[RkWard-devel] RKWard 0.4.5pre1, 0.4.5 release process

[Thomas Friedrichsmeier]

On Wednesday 10 January 2007 16:37, I. Soumpasis wrote:
> I find it good. However I would suppose the contents of the menus to be
> stable.
>
> > Of course besides this, please do some testing on the preview release, so

Yes, some more stability would be nice, here, as well. However, it's not easy 
to do, as all windows are KParts and each defines it own separate menu 
actions.

> Today I played a little with correlation matrix plugin and I found some
> different results than the ones waiting. Seeing the code I saw that there
> is the use of the "pairwise.complete.obs" which means that the correlation
> between each pair of variables is computed using all complete pairs of
> observations on those variables. I have a dataset where na values must not
> be excluded, but replaced with zero which I have done using impute of Hmisc
> package.
>
> Now to come to the suggestion. Sould the user be warned about this? Or can
> the user choose what to do with the na values?

But isn't this the whole point of missing value? That they are missing, and 
you can not assume any specific number, but need to exclude them? Consider an 
example like this:

x <- c (1001, 1002, 1003, NA, 1005, 1006)
x.0 <- c (1001, 1002, 1003, 0, 1005, 1006)
y <- c (1, 2, 3, 4, 5, 6)
cor (x, y, use="pairwise.complete.obs")
cor (x.0, y, use="pairwise.complete.obs")

Well, this is not a real world example, obviously, but I think it does show 
that assuming 0s for NAs is generally *not* a good idea, so excluding them is 
the only real option.

Another thing is, that indeed we should provide a plugin to replace NAs using 
several different methods (such as several modes of interpolation, fixed 
values, mean, ...). However, I don't think, it makes much sense to include 
all these possibilities in every single plugin, but provide them separately.

> I also would like to suggest for the future another method for correlation
> which gives the probabilities too. There are two functions doing this in
> two different packages in my knowledge.
> The one is rcorr from Hmisc and the other one is rcor.test from package
> ltm. However none of them has a good output using rk.print(), or
> rk.results().

True. Hmisc is a rather heavy package with lots of dependencies, so it's 
probably not a good choice for a basic task like this. ltm is less heavy (but 
not exactly light weight, either). rcor.test seems to be a rather simple 
wrapper around cor and cor.test (from stats), so maybe it's easiest to brew 
our own similar to rcor.test.

Regards
Thomas
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
URL: <http://mail.kde.org/pipermail/rkward-devel/attachments/20070111/8bc15875/attachment.sig>