Hi Zhiqiang,
There's actually nothing false about the significant results in B and
C, the two-group t-test is genuinely finding which voxels show the
most significant difference. There is no requirement that significant
differences between two samples have to be large/significant in the
individual samples. Conversely, there is no guarantee that an apparent
difference in significance in two individually-significant results
will correspond to a significant difference.
It isn't correct to look at separate one-sample tests and deduce the
significance of the two-sample test based on their apparent
difference; the two-sample test will give the correct answer (subject
to the usual caveats about the model being appropriate, the data
normal, etc. etc.). Also note that the two-sample t-statistic is not
given simply by the difference in one-sample t-statistics, as your
figure seems to suggest (the contrast image for the two-group test
will be the difference between the contrast images for the one-sample
ones though).
My previous answer about contrast masking shows how you can aid the
interpretation of the two-sample results, by looking at them only
where the smaller (or larger, or average, depending on the contrast
you mask with) one-sample contrast was significant.
I'm afraid I don't know how/if you can go beyond this, for example
upweighting two-sample results based on one-sample significance...
possibly you could do something like this in a permutation-testing
framework, but I couldn't be sure without giving it a lot more thought
(and asking someone who genuinely knows what they're talking about
;-)).
Sorry if this isn't the answer you're looking for! Best,
Ged
On 26/10/2007, Zhang Zhiqiang <[log in to unmask]> wrote:
> Dear Ged,
>
> Thank you for your answer. I think I didn't express my question clearly. I
> just drew a figure to assist to illustrate it. pls see the attachment.
>
> Thank you for your time and effort.
>
> Best Regards,
>
> Zhiqiang Zhang
>
> DRC SPM <[log in to unmask]> дµÀ£º
> Hi Zhiqiang,
>
> If I've understood you correctly, you're saying that you are
> interested in significant differences between the groups when they are
> also significant individually. In which case you probably just want to
> use the single-group contrast to inclusively mask the difference
> contrast. E.g. mask [1 -1] by [0 1]. Does that answer your question?
> If you want a more general up-weighting of two-sample significance
> based on single-sample significance, then I'm not sure whether this is
> possible, hopefully someone else will comment.
>
> Best,
> Ged
>
>
> On 25/10/2007, Zhang Zhiqiang wrote:
> > Dear SPMers,
> >
> > I encounter a problem when performing 2-sample test in group analysis. as
> > it's known the T score in con*.img presents with positive and negative
> > value. Providing there is a brain region, whose T value is +1 in the
> > first(group) subjects, and is -1 in another group, it will be +2 after
> > A-B(2-sample t test). it will present more stronger than the result of
> (area
> > that T=1=+8 minus +7). In practice, the area that T=+8 or +6 is more
> > significative than the area that T=+1 or -1, but after 2-sample t-test
> > comparing, the later area look more significative than the former.
> > obviously it's unreasonable. I attached a sketch map to illustrate my
> > question.
> >
> > anyone who can advice me how to avoid this occurrence? thanks a lot!
> >
> >
> > ---------------------
> > Zhiqiang Zhang MD.
> > Dept. of Med. Img., Nanjing Jinling Hospital,
> > Medical school, Nanjing University. China.
> >
> >
> >
> >
> > ________________________________
> > ÑÅ»¢ÓÊÏ䣬ÖÕÉú»ï°é£¡
> >
> >
>
>
>
> ________________________________
> ÑÅ»¢ÓÊÏ䣬ÖÕÉú»ï°é£¡
>
>
>
|