 |
 |
Dear Dr. Peter Zeidman, Thank you for your explanation and clarification. Regarding the explained variance, I find that in all but 1 subject, the variance explained is less than 10%. I also checked the same with 1. same models but self modulations 2. fully connected intrinsic models without self modulations and I get the same result for explained variance. I would really appreciate if you could suggest some other diagnostics that I can run to understand why the models aren't fitting the data. Thanks, Best regards, Atesh On Mon, Mar 9, 2015 at 6:08 PM, Zeidman, Peter <[log in to unmask]> wrote: > Dear Atesh, > > I see you are using 2-state DCM. Two interpret these results you indeed > need to do exp(DCM.Ep.A) and exp(DCM.Ep.B). Having done that, a value of 1 > in Ep.A or Ep.B means no effect. See > https://en.wikibooks.org/wiki/SPM/Two_State_DCM . > > > > You may wish to only look at those parameters significantly different from > their priors: > > > > exp(DCM.Ep.A) .* (DCM.Pp.A > 0.95) > > > > You’ll then find your parameters with exp(-32) ~= 0 disappear. Doing this > on the first model you sent me shows that no parameter has deviated from > its prior. Furthermore, if you run diagnostic checks on this model via > spm_dcm_fmri_check(DCM), the explained variance is 0% and only 1 parameter > is estimable. Therefore, there is something wrong with this model. If this > is common across your subjects, your post-hoc results will be meaningless. > You’ll need to do some more work to investigate why your individual models > aren’t fitting at all before you return to the post-hoc reduction. > > > > Best, > > Peter > > > > *From:* atesh koul [mailto:[log in to unmask]] > *Sent:* 09 March 2015 16:33 > *To:* Zeidman, Peter > *Cc:* [log in to unmask] > > *Subject:* Re: DCM post-hoc family inferences > > > > Dear Dr. Peter Zeidman, > > Thank you for your email. My confusion was whether all the best models > could have the very same structure. Thank you for your help clearing my > confusion. I just have a couple of questions more. > > > In my results, I find with respect to intrinsic connectivity, the two > regions that were not connected, have a Bayesian parameter average under > selected model value as -32, which as I understand from this post ( > https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1412&L=spm&P=R18218&1=spm&9=A&I=-3&J=on&d=No+Match%3BMatch%3BMatches&z=4) > is log scale parameters (Model posterior value = 0). I would like to > confirm if a high negative value thus would mean an (almost) 0 probability. > > Using the same formula for conversion to rate changes, if I understand > correctly, I would find that all the intrinsic parameters have negative > values (or zero). Also, I get negative values in the modulation matrix and > if I understand correctly from your conversation that negative modulation > and negative intrinsic connections would add up ( > https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1407&L=spm&P=R40113&1=spm&9=A&I=-3&J=on&d=No+Match%3BMatch%3BMatches&z=4). > The thing I am not sure about, is whether I should consider the negative > rate constants or the log scale parameters for such an interpretation. > > Also, it seems that the graph overlay of intrinsic connections produced > by spm_dcm_graph uses the log scale parameters and thus suggests a strong > connection between two regions that have a value of -32. I am not sure what > this might suggest. I would really appreciate help interpreting this result. > > Thanks again, > > Best regards, > > Atesh > > > > > > > On Mon, Mar 9, 2015 at 3:51 PM, Zeidman, Peter <[log in to unmask]> > wrote: > > Dear Atesh, > > I hope you don’t mind me CC’ing the SPM mailing list so that other can > benefit. You emailed regarding your family definition for post-hoc DCM. You > were surprised that whichever family you had first in your if…elseif > statement in your family definition function, your models would get > assigned to the first family. > > > > The role of the family function is to map each individual candidate model > to a single family. The free energies of the models in each family are then > pooled to allow families of models to be compared. I think your confusion > is a simple matter regarding the use of “if” and “elseif” statements. When > the first statement in your family function is: > > > > if B_matrix(regionC,regionA,input_index) ~= 0 > > family = 1; > > Your best models are being assigned to family 1, and no other family, > because your best models have the A->C connection. If you change this to: > > > > if B_matrix(regionD,regionA,input_index) ~= 0 > > family = 1; > > > > I expect that some or all of the same best models as before are getting > assigned to the new family 1, because these models have both the A->C > connection and the A->D connection. > > > > Let me give you another example of where this will happen. See lines 76-79 > in your family function: > > > > elseif B_matrix(regionB,regionE,input_index) ~= 0 > > family = 8; > > elseif (B_matrix(regionB,regionE,input_index) && > B_matrix(regionB,regionE,input_index)) ~= 0 > > family = 9; > > > > The second elseif statement can never be executed, because if > B_matrix(regionB,regionE,input_index) ~= 0 is true, then the assigned > family will be 8, and Matlab will never reach the next line of code. This > is the definition of “elseif”. > > > > So, you need to think about how you can define families such that every > model is only assigned to one family. > > > > Please direct any further queries to the SPM mailing list, where one of my > colleagues or I will reply. > > > > Hope that helps, > > Peter > > > > *From:* atesh koul [mailto:[log in to unmask]] > *Sent:* 09 March 2015 13:26 > *To:* Zeidman, Peter > *Subject:* Re: DCM post-hoc family inferences > > > > Dear Dr. Peter Zeidman, > > Thank you for your email. The concern that you have mentioned is correct. > > The problem is actually the same. However, since the data is from a lesser > number of participants, the winning modulatory connection is different. So, > in the same data that I have sent you, if a modulatory connection from > region A to B is tested first (as in the attached which_family2.m file), > the user defined family is selected to be family 1. Alternatively, if a > modulatory connection from region A to C is checked first, the result is > again family 1 (basically all the connections that have DCM.Pp.B value >0) . > > Thanks again, > > Best regards, > > Atesh > > > > >