Plotting question

On 11-08-01 11:48 AM, Bert Gunter wrote:

IMHO:

On Mon, Aug 1, 2011 at 7:51 AM, Duncan Murdoch<murdoch.duncan at gmail.com>
?wrote:

On 11-08-01 5:44 AM, Andrew McCulloch wrote:

Hi,

I use R to draw my graphs. I have 100 points on a simple xy-plot. The
points are
distinguished by a third variable which is categorical with 10 levels. I
have
been plotting x against y and using gray scales to distinguish the level
of the
categorical variable for each point. It looks ok to me but a journal
reviewer
says this is not any use. I cannot afford to pay for colour prints. Any
ideas on
what is the best way to distinguish 10 groups on an xy scatter plot?

Plot digits or letters or other symbols.

Duncan Murdoch

No, this does not work.

You have amazing perception to know that it doesn't work in Andrew's graph.
?But then you go on to suggest that sometimes it does, and then suggest
using symbols.

Obviously you need to see the graph to know what works. ?If the 10
categories are ordered, then something like thermometer plots would work.
?If they are grouped into a small number of variations on a small number of
groups, then digits or letters combined with shading might work, especially
if the groups are well separated, or there are clear patterns.

I'd agree with the reviewer than 10 levels of shading is probably too many
to distinguish,

in an unstructured scatterplot are probably not a good presentation, but I
wouldn't want to give specific advice about plotting a dataset without
seeing it.

Duncan Murdoch

See Cleveland's books (e.g. "Visualizing

Data"). 10 is too many symbols to constantly refer to a legend to keep
straight, and digits or letters do not allow you to readily perceive
the pattern. (Caveat: If "most" of the data are only 2 or 3 of the
symbols, then these can work).

I think the OP's idea of using gray scales was better. I would dispute
the reviewer and refer them to appropriate references. Alternatively,
thermometer plots (aka "filled rectangle" plots) would be best. Again,
Cleveland's books provide scientific justification rather than merely
the (possibly uninformed) aesthetic opinion of a reviewer. Presumably,
the journal editor would accept hard data and psychological research
in preference to opinions.

If all else fails I can just remove the graph and give them a table of
regression coefficients.

No. I think your attempt to use a graph is a much better way to go.
Try to resist poor practices such as just publishing summary
statistics.

Cheers,
Bert

Thanks.

Yours Sincerely
Andrew McCulloch