Hi there,
Just thought I'd jump in and comment about the two models you
described: M1=paired t test at level 3 and M2=calculating paired
differences at level 2 and then taking the mean of these differences
at level 3.
You are correct, these should be giving the same (or very similar)
results. I think you were saying that M1 was more conservative that
M2, and that leads me to believe that you may not be doing a true
paired t test at the 3rd level or something is going wrong when you
calculated the paired differences at the second level. If your 3rd
level model is actually a 2-sample t test, then your p-values will be
larger, if there is indeed pairing between measurements. Your 3rd
level paired t test should look like what is shown here
http://www.fmrib.ox.ac.uk/fsl/feat5/detail.html#PairedTwoGroupDifference
One last note is that if you are truly interested in looking at paired
comparisons, looking at the copes from each group can be very
misleading, since the correlation between pairs is completely ignored.
Cheers,
Jeanette
On 9/20/07, Vishwadeep Ahluwalia <[log in to unmask]> wrote:
> Thanks Steve! I think i understand the last part. i thought using max or
> mean or median was an arbitrary choice and susceptible to outliers and so i
> chose 90th percentile value.
> However, about the model, i'm not sure what you mean. Like i said in my
> previous post, i ran GLM for all individual runs at the 1st level using same
> time series model. At 2nd level i combined the 2 runs for each session and
> for each subject using fixed effects analysis. At 3rd level, i ran paired
> group difference between all session1's and Session2's. Is this not the
> correct model to see if there are any significant differences between the
> sessions?
> Another model to compare sessions would be to specify Sess1-Sess2 contrast
> at individual level and do a group mean. Are the 2 models conceptually same?
> The latter seems to give less conservative results than the former.
> -Vish
>
|