Data visualization using R

1. Choosing the "right" plot
2. Best practices and deadly sins of data visualization
3. How to create a plot
4. How to create a plot that's not totally ugly
5. Saving plots

• Please open R and, ideally, RStudio.

• Please open the Googledoc: https://tinyurl.com/ya96r5bf

• Please download the toy datasets: https://tinyurl.com/neuchaRtel

• Feel free to interrupt me at any time!
• There's A LOT of code in this presentation...

• Feel free to interrupt me at any time!
• There's A LOT of code in this presentation...

this <- is(what, code) {
  looks, like
}

• Feel free to interrupt me at any time!
• There's A LOT of code in this presentation...

this <- is(what, code) {
  looks, like
}

• You don't have to type all of the code, it's more important that you understand the conceptual background first.
• There will be some hands-on exercises.

• Feel free to interrupt me at any time!
• There's A LOT of code in this presentation...

this <- is(what, code) {
  looks, like
}

• You don't have to type all of the code, it's more important that you understand the conceptual background first.
• There will be some hands-on exercises.
• Also, there will be some slides with more advanced stuff (and yellow background).

• For yourself

  • Exploring your data
  • detecting outliers
  • checking assumptions of statistical tests or models (e.g. are the data normally distributed?)
  • etc.

• For others

  • Showing your findings in a clear and efficient way
  • Graphs tend to be more reader-friendly than tables...
  • and much more reader-friendly than long inline lists!

• What kind of data are you dealing with?
• What is your research question?

• categorical variables:
  • nominal variable: e.g. married, not married, divorced; Swiss, German, French...
    • Subtype: binary variable, e.g. living/dead
  • ordinal variable: e.g. gold medal, silver medal, bronze medal
• metric variables:
  • interval variable: e.g. temperature (Celsius, Fahreneit)
  • ratio variable: e.g. weight, temperature (Kelvin)
  • (absolute / count variable: natural unit, e.g. number of students, age)

• What kind of variable are we dealing with here?
• Which visualization seems most appropriate to you?

• What kind of variable are we dealing with here?
• Which visualization seems most appropriate to you?

• What kind of variable are we dealing with here?
• Which visualization seems most appropriate to you?

• from http://www.stubbornmule.net/2010/10/visualizing-smoking-risk/
• "Risk Characterization Theatre" from Rifkin & Bouwer (2007)

• from http://www.stubbornmule.net/2010/10/visualizing-smoking-risk/

• Who among the expeRts is a native speaker of English?

"The essence of a graphic display is that a set of numbers having both magnitudes and an order are represented by an appropriate visual metaphor - the magnitude and order of the metaphorical representation match the numbers." (Wainer 1984: 139)

• Most importantly: Know your data!
• When reporting percentages, also report the denominator (i.e. the size of your sample)
• Example: "50% of academics are alcoholics" - it makes a difference whether your sample size is 2 or 2,000.
• When reporting comparisons of absolute frequencies, double-check if your samples are comparable.
• Example: "255 women agree that cats are adorable, but only 5 men." - it makes a difference whether your sample consists of 300 women and 300 men or of 300 women and 10 men.
• When reporting means, also report standard deviations.
• Example: [5,5,5,5,995] has the same mean as [1,180,300,223,146].

• Show the data
• Avoid distorting the data
• Keep "Ink-to-data ratio" as low as possible
• Use meaningful x and y labels
• Avoid overplotting (e.g. 3-dimensional plots when only 2 dimensions are displayed)

• If there is no natural order in your data, order them by value
  

• Don't cut the y axis unless there are good conceptual reasons to do so.

• Use "tidy data":
  • One variable per column
  • One observation per row

• wide format: repeated responses in a single row
• long format: repeated responses in different rows

• Golden rule: Don't be sloppy with your data!

• Data are often messy: What's wrong here?

• show / explore correlations between two variables
• metric data on both the x- and the y-axis

• First, let's create some data:

x <- c(1,3,3,4,7,8)
y <- c(1,1,3,9,8,5)

• and create a simple plot:

plot(x,y)

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main")

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = "red")

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = rgb(1, 0, 0, alpha = 0.5))

• keep in mind that what you see on your screen is not always what you get on a printer or projector
• → make sure that your colors are not too similar!
• use color-blind friendly color schemes
• avoid red-green contrasts
• in many cases, it makes sense to combine color with other aesthetics like shape or line type

• If you work with many different colors in a plot, check out R's color palettes, e.g. rainbow, heat.colors, terrain.colors
• Another useful resource is the RColorBrewer package with a number of color-blind friendly palettes (argument colorblindFriendly = TRUE)

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = "red")
text(x = 5, y = 5,"Note the the added text! \n In a different color!", col = "blue")

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = "red",
     pch = 20)

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = "red",
     pch = "\u263A", cex = 2)

• Any single character can be used.

par(mar = c(5, 4, 4, 2) + 0.1)

par_cur <- par() # save default graphics parameters
par(mar = c(1,1,1,1)) # change margins
plot(1:20, rep(10,20), pch = c(1:20), cex=1.5, ylab="", xlab="", yaxt="n", xaxt="n")
text(1:20, rep(8.5,20), labels = 1:20)

par(par_cur) # restore default graphics parameters

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = "red",
     xlim = c(0, max(x)),
     ylim = c(0, max(y)))

plot(x, y, xlab = "xlab", ylab = "ylab", main = "main", col = "red", 
     xlim = c(0, max(x)), ylim = c(0, max(y)))
grid(nx = 0, ny=10)

• Note: default color is "lightgray", which is often invisible in print
• In many cases grids are a waste of ink.

plot(x, y, 
     # cex = 2,
     # cex.axis = 2,
     xlab = "xlab", ylab = "ylab", # cex.lab = 2,
     main = "main", # cex.main = 2,
     xlim = c(0, max(x)), ylim = c(0, max(y)))

plot(x, y, 
     cex = 2,
     # cex.axis = 2,
     xlab = "xlab", ylab = "ylab", # cex.lab = 2,
     main = "main", # cex.main = 2,
     xlim = c(0, max(x)), ylim = c(0, max(y)))

plot(x, y, 
     cex = 2,
     cex.axis = 2,
     xlab = "xlab", ylab = "ylab", # cex.lab = 2,
     main = "main", # cex.main = 2,
     xlim = c(0, max(x)), ylim = c(0, max(y)))

plot(x, y, 
     cex = 2,
     cex.axis = 2,
     xlab = "xlab", ylab = "ylab", cex.lab = 2,
     main = "main", cex.main = 2,
     xlim = c(0, max(x)), ylim = c(0, max(y)))

plot(x, y, col = "red", pch = 20)
points(x = c(4, 5, 6), y = c(2,6,8), col = "green", pch = 2)

plot(x, y, col = "red", pch = 20)
points(x = c(4, 5, 6), y = c(2,6,8), col = "green", pch = 2)
legend ("topleft", 
        inset = c(0.01,0.01),    # distance from the margins
        pch = c(20,2),           # the two point characters we used
        col = c("red", "green"), # the two colors we used
        legend = c("red dots", "green triangles"))

• In scatterplots, you often don't see the wood for the trees
• So you might want to visualize a general trend
• To this end, you can add a regression line
• i.e. the straight line that is closest to all points

• We use the lm function for generating the model and the abline function, which adds straight lines to a plot

plot(x, y)
model <- lm(y ~ x)
abline(model)

• lowess/loess: locally weighted polynomial regression models
• "The basic idea underlying smoothers is to use the observations in a given span (or bin) of values of X to calculate the average increase in Y . You then move this span from left to right along the horizontal axis, each time calculating the new increase in y." (Baayen 2008: 34)

plot(x,y, main = "lowess")
lines(lowess(x, y))
scatter.smooth(x,y, main = "loess")

plot(x, y, type = "l")

plot(x, y, type = "b")

• Line types can be customized using the lty parameter:

plot(x, y, type = "b", lty = 2)
lines(x = c(2:7), y = c(4:9), lty = 3, col = "darkgrey")

• Lineplots are useful for showing e.g. change over time
• Count variable on y-axis, (at least) ordinal variable on x-axis
• Why should we avoid nominal variables on the x-axis?

• Lineplots are useful for showing e.g. change over time
• Count variable on y-axis, (at least) ordinal variable on x-axis
• Why should we avoid nominal variables on the x-axis?

• Use read.csv to read in the dataframe height_weight.csv
• Plot height against weight.

hw <- read.csv("examples/height_weight.csv")
plot(hw$height, hw$weight, 
     xlab = "Height", ylab = "Weight", main = "Height~Weight")
model_hw <- lm(hw$weight~hw$height)
abline(model_hw, col = "darkgrey", lty = 2)

• Use read.csv to read in the dataframe Pokemon.csv
• Plot height_m against weight_kg.
• Use the col parameter to show the color of each Pokémon, as indicated in the "Color" column.
• Use the pch parameter to show the form of each Pokémon, as indicated in the "Form" column.
• Hint: For the pch part, first look what happens when you try as.numeric(pok$Form)
• Finally, add a regression line to the plot.

# read data
pok <- read.csv("examples/Pokemon.csv")

# plot
plot(pok$Height_m, pok$Weight_kg, col = pok$Color, pch = as.numeric(pok$Form),
    xlab="Height", ylab="Weight", main = "Height~Weight, Pokémon")
model_pok <- lm(pok$Weight_kg~pok$Height_m)
abline(model_pok)

• Barplots are useful to show counts of categorical variables (e.g. number of men vs. number of women in parliament)...
• summary statistics (usually: means) of metric variables across different categories (e.g. mean height of humans vs. Klingons)
• But beware: Bar plots can hide information, cf. #barbarplots ("Friends don't let friends do bar charts")

• Main argument of the barplot() function is height
• This can be a vector or a matrix
• Let's try it out:

# define a vector
bar_heights <- c(50, 80)
barplot(bar_heights)

• The labels of the bars can be specified using names.arg:

barplot(bar_heights, names.arg = c("stuff", "more\nstuff"))

• The other arguments are largely the same as in the case of scatterplots:

barplot(bar_heights, names.arg = c("stuff", "more\nstuff"),
        main = "I'm a barplot", xlab = "I'm the x label", ylab = "I'm the y label",
        cex.main = 2, cex.lab = 2, cex.axis = 2, cex.names = 2)

• The space argument defines the space between bars, default is 0.2 if height is a vector

• The space argument defines the space between bars, default is 0.2 if height is a vector

• The space argument defines the space between bars, default is 0.2 if height is a vector

• The space argument defines the space between bars, default is 0.2 if height is a vector

• The space argument defines the space between bars, default is 0.2 if height is a vector

• Knowing the width of bars and the space between them is important if you want to add text
• To simplify the task, you can set space = 0

barplot(bar_heights / sum(bar_heights), # get relative frequencies
        names.arg = c("stuff", "more\nstuff"), space = 0)
text(x = c(0:1)+0.5,
     y = (bar_heights / sum(bar_heights)) - 0.05,
     labels = bar_heights)

• Or you can use the magic of vector addition and multiplication
• 0.5:1.5 yields {0.5,1.5}, 0.2 * 1:2 yields {0.2, 0.4} (= 0.2 * 1, 0.2 * 2)

barplot(bar_heights / sum(bar_heights), # get relative frequencies
        names.arg = c("stuff", "more\nstuff"))
text(x = (0.5:1.5)  + (0.2 * 1:2),
     y = (bar_heights / sum(bar_heights)) - 0.05,
     labels = bar_heights)

bar_matrix <- matrix(c(2,4,5,4,3,3,7,6), nrow = 2)
bar_matrix

##      [,1] [,2] [,3] [,4]
## [1,]    2    5    3    7
## [2,]    4    4    3    6

barplot(bar_matrix)

barplot(bar_matrix, beside = T)

Hands-on example: action-sentence compatibility task

• Task: Read in file actionsentence.csv with read.csv
• Inspect the data using head, str, and View
• Subset the data: Omit rows with direction == "distractor"

rt <- read.csv("examples/actionsentence.csv", fileEncoding = "UTF8")
rt <- subset(rt, direction != "distractor")

• No we want to show the means for "toward" and "away" sentences using a barplot.
• We want to abstract over the individual subjects, so it makes sense to transpose the table from wide to long format first.

library(reshape2)
rt2 <- melt(rt, id.vars = c("ID", "sentence", "direction"))

• No we want to show the means for "toward" and "away" sentences using a barplot.
• We want to abstract over the individual subjects, so it makes sense to transpose the table from wide to long format first.

library(reshape2)
rt2 <- melt(rt, id.vars = c("ID", "sentence", "direction"))

# same result but with (imho) more complicated syntax: "gather" from tidyr package
library(tidyr)
rt3 <- gather(rt, variable, value, -ID, -sentence, -direction)
all(rt2==rt3) # cheks if all values are identical

• No we want to show the means for "toward" and "away" sentences using a barplot.
• We want to abstract over the individual subjects, so it makes sense to transpose the table from wide to long format first.

library(reshape2)
rt2 <- melt(rt, id.vars = c("ID", "sentence", "direction"))

# same result but with (imho) more complicated syntax: "gather" from tidyr package
library(tidyr)
rt3 <- gather(rt, variable, value, -ID, -sentence, -direction)
all(rt2==rt3) # cheks if all values are identical

## [1] TRUE

# get mean values for "away" and "towards" subsets:
rt2_away <- subset(rt2, direction == "away")
rt2_toward <- subset(rt2, direction == "toward")

mean_away <- mean(rt2_away$value)
mean_toward <- mean(rt2_toward$value)

# combine both to one vector
rt_means <- c(mean_away, mean_toward)
rt_means

## [1] 1867.9 1789.5

barplot(rt_means, names.arg = c("away", "toward"))

• The mean alone doesn't say very much
• As mentioned before, [5,5,5,5,995] has the same mean as [1,180,300,223,146]
• This is why researchers tend to add error bars to barplots
• in most cases, these error bars represent 95 % confidence intervals...
• i.e. the interval where we can be 95% confident that it contains the true mean.

• We can obtain the confidence intervals for each of our two means using the t.test() function

t_away <- t.test(rt2_away$value)
str(t_away)

## List of 9
##  $ statistic  : Named num 14.8
##   ..- attr(*, "names")= chr "t"
##  $ parameter  : Named num 29
##   ..- attr(*, "names")= chr "df"
##  $ p.value    : num 4.65e-15
##  $ conf.int   : atomic [1:2] 1610 2126
##   ..- attr(*, "conf.level")= num 0.95
##  $ estimate   : Named num 1868
##   ..- attr(*, "names")= chr "mean of x"
##  $ null.value : Named num 0
##   ..- attr(*, "names")= chr "mean"
##  $ alternative: chr "two.sided"
##  $ method     : chr "One Sample t-test"
##  $ data.name  : chr "rt2_away$value"
##  - attr(*, "class")= chr "htest"

• We can use the arrows function to plot confidence intervals
• arrows is usually used to draw arrows (duh!)
• But these arrows can be customized in very useful ways...

plot(c(1:10), c(1:10), type = "n")
arrows(x0 = 2, x1 = 4,y0 = 5, y1 = 5)
arrows(x0 = 8, x1 = 8, y0 = 5, y1 = 9, 
       angle = 90,  # set angle to 90 degrees = flat arrow head
       code = 3)    # draw arrow head on BOTH ends of the "arrow"

ci_away <- t.test(rt2_away$value)$conf.int
ci_toward <- t.test(rt2_toward$value)$conf.int

barplot(rt_means, names.arg = c("away", "toward"))
par(xpd=T)
arrows(x0 = 0.7, x1 = 0.7, y0 = ci_away[1], y1 = ci_away[2], angle = 90, code = 3, length = .2)
arrows(x0 = 1.9, x1 = 1.9, y0 = ci_toward[1], y1 = ci_toward[2], angle = 90, code = 3, length = .2)

• Read in the file avengers.csv
• Plot the screentimes of Thor and Iron Man in the three Avengers movies as a dodged barplot (i.e. a barplot with side-by-side bars).
• Hint: For a dodged barplot, barplot() needs a matrix as input. This is a bit tricky - toy around with matrix() to create a matrix that looks like this (without the row and column names)

##          Avengers1 Avengers2 Avengers3
## Iron Man        37        45        18
## Thor            25        14        14

• Another hint: It might help to reorder the data. Using avengers[order(avengers$character),] you can sort them by the "character" column.

# read in data
avengers <- read.csv("examples/avengers.csv")

# sort by "character" column
avengers <- avengers[order(avengers$Character),]
avengers_matrix <- matrix(avengers$Screentime, ncol = 3, byrow = T)

# plot
barplot(avengers_matrix, beside = T, names.arg = c("Avengers 1", "Avengers 2", "Avengers 3"),
        legend.text = c("Iron Man", "Thor"))

With the help of graphical parameters, you can change the appearance of your plot. See ?par for more information. Some of the most important ones:

• mar: margins.
• xpd: If TRUE, you can plot outside the plot region. If FALSE (the default), plotting is confined to the plot region.
• mfrow: numer of c(rows, columns)
• bg: background color (or no color if you choose "transparent"; default is white)
• Type par() to see the current settings (= the default values if you haven't changed them). This can come in handy if you want to change parameters and then restore the defaults afterwards.
• You can even store the current values as an object by typing e.g. par_default <- par() and later on restore the current settings via par(par_default).

• In RStudio, you can use the "Export" button in the plot window
• However, the graphics files generated this way have low resolution (72dpi), unless you export an svg image
• This is why you should use png(), tiff(), jpeg(), or bmp() instead.

png(filename = "myplot.jpg")
plot(x,y)
dev.off()

• Saving graphs as vector graphics (SVG) has its advantages...
• but not all programmes can handle SVG files ☹
• Most publishers request PNG or TIFF files (some are also ok with JPG or BMP)

• Often you'll want to arrange plots in rows and/or columns
• You have already encountered par(mfrow=c(nrow,ncol))
• More complex arrangements are possible with layout
• To use this function, you first have to define a matrix

m <- matrix(c(1,2,2,
              1,2,2),
           nrow = 2, byrow = T)

layout(m)
barplot(c(20, 40), names.arg = c("a", "b"))
plot(x = c(7,9,15,24), y = c(8,15, 40, 32), type = "l", ylab="y", xlab="x")

par(par_cur)

• In order to export the plot, we just add the graphics device commands png() (or the like) and dev.off():

png("myfile.png", width = 7, height = 7, un = "in", res = 300)
layout(m)
barplot(c(20, 40), names.arg = c("a", "b"))
plot(x = c(7,9,15,24), y = c(8,15, 40, 32), type = "l", ylab="y", xlab="x")
dev.off()

## quartz_off_screen 
##                 2

par(par_cur)

• lattice graphics
• ggplot2
• plotly / shiny for interactive visualizations
• GoogleVis motion charts
• and a lot more!

• Baayen, R. Harald. 2008. Analyzing Linguistic Data: A Practical Introduction to Statistics Using R. Cambridge: Cambridge University Press.
• Freeman, Jenny V., Stephen John Walters & Michael J. Campbell. 2008. How to display data. Malden, Mass: BMJ Books.
• Rifkin, Erik & Edward Bouwer. 2007. The illusion of certainty: health benefits and risks. New York, NY: Springer.
• Tufte, Edward R. 2001. The visual display of quantitative information. 2nd ed. Cheshire: Graphics Press.
• Wainer, Howard. 1984. How to display data badly. The American Statistician 28(2). 137–147.