it's important to note that the multiple regression/no constant approach
is likely to violate the sphericity assumptions inherent in the general
linear model. we'll have to wait for real mixed-effects modeling before
this problem is fully solved.
cheers
russ
--
Russell A. Poldrack Ph.D.
Assistant Professor of Radiology, Harvard Medical School
Athinoula A. Martinos Center for Biomedical Imaging
149 13th St.
Charlestown, MA 02129
Phone: 617-726-4060
FAX: 617-726-7422
Email: [log in to unmask]
Web: http://www.poldracklab.org
Karsten Specht wrote:
>Dear Ulrich,
>
>indeed, it's usual a problem to combine different groups and conditions in
>one mixed-effects analysis.
>However, I have good experience with the 'Multiple regression without a
>constant' model. This was
>an update by Karl Friston last year:
>http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0104&L=spm&P=R8273&D=0
>
>This model allows in the second level analysis to combine the different
>groups and conditions.
>This model has the big advantage, that it allows you to create single
>contrast , i.e. [1 0 0 0 ...],
>masking, and conjunctions between groups, for example.
>Disadvantage of this model is, you have to create your own design matrix by
>entering
>all columns of the design matrix (use the 'ones' and 'zeros' commands from
>matlab to make the life easier ...)
>
>I hope, this model can solve some of your analysis problems.
>
>Good luck
>
>Karsten
>
>----------------------------------
>Karsten Specht
>
>fMRI Section
>Department of Neuroradiology
>Medical Center Bonn
>Spessartstrasse 9
>53119 Bonn
>Germany
>
>Phone: ++49-(0)228/90 81-178
>Fax: ++49-(0)228/90 81-190
>E-Mail: [log in to unmask]
>WWW: http://www.mcbonn.de/Praxis/praxis15/fmri1.htm
> -----Ursprungliche Nachricht-----
> Von: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]Im
>Auftrag von Ulrich Moeller
> Gesendet: Sonntag, 10. Februar 2002 18:36
> An: [log in to unmask]
> Betreff: model for group comparisons, and masking
>
>
> Dear SPMers,
>
> we are looking for a solution that might also be of interest for other
>people doing imaging studies.
>
> We need an appropriate model for group comparisons in SPM. When going
>through the SPM email archives we found several entries which seem to be
>more or less close to the problem. However, we are still not convinced to
>have a valid solution. Also a discussion with people who are relatively
>familiar with SPM did not yield a definite answer. Therefore, we pose our
>question here.
>
> ==============================
> Our fMRI studies included
> ---------------------------------------------------
> - groups of patients and control subjects (each with n > 12) in different
>ranges of age
> - 4 'activation' conditions (A1, .., A4) and a control condition (C)
>
> We have obtained contrast images for each subject from a first level
>analysis (A1-C, A2-C, ..., A1-A2, ...).
> Now, we are looking for an appropriate model for a second level analysis.
>
> It is clear what a one or two sample t-test or a one way ANOVA do, and it
>is also clear, in principle, how to configure these methods using SPM's
>Basic Models.
> ==============================
>
> It seems that the problems begin when making group comparisons, where some
>kind of masking is required.
> Even in our control condition, the subjects perform a basic task. Hence,
>in all contrasts A-B, there occur voxels with negative values that may need
>to be excluded.
> Some people discussed masking by using the ImCalc tool (e.g. Shelton, 10
>Jan 2001 or Henson, 4 Feb 2000)
> http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0101&L=spm&F=&S=&P=7197
> http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0002&L=spm&F=&S=&P=5299
> Others argued that masking can be achieved at the level of the SPM display
>(e.g. Friston, 3 Feb 2000)
> http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0002&L=spm&F=&S=&P=4409
>
> We have been recommended not to cut the negative contrast values, but to
>mask with another contrast and specifying an appropriate p-value. For
>example, we assume that it may be needed to mask the group contrast (A1-C) -
>(A2-C) by a positive (p-thresholded) group contrast (A1-C) and/or (A2-C).
>
> In order to have an elegant way to model any group comparison for any
>contrast of interest, we have been recommended to use one of the larger
>models which are available under the PET option, e.g. Multi-group conditions
>& covariates. This would also give the opportunity to model covariate(s)
>and/or nuisance variable(s), which will become a relevant topic in our
>study.
>
> However, after entering the images and other parameters, our attempts to
>implement group comparisons provided 'invalid contrast' messages.
>
> ===========================
> Example:
> ----------------------------------------------
> We entered 2 groups, and for each subject 4 contrast images A1-C, .., A4-C
>(now called A1,...A4).
> The design matrix had 8+N columns:
>
> contrast A1 A2 A3 A4 (group 1),
> contrast A1 A2 A3 A4 (group 2),
> and a constant for each of the N subjects.
>
> Now we tried to specify, e.g. where group 1 activated more than group 2
>with respect to A1 by entering
> 1 0 0 0 -1
> However, this provided an 'invalid contrast' message, whereas an
>intra-group comparison of the type
> 1 -1 ...
> was termed valid.
> ===========================
>
> Was there a mistake in our modelling considerations or in the use of the
>SPM tool?
> We would greatly appreciate some advice in choosing an effective and
>efficient model and masking method.
> Best regards
>
> Ulrich
>
> ***************************
> Dr.-Ing. Ulrich Moeller
> Institute of Medical Statistics, Computer Sciences and Documentation,
> Clinic for Child and Adolescents Psychiatry
> Friedrich Schiller University Jena
> Philosophenweg 3/5, D-07740 Jena
> Email: [log in to unmask]
>
|