1 Agenda
2 Resources
3 Preface
4 Learning Objectives
5 Consider Information Density
6 Find the Goldilocks Plot
7 Choose Human-Interpretable Aesthetic Mappings and Geom’s
8 Consider Zero
9 Order factors

Making effective plots can tell you a LOT about data. Its hard! Its an under-rated but very powerful skill to develop.

- Di Cook

suppressPackageStartupMessages(library(tidyverse))
library(gapminder)
knitr::opts_chunk$set(fig.width=5, fig.height=3)

1 Agenda

Tips for effective graphing

At least two exercises related to content and http://viz.wtf/ (see the worksheet).

2 Resources

These resources are listed on the syllabus in the lecture table. They provide a good overview of tips for effective plotting.

Geckoboard’s data vis tips
Jenny’s STAT545 lecture notes: do’s and don’ts

Here are some resources that dive a little deeper:

Di Cook’s Rookie Mistakes.
- Especially focusses on categorical data.
Richard Hollins’ Three reasons why pie charts suck

An entertaining but inspiring resource:

Gallery of poor data vis: http://viz.wtf/

If you want to spend more time on this and/or dig deeper, take a look at the following books:

Visualization Analysis and Design – Tamara Munzner
Creating more effective graphs – Naomi Robbins.

3 Preface

Effectiveness: how well a graph conveys the information conveyed in data.
!= publication quality, although sometimes the line between these two can be blurry (no pun intended).
Take solace in the fact that we can’t make perfectly effective graphics.
Is a topic in human psychology.
Don’t know human psychology? Luckily, you’re a human, and have an innate knowledge of the relative effectiveness of a plot. Use that as your guide!
- From Gelman et al, As a guiding point, ask yourself: how can I modify my graph to better:
  1. faciliate comparisons, and
  2. reveal trends?

Disclaimer: The tips you see here and online hold true for most cases. There might be some rare cases where the tips don’t hold – the key is to be intentional about every component of the graph.

“Let’s Practice What We Preach: Turning Tables into Graphs” by Gelman A, Pasarica C, Dodhia R. The American Statistician, Volume 56, Number 2, 1 May 2002 , pp. 121-130(10).

4 Learning Objectives

From today’s lecture, students are expected to:

be intentional with your choice of graph components.
be able to spot bad graphs, and avoid making them.
internalize some tips of plotting effectiveness provided in class.

For the quiz, you aren’t expected to know/memorize all of the tips.

5 Consider Information Density

Sometimes called overplotting.

Scatterplot too dense?
- Do you need a log transform?
- Try alpha transparency
- Change geom: geom_hex() or geom_bin2d()
- Spread the data into separate panels: facet by a grouping variable or two.

gapxy <- ggplot(gapminder, aes(lifeExp, gdpPercap)) +
    theme_bw()
gapxy + geom_point()

gapxy <- gapxy + scale_y_log10()
gapxy + geom_point()

gapxy + geom_point(alpha=0.2)

gapxy + geom_hex()

gapxy + geom_density2d()

gapxy + facet_wrap(~continent) + geom_point(alpha=0.2)

Using too many geom’s? Don’t.

ggplot(gapminder, aes(continent, lifeExp)) +
    geom_violin(fill="red", alpha=0.2) +
    geom_boxplot(fill="blue", alpha=0.2) +
    geom_jitter(alpha=0.2)

6 Find the Goldilocks Plot

Display just the right amount of content: not too much, not too little.

In particular: reveal as much relevant information as possible; trim irrelevant and redundant information.

6.1 Reveal as much relevant information as possible

Because hiding your data is not effective at conveying information!

jitter + violin, not pinhead plots.
mosaic plots

6.2 Trim Irrelevant Information

Only use as much data as is required for answering a data analytic question.

Di Cook’s example in Rookie Mistakes: reduce complexity section.
Remove the special effects (sorry, Excel). Great demo: Darkhorse analytic’s Less is more gif.
More examples of extraneous information:

map_data("france") %>% 
    ggplot(aes(long, lat)) +
    geom_polygon(aes(group=group), fill=NA, colour="black") +
    theme_bw() +
    ggtitle("Are lat and long really needed?")

## 
## Attaching package: 'maps'

## The following object is masked from 'package:purrr':
## 
##     map

ggplot(gapminder, aes(year, lifeExp)) +
    geom_line(aes(group=country, colour=country), alpha=0.2) +
    guides(colour=FALSE) +
    theme_bw() +
    ggtitle("Is colouring by country really necessary here?\nNevermind fitting the legend!")

6.3 Trim Redundant Information

Don’t redundantly map variables to aesthetics/facets.

Common example: colouring/filling and facetting by the same variable.

HairEyeColor %>% 
    as_tibble() %>% 
    uncount(n) %>% 
    ggplot(aes(Hair)) +
    facet_wrap(~Sex) +
    geom_bar(aes(fill=Sex)) +
    theme_bw() +
    ggtitle("Don't do this.")

Really want to use colour? No problem, colours are fun! Try this:

HairEyeColor %>% 
    as_tibble() %>% 
    uncount(n) %>% 
    ggplot(aes(Hair)) +
    facet_wrap(~Sex) +
    geom_bar(fill="#D95F02") +
    theme_bw() +
    ggtitle("Do this.")

Delegate numeric details to an appendix, not the graph (or omit entirely).

HairEyeColor %>% 
    as_tibble() %>% 
    uncount(n) %>% 
    count(Hair) %>% 
    ggplot(aes(Hair, n)) +
    geom_col() +
    geom_text(aes(label=n), vjust=-0.1) +
    theme_bw() +
    labs(x="Hair colour", y="count", 
         title="Are the bar numbers AND y-axis really needed?")

7 Choose Human-Interpretable Aesthetic Mappings and Geom’s

Here’s an iconic fail in Henrik Lindberg’s tweet: the “depeche plot”.
Don’t use pie charts – use bar charts instead. 3 reasons why they suck.
To bar, or not to bar? Not if zero doesn’t matter! As a general rule, I like to err on the side of points over bars.

plot_beav2 <- bind_rows(
    mutate(beaver1, beaver = "Beaver 1"), 
    mutate(beaver2, beaver = "Beaver 2")
) %>% 
    group_by(beaver) %>% 
    summarize(med = median(temp)) %>% 
    ggplot(aes(beaver, med)) +
    theme_bw() +
    xlab("") +
    ylab("Body Temperature\n(Celsius)")
cowplot::plot_grid(
    plot_beav2 +
        geom_col() +
        ggtitle("Don't do this."),
    plot_beav2 +
        geom_point() +
        ggtitle("Do this.")
)

(Yes, that’s really all the info you’re conveying. Own it.)

Colour your groups so that not one group unintentionally stands out.

plot_iris <- ggplot(iris, aes(Sepal.Width, Sepal.Length)) +
    geom_jitter(aes(colour=Species)) +
    theme_bw() +
    theme(legend.position = "bottom")
cowplot::plot_grid(
    plot_iris +
        scale_colour_manual(values=c("brown", "gray", "yellow")) +
        ggtitle("Don't do this."),
    plot_iris +
        scale_colour_brewer(palette="Dark2") +
        ggtitle("Leave it to an expert.\nDo this.")
)

8 Consider Zero

Are you comparing data across groups? Consider what a meaningful distance measure might be between two groups.

Are differences meaningful, and proportions not? Example: temperature. Zero doesn’t matter.

plot_beav <- bind_rows(
    mutate(beaver1, beaver = "Beaver 1"), 
    mutate(beaver2, beaver = "Beaver 2")
) %>% 
    ggplot(aes(beaver, temp)) +
    geom_violin() +
    geom_jitter(alpha=0.25) +
    theme_bw() +
    xlab("") +
    ylab("Body Temperature\n(Celsius)")
cowplot::plot_grid(
    plot_beav + 
        ggtitle("This."), 
    plot_beav + 
        ylim(c(0,NA)) +
        ggtitle("Not This.")
)

Are proportions meaningful, and differences not? Example: counts.

HairEyeColor %>% 
    as_tibble() %>% 
    uncount(n) %>% 
    ggplot(aes(Hair)) +
    geom_bar() +
    theme_bw() +
    ggtitle("Keep this starting from 0.")

Want to convey absolute life expectancies, in addition to relative life expectancies? Show 0.

ggplot(gapminder, aes(continent, lifeExp)) +
    geom_boxplot() +
    ylim(c(0, NA)) +
    geom_hline(yintercept = 0,
               linetype = "dashed")

9 Order factors

It’s easier to see rankings. See this STAT 545 example by Jenny Bryan. Use forcats!

Lecture 5: Plotting for Humans