Data Visualization

Behavioral Data Science Toolbox 2026

Welcome to BDS Toolbox: Data Visualization

About these slides

Slides made using quarto

Cover photo made in R by Danielle Navarro

• Quick housekeeping: these slides are built with Quarto, from the same repository as the course site.
• The cover art is generative art made in R by Danielle Navarro, a nice reminder that R and ggplot can go far beyond default charts.

📌 Course website

data-visualization-2026

This is our central hub! Access slides, textbook, worksheets, assignments and other resources here.

💡 Tip: press S while viewing the slides to open the speaker notes.

Schedule

When	Topic	Preparation / Worksheets / Assignments
Session 1. Tue 29/9, 15.00–18.00 GS.11	Fundamentals of DV	Read: DV ch. 1; ch. 3 Worksheet: ggplot recap
Session 2. Fri 2/10, 12.00–15.00 HvA	Exploratory viz	Worksheet: ggplotly Worksheet: working with AI, pt. 1 🚨 Sun 4/10, 23.59 Deadline Assignment 1 (individual) + Assignment 2, Part 1 (group)
Session 3. Tue 6/10, 15.00–18.00 GS.11	Explanatory viz	Worksheet: working with AI, pt. 2 Read: DV ch. 8
Session 4. Fri 9/10, 12.00–15.00 HvA	Pitches & Q&A	🚨 Sun 11/10, 23.59 Deadline Assignment 2, Part 2 (group)

Friday sessions: time to work on group project!
Bring a (charged) laptop! 💻

Session 1 Tuesday, Sept. 29 Fundamentals of Data Visualization

Today

• What is Data Visualization?
  
• Why visualize data?
• What makes a plot good (and bad)?
• Guiding principles
• Data Visualization Project

• Roadmap for today: what data visualization is, why we visualize at all, what separates good from bad plots, a set of guiding principles, and finally an introduction to the assignments.
• This session is more conceptual; the hands-on ggplot work comes with the worksheet.

Data visualization = Data thinking

There are many skills involved in producing a good data visualization:

• Technical (how to use particular tools)
• Design (how to make visual representations that are engaging and appeal aesthetically)
• Perception (how do certain notions of visual perception and cognition affect data communication)

• Making a good visualization draws on several different skill sets.
• Technical: knowing your tools, e.g. ggplot. (covered in the worksheets, and a skill you’ll continue to build up over time)
• Design: making it engaging and aesthetically pleasing. (next tuesday’s session)
• Perception: knowing how people actually read visual information (we discuss this a little bit, but more information is in your reading).

Data visualization = Data thinking

Most important is to be able to derive reliable information from data using visual representations.

• But underneath all of these, the most important skill is deriving reliable information from data.
• Walk through the diagram: the world → data → visualization → someone’s understanding.
• Every arrow is a step where information can be lost or distorted: what gets measured, how it’s processed, and how it’s shown and read.

Data visualization = Data thinking

Your data visualization is effective only if it helps you, and your audience, understand the real world better

This requires data thinking skills:

• Understanding of what the data represents (and what it doesn’t)
• Transparency about the transformations and aggregations done to the data to produce the visualization
• Awareness about how different visual representations affect the interpretation of the data, and an ability to justify which choice is most appropriate for your data and audience
• Knowledge about the context from which the data is taken, and later presented.

• The goal is not a pretty plot; the goal is a better understanding of the real world, for you and your audience. Knowledge about data validty and technical choices is saved for other parts of the course. But you need to think about all of these steps when visualizing.

Data visualization = Data thinking

Especially in our age of AI, these skills are invaluable

• AI tools can now generate polished dashboards and charts from a text prompt in seconds.
• That makes producing a visualization easy, but doesnt judge whether it is correct, honest, or appropriate.
• The data thinking skills from the previous slide are exactly what these tools can’t do for you.
• We will come back to working with AI in sessions 2 and 3.

Why visualize data? 👀

• Move to the first big question: why visualize data at all?

💬 Discussion

Why visualize data?

• Ask the room: why visualize data? Collect a few answers.
• Expected answers: see patterns quickly, communicate results, spot errors and outliers, check assumptions.
• Keep the answers in mind; the next slides make the case for “looking at your data”.

Communicating with the public

…or boss, or stakeholder, or client

• Before the technical reasons to visualize, the bigger picture: visualizations are one of the main ways data reaches the public.
• During COVID-19, maps and case curves were on every news site, updated daily. Millions of people who never read a statistics report looked at these charts to decide how worried to be, whether to travel, or whether policies were justified. “Flatten the curve” was itself a visualization.
• Point to the footnotes on both charts: the data come from Johns Hopkins, collated from sources with different methods, and testing rates differ between countries. Good public-facing charts are transparent about what the data can and can’t tell you (link back to the data thinking skills).
• What responsibility comes with making a chart that millions of people will see?

Why visualize data?

• Left: income inequality vs. voter turnout. The negative slope is driven almost entirely by one country, South Africa. Without it, the relationship almost disappears.
• A regression coefficient alone would never tell you that; one look at the plot does.
• (Fragment) Right: sixteen datasets that all produce a similar positive regression line, with very different underlying structures: skew, outliers, non-linear trends, groups, categorical data.
• Message: the same summary statistic can hide very different data. Always look at your data.
• Both figures from Healy, Data Visualization (socviz.co), ch. 1.

Anscombe’s quartet

library("datasauRus")
library(scales)

datasaurus_dozen |>
  ggplot2::ggplot(aes(x = x, y = y, color = dataset)) +
  ggplot2::geom_point() +
  ggplot2::theme_void() +
  ggplot2::geom_smooth(
    method = "lm",
    color = "gray",
    fill = "gray",
    alpha = .5
  ) +
  ggplot2::theme(legend.position = "none", text = element_text(size = 30)) +
  ggplot2::facet_wrap(~dataset, ncol = 4)

• datasets can have with (nearly) identical means, variances, correlations, and regression lines, but they look completely different.

💬 Discussion

What is more important? An eye or an algorithm?

What are the consequences of overrelying on statistical techniques?
What are the consequences of overrelying on visualizations?

• Discussion: what is more important, an eye or an algorithm?
• Over-relying on statistics: you miss outliers, non-linearity, data errors (Anscombe, Datasaurus).
• Over-relying on visualizations: you see patterns in noise, and plots can be designed to mislead.
• You need both; visualization and statistics verify each other.

Bad plots 💩

What makes a bad plot bad?

• Aesthetic (ugly)

• Perceptual (bad)

• Substantive (wrong)

Claus Wilke - Ugly, bad, and wrong figures

• Claus Wilke’s three categories of problems:
  • Ugly: aesthetic problems, but data patterns might still be sound / informative (e.g. illegible colors)
  • Bad: perceptual problems; unclear, confusing, or deceiving (patterns do not lead to correct interpretation).
  • Wrong: objectively, mathematically incorrect (e.g. bars without a meaningful scale).
• We’ll use these categories to critique plots in the next exercise.

💬 Speed dates

• What do you think about this plot?

• What elements can be improved?

• Are the problems aesthetic, perceptual, or substantive?

Ireland is headed straight for zero? 📉

• Speed dates: discuss with your neighbour for 2–3 minutes.
• (Fragment) After the group discussion, click to reveal the pop-up: “Ireland is headed straight for zero?” The line plunges towards 0, but Paris 2024 was actually Ireland’s best result.
• Plot: Ireland’s position in the Olympic medal table. Lower rank number = better, but on this axis the line goes down as Ireland improves, so 14th in Paris looks like a crash.
• Also: does the axis need to include 0? There is no rank 0.
• Classify: mostly a perceptual (bad) problem; you could argue substantive, since it invites the wrong conclusion.
• Fix: reverse the y-axis so up means better, use a barplot or something that better represents discrete values.

💬 Speed dates

reddit.com/r/dataisugly

• Misleading (or missing) information?
• Many data visualizations might have a hidden agenda, due to e.g. marketing strategies.

• Lidl advertisement comparing the price of a basket of groceries at UK supermarkets.
• (Fragment) Missing baseline and scale: the bars don’t start at zero, so a ~£9 difference looks huge; We don’t know what’s on the axis, when/how it was measured
• Many visualizations have an agenda; always ask who made it and why.
• Footer: Huff’s How to Lie with Statistics is a classic on this.

💬 Speed dates

Source: reddit.com/r/dataisbeautiful

• Data to ink ratio?

• School shooting incidents by country, 2009–2018.
• Ask: what do you notice about the design?
• (Fragment) Data-to-ink ratio: the proportion of ink used to show data versus everything else.
• Here, most of the chart is spent on bars for countries with 1 incident; the United States dominates, which is the story, but a lot of ink carries little information.
• Also worth discussing: counts not normalized by population or number of schools.

💬 Speed dates

Data to ink ratio

• Compare the two charts side by side.
• Right: US household gun ownership 1972–2023, one bar per year with a label on each. Lots of ink, and the trend (roughly flat around 40%) is hard to see.
• A simple line chart would show the same information with far less ink.
• Tufte’s principle: maximize the share of ink devoted to data; remove what doesn’t add information.

💬 Speed dates

• Treemap of US COVID-19 relief spending by category (NYT).
• Ask: can you compare the sizes? Is Businesses bigger or smaller than Individuals and families?
• Area is hard to compare precisely, especially across differently shaped rectangles. Treemaps are good for showing hierarchy and part-to-whole, not for precise comparison.
• What perceptual element is working to tell the story, and allow an easy conclusion to be drawn?

💬 Speed dates

• Do the colors make sense?

• Safety index by country, 2023.
• Ask: do the colours make sense?
• (Fragment) The index is an ordered, continuous variable, but the colours are categorical hues in no logical order: red, blue, yellow, purple, green. You can’t tell at a glance whether a country is safer or less safe.
• A sequential palette (light → dark) would map naturally to the ordering.

💬 Speed dates

Do the colors make sense?

• Another colour example: bar chart of total bill by day.
• Each bar has its own colour, but colour here encodes nothing that the x-axis doesn’t already show.
• Colour should map to a relevant attribute; otherwise it adds noise (and a legend people have to decode).

💬 Speed dates

• Are all elements needed?

• Chart might not be wrong, but a person should be able to understand the conclusion within seconds,

💬 Speed dates

• Are the axes correct?

• Years on axes barely noticable, unclear why these teams are on the bars
• Also: the Cincinnati bar (54 points, 9 games left) is compared to completed seasons; the chart shows a record chase that hasn’t happened yet.
• What is the evidence?
• Decoration competes with the data.

💬 Speed dates

Is a zero-point needed?

💬 Speed dates

• Is it simple?

• Stacked bar chart of values by type, 2005–2016.
• Ask: is it simple? Can you tell how type B changed over time?
• (Fragment) Only the bottom category (D) and the total share a common baseline. Every other segment floats, so comparing them across years is very hard.

💬 Speed dates

⚠️ Stacked bar charts ⚠️

• Left: market shares of five companies over three years. Can you see that company A is growing and E is shrinking? Only because they’re at the edges.
• Right: Brazil’s top Olympic medallists. The stacks mix gold, silver, and bronze, with photos and rotated labels on top; comparing e.g. silver medals across athletes is almost impossible.
• Stacked bars work for showing totals, or with very few categories. Otherwise, consider side-by-side bars or small multiples.

💬 Speed dates

⚠️ Pie charts ⚠️

• CBS News: Americans who have tried marijuana.
• A pie chart shows parts of a whole, but 51%, 43%, and 34% add up to 128%. These are three separate measurements from different years, not parts of a whole.
• This is a wrong figure in Wilke’s terms.
• A simple bar or line chart over time would be correct.

💬 Speed dates

Visualizing proportions

• Pie chart, stacked bar, or side-by-side bars?

https://clauswilke.com/dataviz/visualizing-proportions.html#

• Pie charts are not always wrong. Wilke’s table compares pie charts, stacked bars, and side-by-side bars.
• Pie charts: good for simple fractions (1/2, 1/4) and few categories.
• Stacked bars: good for many sets of proportions, e.g. over time.
• Side-by-side bars: best for comparing the relative sizes directly.
• Choose based on what the reader needs to compare.

💬 Speed dates

Elementary Perceptual Tasks

• Cleveland & McGill (1984): ranking of elementary perceptual tasks by how accurately people can read them.
• Most accurate: position along a common scale (dot plots, bar charts, scatter plots). Then position on non-aligned scales, length, direction/angle, area, volume, and least accurate: shading and colour hue.
• This explains many of the problems we saw: pie charts use angle, treemaps use area, the map uses hue.

💬 Speed dates

• Does it support one conclusion?

• Warehouse accuracy rates, stacked bars for 19 warehouses.
• Ask: does this chart support one conclusion? What should the reader take away?
• (Fragment) It shows everything with equal weight: red accuracy bars dominate, and the interesting part (error rates) is squashed at the bottom.

💬 Speed dates

Storytelling with data - show just one conclusion!

Guiding principles 🪄

• Let’s summarize everything we’ve seen into a checklist.

Tips for the best viz

• Is it explaining data?

• Is the information complete and correct?

• Are axes correct? (+ Should they have a zero-point?)

• Do the colors work? ( + Do they map to a relevant attribute?)

• Are all elements needed?

• What is the data to ink ratio?

• Is it understandable & simple?

• Does it portray one conclusion?

• Use this as a checklist for your own plots and for the assignments.
• Each point maps to an example we saw: axes (Ireland), colours (safety map), elements (Cincinnati), data-to-ink (gun ownership), simplicity (stacked bars), one conclusion (warehouses).
• “Explaining data”: is this an explanatory plot with a clear message?
• “Complete and correct”: the pie chart that didn’t add up to 100%.

Some good examples

Distributions are informative

NYT Graphic - Obamacare spending

• Some examples of good visualizations.
• Butterfat by cow breed (Wilke): overlapping density plots with direct labels instead of a legend; distributions are more informative than just means.
• NYT Medicare spending: projected trend vs. actual spending, with the gap shaded and annotated. One clear message.
• NYT smoking rates: a simple line with a declarative title that states the conclusion.
• Common thread: clear message, direct labels, minimal clutter.

Making good plots: the BBC way

The BBC data journalism team makes its published graphics in R with ggplot, using their own package, bbplot.

➡️ BBC Visual and Data Journalism cookbook for R graphics

# devtools::install_github("bbc/bbplot")
library(bbplot)

my_plot <- ggplot(data, aes(x = year, y = value)) +
  geom_line(colour = "#1380A1", linewidth = 1) +
  bbc_style() # BBC house style

finalise_plot(
  plot_name = my_plot,
  source_name = "Source: ...",
  save_filepath = "my_plot.png"
)

NYT: Medicare spending

➡️ How to reproduce a NYT graphic in ggplot

• A real-world example of the principles we just listed: newsrooms that publish charts to millions of people every day, and make them in R.
• BBC: bbplot gives every chart the same house style. bbc_style() is a ggplot theme, and finalise_plot() adds the source and logo and saves the chart at the right size.
• Open the cookbook and scroll through the examples: line charts, bar charts, small multiples, annotations. Point out the pattern: a declarative title and subtitle, few gridlines, direct labels instead of legends, one clear message.
• NYT (right): Medicare spending per beneficiary, projected trend vs. actual spending, with the gap shaded and annotated. The tutorial shows how to rebuild a NYT-style graphic step by step in ggplot.
• Takeaway for the assignments: a consistent theme and a few deliberate choices go a long way. Students can borrow ideas (or code) from both for their own plots.

Making good plots: the BBC way

Break

library("RXKCD")
RXKCD::getXKCD(which = "833")

• Break time.
• XKCD 833 “Convincing”: a reminder to label your axes!

Data Visualization Project 📊

Assignment 📝

Assignment 1

Session 2 Friday, Oct. 2 Exploratory Data Visualization

Today

• Exploratory data visualization
• Brief introduction to interactive plots
• Worksheet: ggplotly
• Using AI for exploratory data viz
• Worksheet: Working with AI, pt. 1
• Work on your group assignment

💬 What makes exploration useful?

Illustration by Allison Horst

Exploratory versus explanatory

Exploratory in the sense of data exploration. Exploratory in the sense of visualization exploration (as in previous example, where the same data can be plotted in several ways). In both cases, you can create ugly plots.

assignment 2 = explanatory data viz (also bridge to next). Here we want beautiful plots!

Source: Natalia Kiseleva

Source: Scott McCloud

• Exploratory: examine the structure of your data.
• Explanatory: tell a story with your data.

Exploratory versus explanatory

Exploratory

Explanatory

Source: storytellingwithdata.com

• Storytelling with data’s metaphor: exploratory analysis is like opening 100 oysters to find a couple of pearls.
• Explanatory: show your audience the pearls, not all the oysters.
• For your assignments: you’ll do both. Exploratory plots to understand the data, and one explanatory plot that makes a clear point.

Using data visualizations for exploration

1. Understand your problem.
2. Look at your data: head()
3. What variables are available? What do they mean, what values do they take on?
4. Identify your target variables.
   
5. Now plot: their distributions, relationships with other variables, explore grouping structures, etc.
6. What patterns do you see? Write them down, save the code.

Using data visualizations for exploration

Exploratory visualizations can be ugly

Use them to guide your understanding of the data, and inspire the final result you wish to present.

One variable → its distribution

Two variables + a group → relationships

Using data visualizations for exploration

Exploratory visualizations can also be beautiful

When the goal is to have the receiver explore and understand the data.

The Guardian: How coronavirus spreads ↗

SVT: Replay the 2026 Swedish election count ↗

Adding interactivity 🎮

💬 Why use interactive visualizations?

Adding interactivity 🎮

➡️ ggplotly()

• Two main ways to creating a plotly object in R:

  1. Convert a ggplot2 object with ggplotly()
  2. Directly initialize a plotly object with plot_ly()

Adding interactivity 🎮

Additional functionality with plot_ly()

• Adding custom controls
• Animations

Adding interactivity 🎮

✅  Advantages	⚠️  Disadvantages
Explore data dynamically	Harder to include in static formats (papers, print)
Zoom, pan, and hover for details	Accessibility issues (screen readers, non colorblind-friendly defaults)
Reveal extra information without clutter	Can overwhelm the audience if overused
Enhance engagement in presentations and reports	Larger file sizes, slower rendering
Useful for exploratory analysis and dashboards	Requires additional packages (e.g., {plotly})

Adding interactivity 🎮

Interactivity enhances, but does not replace, the fundamentals of good data visualization

Using AI for Data Exploration 🤖

💬 What about using AI for data exploration in visualization?

Using AI for Data Exploration 🤖

1. AI lets you explore wider
   Use it for an efficient overview of the available data
2. But, every plot hides decisions
   Ask the AI to list every transformation it made
3. More plots = more chances of false positives
   Treat exploratory findings as a lead to continue checking, not a result
4. You bring the meaning
   what does a variable measure? does the pattern make sense?
5. Give the AI all the context you can
   e.g. codebooks, background information
6. Evaluate, evaluate, evaluate!
   Before prompting, decide how you will verify the results yourself

• You can get a distribution of every variable, and every relationship you’re curious about, in a few minutes. So use that! Look at more than you normally would. And remember, these plots don’t have to be pretty, they’re just for you.
• every plot you get back already has choices baked into it. Which variables did it pick, did it take a mean or a median, did it filter anything out? The AI makes these choices without telling you.
• More plots, more false patterns: If you can make 50 versions of a plot in a minute, one of them is going to show you something interesting, just by chance. So anything you find while exploring is a lead, not a result. You still have to check it.
• You bring the meaning: the AI can make the plot, but it doesn’t know what your variables actually measure, or whether the pattern makes any sense in the real world. That part is your job. If you can’t judge whether the answer is right, the AI’s answer isn’t going to help you.
• Give the AI all the context you can: the more it knows, the less it has to guess. So don’t just give it the column names, give it the codebook, the background of the data, what your research question is. And a nice trick: ask it to describe the data back to you first. Where it gets it wrong, correct it, and now you also know where its understanding stops.
• Evaluate, evaluate, evaluate!: decide before you prompt how you’re going to check the result, otherwise the AI’s output becomes the standard you judge it by. Simple checks: how many schools went in, and how many ended up in the plot? Can you reproduce one number from the plot yourself? Does the pattern still hold with a median instead of a mean? Keeping an AI log will also help with this.

Using AI for Data Exploration 🤖

Try it yourself!

Submitting Assignment 2, Part 1

• The assignment is complete if:

  1. it contains multiple plots: show that you analysed your variable from at least 2 different perspectives.
     
  2. Has a theme and color-scheme. Although not necessary for exploratory viz, this gives me a chance to give feedback on your styling choices for Part 2.
     
  3. it highlights some interesting data points, for example using color or text. While this is also not necessary for exploratory viz, do this to show me the direction you will take your final visualization in, so I can give feedback.
  4. it contains a short description outlining your workflow: what exploration choices led to this visualization? did you use AI? if so, how and with which model? what did you learn about the data and how can you use this to inform the final, explanatory data visualization?
     
  5. your code runs, and the html can be downloaded and viewed on any laptop.
     
  6. your code is lintr-proof.

Now

In your groups:

AI worksheet, part 1. This should help set you up for your assignment. In 30 minutes we have a short discussion about your findings

Continue with Assignment 2, part 1.

Session 3 Tuesday, Oct. 6 Explanatory Data Visualization

Today

Better plots 🥇

• General feedback assignments
• Going from exploratory to explanatory
• Advanced DV features
• Using AI for explanatory DVs

Feedback on Assignment 2, Part 1

• Use informative titles: state your conclusion in the title or subtitle in a declarative manner
• Text: Increase font size, use bold-face in main statements
  
• Visualize reliability: show sample size, confidence intervals, etc. Especially if using bins/groupings.
  
• What is the added value of a boxplot?
• Go from exploratory to explanatory!
  the audience should get the main conclusion without reading the description

Use point size to your advantage

Boxplots

The Anatomy of a Boxplot

Exploratory versus explanatory

Exploratory

Explanatory

Source: storytellingwithdata.com

Making an explanatory DV

1. Start from your conclusion
2. Adapt it to your audience
3. Choose a chart type that fits the conclusion and the data
4. Highlight your main conclusion

• These four steps are the structure for the rest of today’s lecture.
• 1: what is the main thing you learned from the data, that you want to show in the DV?
• 2: how do you say that to your audience, so they understand it?
• 3: different chart types are more or less appropriate depending on context.
• 4: there are more advanced techniques to make the details, and your main point, clear to the audience.

Step 1: Start from your conclusion

What is the main thing you learned from the data?

• Pick one main statement.
• Write down how you got there.
  Which inferential or modeling choices led to this conclusion?
• Justify
  Why is this relevant, and why do you think it is correct?
  How confident are you in the validity, reliability, and robustness of the results?

• Pick one main statement that you want the DV to show. Everything else in the plot is there to support that one statement.
• Write down which inferential/modeling choices led to this conclusion: which schools did you keep, how did you handle missing values, mean or median, which comparison.
• Why do you think it is relevant, and why do you think it is correct? If you can’t answer that, go back to exploring.

Step 2: Adapt it to your audience

Formulate your conclusion so your audience understands it.

• Who are they?
  Each research question has a target audience: a municipal officer, policymakers, newspaper readers, school boards…
• Which details do they need?
  Pick the inferential/modeling details you want to show, and explain how they were derived.
• Can they get it without reading your report?
  Use the title and labels to carry the message.

• Choose a formulation of your conclusion that is understandable to your audience. A municipal officer wants to see at a glance which schools are behind; a newspaper reader needs one clear message.
• Pick which inferential/modeling details you want to display to the audience, and make sure you include an explanation of how those were derived (e.g. in a subtitle or caption).

Step 2: Adapt it to your audience

Use your title and caption to do the explaining

• Title: descriptive or declarative (state your conclusion!)
• Methods: keep it brief, if needed at all
• Results: only if not (fully) captured in the title
• Definitions of colors, line types, error bars, etc.
• Data source if external
• Date and name for crediting

Source: datasketch.es

Step 2: Adapt it to your audience

# lineplot = the original plot from Assignment 1
title_text <- paste(
  "LIB has the highest percentage of students reaching the target level,",
  "and Route8 the lowest. This gap holds even when accounting for the",
  "socio-economic status of schools."
)

caption_text <- paste0(
  "Data: DUO Open Onderwijsdata, doorstroomtoets results and reference ",
  "levels, school year 2024-2025;\n",
  "Inspectie van het Onderwijs, schoolweging 2024-2025.\n",
  "Chart: Annie Johansson | ",
  format(Sys.Date(), "%e %B %Y") |> str_trim()
)

legend_labels_oneline <- setNames(
  paste0(
    provider_n$provider,
    " (n = ", format(provider_n$n_pupils, big.mark = ",", trim = TRUE), ")"
  ),
  provider_n$provider
)

lineplot_titled <- lineplot +
  scale_color_manual(
    values = provider_colors,
    breaks = highlighted_providers, # legend in the same order as the lines
    labels = legend_labels_oneline,
    name = "Test provider"
  ) +
  scale_y_continuous(
    limits = c(0, 100),
    breaks = seq(0, 100, 25),
    labels = \(x) paste0(x, "%"),
    expand = expansion(mult = c(0, 0.02))
  ) +
  labs(
    title = str_wrap(title_text, width = 62),
    y = "Students reaching target level",
    caption = caption_text
  ) +
  theme(
    plot.title = element_text(
     face = "bold", size = 18, lineheight = 1.1,
     margin = margin(b = 12)
    ),
    plot.title.position = "plot",
    plot.caption = element_text(
      color = "grey40", size = 10, hjust = 0, lineheight = 1.2,
      margin = margin(t = 14)
    ),
    plot.caption.position = "plot",
    axis.title = element_text(color = "grey25", size = 13),
    axis.title.x = element_text(margin = margin(t = 8)),
    axis.title.y = element_text(margin = margin(r = 8)),
    axis.text = element_text(color = "grey35", size = 12),
    axis.line.x = element_line(color = "grey60", linewidth = 0.4),
    panel.grid.major.x = element_blank(),
    panel.grid.major.y = element_line(color = "grey90", linewidth = 0.4),
    legend.justification = "left",
    legend.title = element_text(face = "bold", size = 12.5),
    legend.text = element_text(size = 12),
    legend.margin = margin(b = 0),
    plot.margin = margin(16, 20, 12, 16),
    plot.background = element_rect(fill = "white", color = NA)
  )

Step 3: Choose a chart type

One dataset, many visualizations

data("pizzaplace")
pizza_top <- pizzaplace |>
  mutate(size = factor(size, levels = c("S", "M", "L"))) |>
  count(name, type, size, price, sort = TRUE) |>
  top_n(n = 5)
pizza_top |>
  gt() |>
  tab_header(title = "Pizza Top 5", subtitle = "2015") |>
  fmt_currency(columns = price, currency = "USD") |>
  tab_source_note(
    source_note = md(paste0(
      "Source: [pizzaplace dataset]",
      "(https://gt.rstudio.com/articles/gt-datasets.html#pizzaplace)"
    ))
  ) |>
  opt_stylize(style = 6)

Pizza Top 5
2015
name	type	size	price	n
big_meat	classic	S	$12.00	1914
thai_ckn	chicken	L	$20.75	1410
five_cheese	veggie	L	$18.50	1409
four_cheese	veggie	L	$17.95	1316
classic_dlx	classic	M	$16.00	1181
Source: pizzaplace dataset

pizza_top |>
  ggplot(aes(x = reorder(name, n, decreasing = TRUE), y = n)) +
  geom_point(aes(color = type, size = size)) +
  geom_text(aes(label = price), nudge_y = -30) +
  labs(title = "Pizza Top 5", subtitle = "2015", x = "name")

• The same data can be shown as a table or as a plot. Which one is better depends on what you want your audience to take away.

Step 3: Choose a chart type

One dataset, many visualizations

pizza_season <- pizzaplace |>
  mutate(month = lubridate::month(date, label = TRUE)) |>
  group_by(month) |>
  count(type)

fig_season_1 <- pizza_season |>
  ggplot(aes(x = month, y = n, group = type)) +
  geom_bar(aes(fill = type), stat = "identity") +
  labs(
    title = "Pizza Season",
    subtitle = "2015",
    y = "Number of pizzas sold",
    x = "Month"
  )
fig_season_1

fig_season_2 <- pizza_season |>
  ggplot(aes(x = month, y = n, group = type)) +
  geom_line(aes(linetype = type)) +
  labs(
    title = "Pizza Season",
    subtitle = "2015",
    y = "Number of pizzas sold",
    x = "Month"
  )
fig_season_2

• Stacked bars: you can see the total per month, but its hard to compare one pizza type over time (only the bottom one has a common baseline, remember the speed dates!).
• Lines: good for comparing the types over time. Harder to see the total.
• Which one you choose depends on your conclusion.

Step 3: Choose a chart type

Step 3: Choose a chart type

Step 4: Highlight your main conclusion

• Decide where you want the viewer to look first.
• Use visual elements (colors, arrows, highlights) to guide that.

Step 4: Highlight your main conclusion

Advanced DV features

There are more advanced techniques available to you to make certain details, and your main point, clear to the audience.

• Declutter remove what doesn’t add information
• Themes, facets, layouts for structure and consistency
• Color e.g., gray-scale with one accent color
• Typography and annotation guide the eye to your conclusion
• Uncertainty show how sure you are

Chart junk

NYT

Claus Wilke

NYT

Chart junk

You can try it yourself with geom_image() from ggimage

library(ggimage)
# aggregate to totals
pizza_totals <- pizza_season |>
  group_by(type) |>
  summarise(total = sum(n), .groups = "drop") |>
  mutate(n_icons = round(total / 1000)) # 1 pizza = 1000 sales

# expand data: one row per pizza icon
# this is a trick so that each pizza icon can be plotted separately (essentially
# using geom_point but appearing as a bar)
chart_junk_data <- pizza_totals |>
  rowwise() |>
  mutate(icon_id = list(1:n_icons)) |>
  unnest(icon_id)

# plot with 🍕
pizza_plot <- ggplot(chart_junk_data, aes(x = icon_id, y = type)) +
  geom_image(aes(image = "images/pizza.png"), size = 0.2) +
  # add white rectangle to cover up excess icons (taken from pizza_totals data)
  geom_rect(
    data = pizza_totals,
    aes(
      xmin = total / 1000, xmax = total / 1000 + 2,
      ymin = as.numeric(factor(type)) - 0.5,
      ymax = as.numeric(factor(type)) + 0.5
    ),
    inherit.aes = FALSE,
    fill = "white"
  ) +
  labs(
    title = "Total Pizza Sales in 2015",
    x = "Sales (Scale 1:1000)",
    y = NULL
  ) +
  theme_minimal(base_size = 14) +
  theme(
    panel.grid.minor.y = element_blank(),
    panel.grid.major.y = element_blank(),
    panel.grid.minor.x = element_blank(),
    axis.ticks = element_blank()
  )

pizza_plot

Themes

fig_season_2 +
  theme_cowplot()

my_theme <- theme_cowplot() +
  theme(
    panel.grid.major = element_line(color = "gray90"),
    axis.ticks = element_line(color = "gray20"),
    axis.text = element_text(color = "gray20", face = "italic", size = 16),
    axis.title = element_text(color = "gray20", face = "bold", size = 16),
    plot.title = element_text(hjust = 0.5, size = 20, face = "bold"),
    plot.subtitle = element_text(hjust = 0.5, size = 18),
    legend.position = "top",
    legend.title = element_text(size = 16, face = "bold"),
    legend.text = element_text(size = 16)
  )
fig_season_2 + my_theme

Facets

• theme_cowplot and theme_bw() format facets nicely.
• Change facet format manually with theme(strip.text = element_text(...)) and theme(strip.background = element_rect(...)).

fig_season_2 +
  facet_wrap(~type) +
  theme_bw(12) +
  theme(legend.position = "none")

fig_season_2 +
  geom_line(linetype = "solid", color = "gray30", linewidth = 0.5) +
  facet_wrap(~type) +
  theme_minimal(12) +
  theme(
    legend.position = "none",
    strip.text = element_text(face = "bold", color = "plum4"),
    strip.background = element_rect(fill = "thistle2", color = NA)
  ) # fill for background; color for border

Combining plots: patchwork

• Combine plots with +
• Add a shared title, subtitle, and caption with plot_annotation()
• Style them with plot_annotation(theme = theme(...))

library(patchwork)
p1 <- ggplot(mtcars, aes(x = hp, y = mpg)) +
  geom_point(color = "steelblue", size = 3, alpha = 0.7) +
  theme_minimal() +
  labs(title = "Fuel efficiency vs Horsepower")

p2 <- ggplot(mtcars, aes(x = factor(cyl), y = mpg)) +
  geom_boxplot(fill = "steelblue", alpha = 0.7) +
  theme_minimal() +
  labs(title = "MPG by Number of Cylinders")

p1 + p2 +
  plot_annotation(
    title = "Analysis of mtcars dataset",
    caption = "Data source: mtcars",
    theme = theme(plot.title = element_text(face = "bold", size = 16))
  )

Color scales

fig_quarter <- pizza_season |>
  mutate(quarter = case_when(
    month %in% c("Jan", "Feb", "Mar") ~ "Q1",
    month %in% c("Apr", "May", "Jun") ~ "Q2",
    month %in% c("Jul", "Aug", "Sep") ~ "Q3",
    month %in% c("Oct", "Nov", "Dec") ~ "Q4"
  )) |>
  ggplot(aes(x = quarter, y = n, group = type)) +
  geom_bar(aes(fill = type), stat = "identity", position = "dodge") +
  labs(title = "Pizza Season", y = "Number of pizzas sold", x = "")

fig_quarter +
  labs(subtitle = "Qualitative color scale") +
  scale_fill_brewer(type = "qual")

qualitative
categorical data

fig_quarter +
  labs(subtitle = "Sequential color scale") +
  scale_fill_brewer(type = "seq")

sequential
ordered data, from low to high

fig_quarter +
  labs(subtitle = "Diverging color scale") +
  scale_fill_brewer(type = "div")

diverging
ordered data with a critical midpoint, e.g. 0

Color blindness

# colorblindr is installed from GitHub (clauswilke/colorblindr)
colorblindr::cvd_grid(fig_quarter)

The package MetBrewer has many colorblind-friendly palettes, e.g. scale_fill_manual(values = MetBrewer::met.brewer("VanGogh3", n = 4))

Typography

• Learn everything about typography.
• Choose a font for data visualizations.
• Pick good font combinations.
• Or just use arial or helvetica.

Change fonts with package showtext and function font_add_google(). Browse Google Fonts

library(showtext)
sysfonts::font_add_google("Lekton")
showtext::showtext_auto()

fig_season_2 +
  theme_cowplot() +
  theme(text = element_text(family = "Lekton", size = 20))

Show uncertainty

How sure are you of your conclusion? Show it.

• Sample size how many schools (or students) is each group based on?
• Spread: is it informative to show the distribution, or error bars / intervals, not just the mean?
• Missing data how did you handle them, and does it matter? If yes, add a note in the caption.
• Confidence intervals (+ explain them in a way your audience understands.)

• Polished charts tend to hide uncertainty. But your audience needs to know how sure you are.
• This is also part of your report: how robust are your findings?

Putting it together

Consider your audience:

pizza_plot

Putting it together

Consider your audience:

sysfonts::font_add_google("Lobster")
sysfonts::font_add_google("Lexend")
showtext::showtext_auto()

pizza_plot +
  # add a geom_rect element with opacity on top of pizza bars, to highlight the
  # biggest bar
  geom_rect(
    data = pizza_totals |>
      mutate(highlight = ifelse(type == "classic", TRUE, FALSE)),
    aes(
      xmin = 0, xmax = total / 1000,
      ymin = as.numeric(factor(type)) - 0.5,
      ymax = as.numeric(factor(type)) + 0.5,
      alpha = highlight
    ),
    inherit.aes = FALSE,
    fill = "white"
  ) +
  scale_alpha_manual(values = c("TRUE" = 0, "FALSE" = 0.3), guide = "none") +
  labs(
    title = "The classic is a classic for a reason.",
    subtitle = "Total Pizza Sales in 2015",
    caption = "Source: pizzaplace dataset"
  ) +
  theme(
    text = element_text(family = "Lexend", color = "gray30"),
    plot.title = element_text(family = "Lobster", size = 20, color = "tomato3"),
    plot.subtitle = element_text(family = "Lexend", size = 16),
    axis.title.x = element_text(family = "Lobster", size = 16),
    axis.text.y = element_text(family = "Lobster", size = 16, color = "tomato3")
  )

Putting it together

Consider your audience:

max_sales <- max(pizza_totals$total)
best_seller <- pizza_totals$type[pizza_totals$total == max_sales]

plot_data <- pizza_totals |>
  mutate(is_max = total == max_sales) |>
  mutate(type = forcats::fct_reorder(type, total, .desc = TRUE))

# the best seller is colored in the subtitle, like its bar (needs ggtext)
main_conclusion <- paste0(
  "The <span style='color:tomato3;'>", best_seller, " pizza</span> ",
  "is the clear leader with ", scales::comma(max_sales), " units sold."
)

ggplot(plot_data, aes(x = type, y = total, fill = is_max)) +
  geom_col(width = 0.7) +
  geom_text(
    aes(label = scales::comma(total)),
    vjust = -0.5,
    size = 4,
    fontface = "bold"
  ) +
  # highlight color for the best seller, neutral gray for the rest
  scale_fill_manual(
    values = c("TRUE" = "tomato3", "FALSE" = "grey70"),
    guide = "none"
  ) +
  labs(
    title = "Total Sales by Pizza Type",
    subtitle = main_conclusion,
    x = "Pizza Type",
    y = "Total Units Sold"
  ) +
  theme_cowplot() +
  theme(
    plot.title.position = "plot",
    plot.title = element_text(size = 16, face = "bold"),
    plot.subtitle = ggtext::element_markdown(size = 16, color = "grey30"),
    # the values are already on the bars, so the y-axis can go
    axis.line.y = element_blank(),
    axis.ticks.y = element_blank(),
    axis.text.y = element_blank(),
    axis.title.y = element_blank()
  ) +
  scale_y_continuous(expand = expansion(mult = c(0, 0.15)))

• Same data, three plots. The first is chart junk; the second is fun and highlights the winner (fine for a pizza menu!); the third states the conclusion in the subtitle, highlights it with color, and removes everything that doesn’t help.
• Which one is right depends on your audience.

Final considerations

• Are all visual elements necessary? (borders, gridlines, etc.)
• Are names informative and adapted to the audience?
• Is text big enough and legible at all plot sizes?
• Can the plot render nicely on a presentation screen?

Some good examples 🏆

Some good examples 🏆

Continue learning

Viz types and examples: From Data to Viz, The R Graph Gallery, clauswilke.com

Extensions: ggplot2 Extensions Gallery

Books: Fundamentals of Data Visualization

NYT: What’s going on in this graph?

Data: Statistics Netherlands or data()

RTFM: ggplot2 Book

Get inspiration

Podcasts: Data Stories, Explore Explain, Data Viz Today

Blogs: FlowingData

Journals: Nightingale

Comics: Matt-Heun Hong, Martin Telefont, Natalia Kiseleva

Generative art: Clause O. Wilke, Danielle Navarro, Thomas Lin Pedersen

Awards: Information is Beautiful

Interactive visualizations: R Psychologist

Explorable explanations: Nicky Case, Setosa

Thomas Lin Pedersen

Using AI for explanatory DVs 🤖

1. Write your conclusion first, without AI
   The AI should help you show your conclusion, not choose it
2. Let it do the technical work
   Annotations, highlighting, themes, layouts, color palettes
3. Ask for alternatives
   3–4 different designs for the same message; then choose and justify
4. Criticize everything it adds
   Does it make the conclusion clearer, or only decorate? Check the title for causal language
5. Use it as a critic, carefully
   Ask for a skeptical review; you decide which critiques are valid
6. Test with a naïve reader
   What do they conclude? Is that conclusion true?

• For explaining, AI is mostly useful for the technical features: all the things from the previous slides, it can write that code for you quickly.
• But also experiment with using it for feedback, and for alternative visualizations of the same message.

Using AI for explanatory DVs 🤖

Try it yourself!

AI: Final note

Beware the Slop Zone

Pitches: Friday, Oct. 9

Bring a first draft of your final visualization

Render it in an html or pdf file and upload it to Canvas. The draft is not graded, but the further along it is, the more useful the feedback. No slides needed!

• 4-5 minutes pitch
  research question, approach, conclusion, open questions
• 3-4 minutes discussion
  bring the questions you would like input on
• AI use
  how did you use it, what did you learn, which obstacles or risks did you run into?

Handing in Assignment 2, Part 2

Deadline: Sunday, October 11th, 23:59

Hand in, on Canvas:

1. The link to your group’s (public) GitHub repository, with all your code pushed

• Your report, as:
  2. .Rmd
     with all code (lintr-proof!) to produce your final DV and report.
     (A quarto document as .qmd is also fine.)
  3. .html
     compiles only the final DV and report text.
     Hide your code by setting code_folding = TRUE in the preamble.
     make sure it knits correctly; we will not compile your .Rmd for you

And make sure I can clone your Github repo and completely reproduce your results.

Handing in Assignment 2, Part 2

Grading

Aesthetics 30%	Clear and legible, no unnecessary clutter; color, labels and scales enhance readability
Communication 30%	Can the main point be understood quickly, with minimal explanation?
Description 20%	Research question, what you wanted to communicate, why this visualization, what you can conclude. AI statement
Creativity and Complexity 10%	ex. Going beyond a default chart type, where it adds value without obscuring the message; adding consistent themes and styling matching the target audience / context.
Code Quality and Styling 10%	Runs without errors, readable, lintr-proof

Final report formatting

You do not need to display all your code.

Hide or fold all code in your preamble:

---
title: "example rmd"
output:
  html_document:
    code_folding: hide
---

Or per code chunk:

```{r}
#| echo: false

# your code here
```

Now

In your groups:

1. AI worksheet, part 2
2. Continue with Assignment 2, part 2

Ask questions and consult me when needed!