Re: [Gretl-users] Correlation heatplot
I've made a first stab at this (in git and snapshots). I'll
briefly
describe what's there, and pose a few questions for consideration.
Thank you for this, Allin. It's a nice feature I think.
Command line: the "corr" command now has a --plot=whatever
option (along
the same lines as the "freq" command). You can use --plot=none to
suppress a heatmap plot in interactive mode, --plot=display to show it
on-screen, or --plot=somename.pdf, and so on.
That works quite nicely here on Ubuntu.
However, it seems that using the "corr" excludes the possibility to plot
the correlation-matrix using a matrix, right? I think this is a bit
unfortunate. In this case, the user would still need to program it
itself or use an additional external package, right?
For now I'm showing positive correlations in red, of varying
intensity,
and negative correlations in blue, with white in the middle (around
zero). To cut down on visual clutter I've extended the white range to
cover correlations that are not significantly different from zero at the
20% significance level.
Personally, I can live with the 20% sign. level ;-)
However, I would also favor a monochrome option for the heat plot as I
had in my little function. Especially when it comes to publication, one
often needs to provide black-white figures. So I would favor another
option like "--monochrome" or so.
By default, the whole matrix is shown but there's a --triangle
option to
show just the lower triangle.
Also a nice idea, and it works good here.
Some questions:
* Full matrix vs triangle: which should be the default? Is the choice
worth having or should we just fix on one or the other?
I like to have both options. And if it doesn't impose any issues in
gretl, I don't see a reason not to consider these.
* Right now there's a minimum dimension of 3 for doing the
heatmap:
should the minimum be bigger than that? (3 x 3 looks kinda silly, but
maybe it should be available anyway?)
Again, I think it's good if the user can decide on this.
* The plot works reasonably well for up to about 30 series, but is
going
to get quite messy for more than that. Should we set a max? And/or,
should we make a special effort to get the plot working acceptably for
bigger dimensions? (Smaller font, not sure what else could be done.)
Mmmhh, 30 sounds quite large already. I couldn't imagine of a case where
you would need an even larger dimension...
Artur