Hi Peter,
I am not sure why those files crashed your outlook. Please accept my
apologies for that. I will use different mechanism to share files in future.
Perfect, changing from structure to cell array resolved that issue. Thanks
for that.
Lastly, I performed following steps:
M = struct();
M.alpha = 1;
M.beta = 16;
M.hE = 0;
M.hC = 1/16;
M.Q = 'single';
M.X = [1 1
1 -1];
PEB3 = spm_dcm_peb({PEB_A; PEB_B}, M);
where my PEB_A is:
PEB_A =
struct with fields:
Snames: {26×1 cell}
Pnames: {16×1 cell}
Pind: [16×1 double]
Xnames: {'Covariate 1' 'Covariate 2' 'Covariate 3' 'Covariate 4'
'Covariate 5' 'Covariate 6'}
M: [1×1 struct]
Ep: [16×6 double]
Eh: -6.2590
Ch: 0.0041
Cp: [96×96 double]
Ce: [16×16 double]
F: -2.5418e+04
and PEB_B is:
struct with fields:
Snames: {26×1 cell}
Pnames: {16×1 cell}
Pind: [16×1 double]
Xnames: {'Covariate 1' 'Covariate 2' 'Covariate 3' 'Covariate 4'
'Covariate 5' 'Covariate 6'}
M: [1×1 struct]
Ep: [16×6 double]
Eh: -6.2465
Ch: 0.0041
Cp: [96×96 double]
Ce: [16×16 double]
F: -728.5970
Here 'Covariate 1' and 'Covariate 2' for PEB_A and PEB_B represent means
over pre- and post-treatment conditions and difference between pre- and
post-treatment conditions for group A and group B.
Covariates 3 to 6 are other covariates like age, sex, amount of dose and
time since stroke.
My PEB3 looks like:
PEB3 =
struct with fields:
Snames: {2×1 cell}
Pnames: {1×96 cell}
Pind: [96×1 double]
Xnames: {'Covariate 1' 'Covariate 2'}
M: [1×1 struct]
Ep: [96×2 double]
Eh: 0.1460
Ch: 0.0518
Cp: [192×192 double]
Ce: [96×96 double]
F: -2.6137e+04
Then I ran:
BMA = spm_dcm_peb_bmc(PEB3(1));
spm_dcm_peb_review(BMA,GCM_updated)
As you notice, here I am getting just 2 covariates rather than 4 that we
discussed in our previous conversation.
I am not sure what I am doing wrong here, or if something is missing, or I
am misunderstanding something.
Thanks.
On Wed, Mar 27, 2019 at 3:30 AM Zeidman, Peter <[log in to unmask]>
wrote:
> Dear Martin
>
> This turned out to be a simple problem – the function spm_dcm_peb expects
> a cell array of DCMs or PEBs. However, you gave it a structure array. So,
> simply change:
>
>
>
> PEB3 = spm_dcm_peb([PEB_A; PEB_B], M);
>
>
>
> To:
>
>
>
> PEB3 = spm_dcm_peb({PEB_A; PEB_B}, M);
>
>
>
> And it will work fine.
>
>
>
> Best
>
> Peter
>
> PS the email you sent me with the files crashes my Outlook every time I
> click on it! Perhaps it would be better to share files through some other
> mechanism.
>
>
>
> *From:* Zeidman, Peter
> *Sent:* 21 March 2019 09:15
> *To:* 'Martin Juneja' <[log in to unmask]>
> *Cc:* [log in to unmask]
> *Subject:* RE: [SPM] Fwd: Defining covariates in DCM M.X matrix
>
>
>
> Dear Martin
>
> (1). Following your suggestion to model 2x2 design, after I performed
> following steps:
>
>
>
> %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>
> load PEB_G1.mat % from pre-post comparison for group 1
>
> PEB_G1 = PEB;
>
> load PEB_G2.mat % from pre-post comparison for group 2
>
> PEB_G2 = PEB;
>
>
>
> N = 2;
>
> M = struct();
>
> M.alpha = 1;
>
> M.beta = 16;
>
> M.hE = 0;
>
> M.hC = 1/16;
>
> M.Q = 'single';
>
>
>
> M.X = [1 1
>
> 1 -1];
>
>
>
> fields = {'A'};
>
> PEB3 = spm_dcm_peb([PEB_G1; PEB_G2], M,fields);
>
> %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
>
>
>
> I am getting following error:
>
> Undefined function or variable 'pC'.
>
>
>
> Error in spm_find_pC (line 52)
>
> if isstruct(pC)
>
>
>
> Error in spm_dcm_peb (line 144)
>
> q = spm_find_pC(DCM,field); % parameter indices
>
>
>
> Error in A_Run_large_DCM_NextLevel (line 21)
>
> PEB3 = spm_dcm_peb([PEB_A; PEB_B], M,fields);
>
>
>
> Does it resolve it if you remove the ‘fields’ options from the
> specification of PEB3? Otherwise, please can I check whether you are using
> the latest release of SPM12, released 14ht November 2018? There were some
> fixes in there for the PEB-of-PEBs approach.
> https://www.fil.ion.ucl.ac.uk/spm/software/spm12/
>
>
>
> (2). Also, as you mentioned that * "**in this final model, you’ll have
> parameters quantifying the mean connectivity across all subjects, the group
> difference in mean connectivity, the mean effect of treatment, and the
> group difference in the effect of treatment (i.e. the interaction)", *I
> was wondering how we are getting all of these i.e. mean connectivity
> across all subjects, the group difference in mean connectivity, the mean
> effect of treatment, and the group difference in the effect of treatment
> (i.e. the interaction), because M.X is just [1 1 1-1], which should give me
> only first two things (i.e., mean connectivity across all subjects and the
> group difference in mean connectivity only), out of four you mentioned. I
> apologize if I did not understand this correctly.
>
>
>
> I would really appreciate any additional clarification on these two issues.
>
>
>
> Excellent question. To recap, the vector of second level parameters that
> your third level GLM is trying to explain are:
>
>
>
> Group 1 Mean
>
> Group 1 Treatment
>
> Group 2 Mean
>
> Group 2 Treatment
>
>
>
> There are two parts to the level design matrix. You specify the
> between-groups design matrix Xb:
>
>
>
> 1 1
>
> 1 -1
>
>
>
> There’s also a within-groups design matrix, specified automatically by the
> software: Xw = eye(n) where n is the number of effects being brought up
> from the level below (n=2 in this example – mean and treatment):
>
>
>
> 1 0
>
> 0 1
>
>
>
> The actual design matrix is calculated by: X = kron(Xb, Xw):
>
>
>
> 1 0 1 0
>
> 0 1 0 1
>
> 1 0 -1 0
>
> 0 1 0 -1
>
>
>
> As you can see, the first regressor (column) is the mean of the group
> means, the second regressor is the mean of the treatment effects, the third
> regressor is the difference between group means and the fourth regressor is
> the difference in treatment effects between groups. You’ll have one
> parameter estimate for each of these four columns.
>
>
>
> For more about how the design matrices are specified, please see part 2 of
> the tutorial paper at https://github.com/pzeidman/dcm-peb-example .
>
>
>
> Best
>
> Peter
>
>
>
> Thank you so much once again.
>
>
>
> On Tue, Mar 19, 2019 at 3:59 AM Zeidman, Peter <[log in to unmask]>
> wrote:
>
> Dear Martin
>
> On second thoughts, an alternative and potentially more elegant way to
> model your between-subjects 2x2 design would be to have a 3-level PEB
> model: DCMs at the first level, group at the second level and treatment at
> the third level. This would properly account for differences in
> between-subject variability between groups.
>
>
>
> For example, say you have DCMs for each group: GCM1 and GCM2. At the
> second level, model them using:
>
>
>
> PEB2a = spm_dcm_peb(GCM1, M1);
>
> PEB2b = spm_dcm_peb(GCM2, M2);
>
>
>
> Where the design matrix for first group, M1.X, has regressors for mean,
> treatment (pre vs post) and covariates such as age (mean centred within
> this group). Similarly for the second group. Then, take the mean
> connectivity and effect of treatment from each group up to the third level:
>
>
>
> PEB3 = spm_dcm_peb([PEB2a; PEB2b], M3, fields);
>
>
>
> Where the design matrix M3 has regressors for the mean across both groups
> and the difference between groups:
>
>
>
> M3.X = [1 1
>
> 1 -1];
>
>
>
> So, in this final model, you’ll have parameters quantifying the mean
> connectivity across all subjects, the group difference in mean
> connectivity, the mean effect of treatment, and the group difference in the
> effect of treatment (i.e. the interaction). You can use the ‘fields’ option
> in PEB3 to limit which parameters go up to the third level (i.e. no need
> for uninteresting covariates such as age).
>
>
>
> I hope that helps
>
> Peter
>
>
>
> *From:* Zeidman, Peter
> *Sent:* 19 March 2019 10:01
> *To:* 'Martin Juneja' <[log in to unmask]>; [log in to unmask]
> *Subject:* RE: [SPM] Fwd: Defining covariates in DCM M.X matrix
>
>
>
> Dear Martin
>
> You have a 2x2 factorial design at the between-subjects level: treatment
> (pre vs post) and group (subject set). So to fully encode this in your GLM,
> you could model the mean (expected by the PEB system to the first column),
> the two main effects (treatment and group) and their interaction. You could
> code the main effects as 1s and -1s as you suggest, making sure you
> mean-correct them, and to generate the interaction regressor you
> element-wise multiply the mean-corrected main effect regressors.
>
>
>
> I am interested in determining whether its treatment 'A' which improves
> effective connectivity within a network for set 1, or its treatment 'B'
> which improves effective connectivity within a network for set 2.
>
>
>
> Those are the simple effects – an effect of treatment in group 1 and an
> effect of treatment in group 2. If that’s all you want, you could just
> include these in the design matrix as you suggest (and mean correct them)
> but it doesn’t fully cover your factorial design without the interaction.
>
>
>
> *Q1:* Whether I mean center 'age', 'sex', 'amount of dose' and 'time
> since stroke' from overall sample i.e., from both conditions together, or
> calculate mean centered values separately for treatments A and B.
>
>
>
> From the overall sample (for age and sex). Just include non-zero values
> for dose and time since stroke in the patient group.
>
>
>
> *Q2:* For M.X matrix, should I use mean centered 'dose used' values only
> for post-treatment conditions and put 0's for pre-treatment condition,
> because there was no dose used during pre-treatment condition? Correct?
>
> *Q3:* For M.X matrix, should I use mean centered 'time since stroke' only
> for post-treatment condition and put 0's for pre-treatment condition? Or it
> will be same values for pre- and post-treatment condition because 'time
> since stroke' values are same for pre- and post conditions.
>
>
>
> Yes but note, for every covariate you add, there’s a new parameter added
> for every DCM connection. So you’ll be using up your degrees of freedom for
> each new covariate. You might want to keep the number of covariates at a
> minimum.
>
>
>
> *Q4: *If I want to compare connectivity for controls with post-treatment
> condition, then again should I mean center 'age' and 'sex' from overall
> sample or separately for controls and patients?
>
>
>
> I think overall should be fine.
>
>
>
> Also, should I include 'time since stroke' and 'amount of dose' used also
> as covariates, which will of course be zeros for controls.
>
> Or there is no need to include 'time since stroke' and 'amount of dose' as
> covariates because I am comparing controls with patients. If so, I doubt
> then how to incorporate the variance as the 'amount of dose used' and 'time
> since stroke' values differ between treatments A and B.
>
>
>
> See above comments – yes you can model them, but there could be a big cost
> in terms of model complexity.
>
>
>
> Best
>
> Peter
>
>
>
>
