library(ggplot2); theme_set(theme_bw())
library(rainbow)  ## bagplots etc.
library(ggthemes)
library(directlabels)
theme_update(panel.spacing=grid::unit(0,"lines"))
library(cowplot) ## for arranging multiple plots, labeling, etc.
library(Hmisc)
library(dplyr)

• simplicity
• speed
• flexibility
• robustness
• parsimony

Distribution of horsepower:

stem(mtcars$hp)

## 
##   The decimal point is 2 digit(s) to the right of the |
## 
##   0 | 5677799
##   1 | 0011111122
##   1 | 55888888
##   2 | 123
##   2 | 556
##   3 | 4

ggplot(mtcars,aes(cyl,hp,group=cyl))+geom_boxplot()

(Rousseeuw et al. 1999) In ggplot (note hidden code):

ggplot(iris, aes(Sepal.Length, Sepal.Width, colour=Species,
                 shape=Species, fill=Species))+  geom_point()+ geom_bag()

The rainbow package implements functional boxplots, for high-dimensional (functional) data analysis (also fda, roahd packages): it uses various forms of projection or dimension reduction, followed by a bagplot of the first two projected dimensions

rainbow::fboxplot(data = ElNino_ERSST_region_1and2,
                  plot.type = "bivariate",
                  type = "bag", projmethod="PCAproj")

• yes (principles)
• data size/complexity, computing power both increasing

• accuracy of quantitative representation
• visual estimation of differences

• show participants the same data in different formats
• ask them questions about relative magnitudes

• yes!
• Elliott (2016), “39 studies about human perception in 30 minutes”
  • healthy tradition of scientific experiments on graphical perception
    • accuracy
    • memory
    • preference

Heer et al. (2010)

• functional, minimal graphics
• maximize data-ink / minimize non-data-ink
• don’t lie (lie factor)
• small multiples (= trellis/lattice, facets, panels)
• “As for a picture, if it ain’t worth 1000 words, the hell with it” - Ad Reinhardt
• information at the point of need (legends etc.)
• Powerpoint sucks

“Understand that Tufte’s ideas are a good starting point, not a religion” Robert Kosara

• maximize data ink (within reason)

g0 <- ggplot(OrchardSprays,aes(treatment,decrease))+scale_y_log10()
print(plot_grid(g0 + geom_boxplot(),  g0 + geom_tufteboxplot(stat="boxplot")))

ggthemes::geom_tufteboxplot() (stat="boxplot" for Tukey-style definition)

• less eye movement is better
• direct labels > legends > info in caption > info in text

g1 <- ggplot(iris,aes(Sepal.Length,Petal.Length,colour=Species,
                shape=Species))+geom_point()
print(plot_grid(g1,direct.label(g1)))

• manually if necessary
• directlabels package
  • works with lattice and ggplot graphics
  • variety of labeling choices, e.g. last.bumpup: “Label last points, bumping labels up if they collide”
  • documentation
• related: ggrepel (auto-repelling text labels)

• (Continuous) response on the \(y\)-axis
  • assumes we have a single, quantitative/ordered (continuous or discrete) response variable; multivariate responses more challenging
• put most salient predictor on the \(x\)-axis
  • highest value in Cleveland hierarchy
  • if most important predictor is categorical, use most important continuous predictor on \(x\)-axis
  • if most important predictor has few categories, use next most important predictor with many categories

• Put most salient comparisons within the same subplot (distinguished by color/shape), and nearby within the subplot when grouping bars/points
• Facet rows > facet columns

• Use transparency to include important but potentially distracting detail
• Do category levels need to be identified or just distinguished? (Direct labeling, e.g. via directlabels package)
• Order categorical variables meaningfully (“Alabama/Alberta” problem)
• Think about whether to display population variation (standard deviations, boxplots) or estimation uncertainty (standard errors, mean \(\pm\) 2 SE, boxplot notches)
• Try to match graphics to statistical analysis, but not at all costs
• Choose colors carefully (RColorBrewer/ColorBrewer, IWantHue/hues package: respect dichromats and B&W printouts (dichromat & colorblindr & cividis packages: Sciani (2018))

• small show all points, possibly dodged/jittered, with some summary statistics: dotplot, beeswarm. Simple trends (linear/GLM/loess)
• medium boxplots, loess, histograms, GAM (or linear regression)
• large modern nonparametrics: violin plots, hexbin plots, kernel densities: computational burden, and display overlapping problems, relevant
• combinations or overlays where appropriate (beanplot; rugs+scatterplot; pirate plots

Some examples (from a screed on “dynamite plots”):

a. the dreaded “dynamite plot”. Problems:

• bar plot on logarithmic axis is inappropriate (anchors graph to arbitrary zero point)
• assumes distribution is symmetric (although this applies to b,c as well)
• some forms of this plot show only top whisker (makes comparison even harder)

b. inferential (point \(\pm\) 2 SE) plot

• same assumptions as dynamite plot
• less strongly anchored to zero

c. points \(\pm\) 1 and 2 SE

• de-emphasizes approximate 95% CI
• equivalent for Bayesian posterior intervals would typically show both 50% and 95% credible intervals (based on quantiles or highest posterior density)

d. points alone

• true to the data
• description only; provides no inferential help
• can confound sample size and range (larger samples have more extreme values so look more variable)

e. boxplots

• well-established
• “outliers” can be misleading (Dawson 2011)
• can add notches to indicate approximate 95% CI on medians (McGill et al. 1978)

f. violin plots

• mirror-image density plots
• best for large data sets
• may be funky for small/medium data sets
• can be combined with jittered data, segments indicating median/quantiles, etc.

• story
• github repo
• direct link to data