11  ggplot

11.1 Deconstructing a graph

11.1.1 The Grammar of Graphics (gg)

Yau (2013) and Wickham (2014) have come up with a taxonomy and a grammar for thinking about the parts of a figure just like we conceptualize the parts of a body or the parts of a sentence.

One great way of thinking of the new process: it is no longer necessary to talk about the name of the graph (e.g., boxplot). Instead we now think in glyphs (geoms), and so we can put whatever we want on the plot. Note also that the transition leads you from a passive consumer (I need to make plot XXX because everyone else does, so I just plug in the data) into an active participant (what do I want my data to say? and how can I put that information onto my graphic?)

The most important questions you can ask with respect to creating figures are:

  1. What do we want R to do? (What is the goal?)
  2. What does R need to know?

Yau (2013) gives us nine visual cues, and Wickham (2014) translates them into a language using ggplot2. (The items below are from Baumer, Kaplan, and Horton (2021), chapter 2.)

  1. Visual Cues: the aspects of the figure where we should focus.
    Position (numerical) where in relation to other things?
    Length (numerical) how big (in one dimension)?
    Angle (numerical) how wide? parallel to something else?
    Direction (numerical) at what slope? In a time series, going up or down?
    Shape (categorical) belonging to what group?
    Area (numerical) how big (in two dimensions)? Beware of improper scaling!
    Volume (numerical) how big (in three dimensions)? Beware of improper scaling!
    Shade (either) to what extent? how severely?
    Color (either) to what extent? how severely? Beware of red/green color blindness.

  2. Coordinate System: rectangular, polar, geographic, etc.

  3. Scale: numeric (linear? logarithmic?), categorical (ordered?), time

  4. Context: in comparison to what (think back to ideas from Tufte)

Figure 3.3 of Yau’s Data Points: Visualization That Means Something, https://flowingdata.com/data-points/DataPoints-Ch3.pdf, (Yau 2013).

Figure 3.3 of Yau’s Data Points: Visualization That Means Something, https://flowingdata.com/data-points/DataPoints-Ch3.pdf, (Yau 2013).

Order Matters

Figure 3.12 of Yau’s Data Points: Visualization That Means Something, https://flowingdata.com/data-points/DataPoints-Ch3.pdf, (Yau 2013). Original work by Cleveland and McGill (1984).

Figure 3.12 of Yau’s Data Points: Visualization That Means Something, https://flowingdata.com/data-points/DataPoints-Ch3.pdf, (Yau 2013). Original work by Cleveland and McGill (1984).

Cues Together

Figure 3.25 of Yau’s Data Points: Visualization That Means Something, https://flowingdata.com/data-points/DataPoints-Ch3.pdf, (Yau 2013).

Figure 3.25 of Yau’s Data Points: Visualization That Means Something, https://flowingdata.com/data-points/DataPoints-Ch3.pdf, (Yau 2013).
What are the visual cues on the plot?

  • position?
  • length?
  • shape?
  • area/volume?
  • shade/color?
  • coordinate System?
  • scale?
What are the visual cues on the plot?

  • position?
  • length?
  • shape?
  • area/volume?
  • shade/color?
  • coordinate System?
  • scale?
What are the visual cues on the plot?

  • position?
  • length?
  • shape?
  • area/volume?
  • shade/color?
  • coordinate System?
  • scale?

11.1.1.1 The grammar of graphics in ggplot2

geom: the geometric “shape” used to display data

  • bar, point, line, ribbon, text, etc.

aesthetic: an attribute controlling how geom is displayed with respect to variables

  • x position, y position, color, fill, shape, size, etc.

scale: adjust information in the aesthetic to map onto the plot

  • particular assignment of colors, shapes, sizes, etc.; making axes continuous or constrained to a particular range of values.

guide: helps user convert visual data back into raw data (legends, axes)

stat: a transformation applied to data before geom gets it

  • example: histograms work on binned data

11.1.2 ggplot2

In ggplot2, an aesthetic refers to a mapping between a variable and the information it conveys on the plot. Further information about plotting and visualizing information is given in chapter 2 (Data visualization) of Baumer, Kaplan, and Horton (2021). Much of the data in the presentation represents all births from 1978 in the US: the date, the day of the year, and the number of births.

Goals

What I will try to do

  • give a tour of ggplot2

  • explain how to think about plots the ggplot2 way

  • prepare/encourage you to learn more later

What I can’t do in one session

  • show every bell and whistle

  • make you an expert at using ggplot2

Getting help

  1. One of the best ways to get started with ggplot is to Google what you want to do with the word ggplot. Then look through the images that come up. More often than not, the associated code is there. There are also ggplot galleries of images, one of them is here: https://plot.ly/ggplot2/

  2. ggplot2 cheat sheet: https://rstudio.github.io/cheatsheets/html/data-visualization.html

  3. Look at the end of the presentation. More help options there.

require(mosaic)    # where Births78 lives
data(Births78)     # restore fresh version of Births78
head(Births78, 3)
date births wday year month day_of_year day_of_month day_of_week
1978-01-01 7701 Sun 1978 1 1 1 1
1978-01-02 7527 Mon 1978 1 2 2 2
1978-01-03 8825 Tue 1978 1 3 3 3
How can we make the plot?

Two Questions:

  1. What do we want R to do? (What is the goal?)

  2. What does R need to know?

    • data source: Births78

    • aesthetics:

      • date -> x
      • births -> y
      • points (!)

  3. Goal: scatterplot = a plot with points

    • ggplot() + geom_point()

  4. What does R need to know?

    • data source: data = Births78

    • aesthetics: aes(x = date, y = births)

How can we make the plot?

What has changed?

  • new aesthetic: mapping color to day of week
Adding day of week to the data set
ggplot(data = Births78) +
  geom_point(aes(x = date, y = births, color = wday)) +
  labs(title = "US Births in 1978")

How can we make the plot?

Now we use lines instead of dots

ggplot(data = Births78) +
  geom_line(aes(x = date, y = births, color = wday)) +
  labs(title = "US Births in 1978")
How can we make the plot?

Now we have two layers, one with points and one with lines

ggplot(data = Births78, 
       aes(x = date, y = births, color = wday)) + 
  geom_point() +  
  geom_line() +
  labs(title = "US Births in 1978")

  • The layers are placed one on top of the other: the points are below and the lines are above.

  • data and aes specified in ggplot() affect all geoms

Alternative Syntax
ggplot(data = Births78,
         aes(x = date, y = births, color = wday)) + 
  geom_point() + 
  geom_line() +
  labs(title = "US Births in 1978")

What does adding the color argument do?
ggplot(data = Births78, 
       aes(x = date, y = births, color = "navy")) + 
  geom_point()  +
  labs(title = "US Births in 1978")

Because there is no variable, we have mapped the color aesthetic to a new variable with only one value (“navy”). So all the dots get set to the same color, but it’s not navy.

Setting vs. Mapping

If we want to set the color to be navy for all of the dots, we do it outside the aesthetic, without a dataset variable:

ggplot(data = Births78, 
       aes(x = date, y = births)) +   # map x & y 
  geom_point(color = "navy")   +     # set color
  labs(title = "US Births in 1978")

  • Note that color = "navy" is now outside of the aesthetics list. That’s how ggplot2 distinguishes between mapping and setting.
How can we make the plot?

ggplot(data = Births78, 
       aes(x = date, y = births)) + 
  geom_line(aes(color = wday)) +       # map color here
  geom_point(color = "navy") +          # set color here
  labs(title = "US Births in 1978")

  • ggplot() establishes the default data and aesthetics for the geoms, but each geom may change the defaults.

  • good practice: put into ggplot() the things that affect all (or most) of the layers; rest in geom_blah()

Setting vs. Mapping (again)

Information gets passed to the plot via:

  1. map the variable information inside the aes (aesthetic) command

  2. set the non-variable information outside the aes (aesthetic) command

Other geoms
apropos("^geom_")
 [1] "geom_abline"            "geom_area"              "geom_ash"              
 [4] "geom_bar"               "geom_bin_2d"            "geom_bin2d"            
 [7] "geom_blank"             "geom_boxplot"           "geom_col"              
[10] "geom_contour"           "geom_contour_filled"    "geom_count"            
[13] "geom_crossbar"          "geom_curve"             "geom_density"          
[16] "geom_density_2d"        "geom_density_2d_filled" "geom_density2d"        
[19] "geom_density2d_filled"  "geom_dotplot"           "geom_errorbar"         
[22] "geom_errorbarh"         "geom_freqpoly"          "geom_function"         
[25] "geom_hex"               "geom_histogram"         "geom_hline"            
[28] "geom_jitter"            "geom_label"             "geom_line"             
[31] "geom_linerange"         "geom_lm"                "geom_map"              
[34] "geom_path"              "geom_point"             "geom_pointrange"       
[37] "geom_polygon"           "geom_qq"                "geom_qq_line"          
[40] "geom_quantile"          "geom_raster"            "geom_rect"             
[43] "geom_ribbon"            "geom_rug"               "geom_segment"          
[46] "geom_sf"                "geom_sf_label"          "geom_sf_text"          
[49] "geom_smooth"            "geom_spline"            "geom_spoke"            
[52] "geom_step"              "geom_text"              "geom_tile"             
[55] "geom_violin"            "geom_vline"

help pages will tell you their aesthetics, default stats, etc.

?geom_area             # for example
Let’s try geom_area 
ggplot(data = Births78, 
       aes(x = date, y = births, fill = wday)) + 
  geom_area() +
  labs(title = "US Births in 1978")

Using area does not produce a good plot

  • over plotting is hiding much of the data
  • extending y-axis to 0 may or may not be desirable.
Side note: what makes a plot good?

Most (all?) graphics are intended to help us make comparisons

  • How does something change over time?
  • Do my treatments matter? How much?
  • Do men and women respond the same way?

Key plot metric: Does my plot make the comparisons I am interested in

  • easily, and
  • accurately?
Time for some different data

HELPrct: Health Evaluation and Linkage to Primary care randomized clinical trial

head(HELPrct)
age anysubstatus anysub cesd d1 daysanysub dayslink drugrisk e2b female sex g1b homeless i1 i2 id indtot linkstatus link mcs pcs pss_fr racegrp satreat sexrisk substance treat avg_drinks max_drinks hospitalizations
37 1 yes 49 3 177 225 0 NA 0 male yes housed 13 26 1 39 1 yes 25.111990 58.41369 0 black no 4 cocaine yes 13 26 3
37 1 yes 30 22 2 NA 0 NA 0 male yes homeless 56 62 2 43 NA NA 26.670307 36.03694 1 white no 7 alcohol yes 56 62 22
26 1 yes 39 0 3 365 20 NA 0 male no housed 0 0 3 41 0 no 6.762923 74.80633 13 black no 2 heroin no 0 0 0
39 1 yes 15 2 189 343 0 1 1 female no housed 5 5 4 28 0 no 43.967880 61.93168 11 white yes 4 heroin no 5 5 2
32 1 yes 39 12 2 57 0 1 0 male no homeless 10 13 5 38 1 yes 21.675755 37.34558 10 black no 6 cocaine no 10 13 12
47 1 yes 6 1 31 365 0 NA 1 female no housed 4 4 6 29 0 no 55.508991 46.47521 5 black no 5 cocaine yes 4 4 1

Subjects admitted for treatment for addiction to one of three substances.

Who are the people in the study?
ggplot(data = HELPrct,
       aes(x = substance)) + 
  geom_bar() +
  labs(title = "HELP clinical trial at detoxification unit")

  • Hmm. What’s up with y?

    • stat_bin() is being applied to the data before the geom_bar() gets to do its thing. Binning creates the y values.
Who are the people in the study?
ggplot(data = HELPrct,
       aes(x = substance, fill = sex)) + 
  geom_bar() +
  labs(title = "HELP clinical trial at detoxification unit")

Who are the people in the study?
library(scales)
ggplot(data = HELPrct,
       aes(x = substance, fill = sex)) + 
  geom_bar() +
  scale_y_continuous(labels = percent) +
  labs(title = "HELP clinical trial at detoxification unit")

Who are the people in the study?
ggplot(data = HELPrct,
       aes(x = substance, fill = sex)) + 
  geom_bar(position="fill") +
  scale_y_continuous("actually, percent") +
  labs(title = "HELP clinical trial at detoxification unit")

How old are people in the HELP study?
ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_histogram() +
  labs(title = "HELP clinical trial at detoxification unit")
`stat_bin()` using `bins = 30`. Pick better value with `binwidth`.

Notice the messages

  • stat_bin: Histograms are not mapping the raw data but binned data.
    stat_bin() performs the data transformation.

  • binwidth: a default binwidth has been selected, but we should really choose our own.

Setting the binwidth manually
ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_histogram(binwidth = 2) +
  labs(title = "HELP clinical trial at detoxification unit")

How old are people in the HELP study? – Other geoms
ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_freqpoly(binwidth = 2) +
  labs(title = "HELP clinical trial at detoxification unit")

ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_density() +
  labs(title = "HELP clinical trial at detoxification unit")

Selecting stat and geom manually

Every geom comes with a default stat

  • for simple cases, the stat is stat_identity() which does nothing
  • we can mix and match geoms and stats however we like
ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_line(stat = "density") +
  labs(title = "HELP clinical trial at detoxification unit")

Selecting stat and geom manually

Every stat comes with a default geom, every geom with a default stat

  • we can specify stats instead of geom, if we prefer
  • we can mix and match geoms and stats however we like
ggplot(data = HELPrct,
       aes(x = age)) + 
  stat_density(geom = "line") +
  labs(title = "HELP clinical trial at detoxification unit")

More combinations
ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_point(stat = "bin", binwidth = 3) + 
  geom_line(stat = "bin", binwidth = 3)  +
  labs(title = "HELP clinical trial at detoxification unit")

ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_area(stat = "bin", binwidth = 3) +
  labs(title = "HELP clinical trial at detoxification unit") 

ggplot(data = HELPrct,
       aes(x = age)) + 
  geom_point(stat = "bin", binwidth = 3, 
             aes(size=..count..)) +
  geom_line(stat = "bin", binwidth = 3) +
  labs(title = "HELP clinical trial at detoxification unit")

How much do they drink? (i1)
ggplot(data = HELPrct,
       aes(x = i1)) + geom_histogram() +
  labs(title = "HELP clinical trial at detoxification unit")

ggplot(data = HELPrct,
       aes(x = i1)) + 
  geom_density() +
  labs(title = "HELP clinical trial at detoxification unit")

ggplot(data = HELPrct,
       aes(x = i1)) + 
  geom_area(stat = "density") +
  labs(title = "HELP clinical trial at detoxification unit")

Covariates: Adding in more variables

Using color and linetype:

ggplot(data = HELPrct,
       aes(x = i1, color = substance, linetype = sex)) + 
  geom_line(stat = "density") +
  labs(title = "HELP clinical trial at detoxification unit")

Using color and facets

ggplot(data = HELPrct,
       aes(x = i1, color = substance)) + 
  geom_line(stat = "density") + 
  facet_grid( . ~ sex ) +
  labs(title = "HELP clinical trial at detoxification unit")

ggplot(data = HELPrct,
       aes(x = i1, color = substance)) + 
  geom_line(stat = "density") + 
  facet_grid( sex ~ . ) +
  labs(title = "HELP clinical trial at detoxification unit")

Boxplots

Boxplots use stat_quantile() which computes a five-number summary (roughly the five quartiles of the data) and uses them to define a “box” and “whiskers”.

The quantitative variable must be y, and there must be an additional x variable.

ggplot(data = HELPrct,
       aes(x = substance, y = age, color = sex)) + 
  geom_boxplot() +
  labs(title = "HELP clinical trial at detoxification unit")

Horizontal boxplots

Horizontal boxplots are obtained by flipping the coordinate system:

ggplot(data = HELPrct,
       aes(x = substance, y = age, color = sex)) + 
  geom_boxplot() +
  coord_flip() +
  labs(title = "HELP clinical trial at detoxification unit")

  • coord_flip() may be used with other plots as well to reverse the roles of x and y on the plot.
Axes scaling with boxplots

We can scale the continuous axis

ggplot(data = HELPrct,
       aes(x = substance, y = age, color = sex)) + 
  geom_boxplot() +
  coord_trans(y = "log") +
  labs(title = "HELP clinical trial at detoxification unit")

Give me some space

We’ve triggered a new feature: dodge (for dodging things left/right). We can control how much if we set the dodge manually.

ggplot(data = HELPrct,
       aes(x = substance, y = age, color = sex)) + 
  geom_boxplot(position = position_dodge(width = 1)) +
  labs(title = "HELP clinical trial at detoxification unit")

Issues with bigger data
require(NHANES)
dim(NHANES)
[1] 10000    76
ggplot(data = NHANES,
       aes(x = Height, y = Weight)) +
  geom_point() + 
  facet_grid( Gender ~ PregnantNow ) +
  labs(title = "National Health and Nutrition Examination Survey")

  • Although we can see a generally positive association (as we would expect), the over plotting may be hiding information.
Using alpha (opacity)

One way to deal with over plotting is to set the opacity low.

ggplot(data = NHANES,
       aes(x = Height, y = Weight)) +
  geom_point(alpha = 0.01) + 
  facet_grid( Gender ~ PregnantNow ) +
  labs(title = "National Health and Nutrition Examination Survey")

geom_density2d

Alternatively (or simultaneously) we might prefer a different geom altogether.

ggplot(data = NHANES,
       aes(x = Height, y = Weight)) +
  geom_density2d() + 
  facet_grid( Gender ~ PregnantNow ) +
  labs(title = "National Health and Nutrition Examination Survey")

Multiple layers
ggplot(data = HELPrct, 
       aes(x = sex, y = age)) +
  geom_boxplot(outlier.size = 0) +
  geom_jitter(alpha = .6, width = 0.2) + 
  coord_flip() +
  labs(title = "HELP clinical trial at detoxification unit")

Multiple layers
ggplot(data = HELPrct, 
       aes(x = sex, y = age)) +
  geom_boxplot(outlier.size = 0) +
  geom_point(alpha = .6, 
             position = position_jitter(width = .1, height = 0)) +
  coord_flip() +
  labs(title = "HELP clinical trial at detoxification unit")

Things I haven’t mentioned (much)

  • coords (coord_flip() is good to know about)

  • themes (for customizing appearance)

  • position (position_dodge(), position_jitterdodge() (for use with points on top of dodged boxplots), position_stack(), etc.)

  • transforming axes

require(ggthemes)
ggplot(Births78, 
       aes(x = date, y = births)) + 
  geom_point() + 
  theme_wsj()

ggplot(data = HELPrct, 
       aes(x = substance, y = age, color = sex)) +
  geom_boxplot(position = position_dodge(width = 1)) +
  geom_point(aes(color = sex, fill = sex), 
             position = position_jitterdodge(dodge.width = 1, jitter.width = 0.1), 
             size = 0.5) +
  labs(title = "HELP clinical trial at detoxification unit")

A little bit of everything
ggplot(data = HELPrct, 
       aes(x = substance, y = age, color = sex)) +
  geom_boxplot(coef = 10, 
               position = position_dodge(width = 1)) +
  geom_point(aes(fill = sex), alpha = .5, 
             position = position_jitterdodge(dodge.width = 1)) + 
  facet_wrap(~homeless) +
  labs(title = "HELP clinical trial at detoxification unit")

What else can we do?

shiny

  • interactive graphics / modeling

  • https://shiny.posit.co/

plotly

Plotly is an R package for creating interactive web-based graphs via plotly’s JavaScript graphing library, plotly.js. The plotly R library contains the ggplotly function , which will convert ggplot2 figures into a Plotly object. Furthermore, you have the option of manipulating the Plotly object with the style function.

  • https://plot.ly/ggplot2/getting-started/

Dynamic documents

  • combination of Quarto, ggvis, and shiny

  • Really good shiny help: https://shiny.posit.co/r/getstarted/shiny-basics/lesson1/index.html

11.2 Reflection questions

  • What do you want R to do? What does R need to know in order to do that thing?

  • How can you break down a plot into visual cues? What are the visual cues?

  • Which visual cues are most accurate? Does that mean only such cues should be used?

  • Does the order of layers in a ggplot matter? When does it change the plot and when does it not change the plot?

11.3 Ethics considerations

  • Why might you want to avoid overplotting?
Baumer, Ben, Daniel Kaplan, and Nicholas Horton. 2021. Modern Data Science with r. CRC Press. https://mdsr-book.github.io/mdsr2e/.
Cleveland, William S., and Robert McGill. 1984. “Graphical Perception: Theory, Experimentation, and Application to the Development of Graphical Methods.” Journal of the American Statistical Association 79 (387): 531–54. https://doi.org/10.1080/01621459.1984.10478080.
Wickham, Hadley. 2014. “Tidy Data.” Journal of Statistical Software 59 (10). http://www.jstatsoft.org/v59/i10/paper.
Yau, Nathan. 2013. Data Points: Visualization That Means Something. Wiley.