See answers below.
On Thu, Oct 25, 2012 at 11:57 AM, Anja Dietrich <[log in to unmask]> wrote:
> Dear experts,
>
> I'm refering to my last post on the mailing-list (subject: vbm covariate in fmri analysis). Since nobody has commented on it and this problem bothers me a lot, I try it again and hope that somebody can help me.
> Especially because the answer on my previous question confused me quite a lot (see below).
>
> Here are the elements of the design:
>
> subject factor
> interaction group (3 levels)/ condition (2 levels)
> covariate 'smoker' (bivariate coded)
>
> I defined contrast vectors like the following:
>
> condition1/group1 > condition1/group2:
> [ones(1,nSubj1)/nSubj1 -ones(1,nSubj2)/nSubj2 zeros(1,nSubj3) 1 0 -1 0 zeros(1,nConds)]
>
> nSubj: number of subjects in the group
> nConds: number of conditions (which is 2)
>
> Is that contrast valid?
NO. The contrast is not valid. You are comparing groups. You can only
compare conditions in a repeated-measures design in the standard GLM
framework that is used in SPM. Unless you are contrasting at least 2
conditions, your contrast is not statistically valid. For example,
condition1/group1 > condition1/group2 is a group comparison and is not
valid; condition1/group1>condition2/group1 would be valid because you
are contrasting two conditions; and condition1/group1
-condition2/group1 > condition1/group2 - condition2/group2 is a
group*condition interaction and is valid because you are contrasting
two conditions. If you want to compare groups, condition1/group1 >
condition1/group2, then you need to use GLM Flex OR a two-sample
t-test where you only enter the scans for condition1. There have been
numerous posts on the list describing the reasons behind this issue.
In brief, the degrees of freedom and error term are incorrect for
group/between-subject effects of a standard GLM assessing
repeated-measures. GLM Flex generates multiple error terms and the
correct error term can be used for each effect you want to test.
And is it possible to insert a covariate in a within-subject design?
NO. You cannot add covariates to the design as the covariates. You
need more complicated models to get the correct effects for
repeated-measure ANCOVAs.
Hopefully that clears up your confusion.
>
>
> Thank's a lot in advance!
>
> Best
>
> Anja
>
>
> previos post and answer:
>
>>> I performed a 2nd level fmri analysis (flexible factorial design) with a
>>> subject factor (24 subs) and a group factor (3 groups). Regardig this I
>>> found functional differences between the 3 groups. Further a colleague of
>>> mine found differences in gray matter volume between the 3 groups and
>>> relating to some functional relevant areas.
>>
>>>>> Group comparisons (unless its the group*condition interaction) are not
>>>>> valid in the flexible factorial (or full factorial) or any GLM which
>>>>> only has a single error term when you have repeated-measures. In my
>>>>> previous posts, I've stated that between-subject effects are not
>>>>> statistically valid in within-subject designs. If you only have one
>>>>> condition per subject, then you don't need to enter subjects as a
>>>>> factor.
>>
>>> Now I would like to analyze, if the functional differences between the
>>> groups are a consequence of gray matter volume differences.
>>> Regarding this I thought it should be a good idea to integrate the vbm
>>> data as a covariate in the fmri analysis. I have heard of the bpm toolbox
>>> (biological parametric mapping) which offers the opportunity to integrate
>>> voxel-wise covariates. But I'm not sure if it is possible to realize the
>>> flexible factorial design in the environment of this toolbox (the subject
>>> factor seems to be important for my fmri analysis). I would be very
>>> thankful if somebody could give me some advice. Maybe there are also other
>>> opportunities in solving the problem.
>>
>>>>> Covariates should not be included in within-subject designs.
>
>
>
>
> l
|