Graphics and Web Design Based on Edward Tufte's Principles
This is an outline of Edward Tufte's pioneering work on the
use of graphics to display quantitative information. It mainly
consists of text and ideas taken from his three books on the
subject along with some additional material of my own. This
page is in text only format: in order to understand the
concepts you need to read the books because the concepts cannot
really be grasped without the illustrations, and current video
monitor technology is too low in resolution to do them justice.
His work has been described as "a visual Strunk and
White" (here is a German
translation of this article).
Throughout this outline I have included references to the
illustrations in his books that are labeled with the abbreviations
VD-pp, VE-pp, and EI-pp, where "pp" is a page number and:
- VD is "the Visual Display of Quantitative Information"
- VE is "Visual Explanations"
- EI is "Envisioning Information"
Outline
- #Introduction
- #History
of Plots
- #The
Explanatory Power of Graphics
- #Basic
Philosophy of Approach
- #Graphical
Integrity
- #Data
Densities
- #Data
Compression
- #Multifunctioning
Graphical Elements
- #Maximize
data-ink; minimize non-data ink
- #Small
Multiples
- #Chartjunk
- #Colors
- #General
Philosophy for Increasing Data Comprehension
- #Techniques
for Increasing Data Comprehension
- #When
NOT to Use Graphics
- #Aesthetics
Tufte's works address the following issues:
- The Problem: The problem is that of
presenting large amounts of information in a way that is compact,
accurate, adequate for the purpose, and easy to understand.
Specifically, to show cause and effect, to insure that the proper
comparisons are made, and to achieve the (valid) goals that are
desired.
- Its Importance: Printed and
graphical information is now the driving force behind all of our
lives. It no longer is confined to specialized workers in selected
fields but impacts nearly all people through the widespread use of
computing and the Internet. Rapid and accurate transfers of
information can be a life and death matter for many people (an
example being the Challenger disaster). The extent to which
symbols and graphics affect our lives can be seen by the dramatic
increase in IQ scores in all cultures which have acquired
information technology: in the United States there has been an
average increase of 3 IQ points per decade over the last 60 years,
for a total of an 18 IQ point increase. There is no known
biological explanation for this increase and the most likely cause
is widespread exposure to text, symbols, and graphics that
accompany modern life. As mentioned above, this increase
has been seen in all cultures exposed to information technology.
- Its Application: Some of the
information relates to the displays of statistical information,
but much applies to any type of display, even plain text.
- The Solution: To develop a
consistent approach to the display of graphics which enhances its
dissemination, accuracy, and ease of comprehension.
The very first known plot dates back to the 10-th century (VD-28:
first known graph). This was about the same time that Guido of
Arezzo was developing the two-dimensional musical staff notation
very similar to the one we use today. In the 15-th century Nicolas
of Cusa developed graphs of distance versus speed. In the 17th
century Rene Descartes established analytic geometry which was used
only for the display of mathematical functions. But the main
initiator for informative graphics was William Playfair (1759-1823)
who developed the line, bar, and pie charts as we know them today.
The importance and explanatory power of graphics can be seen in
these examples:
- Illustration VD-13/14 shows 4 plots which have a large number
of absolutely identical statistical measures and properties and
yet are very different, as can be immediately seen from their
graphs.
- The Challenger disaster: the data graphs shown to NASA did not
convey the real information which was needed (VE-47 versus VE-45).
If NASA had seen the appropriate, but very simple, graphics
which showed the effects of low temperature and damage to the
solid rocket boosters, the Challenger would not have been launched
that (very cold) day.
- The Broad Street Pump cholera epidemic in 1854 in London, as
displayed by John Snow (VE-31: cholera deaths). This graph showed
clusters of cholera deaths around the site of the pump.
- Illustration VD-166: "communes in France" shows an extremely
dense plot which displays the boundaries of more than 30,000
communes in France.
Important rules and themes to use when presenting graphics:
- Assume that the audience is intelligent (a paraphrase from
E.B. White). Even publications, such as NY Times, assume that
people are intelligent enough to read complex prose, but too
stupid to read complex graphics.
- Don't limit people by "dumbing" the data -- allow people to
use their abilities to get the most out of it.
- To clarify -- add detail (don't omit important detail; e.g.,
serif fonts are more "detailed" than san serif fonts but are
actually easier to read). And Einstein once said that "an
explanation should be as simple as possible, but no simpler".
- Above all else, show the data. Graphics is "intelligence made
visible"
- Data rich plots can show huge amounts of information from many
different perspectives: cause & effect, relationships,
parallels, etc. (VD-31: train schedule, VD-17: Chloroplethic map,
VD-41: Napoleon's campaign, EI-49: space junk)
- Plots need annotation to show data, data limitations,
authentication, and exceptions (VE-32: text of exceptions)
- Don't use graphics to decorate a few numbers
In addition to "lies, damn lies, and statistics", graphics
can also be used to deceive. For example, deceptive graphics may:
- Compare full time periods with smaller time periods
(VD-60: Nobel prizes, which compares 10 year time periods with one
5 year period)
- Use a "lie factor" [= (size of graphic)/(size of data)]
to exaggerate differences or similarities
- Use area or volume representations instead of linear scales to
exaggerate differences. See VD-69: "Shrinking family
doctor" as an example of how to confuse people using 1
versus 2- and 3- dimensional size comparisons. Area and volume
representations fool people with the square/cube law: an increase
in linear size leads to a square of the increase for areas and a
cube of the increase for volumes.
- Fail to adjust for population growth or inflation in financial
graphs
- Make use of design variation to obscure or exaggerate data
variation (VD-61: exaggeration of OPEC prices)
- Exaggerate the vertical scale
- Show only a part of a cycle so that data from other parts of
the cycle cannot be used for proper comparison
Graphical errors may be more common today than in the past due to
the easy and frequent use of computers. Guidelines to help
insure graphical integrity include the following:
- Avoid chartjunk
- Don't dequantify: provide real data as accurately as is
reasonable. For example, ranking products as better or worse
according to one criteria when several factors are involved is
often not useful unless the magnitudes of the differences are
indicated.
- Don't exaggerate for visual effects, unless it is needed to
convey the information. Sometimes such exaggerations are
essential: for example, it is virtually impossible to show
both the size and the orbits of planets at the right scale on the
same chart. On the other hand, illustration VE-24: "Exaggerated
vertical Venus scale", shows such dramatic mis-information, that
one researcher called for the formation of "a flat Venus
society".
- Avoid dis-information: thick surrounding boxes and underlined
san serif text make reading more difficult
- Watch out for effects of aggregation: e.g., dot maps are often
more honest in this respect than chloroplethic maps which group
results based on (sometimes arbitrary) boundaries.
- Ask the right questions:
- Does the display tell the truth
- Is the representation accurate
- Are the data documented
- Do the display methods tell the truth
- Are appropriate comparisons, contrasts, and contexts shown
Graphics are at their best when they represents very dense and
rich datasets. Tufte defines data density as follows:
Data density = (no. of
entries in data matrix)/(area of graphic)
Note that low data densities on computer displays
force us to view information sequentially, rather than spatially,
which is bad for comprehension. Good quality graphics are:
- Comparative
- Multivariate
- High density
- Able to reveal interactions, comparisons, etc
- And where nearly all of the ink is actual data ink
Example data densities include:
- 110,000 numbers/sq-inch for an astronomical graph. This is the
maximum known density for a graph. For most scientific
journals we get about 50-200 numbers/sq-inch
- 150 Mbits = human eye
8 Mbits = typical computer screen
25 Mbits = color slide 150 Mbits = large foldout map
28,000 Characters = Reference book 18,000 Characters =
phone book 15,000 Characters = non-fiction
An excellent example of a data rich plot is a graphical train
schedule (VD-31: train schedule) which shows start and stop times,
locations, directions, routes, transfers, and speeds all on one
sheet of paper.
- Use data compression to reveal (not hide) data . For example,
EI-22: "Sun Spot cycles" displays sunspots as thin vertical lines
in the y-axis direction only in order to present many such spots
over a period of time on a single graph
- Use compression to show lots of information in a single graph,
such as a plot that shows x-axis, y-axis, and x/y interactions.
(VD-134: Pulsar signals; VE-111)
- Exclude bi-lateral symmetry when it is redundant (e.g.,
charnoff faces) or extend it when it aids comprehension (50% more
view of the world on a world map provides a wrap-around context
that aids understanding). Studies show that we often concentrate
on one side of a symmetrical figure and only glance at the other
side.
Graphical structures can often serve several purposes once.
For example,
- Stem and leaf plots display sequences of numbers which
directly portray structure by the physical length of each
sequence. (VD-140: stem/leaf; VD-141: army divisions; VD-143:
Normal curve)
- The Consumer Reports listing of automobile defects (VD-174:
Consumer Reports) reveal a micro/macro structure: the overall
display of black ink immediately reveals which cars are most
troublesome, whereas each individual element in the display
identifies a particular weakness.
- The data grid itself may be the data, revealing both the
values and the coordinate system at the same time (VD-152:
data-based markers)
Tufte defines the data ink ratio as:
Data Ink Ratio =
(data-ink)/(total ink in the plot)
The goal is to make this as large as is
reasonable. To do this you:
- Avoid heavy grids
- Replace box plots with interrupted lines (VD-125: reduced box
plot)
- Replace enclosing box with an x/y grid
- Use white space to indicate grid lines in bar charts (VD-128:
white spaces)
- Use tics (w/o line) to show actual locations of x and y data
- Prune graphics by: replacing bars with single lines, erasing
non-data ink; eliminating lines from axes; starting x/y
axes at the data values [range frames])
- Avoid over busy grids, excess ticks, redundant representation
of simple data, boxes, shadows, pointers, legends. Concentrate on
the data and NOT the data containers.
- Always provide as much scale information (but in muted form)
as is needed
Small multiples are sets of thumbnail sized graphics on a single
page that represent aspects of a single phenomenon. They:
- Depict comparison, enhance dimensionality, motion, and are
good for multivariate displays (VD-114: particle momentum)
- Invite comparison, contrasts, and show the scope of
alternatives or range of options (VE-111: medical charts)
- Must use the same measures and scale.
- Can represent motion through ghosting of multiple images
- Are particularly useful in computers because they often
permit the actual overlay of images, and rapid cycling.
Chartjunk consists of decorative elements that provide no data
and cause confusion.
- Tufte discusses the rule of 1+1=3 (or more): 2 elements in
close proximity cause a visible interaction. Such interactions can
be very fatiguing (e.g., moiré patterns, optical vibration) and
can show information that is not really there (EI-60: data that is
not there, VD-111: chart junk)
- In major science publications we see 2% to 20% moiré
vibration. For example, in recent statistical and computer
publications chartjunk ranges from 12% to 68%
- Techniques to avoid chartjunk include replacing crosshatching
with (pastel) solids or gray, using direct labeling as opposed to
legends, and avoiding heavy data containers
Colors can often greatly enhance data comprehension.
- Layering with colors is often effective
- Color grids are a form of layer which provides context but
which should be unobtrusive and muted
- Pure bright colors should be reserved for small highlight
areas and almost never used as backgrounds.
- Use color as the main identifier on computer screens as
different objects are often considered the same if they have the
same color regardless of their shape, size , or purpose
- Contour lines that change color based on the background
standout without producing the 1+1=3 effects
- Colors can be used as labels, as measures, and to imitate
reality (e.g., blue lakes in maps).
- Don't place bright colors mixed with White next to each other.
- Color spots against a light gray are effective
- Colors can convey multi-dimensional values
- Scroll bars should be solid pastel colors
- Note that surrounding colors can make two different colors
look alike, and two similar colors look very different (EI-92/93:
effects of context on colors).
- Subtle shades of color or gray scale are best if they are
delimited with fine contour lines (EI-94: shades with contours)
- Be aware that 5-10% of people are color blind to some degree
(red-green is the most common type followed by blue-yellow, which
usually includes blue-green)
- High density is good: the human eye/brain can select, filter,
edit, group, structure, highlight, focus, blend, outline, cluster,
itemize, winnow, sort, abstract, smooth, isolate, idealize,
summarize, etc. Give people the data so they can exercise their
full powers -- don't limit them.
- Clutter/confusion are failures of design and not complexity
- Information consists of differences that make a difference: so
you can "hide" that data which does not make a difference in what
you are trying to depict
- In showing parallels, only the relevant differences should appear
- Value and power of parallelism: once you have seen one element
all the others are accessible
- Important concepts in good design: separating figure and
background (for example, a blurry background often brings the
foreground into sharper focus), layering & separation, use of
white space (e.g., Chinese landscapes emphasize space, as in the
painter known as "one corner Ma"; oriental music is often there to
emphasize the silence and not the sound).
- Graphics should emphasize the horizontal direction
To increase data comprehension you:
- Make marks or labels as small as possible, but as small as
possible to still be clear.
- Avoid pie charts as they are low density and fail to order
values along a visual dimension
- Usually use dot maps in place of chloroplethic maps
because they show more exact detail
- Closely interweave text and graphics: attach names
directly to parts, place small messages next to the data, avoid
legends if possible and annotate the data directly on the graph
(VE-99: anatomy of a font)
- Avoid abbreviations if possible, and use horizontal text
- Use serif fonts in upper/lower case
- Use transforms of scaling if they (honestly) can reveal
information which might otherwise be overlooked.
- Use different structures to reveal 3D and motion, such as
the exploded hexagon, true stereo, and extreme
foreshortening (as on the edge of a sphere: see EI-15 "exploded
hexagon")
- Often text tables can replace graphs for simple data; you can
also use 2D text tables, where row and column summaries are
useful. Non-comparative data sets usually belong in tables, not
charts
- Poster designs are meant just to capture attention, as opposed
to conveying information -- generally they are not good designs
for graphs.
- If a picture is not worth a 1000 words, to hell with it (quote
from Ad Reinhardt -- note this is from the original Chinese quote
that "a picture is worth 10,000 words).
Graphical excellence consists of simplicity of design and
complexity and truth of data. To achieve this
- Use words, numbers, drawings in close proximity
- Display an accessible complexity of data
- Let the graphics tell the story
- Avoid context-free decoration
- Use lines of different weights as an attractive and compact
way to display data (VD-185: Mondrian)
- Make use of symmetry to add beauty (although someone once said
that "all true beauty requires some degree of asymmetry")
| |
|
|
|