27.15	79.6	-6.4	5.4
1997	0	-110	27.55	27.02	75.8	-5.3	5.3
1997	0	-110	27.57	27.00	76.5	-5.1	4.5
1997	0	-110	27.62	26.93	76.2	-4.9	2.5
1997	0	-110	27.65	26.84	76.4	-3.5	4.1
1997	0	-110	27.83	26.94	76.7	-4.4	1.6

Each row represents recordings for a single oceanbuoy:

• The year it was taken
• Latitude and longitude
• Temperature (sea temp and air temp, celcius)
• Humidity (relative humidity)
• Wind (east-west and north-south).

1.2 Practice sketching your plots

I think sketching out your plots is a really useful first step. Especially so when you are starting out with crafting data visualisations.

Doing these kinds of sketches is something that Nicola Rennie promotes in her book, The art of data visualisation with ggplot2. It is free to read online, and is a great read, and goes into more advanced concepts than this book.

The sketches don’t have to be perfect! They can just represent your intentions for what you want to explore in the data. I find if I can draw the general shape of your plot, then it becomes easier to imagine other extensions. What is also particularly useful about this is that it doesn’t require code. So, if you don’t know how to code something up, your limitation is what you can draw.

This is the standard of sketch I am wanting you to think about doing:

Scatterplot

Line plot

barplot

boxplot

facet/subplot

Once you are comfortable identifying the parts you need to draw, it will be easier to link these back to a plotting system like ggplot2. Which, in turn forces you to think about how your data correspond to the plot.

It also reminds me of the feeling of asking a “good question” in research: Spend time thinking about a good question, and you’ll do good work.

Have a vision of what you want to create, then you know what you want to work towards!

Let’s practice doing some sketches of data based on a few snippets of data.

I’ll show my drawings of a couple of these, and you can then practice on some other data.

NoteMy sketching process

Here’s what I do:

Draw axis

Label axis

Sketch the data

Say it in a sentence

Plot the oceanbuoys data, mapping air_temp_c to the x axis and sea_temp_c to the y axis. Display using points.

NoteYour Turn: Sketching plots

Here are 6 rows of oceanbuoys, and three of its columns.

Adding missing grouping variables: `year`

year	air_temp_c	sea_temp_c
1993	23.14	23.71
1993	23.93	23.84
1993	24.24	24.01
1997	27.68	28.83
1997	26.39	26.79
1997	27.51	29.10

I want you to practice drawing this out as a scatterplot:

• For the oceanbuoys data
• air temp on the x axis
• sea temp on the y axis

Sketching your data isn’t the prettiest thing, but the point is that you practice visualising what is in your head, and putting it on the page.

There is something special about this. The better you are at putting these on the page, the easiest it will be to see them in your code.

1.3 Use ggplot2 to make a plot

Here’s one I prepared earlier.

I’ve created a data visualisation of this data using ggplot2:

Figure 1.1

Each point is a reading from a buoy floating in the tropical Pacific.

• The X axis (horizontal) is the air temperature
• the Y axis (vertical) is the temperature of the sea underneath it

NoteYour Turn: What do you see?

Take a couple of minutes to assess what you see in the plot above, and answer these questions:

1. How is the ggplot different to your sketch?
2. What do you see in the data - what is the story?
3. Is there something in here you would want to explore?

1.4 Creating plots in ggplot

OK, let’s walk through how to create the plot using ggplot2. I like to use this template:

ggplot(data = <<DATA>>,
       mapping = aes(x = <<X AXIS VARIABLE>>,
           y = <<Y AXIS VARIABLE>>)) +
  geom_<<GEOM>>(<<OPTIONS>>)

Here is the code that produced the above plot, with that template:

ggplot(data = oceanbuoys,
       mapping = aes(x = air_temp_c,
           y = sea_temp_c)) + 
  geom_point()

What is it made of?

Let’s take another look at the code that creates Figure 1.1:

library(naniar)
library(ggplot2)
ggplot(
1  data = oceanbuoys,
2  mapping = aes(
    x = air_temp_c,
    y = sea_temp_c,
  )
3) +
4  geom_point()

1

We see data, oceanbuoys, which is from the {naniar} package.

2

There is a set of instructions mapping which column goes where - Those live inside aes(): air_temp_c across the x axis, sea_temp_c up the axis

3

There is a + between them, which is how you glue the parts together.

4

There is something saying draw a point for each row: is geom_point().

1.5 Building it back up

Three lines, three pieces. Data, aesthetics, geom.

Let’s show this as a set of panel tabs, to get a bit more of a sense of the plot being built up.

grey

ggplot(oceanbuoys)

axes

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c
  )
)

geom

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c
  )
) +
  geom_point()

Warning: Removed 81 rows containing missing values or values outside the scale range
(`geom_point()`).

NoteYour Turn

Look at Figure 1.1 and answer these without running any code.

1. How many rows do you think oceanbuoys has? Have a guess, then check with nrow(oceanbuoys).
2. If you mapped humidity onto the y axis instead of sea_temp_c, which part of the code would change?
3. Using ggplot, can you draw this plot, with humidity on the y axis? See the above code for a template.

NoteYour Turn

Let’s talk about adding colour - draw a scatterplot of this data:

oceanbuoys |>
  drop_na() |> 
  group_by(year) |> 
  slice_sample(n = 3) |> 
  select(year, air_temp_c, sea_temp_c) |> 
  knitr::kable()

year	air_temp_c	sea_temp_c
1993	22.99	23.29
1993	23.90	23.77
1993	22.35	22.65
1997	26.79	27.73
1997	26.12	28.05
1997	28.28	28.59

• Draw the axis
• air temp
• sea temp
• Colour/shade the points by year

Here it is in ggplot:

• What do you notice?
• How would you explore differences between years?
• What would you search for?

1.6 The components of the grammar of graphics

The grammar of graphics is an idea from Leland Wilkinson, and ggplot2 is Hadley Wickham’s implementation of it. It is described as a grammar, in the sense that you can combine components together to express something complex.

There is more detail to the differences between Wilkinson’s grammar of graphics, and the grammar implemented in ggplot2. If you want to learn more about the specific differences, the ggplot2 book describes these, but I also would recommend reading the chapter in Hadley Wickham’s PhD thesis, which provides other detail, too.

There are seven parts to the grammar of graphics as implemented in ggplot. I don’t know them off by heart, and I don’t expect you to. But they do provide a useful outline for the different features we will discuss in this course:

Data: Data frame you are plotting

year	latitude	longitude	sea_temp_c	air_temp_c	humidity	wind_ew	wind_ns
1997	0	-110	27.59	27.15	79.6	-6.4	5.4
1997	0	-110	27.55	27.02	75.8	-5.3	5.3
1997	0	-110	27.57	27.00	76.5	-5.1	4.5
1997	0	-110	27.62	26.93	76.2	-4.9	2.5
1997	0	-110	27.65	26.84	76.4	-3.5	4.1
1997	0	-110	27.83	26.94	76.7	-4.4	1.6

Aesthetics: How columns map onto visual features

ggplot(oceanbuoys,
       aes(x = air_temp_c,
           y = sea_temp_c,
           colour = factor(year))) + 
  geom_point()

Geoms: The mark drawn

Stats: The computations that happen before drawing

Scales: How data values change into different positions, and colours

Coords: The coordinate system that the marks (geom) are placed

CautionSome history: the most famous polar plot

coord_polar() looks like a novelty, and one of the most consequential statistical graphics ever drawn is a polar bar chart.

Florence Nightingale is remembered as a nurse. She was also a statistician, and during the Crimean War she drew what she called a coxcomb: a circle divided into months, with the area of each wedge showing how many soldiers died, split by cause. The picture made one thing impossible to miss, which is that far more men were dying of preventable disease than of their wounds.

She was arguing for sanitary reform, and she reached for a plot because a table would not have carried it. She was elected the first woman member of the Royal Statistical Society.

A few years earlier, in 1854, John Snow mapped cholera deaths in Soho as marks on a street map, and the marks clustered around one water pump on Broad Street. Same move: put the data where the thing happened, and the pattern argues for you.

If you want to follow the modern thread, Di Cook, Heike Hofmann and Debbie Swayne have spent careers on how we read statistical graphics, and there is a good starting point in Di Cook’s JSM materials.

Facets: How the plot is split into smaller subplots

In this chapter we focus on: Data, aesthetics, and geoms.

We will come back to these components as we go in the course. But I find knowing the names of these features of the plot provide useful features of a mental model to focus on and understand.

To give you a visceral sense of some of these features, I would like you to explore plotting another dataset. There are only 9 values, so you should be able to draw these.

species	year	bill_len	bill_dep
Adelie	2007	39.1	18.7
Adelie	2008	39.6	17.7
Adelie	2009	35.0	17.9
Chinstrap	2007	46.5	17.9
Chinstrap	2008	50.5	18.4
Chinstrap	2009	50.9	17.9
Gentoo	2007	46.1	13.2
Gentoo	2008	49.1	14.8
Gentoo	2009	49.1	14.5

NoteYour Turn

Sketch!

Here’s the data again:

pen_short

# A tibble: 9 × 4
  species    year bill_len bill_dep
  <fct>     <int>    <dbl>    <dbl>
1 Adelie     2007     39.1     18.7
2 Adelie     2008     39.6     17.7
3 Adelie     2009     35       17.9
4 Chinstrap  2007     46.5     17.9
5 Chinstrap  2008     50.5     18.4
6 Chinstrap  2009     50.9     17.9
7 Gentoo     2007     46.1     13.2
8 Gentoo     2008     49.1     14.8
9 Gentoo     2009     49.1     14.5

I want you to draw a rough sketch of the penguins data, as a scatterplot:

• Bill Length (bill_len) on the x axis
• Bill Depth (bill_dep) on the y axis
• Each point should be coloured

Some things to consider as you draw:

• You can use symbols for the species if you don’t have multiple colour options
  • How will you know which species is which (you will need a legend!)
• What should the range of the x axis be?
• What should the range of the y axis be?

Let’s take 5 minutes to do this - I’ll draw this too.

Here’s one I prepared earlier:

NoteMy plot

Here’s what ggplot2 does:

library(ggplot2)
ggplot(pen_short,
       aes(x = bill_len,
           y = bill_dep,
           colour = species)) + 
        geom_point(size = 5)

NoteYour Turn

• What did you notice about drawing your plot?
• Did you notice the extra steps for drawing a legend?
• The steps for determining the x and y axis?

There are many decisions you need to make when making a plot. I think ggplot really does a great job of breaking down those components in a way that make you think about the data, at the right level of detail.

1.7 Building the plot back up

Let’s show the data and the plots alongside ach other again, so you can see the impact of colour:

data

ggplot(data = oceanbuoys)

Figure 1.2: Data, and nothing else.

axes

ggplot(
  data = oceanbuoys,
  mapping = aes(
    x = air_temp_c,
    y = sea_temp_c
  )
)

Figure 1.3: Now with aesthetics. Axes appear, but nothing is drawn.

geom

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c
  )
) +
  geom_point()

Figure 1.4: And now a geom. This is a plot.

colour

fig_ob_col <- ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c,
    colour = year
  )
) +
  geom_point()

fig_ob_col

Figure 1.5: And now a geom. This is a plot.

TipA digression into missing data

You might have noticed a warning in the recent plots:

Removed 81 rows containing missing values or values outside the scale range

That is no mistake! There are 81 readings in this data where the air temperature was never recorded, and ggplot2 cannot draw a point when it does not know where to put it, so it drops them and tells you.

This is a good thing, I think. For two reasons:

1. The fact it told you at all! Missing values are often omitted. It’s nice ggplot2 is honest.

2. The plot you are looking at is missing 81 readings, and nothing on the plot itself shows you that. You would have to read the warning.

I’ve spent a lot of time looking at missing data, and so I wrote two R packages to help explore data, and missing data: {visdat}, and {naniar}.

A plot that I think is worthwhile showing is the vis_miss() plot from {visdat}:

library(visdat)
vis_miss(oceanbuoys)

This shows the overall missingness in the dataset.

We come back to missing data properly later in the course, where we draw those rows instead of throwing them away. For now, just know that they are missing.

1.8 Mapping more variables

We have mapped x and y axis to data columns, and then demonstrated how colour changes the impact of the story.

There are more variables beyond “colour”, such as “shape”, and “size”.

To use these, you map a variable onto one the same way every time, inside aes().

Let’s try shape, instead of colour:

fig_ob_shp <- ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c,
    shape = factor(year)
  )
) +
  geom_point()

fig_ob_shp

Figure 1.6: Year mapped onto shape instead of colour.

It works, and I think it is clearly worse than the colour version - see them side by side below to illustrate this:

Warning: Removed 81 rows containing missing values or values outside the scale range
(`geom_point()`).
Removed 81 rows containing missing values or values outside the scale range
(`geom_point()`).

The two groups are separated well enough here that you can still see the difference in shape, but you are now doing shape recognition instead of colour recognition. Humans are naturally much faster at identifying colour.

Now let’s try size - we can map it onto the same “year” variable:

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c,
    size = year
  )
) +
  geom_point()

Figure 1.7: Year mapped onto size. This is not a good plot.

While more obvious than shape, I do not thinkg this is a great usage of size. We can try mapping size to another variable, such as “humidity”:

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c,
    size = humidity
  )
) +
  geom_point(alpha = 0.4)

Figure 1.8: Humidity mapped onto size. This is not a good plot.

Hmm? I don’t love it. There is variation in size, but a lot of the points cover each other, and it is actually kind of hard to compare the sizes. I don’t see a clear story, beyond the fact that air temperature and sea temperature are correlated.

So, while it is possible, mapping to aesthetics like shape and size aren’t always the best choices.

We come back to this properly later in the course, where we discuss choosing between different aesthetic mappings.

NoteYour Turn

1. Using the oceanbuoys data, make a scatterplot in ggplot2 of wind_ew against wind_ns, coloured by year (as a factor).

ggplot(oceanbuoys,
       aes(x = wind_ew,
           y = wind_ns,
           colour = factor(year))) +
  geom_point()

2. Now map sea_temp_c onto colour instead. What is different about the legend?

ggplot(oceanbuoys,
       aes(x = wind_ew,
           y = wind_ns,
           colour = sea_temp_c)) +
  geom_point()

3. Try mapping year onto colour without wrapping it in factor(). What happens, and why do you think that is?

ggplot(oceanbuoys,
       aes(x = wind_ew,
           y = wind_ns,
           colour = year)) +
  geom_point()

1.9 Using aes inside and outside

What to put inside and outside aes() ?

This is a common gotcha, it catches everybody, including me, sometimes.

If you want some nice “purple” points - a reasonable request - you write might write:

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c,
    colour = "purple"
  )
) +
  geom_point()

Figure 1.9: Asking for purple Getting salmon, and a legend.

We now have a legend with an entry, “purple” - a bit strange, and unexpected?

So, what happened?

• Inside aes() are mappings to a variable in your data.
• “purple” is not a variable in the data.
• ggplot treated “purple” as a variable of a single category

The code above says:

“take this variable and use it to decide the colour”

ggplot2 did exactly that. It took the value "blue", and treated it as a variable with one single category, and gave that category the first colour from its default palette, which happens to be salmon.

If you want to set a colour rather than map one, it goes outside aes(), in the geom.

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c
  )
) +
  geom_point(colour = "purple")

Figure 1.10: Setting the colour instead of mapping it.

Purple points, no legend.

ImportantThis will bite you

The rule is short, and it is worth memorising.

Inside aes() maps a variable.

Outside aes() sets a value.

Unexpected legends and colours can be clues for this kind of error!

NoteYour Turn

1. Make the points in your wind plot dark green and slightly transparent.
2. Now deliberately put colour = "darkgreen" inside aes() and look at what you get.
3. size = 3 and alpha = 0.5 follow exactly the same rule. Try setting each of them in both places.

1.10 Layers

The + is doing a lot! It is worth being explicit about what.

Each thing you add with + is a layer, and layers draw on top of each other in the order you write them. So let’s add a second one.

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c,
    colour = factor(year)
  )
) +
  geom_point(alpha = 0.4) +
  geom_smooth(method = "lm")

Figure 1.11: Two layers. Points, then a smooth on top.

Two trend lines, one per year.

I never told geom_smooth() about the years. It inherited the whole aes() from the ggplot() call, including the colour mapping, and a colour mapping splits the data into groups. So it fitted one line per group.

That inheritance is the useful default most of the time. When you do not want it, give the layer its own aes() and it will use that one instead.

ggplot(
  oceanbuoys,
  aes(
    x = air_temp_c,
    y = sea_temp_c
  )
) +
  geom_point(aes(colour = factor(year)), alpha = 0.4) +
  geom_smooth(method = "lm", colour = "black")

Figure 1.12: Overriding the inherited mapping for one layer only.

Now the colour mapping lives on the points, so the smooth doesn’t see it, and you get one line across everything.

Those two plots say quite different things, and the only difference is where the aes() sits. Worth remembering when a smooth does something you didn’t expect.

NoteYour Turn

1. Add geom_smooth() to your humidity plot.
2. Swap the order of geom_point() and geom_smooth(). Can you see the difference? Look closely at where the lines and points overlap.
3. Try geom_smooth(se = FALSE). What went away?

1.11 When it doesn’t work

You are going to hit errors, and most of them in ggplot2 are one of a small number of things. This is the list I actually run through.

Count your brackets and quotes. aes(x = hour, y = mean_count will not run, and neither will colour = "purple. Note that ' and " are different characters, and they have to match each other.

Look for a + in the console. If R is showing you a + instead of a >, it is still waiting for you to finish something. Press ESCAPE and start again.

Check where your + signs are. In ggplot2 the + goes at the end of a line, never the start. This works:

ggplot(oceanbuoys, aes(x = air_temp_c, y = sea_temp_c)) +
  geom_point()

and this quietly does not:

ggplot(oceanbuoys, aes(x = air_temp_c, y = sea_temp_c))
  + geom_point()

The second one runs the ggplot() call on its own, draws a blank panel, and then hits the + as a separate statement:

Error: Cannot use `+` with a single argument.

Which is true, and gives you no clue that the actual problem is a line break.

Copy the error message and search it. All of it, not your description of it. Somebody has had it before.

TipRead more

The ggplot2 cheatsheet is two pages of every geom, scale and theme, laid out visually. It is genuinely the fastest way to find the geom you half remember.

There is also the ggplot2 extension gallery for when base ggplot2 does not have the thing you want.

1.12 To summarise

Three things to take out of this chapter.

1. A ggplot is data, aesthetics, and a geom, glued together with +. Everything else is detail hanging off those three.
2. Inside aes() maps a variable, outside aes() sets a value. An unexpected legend means you mapped something by accident.
3. The channels are not equal. Position first, then colour, then shape, and size only when you have run out of better ideas.

And one habit, which is the one I actually care about.

When a plot surprises you, do not reach for a different example to copy. Ask which of the seven pieces is doing something you did not ask it to do. Nearly every ggplot2 problem I have ever had turned out to be one piece behaving exactly as documented, in a way I had not thought about.

Next up is the shape of your data, which is where most of the remaining surprises come from.

Links

• Nicola Rennie
• The art of data visualisation with ggplot2
• Leland Wilkinson
• ggplot2
• Hadley Wickham
• ggplot2 book
• Hadley Wickham’s PhD thesis
• Di Cook’s JSM materials
• {visdat}
• {naniar}
• ggplot2 cheatsheet
• ggplot2 extension gallery