On Mon, 6 Apr 2009 18:08:21 +0200, Thomas Stephan
<[log in to unmask]> wrote:
> the Bcov matrix contains a lot of negative elements. I tried to follow your
> suggestion and set independence and variance for side and condition
> to 'no/unequal'.
OK.
> Now I can not find any effects any more, so I can not reproduce the error.
> But SPM.xX.Bcov still has a lot of negative elements, so there seems to be
> some problem with the calculation of the covariance structure. The
> spmF-image for the effects of group contains only negative values.
If it's a diagonal matrix, then yes that's bad.
> Regarding the settings for independence and variance:
> My understanding and the rule of thumb I use to know how to set these
> parameters is (I am surely no statistician):
> Independence: 'no' if the levels of this factor have been collected from one
> subject, 'yes' if they are collected from more than one subject.
> Variance: 'equal' if the levels of this factor have been collected from one
> group, 'unequal' if from more than one group.
> This is my interpretation of the Gläscher/Gitelman manuscript, pg. 2-3.
> Following this, it should not be appropriate to use the combination
> no/unequal
> because it would mean to have data from one subject that belongs to more
> than one group.
The best way to think about these things is from a more abstract point of
view. You use "independent" if the variables are statistically independent.
You use "unequal variance" if there's reason to believe that the variables don't
have equal variance. (If the variables are "close" to having equal variance,
you should use "equal", not "unequal".)
Perhaps the most common situation where you'd get independence is the case
of different subjects. But there are many cases, not just subject groups,
where variance would be different, even within the same subject. That is,
different groups might imply unequal variances
BUT
unequal variances does not imply a situation of different groups
Best,
S
<snip>
|