Dear Andy,
You should either have one model defining each hypothesis, or one family of models defining each hypothesis. So you could one have
model which is bottom-up, one which is top-down (plus you may want a ‘null’ model with neither, depending on whether that’s relevant to you).
If you’d like to use the family approach and have lots of models in each family automatically generated, you want Bayesian Model Reduction
(post-hoc DCM). To do this, define a ‘full’ model (both top-down and bottom up connections). Then run the spm_dcm_post_hoc function. If you look at the help, it also accepts a ‘family’ function which defines how to tell which family a model should be assigned
to. Let us know if you need more help with that.
Best,
Peter.
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
On Behalf Of Andy Yeung
Sent: 05 February 2015 01:14
To: [log in to unmask]
Subject: Re: [SPM] DCM parameters per subject extraction
Dear Peter and Stefan,
Thank you for both of your guidance. If ideal method to test presence or absence of connections is at the model comparison step, ie BMS, does it mean we have to define many models?
If so, is there a way to do it systematically rather than using DCM GUI to manually define every model for every session for every subject?
And my primary objective is to choose the best model/ family to see if modulatory effect acts on bottom-up or top-down connection between 2 VOIs.
What information do you think should be reported? Eg, I commonly see exceedance probability of each model and family, winning family's model structure, VOIs MNI coordinates.
Best,
Andy
On Wed, Feb 4, 2015 at 5:34 PM, Stefan Frässle <[log in to unmask]> wrote:
Hi Andy,
Just as a small note: If you are interested in the value of the posterior probabilities of the parameters after BMA, you could estimate them as follows:
Pp = 1 - spm_Ncdf(0,abs(mEp-pE),sEp^2)
where mEp is the posterior mean, pE is the prior mean and sEp is the standard deviation.
Hope that helps.
Best,
Stefan
Quoting "Zeidman, Peter" <[log in to unmask]>:Hi Andy,
Posterior probabilities aren’t stored in the BMA. You could use the posterior means and SDs to perform t-tests against zero (although the priors aren't exactly zero, so the test wouldn't be perfect). And you could perform paired t-tests to compare parameters if that's of interest. Note that the ideal method for testing the presence or absence of a connection is via model comparison - i.e. the step before the BMA.
Best,
Peter.
From: Andy Yeung [mailto:[log in to unmask]]
Sent: 30 January 2015 01:57
To: Zeidman, Peter
Cc: SPM
Subject: Re: [SPM] DCM parameters per subject extraction
Dear Peter,
Thank you for your guidance. May I know which item in the BMS.mat is storing data for probability of posterior estimates?
I know how to extract mean and SD of the PEs.
Best,
Andy
On Wed, Jan 28, 2015 at 10:45 PM, Zeidman, Peter <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear Andy,
In general it’s better to divide your models into families, if that’s compatible with your experimental design. I would not base conclusions on that one model alone.
Best,
Peter.
From: Andy Yeung [mailto:[log in to unmask]<mailto:[log in to unmask]>]
Sent: 28 January 2015 00:35
To: Zeidman, Peter
Cc: SPM
Subject: Re: [SPM] DCM parameters per subject extraction
Dear Peter and all,
Thanks for guiding me to use BMA. Should I use BMA at model level or at family level? Because if I combine my 6 models into 3 families, this 'strong trend' model belongs to a familiy with exceedance probability of 1.0. Or you think it's better to report posterior estimates for only the 'strong trend' model?
Best,
Andy
On Tue, Jan 27, 2015 at 10:46 PM, Zeidman, Peter <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hi Andy,
The exceedance probability *is* the test of which model stands out statistically. Personally, I don’t think I would accept an XP of 0.8045 to be a “winning” model. It’s a strong trend, but I think I’d want 0.90 or 0.95. Therefore, to look at parameters, I would recommend taking a weighted average of your models (weighted by the posterior probability of each model). This is called Bayesian Model Averaging, which can be switched on in the batch editor when you specify the BMS. You’ll then find posterior estimates and probabilities of each parameter in the BMS.DCM.rfx.bma structure (or BMS.DCM.ffx.bma). The fields in this structure are documented in the SPM manual.
Best,
Peter.
From: Andy Yeung [mailto:[log in to unmask]<mailto:[log in to unmask]>]
Sent: 27 January 2015 03:33
To: Zeidman, Peter; SPM
Subject: Re: [SPM] DCM parameters per subject extraction
Dear Peter and SPM users,
I had 34 subjects and used BMS at the group level analysis to see which model has highest exceedance probability. The best one has 0.8045 and the second one has 0.1955. Do I need any test to claim the best one stands out statistically? And then I want to extract the parameter estimates and standard deviation of the best model. Where are they stored? I can't find them in the BMS.mat generated by group analysis.
Thanks for help,
Andy
On Tue, Jan 20, 2015 at 8:05 PM, Andy Yeung <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Dear Peter,
Thank you very much for your help! Thanks for your wikibook link because I was only reading SPM8 manual DCM practical example section... and could not find instructions about parameter estimates extraction.
Best,
Andy
On Tue, Jan 20, 2015 at 4:47 PM, Zeidman, Peter <[log in to unmask]<mailto:[log in to unmask]>> wrote:
Hi Andy,
For the meaning of A,B and C please see the original DCM paper or this tutorial:
http://en.wikibooks.org/wiki/SPM/The_DCM_Equation._3._Networks_and_Matrices
http://en.wikibooks.org/wiki/SPM/The_DCM_Equation._4._The_State_Equation
The matrix ‘D’ is for non-linear DCMs, which allow regions to directly modulate the connections between other regions. Epsilon, transit and decay are haemodynamic parameters.
Best,
Peter.
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]<mailto:[log in to unmask]>] On Behalf Of Andy Yeung
Sent: 20 January 2015 07:48
To: [log in to unmask]<mailto:[log in to unmask]>
Subject: [SPM] DCM parameters per subject extraction
Dear all SPM and DCM users,
I'd like to ask after loading BMS.mat and going to .mEps and .sEps inside, what does those A, B, C, D, epsilon, transition etc mean? I want to extract intrinsic connections, modulatory parameters and driving input values to check if they are significantly different from 0. Thank you very much in advance.
Best,
Andy