Subject: | | Re: parametric modulations - again |
From: | | [log in to unmask][log in to unmask]<mailto:[log in to unmask]>> wrote: Hi Adnan, Presumably for subject 1 the only model(s) contributing to the average had connection A(1,3) enabled whereas for subject 2 the only model(s) contributing to the average had connection A(3,3) enabled. Does that fit with what you expect?
Best, Peter
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]<mailto:[log in to unmask]>] On Behalf Of adnan alahmadi Sent: 04 November 2015 13:14 To: [log in to unmask]<mailto:[log in to unmask]> Subject: [SPM] DCM BMA
Hi all.
What I understand the best approach to compare or investigate DCM is the use of BMA averaging. I have however couple of questions regarding this option.
Say for example I have two groups (or even one group).
When I compare the models using compare, I should switch on the BMA option and then say choose family: winning family.
Then the models are compared, no problem so far and I did this for both groups separately.
So if I want to compare connections parameters between say two groups or within a group I should go to Group1.BMS.DCM.rfx.bma.mEps where I can find parameters of each subjects for all matrices.
Say for example I look at the A matrix for subjects within a group or between groups. Sometimes I get missing values or zeros in some subjects and sometimes they are not. or in other words,
Sometimes A matrix for subjects 1 is
0 0 -0.4 0 0 0 0 0 0
While for subject 2 is 0 0 0 0 0 0 0 0 0.3
and so on.
What is confusing is that why sometimes in the BMA there are zeros and this is dependent on subjects?
When I tried to investigate the winning model and force DCM to do BMA of the winning model, I can see the values in the right place of connections, or in other words homogeneous among subjects.
Maybe it is my understanding but is it correct to compare A connection in the A matrix with zeros value in some subjects ? Or what is the right procedure to the comparison between groups when I want each group to preserve its winning model?
Many thanks
[log in to unmask] |
Reply-To: | | [log in to unmask][log in to unmask]] >Sent: 01 November 2015 14:19 >To: Zeidman, Peter < [log in to unmask] >; spm < [log in to unmask] > >Subject: DCM model evidence and free parameters > >Dear Peter and SPM team! > >Thanks a lot for your previous answers! I really do appreciate it! > >A have two unclear moments about DCM output and BMS, maybe >you could help me? > >1)I've inverted about 90 DCMs for 20 subjects, performed BMS >and got sums of F-values per model (for all subjects). >Then I plotted these sums (rescaled, blue color) together with the number >of connections (free parameters, red color) per model and got the following picture: >http://tinypic.com/r/2cpc37k/9 >(Sorry, I don't know whether it is possible to send pictures directly to common mailbox) >The correlation between sums of F-values and number of free parameters is very high and >could be seen from attached picture. How could this be explaind? It seems that a model >has a bigger evidence only when the number of free parameters increases? How >to better report the results, because there are some peaks on the F-value plot, these peaks >differ not much in % and the winning model is not common for all subjects. >Maybe its better to report the mean connection strengths (from BMA) and not the model itself? > >2)It's not clear for me, how can I compare with each other mean connection strengths from BMA output? >When for example I can say that 0.36 is significantly bigger than 0.3? > > > >Thanks in advance, >Alex. [log in to unmask] |
Date: | | Fri, 20 Nov 2015 17:34:40 -0500 |
Content-Type: | | text/plain |
Parts/Attachments: |
|
|
|
|
It may be helpful for clarity if you paste an image of your design matrix. Also some explanation of event timings may be a bit helpful in this context. For example you are modelling responses and fixation. Are they very close in time with your events of interest? If you model two nearly overlapping events (e.g. less than 2 seconds apart) this can cause your effect size to be split across the two 'conditions' in an HRF analysis.
My understanding is that your contrast would be correct (main effect of the pmod), and bring the con files into the second level. But it would likewise be helpful to paste an image of the contrast manager output from the SPM results window along with the design matrix. If you are less experienced with SPM and the design matrices it is easy to make a mistake at the contrast specification scale.
Cheers,
Colin Hawco, PhD
Neuranalysis Consulting
Neuroimaging analysis and consultation
www.neuranalysis.com
[log in to unmask]
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Alexandre Obert
Sent: November-20-15 4:55 PM
To: [log in to unmask]
Subject: [SPM] parametric modulations - again
Dear all,
I hope someone will be able to help me...
I'm trying to make a parametric modulation analysis but I've got very strange results.
It remains possible that I incorrectly set my design - a technical mistake rather than a theoretical one...
Thus, maybe someone could help me and say to me if I there is an error in my design ?
First, my experiment :
My experiment contains 2 experimental conditions (let's name them A and
B) each of these conditions contains sentences stimuli preceded by a Context.
I want to test the linear effect of 2 parametric modulators for each condition, Beside them, there is some variables from my experiment that I modeled (I think they could have an effect but it doesn't interest me) :
fixation crosses and the response of the participants (I think it could generate 'motor' activations).
I set my design as follow:
I modeled the onsets of the Context, then the condition A and the 2 parametric modulators and in the same way the condition B and its 2 PM ( the PM are correlated for each condition but since SPM8 orthogonalizes the PM at 1st level, it's not a problem to model them in the same design, right ?).
I also modeled the onsets of the fixation crosses and responses of the participants.
Finally, in order to test the linear effects, I set contrasts such as [0
0 1] with "1" on my parametric modulators at first level and use the
con* files into a one sample 2nd level design for each PM.
The results for at least one PM are very strange and unexpected. It's possible that it comes from an unexpected effect of the PM but I would like to be sure that I did not make any error in my design....
I'm ready for any question and advice ;)
Regards,
Alexandre
|
|
|
|