Skip to content

Suggested color schemes for points, not regions?

3 messages · Kevin Wright, Martin Maechler, Ross Ihaka

Original
Kevin Wright
# 
I have seen a couple of posts about color schemes like those at the
ColorBrewer site.  Most recently:
http://geography.uoregon.edu/datagraphics/color_scales.htm

These color schemes can work very well for regions (bars, polygons,
images, etc.) but are not very suitable for points and/or lines.

Is anyone aware of research/suggestions for a color scheme to use for
scatter plots?  I've looked at great length and have found little on this
topic.

My current scheme of choice is a set of fairly saturated colors along the
lines of:
navy
brown/orange
black
purple
red
medium green
This is similar to the 'paired' color scheme, but using only the saturated
colors and substituting black for yellow.  Depending on circumstances, I
sometimes use a different glyph for each color.  The hard part about all
this is to make sure that each color/glyph combination has the same
'attention-getting' power.

Any discussion or comments are welcome.

Kevin Wright
Martin Maechler
# 
KevinW> I have seen a couple of posts about color schemes like those at the
    KevinW> ColorBrewer site.  Most recently:
    KevinW> http://geography.uoregon.edu/datagraphics/color_scales.htm

    KevinW> These color schemes can work very well for regions (bars, polygons,
    KevinW> images, etc.) but are not very suitable for points and/or lines.

Are you sure?

Currently, for R users, the most accessible of the "new" schemes is the
ColorBrewer (http://colorbrewer.org) one, since Erich Neuwirth
and others have been providing the RColorBrewer package on CRAN.

 install.packages("RColorBrewer")
 library("RColorBrewer")
 example(brewer.pal) # a 'show'

not only has schemes to be used for images/maps,
but also "Set"s of (pairwise well distinguishable) colors well
suited for scatter plots.   In Cynthia Brewer's terminology
these are the "qualitative" schemes.  

    KevinW> Is anyone aware of research/suggestions for a color
    KevinW> scheme to use for scatter plots?  I've looked at
    KevinW> great length and have found little on this topic.

    KevinW> My current scheme of choice is a set of fairly
    KevinW> saturated colors along the lines of:

    KevinW> navy
    KevinW> brown/orange
    KevinW> black
    KevinW> purple
    KevinW> red
    KevinW> medium green

    KevinW> This is similar to the 'paired' color scheme, but
    KevinW> using only the saturated colors and substituting
    KevinW> black for yellow.  Depending on circumstances, I
    KevinW> sometimes use a different glyph for each color.  The
    KevinW> hard part about all this is to make sure that each
    KevinW> color/glyph combination has the same
    KevinW> 'attention-getting' power.

hmm, from the above I tend to read that you've already looked at
(e.g.) 'Set1' or 'Dark2' colors from ColorBrewer but didn't like
them? 

Something that hasn't been properly considered by statisticians
AFAIK is the situation for color blind people, or at least the
most common one. This has just been brought up here two days
ago, see e.g.
  https://stat.ethz.ch/pipermail/r-help/2004-November/058724.html
which mentions new "color ramp" facilities in the future 
R 2.1.0 (2005-04-0x).
Cynthia Brewer and many others also mention and recommend
http://www.vischeck/com/ which allows checking your image/drawing.
There's a simulator trying show how a given picture is seen by
(even different kinds of) color deficient persons.

The colorbrewer.org site has a very nice tool (based on flash 5
plugin), which shows you the pros and cons of a color scheme you
chose (interactively).  She (Cynthia Brewer) uses 6 criteria
(with 3 levels  "ok", "doubtful/unknown", "not ok"):

 1) Color blind friendly   [the "red-green" deficiency]
 2) Photocopy friendly	   [for B&W photocopying: are differences preserved?]
 3) LCD Projector friendly [pastel colors may be problematic]
 4) Laptop (LCD) friendly
 5) CRT-(screen) friendly
 6) Color Printing friendly
- 5 (of 8) were "not ok" for color blinded.
- each had at least one "not ok";
  i.e., there's no "optimal color scheme" that works everywhere,
  but you have to change color schemes depending on the intended
  medium.  

    KevinW> Any discussion or comments are welcome.

I'm pretty sure Ross Ihaka will also chime in here.
(http://www.stat.auckland.ac.nz/~ihaka/colour/ is a first start).

As you see, I'm quite interested also.
Martin Maechler <maechler at stat.math.ethz.ch>	http://stat.ethz.ch/~maechler/
Seminar fuer Statistik, ETH-Zentrum  LEO C16	Leonhardstr. 27
ETH (Federal Inst. Technology)	8092 Zurich	SWITZERLAND
phone: x-41-1-632-3408		fax: ...-1228			<><
3 days later
Ross Ihaka
#
kwright at eskimo.com wrote:
First a warning.  The use of ColorBrewer type color schemes is
inappropriate for many statistical displays -- eg. barplots
piecharts, mosaic plots.  The problem is that the colors in
the schemes vary a lot in luminance and there is a size illusion
associated with luminance variation.  (See Cleveland and McGill
(1983). "A Color-Caused Optical Illusion on a Statistical Graph,"
The American Statistician, 37:2 101-105.)  Varying luminance
in graphs which represent values as length or area can distort
the perception of the encoded values.

For line and glyph colors the size illusion is not as much
of a problem and it's probably best to concentrate on visibility.
The ISO 9241 standard recommends a luminance difference of at
least 3:1 and preferably 10:1 between text and its background.
Something similar probably applies here, and that severely limits
the color choices available (you need quite dark colors on
a white background).

One way around this is to draw a black border around
the line or glyph.  You can do this for glyphs by using a pch
value between 21 and 25 and using bg= in the base graphics or
fill= in grid.  Alternatively you can roll your own with
polygon (we need a better alternative to this).  For lines
you can draw them twice, superimposing a colored line on top
of a black one - e.g.

	lines(x,y,lwd=3)
	lines(x,y,lwd=1,col="yellow")

There are simultaneous contrast issues with this, but mucking
about with the lwd values will generally get you a reasonable
result.

I think this is a pretty interesting area and there is an
opportunity for someone to look at the perceptual questions
involved.