help > gPPI versus weighted GLM for Block design
Showing 1-4 of 4 posts
Feb 7, 2020 03:02 PM | Emily Belleau - McLean Hospital/Harvard Medical School
gPPI versus weighted GLM for Block design
Hello,
Thanks so much for the excellent posts explaining gPPI versus weighted GLM for use with block task designs! They have been very helpful! I had some further follow-up questions that have arisen from a reviewer for a manuscript for publication.
1. I chose to do a weighted GLM but they were insisting that I do a gPPI analysis. However, I noticed in one of your posts, weighted GLM and gPPI are both perfectly valid approaches for task block designs and are about equally as prevalent in the literature. I was wondering if you know of any relevant literature that may help back up these statements?
2. I wanted to ensure that I have correctly set up my weighted GLM.
My task consists of a "trauma" block that included in each trial a a) 60 fixation period b.) a 30 second presentation of a trauma narrative, c. a 30 second recall period, and d.) a 60 second relax period where they were asked to let go of the traumatic event
*The same structure was present in the neutral block, except the trauma narrative was replaced with a neutral narrative
In my analysis I modeled the 30 second narrative presentation, the 30 second recall period, and the 60 second recovery period for both trauma and neutral blocks. I did not model the initial 60 second fixation period. Does this seem correct?
3. I examined both "absolute" and "relative connectivity" measures. We were specifically interested in the "recall" period so I examined amygdala-whole brain resting state connectivity in relation to PTSD symptoms during both trauma and neutral recall period separately (absolute connectivity) and also looked at PTSD symptoms in relation to trauma recall relative to neutral recall resting state connectivity (1, -1 contrast; relative connectivity). Is their merits to looking at both relative and absolute measures of connectivity, or perhaps just relative?
I really appreciate any guidance! And my apologies for the long winded post!
Emily
Thanks so much for the excellent posts explaining gPPI versus weighted GLM for use with block task designs! They have been very helpful! I had some further follow-up questions that have arisen from a reviewer for a manuscript for publication.
1. I chose to do a weighted GLM but they were insisting that I do a gPPI analysis. However, I noticed in one of your posts, weighted GLM and gPPI are both perfectly valid approaches for task block designs and are about equally as prevalent in the literature. I was wondering if you know of any relevant literature that may help back up these statements?
2. I wanted to ensure that I have correctly set up my weighted GLM.
My task consists of a "trauma" block that included in each trial a a) 60 fixation period b.) a 30 second presentation of a trauma narrative, c. a 30 second recall period, and d.) a 60 second relax period where they were asked to let go of the traumatic event
*The same structure was present in the neutral block, except the trauma narrative was replaced with a neutral narrative
In my analysis I modeled the 30 second narrative presentation, the 30 second recall period, and the 60 second recovery period for both trauma and neutral blocks. I did not model the initial 60 second fixation period. Does this seem correct?
3. I examined both "absolute" and "relative connectivity" measures. We were specifically interested in the "recall" period so I examined amygdala-whole brain resting state connectivity in relation to PTSD symptoms during both trauma and neutral recall period separately (absolute connectivity) and also looked at PTSD symptoms in relation to trauma recall relative to neutral recall resting state connectivity (1, -1 contrast; relative connectivity). Is their merits to looking at both relative and absolute measures of connectivity, or perhaps just relative?
I really appreciate any guidance! And my apologies for the long winded post!
Emily
Feb 9, 2020 03:02 PM | Alfonso Nieto-Castanon - Boston University
RE: gPPI versus weighted GLM for Block design
Hi Emily,
Regarding (1) I am sorry I do not know of any reference explicitly comparing wGLM and gPPI approaches. It is nevertheless rather straightforward to show that, in your case with block lengths of 30s or above, the two approaches will be almost identical. The image attached shows, for example, the difference in the task-related effects estimated using wGLM and gPPI in a simple simulated timeseries with TR=2s, and an alternating block-design with two-conditions and block lengths ranging from 4s to 60s. The point of this plot is to show that for block lengths of around 15s or above, not only the two approaches provide unbiased estimates of the task-related connectivity differences (i.e. mean error is zero consistently for both methods and across all block-length scenarios evaluated) and offer almost identical level of precision of these estimates (i.e. the standard errors of the two methods are also very similar for all N>15s scenarios), but also that the estimates that the two methods provide are themselves very similar (i.e. the standard error of the difference between the gPPI and the wGLM estimates is very low compared to the standard error of each of these individual estimates, again for all N>15s scenarios). For block lengths below 15 or 10s the comparison becomes more complicated, but that is the topic for another day. In your experiment you seem to be mixing blocks of 30s and 60s, so that puts you well beyond the point where the two methods become effectively equivalent. Please feel free to use any of the above in your response to the reviewers if you wish to do so. I am also including below the code for replicating running these simulations and create the attached figure in case you or anybody else wants to play with it and/or help explore this issue further.
Regarding (2), yes, that looks perfectly correct, no issues at all.
Regarding (3), yes, it is almost always the case that exploring absolute measures of connectivity (in your case connectivity during recall period, either separately or jointly across both the trauma and neutral narrative conditions) has the potential of providing complementary and useful information beyond that which can be gained by looking only at relative measures of connectivity (e.g. comparing differences in connectivity during recall after trauma vs. neutral narratives). That is in fact one of the main reasons why, despite gPPI and wGLM being almost equivalent for block designs, I generally recommend using wGLM in these scenarios, because of its increased facility to estimate and evaluate this sort of absolute connectivity measures
Hope this helps
Alfonso
------ simulation code for reference ---
N=300; % length (in samples) of timeseries
TR=2; % TR (s)
Nrepeat=1e5; % number of simulations
DOCOMP=true; % true/false compute simulations
DOPLOT=true; % true/false create plot
if DOCOMP
Rall=[];
Dall=[];
Dref=[];
Lall=2*TR+(60-2*TR)*linspace(0,1).^2; % block length (s)
for nL=1:numel(Lall),
L=Lall(nL)/2;
R=[];
h=spm_hrf(2);
for nrepeat=1:Nrepeat
y=convn(randn(N+numel(h)-1,2),h,'valid'); % random data with reasonable temporal crosscorrelation
x1=sin(pi*cumsum(2*rand(N+numel(h)-1,1))/L)>=0; % blocks for condition 1 (note: added random jitter)
x2=1-x1; % blocks for condition 2 (note: added random jitter)
x1=max(0,convn(x1,h,'valid')); % main effects for condition 1
x2=max(0,convn(x2,h,'valid')); % main effects for condition 2
B0=randn; % task-specific association in this simulation (true level)
y(:,2)=y(:,2)+B0*y(:,1).*x1; % only present in condition 1
% remove session and main task effects
x0=[ones(N,1) x1 x2];
y=y-x0*(pinv(x0'*x0)*x0'*y);
% compute ppi parameters (condition 2 = baseline)
X=[x1 y(:,1) x1.*y(:,1)];
B=X\y(:,2);
B1=B(3); % task-specific association (estimated by gPPI)
% computed wGLM parameters for condition 1
X=x1.*[ones(N,1) y(:,1)];
Y=x1.*y(:,2);
B=X\Y;
B2a=B(2);
% computed wGLM parameters for condition 2
X=x2.*[ones(N,1) y(:,1)];
Y=x2.*y(:,2);
B=X\Y;
B2b=B(2);
B2=B2a-B2b; % task-specific association (estimated by wGLM)
R=[R; [B1 B2 B0 B2a B2b]];
end
r=corrcoef(R);
disp(r);
r1=r(1,3);
r2=r(2,3);
disp([L r1 r2]);
Dref(nL,:)=[R(:,1);R(:,2);R(:,3)]';
Rall(nL,:)=[r1 r2];
Dall(nL,:)=(R(:,2)-R(:,1))';
if 1
temp=Dref(:,1:Nrepeat)-Dref(:,2*Nrepeat+(1:Nrepeat));
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'-','linewidth',3); hold on;
temp=Dref(:,1*Nrepeat+(1:Nrepeat))-Dref(:,2*Nrepeat+(1:Nrepeat));
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'-','linewidth',3); hold off
drawnow
end
end
Dref=reshape(Dref,size(Dref,1),[],3);
end
if DOPLOT
clf
hold on;
temp=Dref(:,:,1)-Dref(:,:,3); % gPPI - true
h1=plot(Lall(1:nL),(mean((temp),2)),'b-','linewidth',3);
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'b:','linewidth',3);
plot(Lall(1:nL),-sqrt(mean(abs(temp).^2,2)),'b:','linewidth',3);
temp=Dref(:,:,2)-Dref(:,:,3); % wGLM - true
h2=plot(Lall(1:nL),(mean((temp),2)),'r-','linewidth',3);
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'r:','linewidth',3);
plot(Lall(1:nL),-sqrt(mean(abs(temp).^2,2)),'r:','linewidth',3);
temp=Dref(:,:,1)-Dref(:,:,2); % gPPI - wGLM
h3=plot(Lall(1:nL),(mean((temp),2)),'g-','linewidth',3);
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'g:','linewidth',3);
plot(Lall(1:nL),-sqrt(mean(abs(temp).^2,2)),'g:','linewidth',3);
hold off;
set(gca,'xlim',Lall([1 end]));
grid on
xlabel('Block length (s)')
set(gca,'units','norm','position',[.2 .2 .6 .6])
ylabel({'Estimation Error',sprintf('(mean %c standard error)',177)})
set(gcf,'color','w')
legend([h1,h2,h3],{'gPPI','wGLM','gPPI - wGLM'})
end
Originally posted by Emily Belleau:
Regarding (1) I am sorry I do not know of any reference explicitly comparing wGLM and gPPI approaches. It is nevertheless rather straightforward to show that, in your case with block lengths of 30s or above, the two approaches will be almost identical. The image attached shows, for example, the difference in the task-related effects estimated using wGLM and gPPI in a simple simulated timeseries with TR=2s, and an alternating block-design with two-conditions and block lengths ranging from 4s to 60s. The point of this plot is to show that for block lengths of around 15s or above, not only the two approaches provide unbiased estimates of the task-related connectivity differences (i.e. mean error is zero consistently for both methods and across all block-length scenarios evaluated) and offer almost identical level of precision of these estimates (i.e. the standard errors of the two methods are also very similar for all N>15s scenarios), but also that the estimates that the two methods provide are themselves very similar (i.e. the standard error of the difference between the gPPI and the wGLM estimates is very low compared to the standard error of each of these individual estimates, again for all N>15s scenarios). For block lengths below 15 or 10s the comparison becomes more complicated, but that is the topic for another day. In your experiment you seem to be mixing blocks of 30s and 60s, so that puts you well beyond the point where the two methods become effectively equivalent. Please feel free to use any of the above in your response to the reviewers if you wish to do so. I am also including below the code for replicating running these simulations and create the attached figure in case you or anybody else wants to play with it and/or help explore this issue further.
Regarding (2), yes, that looks perfectly correct, no issues at all.
Regarding (3), yes, it is almost always the case that exploring absolute measures of connectivity (in your case connectivity during recall period, either separately or jointly across both the trauma and neutral narrative conditions) has the potential of providing complementary and useful information beyond that which can be gained by looking only at relative measures of connectivity (e.g. comparing differences in connectivity during recall after trauma vs. neutral narratives). That is in fact one of the main reasons why, despite gPPI and wGLM being almost equivalent for block designs, I generally recommend using wGLM in these scenarios, because of its increased facility to estimate and evaluate this sort of absolute connectivity measures
Hope this helps
Alfonso
------ simulation code for reference ---
N=300; % length (in samples) of timeseries
TR=2; % TR (s)
Nrepeat=1e5; % number of simulations
DOCOMP=true; % true/false compute simulations
DOPLOT=true; % true/false create plot
if DOCOMP
Rall=[];
Dall=[];
Dref=[];
Lall=2*TR+(60-2*TR)*linspace(0,1).^2; % block length (s)
for nL=1:numel(Lall),
L=Lall(nL)/2;
R=[];
h=spm_hrf(2);
for nrepeat=1:Nrepeat
y=convn(randn(N+numel(h)-1,2),h,'valid'); % random data with reasonable temporal crosscorrelation
x1=sin(pi*cumsum(2*rand(N+numel(h)-1,1))/L)>=0; % blocks for condition 1 (note: added random jitter)
x2=1-x1; % blocks for condition 2 (note: added random jitter)
x1=max(0,convn(x1,h,'valid')); % main effects for condition 1
x2=max(0,convn(x2,h,'valid')); % main effects for condition 2
B0=randn; % task-specific association in this simulation (true level)
y(:,2)=y(:,2)+B0*y(:,1).*x1; % only present in condition 1
% remove session and main task effects
x0=[ones(N,1) x1 x2];
y=y-x0*(pinv(x0'*x0)*x0'*y);
% compute ppi parameters (condition 2 = baseline)
X=[x1 y(:,1) x1.*y(:,1)];
B=X\y(:,2);
B1=B(3); % task-specific association (estimated by gPPI)
% computed wGLM parameters for condition 1
X=x1.*[ones(N,1) y(:,1)];
Y=x1.*y(:,2);
B=X\Y;
B2a=B(2);
% computed wGLM parameters for condition 2
X=x2.*[ones(N,1) y(:,1)];
Y=x2.*y(:,2);
B=X\Y;
B2b=B(2);
B2=B2a-B2b; % task-specific association (estimated by wGLM)
R=[R; [B1 B2 B0 B2a B2b]];
end
r=corrcoef(R);
disp(r);
r1=r(1,3);
r2=r(2,3);
disp([L r1 r2]);
Dref(nL,:)=[R(:,1);R(:,2);R(:,3)]';
Rall(nL,:)=[r1 r2];
Dall(nL,:)=(R(:,2)-R(:,1))';
if 1
temp=Dref(:,1:Nrepeat)-Dref(:,2*Nrepeat+(1:Nrepeat));
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'-','linewidth',3); hold on;
temp=Dref(:,1*Nrepeat+(1:Nrepeat))-Dref(:,2*Nrepeat+(1:Nrepeat));
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'-','linewidth',3); hold off
drawnow
end
end
Dref=reshape(Dref,size(Dref,1),[],3);
end
if DOPLOT
clf
hold on;
temp=Dref(:,:,1)-Dref(:,:,3); % gPPI - true
h1=plot(Lall(1:nL),(mean((temp),2)),'b-','linewidth',3);
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'b:','linewidth',3);
plot(Lall(1:nL),-sqrt(mean(abs(temp).^2,2)),'b:','linewidth',3);
temp=Dref(:,:,2)-Dref(:,:,3); % wGLM - true
h2=plot(Lall(1:nL),(mean((temp),2)),'r-','linewidth',3);
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'r:','linewidth',3);
plot(Lall(1:nL),-sqrt(mean(abs(temp).^2,2)),'r:','linewidth',3);
temp=Dref(:,:,1)-Dref(:,:,2); % gPPI - wGLM
h3=plot(Lall(1:nL),(mean((temp),2)),'g-','linewidth',3);
plot(Lall(1:nL),sqrt(mean(abs(temp).^2,2)),'g:','linewidth',3);
plot(Lall(1:nL),-sqrt(mean(abs(temp).^2,2)),'g:','linewidth',3);
hold off;
set(gca,'xlim',Lall([1 end]));
grid on
xlabel('Block length (s)')
set(gca,'units','norm','position',[.2 .2 .6 .6])
ylabel({'Estimation Error',sprintf('(mean %c standard error)',177)})
set(gcf,'color','w')
legend([h1,h2,h3],{'gPPI','wGLM','gPPI - wGLM'})
end
Originally posted by Emily Belleau:
Hello,
Thanks so much for the excellent posts explaining gPPI versus weighted GLM for use with block task designs! They have been very helpful! I had some further follow-up questions that have arisen from a reviewer for a manuscript for publication.
1. I chose to do a weighted GLM but they were insisting that I do a gPPI analysis. However, I noticed in one of your posts, weighted GLM and gPPI are both perfectly valid approaches for task block designs and are about equally as prevalent in the literature. I was wondering if you know of any relevant literature that may help back up these statements?
2. I wanted to ensure that I have correctly set up my weighted GLM.
My task consists of a "trauma" block that included in each trial a a) 60 fixation period b.) a 30 second presentation of a trauma narrative, c. a 30 second recall period, and d.) a 60 second relax period where they were asked to let go of the traumatic event
*The same structure was present in the neutral block, except the trauma narrative was replaced with a neutral narrative
In my analysis I modeled the 30 second narrative presentation, the 30 second recall period, and the 60 second recovery period for both trauma and neutral blocks. I did not model the initial 60 second fixation period. Does this seem correct?
3. I examined both "absolute" and "relative connectivity" measures. We were specifically interested in the "recall" period so I examined amygdala-whole brain resting state connectivity in relation to PTSD symptoms during both trauma and neutral recall period separately (absolute connectivity) and also looked at PTSD symptoms in relation to trauma recall relative to neutral recall resting state connectivity (1, -1 contrast; relative connectivity). Is their merits to looking at both relative and absolute measures of connectivity, or perhaps just relative?
I really appreciate any guidance! And my apologies for the long winded post!
Emily
Thanks so much for the excellent posts explaining gPPI versus weighted GLM for use with block task designs! They have been very helpful! I had some further follow-up questions that have arisen from a reviewer for a manuscript for publication.
1. I chose to do a weighted GLM but they were insisting that I do a gPPI analysis. However, I noticed in one of your posts, weighted GLM and gPPI are both perfectly valid approaches for task block designs and are about equally as prevalent in the literature. I was wondering if you know of any relevant literature that may help back up these statements?
2. I wanted to ensure that I have correctly set up my weighted GLM.
My task consists of a "trauma" block that included in each trial a a) 60 fixation period b.) a 30 second presentation of a trauma narrative, c. a 30 second recall period, and d.) a 60 second relax period where they were asked to let go of the traumatic event
*The same structure was present in the neutral block, except the trauma narrative was replaced with a neutral narrative
In my analysis I modeled the 30 second narrative presentation, the 30 second recall period, and the 60 second recovery period for both trauma and neutral blocks. I did not model the initial 60 second fixation period. Does this seem correct?
3. I examined both "absolute" and "relative connectivity" measures. We were specifically interested in the "recall" period so I examined amygdala-whole brain resting state connectivity in relation to PTSD symptoms during both trauma and neutral recall period separately (absolute connectivity) and also looked at PTSD symptoms in relation to trauma recall relative to neutral recall resting state connectivity (1, -1 contrast; relative connectivity). Is their merits to looking at both relative and absolute measures of connectivity, or perhaps just relative?
I really appreciate any guidance! And my apologies for the long winded post!
Emily
Feb 12, 2020 02:02 AM | Emily Belleau - McLean Hospital/Harvard Medical School
RE: gPPI versus weighted GLM for Block design
Thank you so much Alfonso! This was incredibly helpful and
thorough! And thank you for running those simulations. I am very
grateful.
I went ahead and ran a gPPI on my data to compare the results with the weighted GLM. I then looked at the trauma recall versus neutral recall contrast in relation to PTSD symptoms. For both gPPI and weighted GLM, Both analyses at p <0.0001 FWE corrected to p <0.05, nothing was significant (what I had expected based on my prior analysis). For further exploration for comparison purposes, I set the height threshold to p <0.05 cluster corrected to FWE p <0.05. For both gPPI and weighted GLM, I got the same cluster in the postcentral gyrus, but with the weighted GLM, I got a few additional clusters.
Also, is examining one single condition (e.g., trauma recall) the same for gPPI versus weighted GLM? Here is where I saw things start to diverge more. I had findings when doing weighted GLM but no findings with gPPI. Is this because with weighted GLM you can/are looking at absolute connectivity but with gPPI you can still only look at relative connectivity (trauma recall relative to the implicit baseline)?
I am still trying to wrap my head around absolute connectivity. What exactly are we looking at in the context of a weighted GLM? Is it not still looking at the single condition relative to baseline (or what you did not model)?
Thanks again for your guidance, it has been very helpful!
Emily
I went ahead and ran a gPPI on my data to compare the results with the weighted GLM. I then looked at the trauma recall versus neutral recall contrast in relation to PTSD symptoms. For both gPPI and weighted GLM, Both analyses at p <0.0001 FWE corrected to p <0.05, nothing was significant (what I had expected based on my prior analysis). For further exploration for comparison purposes, I set the height threshold to p <0.05 cluster corrected to FWE p <0.05. For both gPPI and weighted GLM, I got the same cluster in the postcentral gyrus, but with the weighted GLM, I got a few additional clusters.
Also, is examining one single condition (e.g., trauma recall) the same for gPPI versus weighted GLM? Here is where I saw things start to diverge more. I had findings when doing weighted GLM but no findings with gPPI. Is this because with weighted GLM you can/are looking at absolute connectivity but with gPPI you can still only look at relative connectivity (trauma recall relative to the implicit baseline)?
I am still trying to wrap my head around absolute connectivity. What exactly are we looking at in the context of a weighted GLM? Is it not still looking at the single condition relative to baseline (or what you did not model)?
Thanks again for your guidance, it has been very helpful!
Emily
Feb 12, 2020 05:02 PM | Alfonso Nieto-Castanon - Boston University
RE: gPPI versus weighted GLM for Block design
Hi Emily,
Right, precisely. When looking at one single condition (e.g. trauma recall) in gPPI you are effectively measuring the difference in absolute connectivity between that condition and the implicit baseline (the initial 60s fixation period). If you want to explore the same effects using wGLM you would need to first explicitly define the "baseline" condition, and then simply look at "trauma_recall - baseline" differences in your wGLM results. The general rule when comparing gPPI to wGLM is that individual-condition gPPI effects (e.g. "conditionA") are equivalent/similar to between-condition contrasts in wGLM (e.g. "conditionA-baseline"), i.e.
gPPI("conditionA") ~= wGLM("conditionA") - wGLM("baseline")
That is why "conditionA-conditionB" results are almost identical in both gPPI and wGLM, while "conditionA" results are not, i.e.
gPPI("conditionA") - gPPI("conditionB") ~= wGLM("conditionA") - wGLM("baseline") - wGLM("conditionB") + wGLM("baseline") = ...
... = wGLM("conditionA") - wGLM("conditionB")
Hope this helps
Alfonso
Originally posted by Emily Belleau:
Right, precisely. When looking at one single condition (e.g. trauma recall) in gPPI you are effectively measuring the difference in absolute connectivity between that condition and the implicit baseline (the initial 60s fixation period). If you want to explore the same effects using wGLM you would need to first explicitly define the "baseline" condition, and then simply look at "trauma_recall - baseline" differences in your wGLM results. The general rule when comparing gPPI to wGLM is that individual-condition gPPI effects (e.g. "conditionA") are equivalent/similar to between-condition contrasts in wGLM (e.g. "conditionA-baseline"), i.e.
gPPI("conditionA") ~= wGLM("conditionA") - wGLM("baseline")
That is why "conditionA-conditionB" results are almost identical in both gPPI and wGLM, while "conditionA" results are not, i.e.
gPPI("conditionA") - gPPI("conditionB") ~= wGLM("conditionA") - wGLM("baseline") - wGLM("conditionB") + wGLM("baseline") = ...
... = wGLM("conditionA") - wGLM("conditionB")
Hope this helps
Alfonso
Originally posted by Emily Belleau:
Thank you so much Alfonso! This was incredibly
helpful and thorough! And thank you for running those simulations.
I am very grateful.
I went ahead and ran a gPPI on my data to compare the results with the weighted GLM. I then looked at the trauma recall versus neutral recall contrast in relation to PTSD symptoms. For both gPPI and weighted GLM, Both analyses at p <0.0001 FWE corrected to p <0.05, nothing was significant (what I had expected based on my prior analysis). For further exploration for comparison purposes, I set the height threshold to p <0.05 cluster corrected to FWE p <0.05. For both gPPI and weighted GLM, I got the same cluster in the postcentral gyrus, but with the weighted GLM, I got a few additional clusters.
Also, is examining one single condition (e.g., trauma recall) the same for gPPI versus weighted GLM? Here is where I saw things start to diverge more. I had findings when doing weighted GLM but no findings with gPPI. Is this because with weighted GLM you can/are looking at absolute connectivity but with gPPI you can still only look at relative connectivity (trauma recall relative to the implicit baseline)?
I am still trying to wrap my head around absolute connectivity. What exactly are we looking at in the context of a weighted GLM? Is it not still looking at the single condition relative to baseline (or what you did not model)?
Thanks again for your guidance, it has been very helpful!
Emily
I went ahead and ran a gPPI on my data to compare the results with the weighted GLM. I then looked at the trauma recall versus neutral recall contrast in relation to PTSD symptoms. For both gPPI and weighted GLM, Both analyses at p <0.0001 FWE corrected to p <0.05, nothing was significant (what I had expected based on my prior analysis). For further exploration for comparison purposes, I set the height threshold to p <0.05 cluster corrected to FWE p <0.05. For both gPPI and weighted GLM, I got the same cluster in the postcentral gyrus, but with the weighted GLM, I got a few additional clusters.
Also, is examining one single condition (e.g., trauma recall) the same for gPPI versus weighted GLM? Here is where I saw things start to diverge more. I had findings when doing weighted GLM but no findings with gPPI. Is this because with weighted GLM you can/are looking at absolute connectivity but with gPPI you can still only look at relative connectivity (trauma recall relative to the implicit baseline)?
I am still trying to wrap my head around absolute connectivity. What exactly are we looking at in the context of a weighted GLM? Is it not still looking at the single condition relative to baseline (or what you did not model)?
Thanks again for your guidance, it has been very helpful!
Emily