Dear Narges,

I hope you don’t mind me CC’ing the SPM mailing list so others can benefit. As you point out, the parameters in your model barely move away from zero, so the model probably hasn’t fitted very well. You can confirm this by running the function spm_dcm_fmri_check() and looking at the explained variance of the model.

 

So, either there is no experimental effect to explain in your ROI, or there is an effect but the first eigenvariate (principal component) isn’t capturing it. As you only have a few subjects, I would drop the threshold to a sufficiently low value so that you get blobs in every subject  (but using the same threshold for every subject, for clarity), and look to see how the location of the peak differs across subjects. Use these blobs for your ROIs. Bear in mind that if you don’t get strong effect sizes estimated by SPM, you are unlikely to get strong effect sizes estimated by DCM.

 

You may also wish to plot which subjects are contributing to your group-level activation – your experimental effect may only be present in subset of subjects. To do this, position the SPM cursor at the peak of relevant blob in the second level results, then in the Results window, click Plot -> Fitted responses -> Adjusted -> Plot against scan or time.

 

Best,

Peter

 

From: RADMAN Narges [mailto:[log in to unmask]]
Sent: 28 September 2015 14:10
To: Zeidman, Peter <[log in to unmask]>
Subject: Re: ROIs for DCM

 

Dear Peter, 

 

I am sorry for asking several questions on the same topic. 

Following your suggestion o​n using threshold p=1, I have prepared the same model with the timeseries coming from p=1. Here an example of values of connections between ROIs for the same model:

​

Connection

p=1 (5ROIs)

p=0.05(3ROIs) , p=1 (2ROIs)

ACC to BA45

0.007812

0.016548

ACC to BA47

0.007812

0.012369

ACC to Lc

0.007812

0.013379

LC to ACC

0.007812

-0.00357

LC to BA45

0.007812

0.001077

LC to BA47

0.007812

-0.00579

BA45 to ACC

0.007812

0.035466

BA45 to LC

0.007812

0.034333

BA47 to ACC_

0.007812

0.060154

BA47 to LC

0.007812

0.059107

BA45 to BA47

0.007812

0.034647

BA47 to BA45

0.007812

0.039769

BA37/19 to BA45

0.007812

0.053996

BA37/19 to BA47

0.007812

-0.06729

BA37/19 to LC

0.007812

-0.1441

 

The value of connection between ROIs change importantly according to the p values of extraction, and strangely, I got quite the same value of connectivity all over the first model.  

When I see the pattern of the first column, I thinks the model did not reach a satisfying equilibrium and results seem meaningless. Do you agree with that? 

Would it worth to use the version with different thresholds for extracting timeseries (although as you have previously said, it will be hard to justify).

 

Sincerely, 

 

Narges Radman

 

Laboratory for Cognitive and Neurological Sciences

Neurology Unit, Medicine Department, University of Fribourg

Ch. du Musée 5, 1700 Fribourg, Switzerland
http://www.unifr.ch/neurology/
http://www.unifr.ch/med/nibsi​

From: Zeidman, Peter <[log in to unmask]>
Sent: Friday, September 11, 2015 4:46 PM
To: RADMAN Narges; [log in to unmask]
Subject: RE: ROIs for DCM

 

Dear Narges,

Different thresholds are confusing and difficult to justify (although theoretically not a problem). Assuming that the subjects are in a group space (e.g. MNI), have you found a group-level SPM result, that passes some reasonable level of significance? If so, you can simplify things. Constrain your ROI extraction to within the bounds of the group-level activation. Then, at the individual subject level, don’t bother with a threshold – use p=1. You may get the best results by moving each individual’s ROI to the nearest peak (but keep it within the bounds of the group-level activation).

 

Best,

Peter

 

From: RADMAN Narges [mailto:[log in to unmask]]
Sent: 11 September 2015 14:25
To: Zeidman, Peter <[log in to unmask]>; [log in to unmask]
Subject: Re: ROIs for DCM

 

Dear Peter, thanks again for your response,

 

I have agian a naïve question,  

I encounter a problem while extracting the eigenvariate of different ROIs. Actually when I use even e liberal threshold there are some régions which don't show activation (I use masks for the ROIs). So I have used different thresholds for different regions. Now my concern is that by choosing different thresholds, different activation types will be seen. So will it affect the eigenvariate too much?

Should I use all the time the same threshold for extracting the ROIs?  

Thanks in advance,

 

Narges Radman

 

Laboratory for Cognitive and Neurological Sciences

Neurology Unit, Medicine Department, University of Fribourg

Ch. du Musée 5, 1700 Fribourg, Switzerland
http://www.unifr.ch/neurology/
http://www.unifr.ch/med/nibsi​

From: Zeidman, Peter <[log in to unmask]>
Sent: Wednesday, August 19, 2015 8:33 AM
To: RADMAN Narges; [log in to unmask]
Subject: RE: ROIs for DCM

 

Hi Narges,

There needs to be some experimental effect in your ROIs for the DCM to explain. If there isn’t, then the DCM won’t fit the signal. There’s nothing special about any particular statistical threshold – you just need to convince yourself there’s something interesting going on that is worth applying DCM to. Often, people look for a group-level effect and extract the individual subjects’ ROIs in the corresponding locations at a very liberal threshold.

 

So, try using masks from the literature, and see if you get significant results. You can sanity check your models by checking the explained variance (spm_dcm_fmri_check()). Note that when you extract ROIs using the ‘eigenvariate’ button, the cursor will jump to the nearest peak which could be in another brain region. So make sure to set the threshold significantly liberally that you activated voxels in your ROI.

 

Best,

Peter

 

From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of RADMAN Narges
Sent: 18 August 2015 14:35
To: [log in to unmask]
Subject: [SPM] ROIs for DCM

 

Dear DCM experts, 

 

I am using DCM on a small number of patients (at single subject level). Regarding the regions of interest I have two questions : 

1. To define my regions of interest, should I necessarily select the regions which are commonly activated in all participants/all conditions of interest or I can select the regions based on the current literature even if they are not significantly activated in the effect of interest?

 

2. To extract the time series of the ROIs from the t-contrast of the “condition of interest” in a sphere around the maxima, should the maxima be significant? 

 

Many thanks in advance, 

 

Narges