16  Advanced ggplot2

16.1 Prerequisites

Answer the following questions to see if you can bypass this chapter. You can find the answers at the end of the chapter in Section 16.18.

  1. When choosing a colour scale for a visualisation, what property of the data should guide your choice among sequential, diverging, and qualitative scales?
  2. Name two appropriate ways to display uncertainty in a visualisation.
  3. When designing multi-panel figures (for example, with facet_wrap() or patchwork), what is the key consideration that supports comparison across panels?

16.2 Learning objectives

By the end of this chapter you should be able to:

  • Compose multi-panel figures with patchwork and cowplot, with shared legends and panel labels.
  • Annotate plots with ggrepel, geom_text, mathematical expressions, and annotate().
  • Build custom scales, themes, and (briefly) custom geoms.
  • Produce publication-quality figures with consistent fonts, sizes, and export formats appropriate to the destination (paper, slide deck, web).
  • Animate a plot with gganimate for teaching and exploratory uses.
  • Display uncertainty with error bars, confidence bands, density plots, ribbon plots, and bootstrap distributions.

16.3 Orientation

Once the grammar of graphics is in hand, the remaining work in making a finished figure is largely aesthetic: composition, typography, colour, and layout. ggplot2 and its ecosystem give you fine-grained control over each. This chapter covers the tools used in every real paper or report, the ones that turn ‘a plot that conveys the analysis’ into ‘a plot a journal will accept’.

The pieces fall into four groups: composition (combining multiple plots), annotation (adding labels and highlights), styling (themes and colour palettes), and output (exporting at the right resolution and format).

16.4 The statistician’s contribution

ggplot2 defaults are reasonable for exploration; they are not the right choice for a published figure. The adjustments needed to make a figure publication-ready are small in code but consequential in clarity.

Match the polish to the audience. A figure for an internal slide deck does not need 600-DPI vector output. A figure for a clinical journal does. Spending an hour fine-tuning a figure for an audience of three is misuse of time; not spending the hour on a figure for a peer-reviewed paper is misuse of opportunity.

Display uncertainty. A point estimate without a confidence interval invites the reader to over-trust the estimate. A regression line without a confidence band is a similar invitation. The remedy is small: geom_errorbar, geom_ribbon, the se = TRUE argument to geom_smooth. Refusing to show uncertainty because the band looks ‘noisy’ is wrong; if the data are noisy, the figure should show that.

Compose deliberately. Multi-panel figures should make comparisons easy. Shared axes, shared legends, consistent scales, panel labels in the same position. Inconsistent panels create work for the reader; that work is what your figure should be doing for them.

Choose typography that respects the medium. Default ggplot2 text is fine on a slide; on a printed page it often looks small and grey. Customising the theme is a ten-line investment that pays off across an entire paper.

These judgements determine whether figures get accepted on the first review or send the manuscript back for revisions.

16.5 Composing multi-panel figures with patchwork

patchwork provides an arithmetic-style operator interface for combining ggplots:

library(patchwork)
library(ggplot2)
library(palmerpenguins)

p1 <- ggplot(penguins, aes(flipper_length_mm, body_mass_g, colour = species)) +
        geom_point() +
        labs(title = "A. Body mass vs. flipper length")

p2 <- ggplot(penguins, aes(species, body_mass_g, fill = species)) +
        geom_boxplot() +
        labs(title = "B. Body mass distribution")

p3 <- ggplot(penguins, aes(bill_length_mm, bill_depth_mm, colour = species)) +
        geom_point() +
        labs(title = "C. Bill morphology")

# horizontal: side by side
p1 + p2

# vertical: stacked
p1 / p2

# 2x2 grid with shared legend
(p1 + p2) / (p3 + p2) +
  plot_layout(guides = "collect") &
  theme(legend.position = "bottom")

Operators:

  • + places plots side by side.
  • / stacks plots vertically.
  • | is equivalent to + (horizontal).
  • Parentheses group plots into sub-arrangements.
  • & applies a theme or scale to all plots.
  • plot_layout(guides = "collect") collects shared legends into one location.

For more control, cowplot::plot_grid() provides similar functionality with finer-grained alignment options. For inset plots (a small panel inside a larger one), use patchwork::inset_element().

Panel labels (A, B, C) are best added in the labs() call of each individual plot, prepended to the title, rather than via patchwork’s plot_annotation(tag_levels = "A"). The latter places labels in the corners but does not guarantee they appear in the order you want for a non-rectangular layout.

16.6 Annotation: text, arrows, and highlights

library(ggrepel)

# label specific points (one per row)
ggplot(mtcars, aes(wt, mpg, label = rownames(mtcars))) +
  geom_point() +
  geom_text_repel(size = 3, max.overlaps = 10)

# annotate a single text string at fixed coordinates
ggplot(penguins, aes(flipper_length_mm, body_mass_g)) +
  geom_point() +
  annotate("text", x = 200, y = 6000, label = "Larger species",
           hjust = 0)

# math expressions
ggplot(data.frame(x = 1:10, y = (1:10)^2), aes(x, y)) +
  geom_line() +
  labs(title = expression(paste("Quadratic: ", y == x^2)),
       y = expression(y ~ "(units)"))

# arrows pointing at features
ggplot(penguins, aes(flipper_length_mm, body_mass_g, colour = species)) +
  geom_point() +
  annotate("segment", x = 175, xend = 180, y = 5500, yend = 5000,
           arrow = arrow(length = unit(0.3, "cm"))) +
  annotate("text", x = 175, y = 5600, label = "Outlier",
           hjust = 0)

ggrepel produces non-overlapping text labels by optimising their positions. max.overlaps controls how many labels are allowed to overlap before some are dropped. For dense plots, label only a few key points.

16.7 Custom themes

A custom theme keeps figures consistent across a paper or a package. Build it once, apply everywhere.

theme_phb228 <- function(base_size = 11) {
  theme_minimal(base_size = base_size) +
    theme(
      plot.title       = element_text(face = "bold", size = base_size + 2),
      plot.subtitle    = element_text(colour = "grey40"),
      axis.title       = element_text(face = "bold"),
      axis.text        = element_text(colour = "grey20"),
      panel.grid.minor = element_blank(),
      panel.grid.major.x = element_line(linewidth = 0.2, colour = "grey90"),
      panel.grid.major.y = element_line(linewidth = 0.2, colour = "grey90"),
      legend.position  = "bottom",
      strip.background = element_rect(fill = "grey95", colour = NA),
      strip.text       = element_text(face = "bold")
    )
}

ggplot(penguins, aes(flipper_length_mm, body_mass_g, colour = species)) +
  geom_point() +
  scale_colour_brewer(palette = "Dark2") +
  theme_phb228()

Set defaults at the top of an analysis script:

theme_set(theme_phb228())

update_geom_defaults() changes default geom parameters (default geom_point size, default geom_line width):

update_geom_defaults("point", list(size = 1.5, alpha = 0.7))
update_geom_defaults("line",  list(linewidth = 0.7))

For a project’s typography, declare a font family and use it consistently:

library(showtext)
font_add_google("Source Sans 3", "source")
showtext_auto()

theme_set(theme_phb228() + theme(text = element_text(family = "source")))

showtext makes Google Fonts available to ggplot2 and ensures they render correctly when exporting.

16.8 Custom colour palettes

For project-wide colour consistency, define your palette once:

phb228_palette <- c(
  "Adelie"    = "#1f4e79",
  "Chinstrap" = "#9d2235",
  "Gentoo"    = "#2e8b57"
)

scale_colour_phb228 <- function(...)
  scale_colour_manual(values = phb228_palette, ...)
scale_fill_phb228 <- function(...)
  scale_fill_manual(values = phb228_palette, ...)

ggplot(penguins, aes(flipper_length_mm, body_mass_g, colour = species)) +
  geom_point() +
  scale_colour_phb228()

For sequential or diverging continuous palettes, viridis::scale_colour_viridis_c() and scico::scale_colour_scico(palette = "vik") are perceptually uniform and colour-blind-safe.

Question. You are visualising a heatmap of correlations between predictors. Correlations range from -0.9 to +0.9. Which colour scale is most appropriate?

Answer.

A diverging palette centred on zero. Correlations have a meaningful midpoint (zero correlation), and positive vs. negative correlations should be visually distinguishable. Standard choices: RdBu (red-to-blue, ColorBrewer), PuOr (purple-orange), BrBG (brown-bluegreen). Set the midpoint of the scale to zero (midpoint = 0 in scale_fill_gradient2) and the limits symmetric around it. A sequential palette would map zero to a particular colour without visual reflection of the sign change. A qualitative palette would imply discrete categories, losing the continuous correlation magnitude.

16.9 Displaying uncertainty

Multiple geoms exist for showing uncertainty alongside an estimate.

# error bars on a categorical x
ggplot(summarised, aes(group, mean)) +
  geom_point() +
  geom_errorbar(aes(ymin = mean - 2 * se, ymax = mean + 2 * se),
                width = 0.2)

# confidence band on a regression line
ggplot(d, aes(x, y)) +
  geom_point() +
  geom_smooth(method = "lm", se = TRUE)

# ribbon for a manual uncertainty range
ggplot(d, aes(x)) +
  geom_ribbon(aes(ymin = lower, ymax = upper), alpha = 0.3) +
  geom_line(aes(y = mean))

# half-eye plot from posterior or bootstrap samples
library(ggdist)
ggplot(samples, aes(x = group, y = posterior)) +
  stat_halfeye()

ggdist is particularly worth knowing for Bayesian work or any analysis with full uncertainty distributions: stat_halfeye, stat_dotsinterval, stat_lineribbon make rich uncertainty representations one line each.

For point estimates with CIs in a forest plot:

forest_data |>
  ggplot(aes(estimate, term)) +
    geom_pointrange(aes(xmin = conf.low, xmax = conf.high)) +
    geom_vline(xintercept = 0, linetype = "dashed") +
    labs(x = "Effect estimate (95% CI)", y = NULL)

For posterior distributions from MCMC, bayesplot::mcmc_areas produces a similar interval plot directly from posterior samples.

16.10 Exporting for publication

# raster: PNG at high DPI for slides, web
ggsave("figure.png", plot = p, width = 6, height = 4,
       dpi = 300, units = "in")

# vector: PDF for LaTeX submissions
ggsave("figure.pdf", plot = p, width = 6, height = 4,
       device = cairo_pdf)

# vector: SVG for the web
ggsave("figure.svg", plot = p, width = 6, height = 4)

# specific journal sizing (single column, double column)
ggsave("figure_singlecol.pdf", plot = p,
       width = 90 / 25.4, height = 60 / 25.4, units = "in",
       device = cairo_pdf)

Three things to know about export:

  1. Use vector formats (PDF, SVG, EPS) for line art and typography. Raster formats (PNG, JPG) lose quality on zoom. JPG additionally adds compression artefacts; never use JPG for plots.
  2. DPI matters for raster. 300 DPI is the standard for print; 600 for high-quality scientific figures. 72 DPI is web display only.
  3. Embed fonts in PDFs. device = cairo_pdf ensures non-default fonts are embedded so the figure looks the same on systems that lack the font. Without this, the recipient may see a fallback font that looks wrong.

For journal submissions, check the figure size requirements (typically single-column 85–90 mm, double-column 170–180 mm) and produce figures at exactly the target size, not larger. Plots designed at 6×4 inches and submitted at 3 inches wide have illegible labels.

16.11 Animation with gganimate

library(gganimate)

# bootstrap convergence: posterior of mean as n grows
boot_data <- expand_grid(rep = 1:100, n = c(10, 50, 100, 500)) |>
  mutate(mean_est = map2_dbl(rep, n, \(r, k) mean(rnorm(k))))

p <- ggplot(boot_data, aes(mean_est)) +
  geom_histogram(bins = 30) +
  labs(title = "Sampling distribution of the mean, n = {closest_state}",
       x = "Sample mean") +
  transition_states(n, transition_length = 2, state_length = 1)

animate(p, nframes = 100, fps = 10)
anim_save("convergence.gif")

gganimate supports several transitions: transition_states (discrete steps), transition_time (continuous), transition_reveal (incrementally reveal a line), and others. For teaching and exploratory work, animation is often the clearest way to show how a distribution evolves with sample size, iterations, or parameter changes.

For papers, animation is rarely useful: figures are static. For slide decks, blog posts, and supplementary materials, animation can convey what static figures cannot.

16.12 Worked example: regression diagnostics in three panels

library(ggplot2)
library(patchwork)
library(broom)

fit <- lm(body_mass_g ~ flipper_length_mm + species, data = na.omit(penguins))
diag_data <- augment(fit)

p_resid <- ggplot(diag_data, aes(.fitted, .resid)) +
  geom_point(alpha = 0.5) +
  geom_smooth(method = "loess", se = FALSE, colour = "red") +
  labs(title = "A. Residuals vs. fitted",
       x = "Fitted values", y = "Residuals")

p_qq <- ggplot(diag_data, aes(sample = .std.resid)) +
  geom_qq(alpha = 0.5) +
  geom_qq_line(colour = "red") +
  labs(title = "B. Normal Q-Q",
       x = "Theoretical quantiles",
       y = "Standardised residuals")

p_cook <- ggplot(diag_data, aes(seq_len(nrow(diag_data)), .cooksd)) +
  geom_col(width = 0.5) +
  geom_hline(yintercept = 4 / nrow(diag_data),
             linetype = "dashed", colour = "red") +
  labs(title = "C. Cook's distance",
       x = "Observation index",
       y = "Cook's distance")

(p_resid + p_qq) / p_cook + plot_layout(heights = c(1, 0.7))

This composition reads naturally: top row two related diagnostics (residual structure and Q-Q), bottom row a single observation-level diagnostic. Panel labels carry the reader through.

16.13 Collaborating with an LLM on advanced ggplot

LLMs handle ggplot composition reasonably well; they handle typography and journal conventions less reliably.

Prompt 1: matching a journal style. Paste the journal’s figure guidelines (or the URL) and ask: ‘write a theme_journal() function that matches.’

What to watch for. The output theme is a starting point. Specific journal requirements (font family, font size at print, panel border vs. axis lines) often need to be verified by hand against the actual published figures.

Verification. Generate a sample figure in your theme and a sample figure from a recent published paper. Compare side by side. Adjust until indistinguishable.

Prompt 2: combining plots. Describe four plots and ask: ‘combine these into a 2x2 grid with shared legend and panel labels A, B, C, D.’

What to watch for. Most LLM solutions use patchwork, which is correct. Watch for legend handling: the LLM may forget plot_layout(guides = "collect") or place legends outside the plot area in unexpected ways.

Verification. Render the combined plot at the target size. Are the legends in a single shared location? Are the axes consistent? Are the panel labels in the right position?

Prompt 3: animation. Describe what you want to animate (e.g., ‘how does the bootstrap distribution converge as \(n\) grows?’) and ask the LLM to produce gganimate code.

What to watch for. The animation may run too fast (no time to read each frame) or too slow (boring). Adjust nframes and fps.

Verification. Watch the animation. Does it tell the story you wanted? If the message gets lost in the movement, an animated plot is not the right medium.

16.14 Principle in use

Three habits define defensible advanced visualisation:

  1. Customise once, apply everywhere. A theme_phb228() function and a project palette make every figure consistent for free.
  2. Show uncertainty. Confidence bands, error bars, posterior intervals, whatever the analysis produces. Never report point estimates without their uncertainty.
  3. Export for the destination. PDF for LaTeX, PNG for slides, SVG for web. Embed fonts. Match journal sizing.

16.15 Exercises

  1. Build a three-panel figure: (a) raw data scatter; (b) residuals-vs-fitted; (c) QQ plot. Combine with patchwork and add panel labels ‘A’, ‘B’, ‘C’.
  2. Write a custom theme_phb228() function with serif body text, sans-serif axis titles, and a colour-blind- safe default palette. Apply it to three plots.
  3. Export a figure at 300 DPI as both PDF (for LaTeX) and PNG (for Word). Open both and verify the fonts are embedded correctly.
  4. Make a forest plot of effect estimates with 95% CIs for ten coefficients from a regression. Use geom_pointrange and a vertical reference line at zero.
  5. Animate a sampling distribution converging as \(n\) grows. Use gganimate::transition_states. Save as GIF and verify the animation tells the story.

16.16 Further reading

  • (Wickham, 2016), chapters on scales, themes, and extending ggplot2. The 3rd edition is online at ggplot2-book.org.
  • (Wilke, 2019), the source of many of the design principles this chapter invokes.
  • (Healy, 2018), shorter, applied, with R code.
  • The gganimate and patchwork package vignettes are excellent and concise.

16.17 Practice test

The following multiple-choice questions exercise the chapter’s content. Attempt each question before expanding the answer.

16.17.1 Question 1

Which of the following is MOST important when choosing a colour scale for a data visualisation?

    1. Using the widest possible range of different colours
    1. Matching the colour scale to the type of data being visualised (sequential, diverging, or qualitative)
    1. Always using the same colour scheme across all visualisations in a project
    1. Prioritising aesthetically pleasing colour combinations over all other considerations

B. Sequential, diverging, and qualitative scales each encode a different kind of variable structure; mismatching distorts interpretation.

16.17.2 Question 2

Which approach is recommended when displaying uncertainty in your data?

    1. Omit uncertainty information to avoid confusing the audience
    1. Only show the mean or median values as single points
    1. Always show exact numerical values for uncertainty in a caption rather than visualising it
    1. Use visual elements like error bars, confidence bands, or density plots to represent uncertainty

D. Uncertainty should be visualised alongside the estimate rather than omitted or relegated to captions.

16.17.3 Question 3

When designing multi-panel figures, which is the most important design consideration?

    1. Always arrange panels in a perfect grid with equal dimensions regardless of the data
    1. Use as many panels as possible to show every possible data combination
    1. Maintain consistent scale and layout across panels to facilitate comparisons
    1. Avoid panels altogether and instead create a single complex figure

C. Consistent scales and layouts mean that visual differences across panels reflect data differences, not display differences.

16.17.4 Question 4

You want to combine four ggplots into a 2x2 grid with a shared legend. Which approach uses the canonical R tool?

    1. gridExtra::grid.arrange(plot1, plot2, ...) then manually add the legend.
    1. patchwork::wrap_plots(p1, p2, p3, p4) + plot_layout(guides = "collect").
    1. Save each plot, open in Photoshop, manually combine.
    1. cowplot::plot_grid() with default arguments.

B. patchwork is the canonical modern composition tool; guides = "collect" is what shares the legend. cowplot::plot_grid works similarly but with a different syntax.

16.17.5 Question 5

For a journal submission, you should export figures as:

    1. JPG at 72 DPI.
    1. PNG at 96 DPI.
    1. PDF (vector) with embedded fonts, sized to the journal’s column width.
    1. PowerPoint (.pptx) with editable layers.

C. Vector PDFs scale to any size without quality loss. Embedding fonts (device = cairo_pdf) ensures the figure renders correctly on any system. Sizing at the column width avoids the journal scaling your figure down and making labels illegible.

16.18 Prerequisites answers

  1. Match the colour scale to the type of data: sequential for ordered or continuous data, diverging for data with a meaningful midpoint, qualitative for unordered categorical data. Using the wrong family distorts perception. Sequential and diverging scales should be perceptually uniform (viridis, ColorBrewer) and colour-blind-safe.
  2. Error bars, confidence bands, density plots, violin plots, point clouds, ribbons. Do not omit uncertainty information or replace it with a numerical value in a caption. The figure should encode estimate and uncertainty together.
  3. Maintain consistent axis scales and panel layout so that apparent differences across panels reflect real differences in the data, not differences in the display. Shared scales (facet_* or patchwork::plot_layout(axis_titles = "collect")) are the standard mechanism.