Hi,
Thank you for your help. I have a few follow-up questions:
Quoting Jan Gläscher <[log in to unmask]>:
> Dear Christine,
>
> On 2005-05-26 (Thu) at 20:00:44 +0100, Christine Cox
> <[log in to unmask]> wrote:
>>
> I have encountered a similar problem earlier and I suspect that Marsbar
> renders c'*beta as a plot (SPM's plot button does this for sure). Thus,
> what you see in the plots is a scaled version of your parameter
> estimate, NOT the real values from the images. In some case, the
> effects of interest contrast matrix contains some unreasonably high
> values in the range of several hunderds to tousands. Will Penny has
> once explained to a collegue of mine this could be an effect of
> a "non-settled" non-spericity correction and suggested that to
> re-estimate the design without correction for non-spherical errors.
>
> So, I suggest that you have a look at the contrast matrix of the
> contrast that you are plotting against (usually an effects of interest).
> The contrast matrix of the effects of interest is stored in
> SPM.xCon(1).c. If you find very high value there, they a most likely
> causing these high contrast values that you are see.
>
Plotting the effects of interest did result in similarly high values.
I re-ran
an anlysis in SPM2 choosing "none" for a high-pass filter and "none" for
correct for serial correlations. As far as I could tell, these were the two
estimation options which would eliminate the correction for non-sphericity.
The design description showed "Serial Correlations: i.d.d." The contrast
estimation plots now showed values in the range of -10 to 60, which is much
lower than originally, but still not in the range which I have been told is
normal.
Since I want to get the effect sizes for specific ROIs, I did the same type of
analysis in marsbar, choosing no high-pass filter and no correction for serial
correlations. The output in the stats table was in the same lower range of
values. Below is an example of the before and after stats table output:
ROI name: Contrast value: t statistic: Uncorrected P:
Corrected P
Before:
memoriesvsrestRhipp: 186.42: 2.51: 0.006095:
0.006095
After:
memoriesvsrestRhipp: 10.21: 2.47: 0.006914:
0.006914
Do the steps in this estimation as well as the output make sense if
"non-settled" non-sphericity correction was the problem? Do the contrast
estimations and contrast values provide a real representation of the effect
sizes of my different conditions?
Thank you again for your time and help.
Christine Cox
|