Re: [guppi-list] Re: Problems with cvs goose and a new commit

["Asger K. Alstrup Nielsen" <alstrup diku dk>]

> This seems like a reasonable approach.  One thing that struck me as
> odd, though, was how in tests like Kruskal-Wallis and Spearman, where
> special methods are required for the case of ties, the routines go
> ahead and calculate the various values both with and without the
> ties-corrections.

As it turns out, we need both the corrected and non-corrected values
in our code (for didatic purposes), and that's the reason Mikkel 
implemented it like that.

Of course, it does seem to be wasted work in the common case, but if 
I'm not mistaken (I could be, I didn't check in detail), the extra 
overhead is managable.  So, unless you think we should change it for 
the sake of more detailed functionality, I would say that we can just 
keep it as it is.

If you think we should change it, of course we will, but I think we
should keep the current ResultStruct which has place-holders for both
the corrected and non-corrected results.  Then, we could add a bool
to decide whether to use the "intelligent automatic" approach that
will only fill in the relevant values (and a flag to describe what it
did), or the brute-force approach which will calculate everything 
regardless of whether we really need to or not.

(The difference to your proposal is that we keep the possibility of
getting both kinds back in one go -- this is to avoid doing the same
shared work twice in our case.)

Greets,

Asger
P.S. I just checked in a small fix to McNemar, and the Spearman test
will probably be improved soon because right now it's only an 
approximation.