Thanks Cyril & Dionyssios for your answers,
I think what I am actually going to do is either calculate the residuals and
not just the mean squre of the residuals, and then there are a number of
goodness of fit parameters that I can look at for each subject, and then
we'll see where I go from there. I believe I will then be able to
specifically look at differences between the learning strategies/subject
groups.
Best,
Virginia
-----Ursprüngliche Nachricht-----
Von: Cyril Pernet [mailto:[log in to unmask]]
Gesendet: 24 November 2010 16:21
An: Virginia Flanagin
Cc: [log in to unmask]
Betreff: Re: [SPM] goodness of fit in fMRI - unresolved but ..
Hi Virginia
I'm sure if it will answer your question directly but here is what I think
if you do a F test spanning all of the columns but the column(s) coding the
cst of your design matrix then the resulting F and p values are the same as
the F and p values of the R^2 in multiple regression ie it tells you where
your model significantly Fits the data. (that use to be in SPM the effect of
interest contrast) - now I know it doesn't do what you wanted. I'm not sure
how to compute the gof from there? (the error of the model is in the
res.img)
Cyril
> Hello,
>
> I would like to address the question whether or not my first level GLM
fits better in certain subjects and not in others. I tested 3 different
learning strategies in 3 different groups of normal subjects (each subject
had 1 strategy). I have found some interesting differences between the
groups/strategies, but I would like to make sure that the results that I
find are based upon the differences in the experimental manipulation and not
differences in the model fit. In particular, the different strategies vary
slightly in their reaction times.
>
> What I would like is something that tells me whether the fit of
modelA(group1) = fit modelA(group2) = fit modelA(group3).
>
> I have read the Rosa et al. NeuroImage 2010 paper on Bayesian model
selection maps for group studies, and looked here in the Listserve both at
the Bayesian model selection and at other measures of goodness of fit (e.g.
Bullmore et al NeuroIamge 2000). In the Bayesian model selection an
alternative model would be the null model but I am not performing a
subtraction so I cannot use both regressors together as the null model. And
it seems like I should be able to do something akin to a chi-squared but
what data should I use? The Bullmore paper and the listserve messages
regarding this method talk about SEM and not directly fMRI. I have looked at
PSC in specific regions that I am interested in and at the time courses of
the data, but have not done a direct comparison of the models.
>
> If anyone has any suggestions I would greatly appreciate it.
>
> Thanks,
>
> Virginia
>
>
--
Dr Cyril Pernet,
SBIRC fMRI Manager
Lecturer in cognitive neuroimaging
SFC Brain Imaging Research Center
Division of Clinical Neurosciences
University of Edinburgh
Western General Hospital
Crewe Road
Edinburgh
EH4 2XU
Scotland, UK
[log in to unmask]
tel: +44(0)1315373661
http://www.sbirc.ed.ac.uk/cyril
http://www.sinapse.ac.uk/
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.
|