Luigi Ponti wrote:
I found the refs I posted while browsing the net in search for good
color schemes: it puzzled me that the majority of the maps (including
mine) would use a rainbow, continuous color bar despite good arguments
[for and ...]
against this technique were available in the literature.
I would not say the matter that is fully "solved"; no more than any
other matter of human cognition. (I am just yet another opinion; but
that's sort of my point)
I have no bias since my experience in very limited
we all have bias on some level. I did not mean intentional bias or
bias about the topic we speak of.
(nor do I want to lie with maps).
The point of "How to lie with statistics" etc is not a guide on how to
become a better villain, but rather how to spot when someone is trying to
sneak something past you, or how to make sure you do not accidentally
fool yourself or miss some important meaning because you are looking at
the data in a bad light. ... a lesson in avoiding common traps.
The argument I read is that color bands act as contours lines do in
elevation maps. It is true that choosing arbitrary band limits on a
continuous value dataset is simply... ...arbitrary.
exactly, as is splitting contour lines by arbitrary human-created units
on a natural landscape. So the matter becomes which devil to choose?
However, when you have a continuously varying color bar in the map,
you will have label numbers next to the bar and you will try to
associate the bar color next to the label value with a color in the
map. The argument Borland & Taylor (2007) make is that if each label
of the color bar is located between two constant color bands in the
bar, you will have that value easily located in the map along the
boundary between the two contiguous areas of constant color. What do
you think about this?
So they argue that (filled) contour lines are better than continuous
rainbow legend. Well, why not draw contour lines over the top of a
continuous rainbow then? It is a false dichotomy, as you can do both
together. My feeling is that there is more room for danger with
contour lines than with the reader squinting to see what orangish-
yellow means. Contour lines are dangerous because they state loudly
where a transition is, regardless of if the data has that confidence or
not. It is like the difference between politicial speak and scientist
speak in a way, one gives black and white options, the other gives some
probability that a theory is correct.
The other interesting physical tangent of this is the way the human eye
sees a (sky) rainbow as bands of colors, and yet in a spectrograph it
is really a steady continuous spectrum. The distinct color bands are
entirely "in our heads" and the breaks are in slightly different places
to all viewers depending on our individual optical/brain physiology.
I have no bias given my very limited experience on the matter -- just
trying to make an informed decision. Thanks again for discussing this
and sorry if this is way off topic for the list (please advice).
I veer way offtopic as well, and again it reminds me about the next book
I need to borrow from the library (www.edwardtufte.com).
To claw back on topic I will mention that you can have a rainbow
colorscale with crossing band lines in ps.map: give the "tickbar"
instruction to the colortable command. (then create and overlay contours
at those levels)
FWIW Matlab has a compromise -- their colorbar legend has little inward
ticks at the labeled numbers.
> The statistical equivalent is to slowly vary the number
> bins in a histogram while the peaks seemingly double and halve.
Yes, you lose information that way.
And worse, it is very easy to do so without noticing that you've done it.
Hamish
