Hi Dan,
you are right, of course, with regard to the degrees of freedom.
Degrees of freedom are only relevant for fixed effects analysis.
Only the con images are entered into the random effects model so
the error term is calculated differently (over all con images instead
of over all scans) and the degrees of freedom are determined by the
number of con-images alone.
However, this does not mean that entering motion parameters or not
does not affect also the random effect analysis. It does, on the
level of fitting the model (getting the con-values).
The random effects analysis will be affected if entering
additional regressors also changes the fit of the model for the
predictors
of interest. This is not very likely for motion perfectly uncorrelated
to the
predictors of interest.
The case is very different, however, for motion correlated
with your paradigm. Consider the extreme case, (e.g. a patient overtly
naming picture in the scanner alternating with rest periods, testing
for task vs. rest or only task)
In this case motion caused by articulation is highly likely and
will be correlated with the predictor of interest (onset of
task-blocks).
If you enter mps into this model, statistics will suffer, because
mps will regress out (too) much of the signal (task). This will then
affect the random effects analyis.
I hope this helps,
Best,
Anja
Dr. Anja Ischebeck
Innsbruck Medical University
Clinical Department of Neurology
Anichstrasse 35
A-6020 Innsbruck - Austria
tel.: +43 (0) 512 504 23661
>>> "Daniel H. Mathalon" <[log in to unmask]> 06.10.2005 15:37
>>>
Dear Anja,
I am not sure why the degree of freedom costs associated with
entering MPs into the first level model matters in the context of
second level random effects analysis. If one is only passing betas
or con images to the second level, the costs of spending degrees of
freedom at the first level are not relevant. Wonder what you think
about this.
Dan
>Hi,
>
>just my two pennies (is that the right expression?)
>to the discussion:
>
>To my experience
>(I tried out different designs with or without MPs modelled for
years:
>
>in this case the simple 6 parameter model - translation/rotation)
>it is not very easy to compare methods
>if one does not take into account the amount of motion in total
>and (very likely to be of importance:) the amount of motion
>correlated to the paradigm or event regressors.
>
>If there is nearly no motion, entering the Mps will conly cost
>degrees of freedom. On the other hand, if there is moderate
>motion, entering the 6 MPs (or the more complex models)
>will explain a lot of variance.
>However, the 6 parameter MP model has its limitaitons:
>With high T-field scanners or excessive (instead of moderate) motion
>as in the case of children or patients
>excess motion cannot be modelled satifactorily.
>At least my results were abysmally bad in these cases and I
>excluded the subjects in question - which is a darn pity.
>
>Excessive motion means (in a 3T field): movement of 0.5 mm
>between two consecutive scans (twiching) - this seems to
>be bad enough for statistics.
>
>I missed in Brett's abstract a mention of the degree of motion
>in the data used for his comparison.
>
>I would therefore like to hear of people who have experience with
>this (excessive motion or moderate motion and high field scanners).
>Maybe the mode complex models of motion are doing
>much better than the simple one in these special case.
>
>Best,
>
>Anja
>
>
>
>Dr. Anja Ischebeck
>Innsbruck Medical University
>Clinical Department of Neurology
>Anichstrasse 35
>A-6020 Innsbruck - Austria
>tel.: +43 (0) 512 504 23661
>
>>>> Matthew Brett <[log in to unmask]> 05.10.2005 17:45 >>>
>Hi,
>
>> I don't think it is the speed of the movement that increases the
>effect
>> of artefacts, it really is the displacement. The artefact is due
to
>the
>> spin excitation history of the voxel, i.e. energy transmitted to a
>brain
>> region before thatt region was at its current voxel location
(maybe
>you
>> should ask an MR physicist instead of reading this...).
>
>Just a question - my impression was that the current thinking is that
>spin-history is a rather minor factor in the motion-related variance.
>Is that true? Certainly there can be quite large effects from motion
>by distortion interactions - this is stuff Chloe Hutton and Jesper
>Andersson have worked on.
>
>I did a very tiny study of including movement parameters up to the 24
>regressor spin-history model (which will include the effects modeled
>by the difference of the parameters) and found, like Tom Johnstone,
>that only the movement parameters themselves seemed to be robustly
>useful:
>
>http://www.mrc-cbu.cam.ac.uk/~matthew/abstracts/Moves/moves.html
>
>The link points to my HBM2005 abstract.
>
>Best,
>
>Matthew
--
Daniel H. Mathalon, Ph.D., M.D.
Assistant Professor of Psychiatry
Yale University School of Medicine
Mail Address: Psychiatry Service 116a
VA Connecticut Healthcare System
950 Campbell Ave
West Haven, CT 06516
Fax: (203) 937-3886
Office Phone: (203) 932-5711, ext. 5539
Pager: (203) 867-7756
|