1. Clarity of Data
2. Trellis Displays
3. Clarity of Plot Elements
4. Clarity of Understanding
5. Scales & Axes
6. Summary of Principles

Good plots should:

• Make the data stand out
• Make it easy for the reader to decode the data
• Avoid unnecessary “chart junk”
• Use large enough plot elements to see and distinguish data

MyData = data.frame(x=rnorm(15), y=rnorm(15))
ggplot(MyData,aes(x,y)) + geom_point(size=5) + 
  theme_bw() + 
  theme(panel.grid.major=element_blank(),
        panel.grid.minor=element_blank(),
        text = element_text(size=20, family="Times")) +
  geom_rug(color="gray")

library(dplyr)

MyData = data.frame(state.x77[3:8,])
MyData$Name = rownames(MyData)

ggplot(MyData, aes(x=reorder(Name,-Area),y=Area)) + 
  geom_bar(stat="identity", fill="darkred") +
  geom_hline(yintercept=seq(from=0,to=150000,by=25000), color="white", size=1.25) +
  theme_bw() + 
  theme(panel.grid.major=element_blank(),
        panel.grid.minor=element_blank(),
        text = element_text(size=20, family="Times")) +
  xlab("State") +
  ylab("Area (square miles)")

• When there are many, overlapping plot elements (points, lines, etc.) plots can become difficult to read
• Can reduce the size of elements, but only so much
• Can remove fill color (or use transparency), but this can make distinctions harder
• Can use different colors or shapes, but again only so much
• Sometimes we can “jitter” points to make them easier to see, but can be harder to understand
• Sometimes we must separate the data
• Often it is better to summarize in some way if there’s too much data to make sense of

WineData<-read.csv ("http://eecs.ucf.edu/~wiegand/ids6938/datasets/r-lab1.dat")
ggplot(WineData, aes(Alcohol, WineType)) + 
  geom_point(shape=21, size=5) +
  theme(text = element_text(size=20, family="Times"))

ggplot(WineData, aes(Alcohol, WineType)) +  
  geom_jitter(shape=21, size=5) +
  theme(text = element_text(size=20, family="Times"))

ggplot(WineData, aes(Alcohol, WineType, fill=WineType)) +  
  geom_jitter(shape=21, size=5) +
  theme(text = element_text(size=20, family="Times"))

ggplot(WineData, aes(y=Alcohol, x=WineType, fill=WineType)) +  
  geom_boxplot() + 
  coord_flip() +
  theme(text = element_text(size=20, family="Times"))

MyData = data.frame(x=rnorm(1500), y=rnorm(1500))
ggplot(MyData, aes(x,y)) +
  geom_point(size=5, alpha=0.4, color="steelblue") +
  theme(text = element_text(size=20, family="Times"))

ggplot(MyData, aes(x,y)) +
  geom_point(size=2, color="steelblue") +
  theme(text = element_text(size=20, family="Times"))

n=40
MyData = data.frame(x=c(rnorm(n),0.6*rnorm(n)+1,1.1*rnorm(n)-1),
                    y=c(rnorm(n),0.6*rnorm(n)+1,1.1*rnorm(n)-1),
                    Type=c(rep("Thing A",n),rep("Thing B",n),rep("Thing C",n)))
ggplot(MyData, aes(x,y,color=Type,shape=Type)) + geom_point(size=5)

• Separating data on multiple plots creates a “lookup” problem for the reader when decoding
• Putting too much data on one plot creates a “clutter” problem for the reader when decoding
• So sometmes we try to separate plots, but organize the plots so that their scales and axes are the same, well aligned and close together
• These called trellis displays

In ggplot we create a trellis using a facet and a model

bwt = birthwt$bwt
smoke = factor(c("Mother Didn't Smoke","Mother Smoked")[birthwt$smoke+1])
MyData = data.frame(bwt,smoke)

ggplot(MyData, aes(x=bwt)) +
  geom_histogram(fill="white",color="black", binwidth=500) +
  facet_grid(smoke ~ .) +
  xlab("Birth Weight") +
  ylab("Count") + 
  theme(text=element_text(size=20, family="Times"))

ggplot(mpg, aes(x=displ, y=hwy)) +
  geom_point(size=3) +
  facet_grid(drv ~ class) +
  xlab("Engine Displacement (liters)") +
  ylab("MPG on Highway") + 
  theme(text=element_text(size=20, family="Times"))

ggplot(mpg, aes(x=manufacturer,y=hwy)) + 
  geom_bar(stat="identity") + 
  facet_grid(year ~ .) +
  xlab("Car Manufacturer") +
  ylab("MPG on the Highway") + 
  theme(text=element_text(size=20, family="Times"))

Cylinders=mtcars$cyl
MPG=mtcars$mpg
Gears=factor(paste(mtcars$gear,"Gears"))
MyData = data.frame(MPG,Cylinders,Gears)

ggplot(MyData, aes(x=factor(Cylinders), y=MPG)) + 
  geom_boxplot() + 
  facet_grid(Gears ~ .) +
  xlab("Number of Engine Cylinders") +
  ylab("Average MPG") + 
  theme(text=element_text(size=20, family="Times"))

• Place the things you most wish to compare close together
• Encode data that requires the most nuanced differentiation with the way we perceive best (e.g., position)
• Save other harder to perceive elments for numeric data that requires less nuanced differentiation (e.g., area, color)
• Label all axes, include units where relevant
• But don’t clutter the plot with labels
• And don’t let data elements obscure labels
• Annotate important elements
• Use legends & keys to clarify distinctions
• Plots look different when reduced / typeset / projected
• Proofread your graphs just as you do your prose

ggplot(data.frame(state.x77), aes(x=Population, y=Income)) + 
  theme(text = element_text(size=20, family="Times")) +  
  geom_text(aes(label=state.abb),hjust=-0.25,size=4) + 
  ggtitle("State Income vs. Population")

• Adding reference lines, points, or regions can help contextualize your plot
• Adding annotations provide explicit context
• All of these mechanisms help focus the reader’s attention on your narrative

ggplot(longley, aes(x=Year, y=Armed.Forces)) + 
  geom_rect(xmin=1950.2,xmax=1953.5,ymin=140,ymax=370,fill="lightgreen",alpha=0.03) +
  geom_text(x=1951.55,y=366,label="Korean War", color="darkgreen") +
  geom_line(size=1.5) + 
  theme(text=element_text(size=20, family="Times")) + 
  ylab("Number of People in Armed Forces (thousands)") +
  ylim(c(150,360))

library(dplyr)
library(ggplot2)
crime <- read.csv('Crime/fbi-crime-1996-2015.csv', header=T)
ggplot(arrange(crime, Year), aes(x=Year, y=Murder.and..nonnegligent..manslaughter..rate.)) +
    geom_line(size=1.35, col="darkblue") +
    geom_point(size=4, shape=21, fill="white", color="darkblue") +
    ylab("Murder & Manslaughter Rate in U.S (per 100K people)") +
    theme(text=element_text(family="Times", size=16)) +
    geom_hline(yintercept=7.7, linetype="dashed", color="black") +
    annotate("text",2005,7.75,vjust=0,label="Murder Rate of Bolivia, 2011", color="black") +        
    geom_hline(yintercept=3.7, linetype="dashed", color="black") +
    annotate("text",2005,3.75,vjust=0,label="Murder Rate of Chile, 2011", color="black")  +
    annotate("text", 1996, 4.5, hjust=0, label="The U.S. Murder Rate was the lowest in the last\n100 years in 1955, when it was 4.5 per 100K people") +
    ggtitle("Murder in the U.S. is at an Historic Low")

• Draw data to the correct / consistent scale
• Make sure there are no computational innaccuracies
• Don’t make the reader do math your computer could have done for you
• Properly baseline viusalizations when possible
• When grouping, be consistent with order, color, and other elements

• When using a bar graph, begin the scale at zero and end the scale a little above the highest value
  • Recall that bar graphs use both length and position to encode a number
  • When it isn’t properly baselined, these do not agree, which creates a cognitive mismatch
  • Can mislead readers because pre-attentively, they fixate on the differences between lengths … though length is no longer an accurate encoding

• With other types of graph, begin and end a little above or below the extrema

• Make intervals on the scale easy numbers to understand (e.g., round numbers)

• Good visualization software will make following such rules easy

Pie Chart of Loans

Positive, but negative bar chart

• WTF Visualizations tumblr feed
• Accidental aRt

• Typically bar plots should almost always be baselined at zero
  (length / position perceptions are similar only when the baseline is zero)
• Unless there is a good reason, baseline at zero for other plots, as well
• Good reasons to baseline at a different value (for non-bar plots):
  • The range of the data hides relevant changes because of its magnitude
  • Length encodings (e.g., confidence intervals or error bars) are obscured because of data magnitude
  • Baseline value has some contextual meaning in the narrative (e.g., there is a “natural” baseline in the data)
• Always make it clear to the reader when a graph is not baselined at zero

When data contains a few very extreme values, it can be difficult to display data appropriately. Things to consider:

• Consider separating long-tail values into a separate plot
• Consider using a logarithmic scale instead
• Try not to break the scale because it can lead to data misrepresentation
• If you must, then make the break as easy as possible for the reader to decode

• “Unskews” data with large values
• Helps spread out data with long-tails by bringing the tails in and using more of the plotting space for the high-density data
• But requires a bit more sophisticated understanding from the reader

• Can be confusing to decode
• It’s easy to mislead reader about the scale
• Difficult to accomplish in most plotting tools
• Typically it is easier to separate data or repeat a plot in a different scale

• Good data visualization is not an afterthought, but takes thought and attention
• Producing a graph is an iterative process, proofread and correct
• Try to create graphs such that the main message supporting the narrative is very easy to decode
• Use visual elements to maximize data content
• Choose grouping and encoding mechanisms that focus on your narative and maximize differentiation
• Avoid misrepresenting the data numerically or visually