Types of Charts

Three principles for visualization

1. be true to your research – design your display to illustrate a particular point
2. maximize information, minimize ink –use the simplest possible representation for the bits you want to convey
3. organize hierarchically – what should a viewer see first? what if they look deeper?

Worked example

• Participants heard examples from an artificial language 
• Three different presentation methods for examples – index cards, list of sentences, mp3 files on ipod
•  Task was to spread $100 of "bets" across different continuations for a new example
• Dependent measure was bet on the correct answer

bar graph

bar graph with standard errors

bar graph with 95% CIs

box plot

viola plot

strip chart

strip chart with means

strip chart with means

Morals of the example

• Summary statistics 
  • almost always necessary
  • but at what level of analysis? 
• Distribution is important
  • what is the form of the data?
  • is your summary misleading?
• Fancier is not always better
  • pretty pictures are awesome
  • but not if they obscure the data

Source: Mike Frank and Ed Vul, https://ocw.mit.edu/courses/res-9-0002-statistics-and-visualization-for-data-analysis-and-inference-january-iap-2009/96df7f49cd50ed9bb0feb4793b9c8d89_lec1_visulzatn.pdf
This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 License.

Conventional visualizations

Histogram, Pie Chart, Stacked bar graph & Venn diagram

Histogram

• Important first way of looking at your data
• One dimensional
• Shows shape by binning a continuous distribution

Grouped histogram

Pie chart

• A whole split into parts
• Emphasizes that all parts sum to a constant
• Single dimension with discrete categories

Stacked bar graph

• Wholes split into parts
• Easy to compare
  • often better than pie chart
• Can have multiple discrete dimensions

Venn diagram

• Shows overlap between discrete groups
• Sometimes the only way to display overlapping sets
• Unintuitive
  • no "popout"

Scatter plot, Line graph & Bar graph

Scatter plot

• Relationship between observations on two continuous dimensions
• Can show multiple groups 
• Can show trend lines etc.
• Uninformative with too much data

Line graph

• Also ubiquitous!
• Good for showing one variable (e.g., time) as continuous even though you have discrete measures
• Can compare several discrete groups

Bar graph

• aka "dynamite plot"
• Ubiquitous!
• Can be used for lots of discrete grouping factors
• Natural semantics of grouping 
• Conceals data

More bar graphs

Heat map & Bubble plot

Heat map

• Works very well when there are natural semantics
• Color mapping can be problematic 
  • grayscale usually fine 
• Can be unintuitive

Bubble plot 

• Can be very intuitive
• Size is not perfectly quantitative

Trellis plots

Tips and Tricks

Three tricks for doing more with less

• Multiple plots
  • simple, easily interpretable subplots
  • can be beautiful but overwhelming
• Hybrid plots
  • a scatter plot of histograms
  • or a venn-diagram of histograms, etc. 
• Multiple axes
  • plot two (or more) different things on one graph

Hybrid plots

Multiple plots

Multiple axes

Two tradeoffs

Two tradeoffs

• Informativeness vs. readability
  • Too little information can conceal data
  • But too much information can be overwhelming
  • Possible solution: hierarchical organization? 
• Data-centric vs. viewer-centric 
  • Viewers are accustomed to certain types of visualization 
  • But novel visualizations can be truer to data

Information vs. readability

• Pirahã people of Brazil
  • Isolated indigenous group
  • No words for numbers 
• Previous research suggested that they were unable to do simple matching games
• Five matching games, 14 participants, quantities 4-10 (split among participants)

Data-centric vs. viewer-centric

• Web study of word learning 
  • n=700 
  • lots of noise
• varied number of objects with different properties
  • asked for bets
• had a model that predicted performance

Trouble

• High Dimensionality Doesn't Guarantee Excellence
  
  

• Messy bar graphs
  •  Sometimes you can discretize way too many variables
    
    
    
    
    
    
    
    
    
    
    
• Too much data for one plot
  
  
  
  
  
• Difficulty of comparison
  
  
  
  
  
• Bad semantics