help
48 Subscribers 
help > gPPI: contrasting to baseline, modeling stimuli duration, and temporal derivatives
Showing 1-2 of 2 posts
May 23, 2018 09:05 PM | Jessica Hayes - Wayne State University
gPPI: contrasting to baseline, modeling stimuli duration, and temporal derivatives
Hello,
I’m a graduate student with experience running task GLM
analyses, but my lab is now interested in performing gPPI analyses on fMRI data
collected during a visual encoding task.
We are interested in whether the connectivity of our regions of interest
during encoding is modulated by subsequent memory (hit or miss during retrieval). During retrieval (out of scanner), participants
also indicate their confidence in their decision (high confidence / low
confidence).
Therefore we have three task conditions (Subsequent Hit-High Confidence [Hit-HC],
Subsequent Hit-Low Confidence [Hit-LC], Subsequent Miss), with implicit baseline
as a fourth condition (per [url=https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1312&L=spm&D=0&1=spm&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4&P=414240)]listserv[/url]
post). Our task GLM also includes
temporal derivatives of the three task conditions, as well as six motion
parameters and regressors for motion outliers.
I have a few questions about gPPI:
I’m happy to provide additional detail where helpful. Thank you!
A) “Also note that PPI effects
(interaction terms in PPI model) are always relative to the baseline state (the
baseline state is defined by the zero values of the interaction term), so they
provide a relative measure of connectivity characterizing differential
task-specific effects, rather than an absolute measure of connectivity such as
that estimated using standard functional connectivity analyses (e.g. weighted
correlation measures).” (CONN toolbox manual)
I’m a graduate student with experience running task GLM
analyses, but my lab is now interested in performing gPPI analyses on fMRI data
collected during a visual encoding task.
We are interested in whether the connectivity of our regions of interest
during encoding is modulated by subsequent memory (hit or miss during retrieval). During retrieval (out of scanner), participants
also indicate their confidence in their decision (high confidence / low
confidence).
Therefore we have three task conditions (Subsequent Hit-High Confidence [Hit-HC],
Subsequent Hit-Low Confidence [Hit-LC], Subsequent Miss), with implicit baseline
as a fourth condition (per [url=https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1312&L=spm&D=0&1=spm&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4&P=414240)]listserv[/url]
post). Our task GLM also includes
temporal derivatives of the three task conditions, as well as six motion
parameters and regressors for motion outliers.
I have a few questions about gPPI:
- My understanding of PPI
involves change in coupling across task conditions relative to the baseline
(see footnote A). However, can this approach
be used to investigate modulation of connectivity between task (i.e. encoding)
and baseline? The gPPI script creates
contrasts for ‘none minus Hit-HC’, ‘Hit-HC minus none’, etc. Can the results of these contrasts be
reasonably interpreted as voxels where coupling with the seed region is
modulated between Hit-HC and implicit baseline? This
post would seem to indicate it can, but then what are the task and baseline
connectivity relative to? If it's relative to the other two task conditions
(Hit-LC and Miss), is it correct to assume that a general encoding (Hit-HC +
Hit-LC + Miss) compared to baseline is not a possibility? - Visual stimuli are
presented for four seconds each – are there arguments to be made for modeling
the stimuli as events, or giving them their 4s durations? This [url=https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1612&L=SPM&P=R924&1=SPM&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4]answer[/url]
seems to suggest we should give the 4s duration, but the linked website
primarily focuses on the need to give the duration when it’s variable (which
ours isn’t). - Related to #2, if we do
give the duration of 4s, should this be done for both the individual
first-level analysis and the PPI? - As noted above, our task
GLM includes temporal derivatives for each of the three task conditions, which
the gPPI toolbox also includes in the PPI GLM.
I’m only creating interaction regressors for the three task conditions
and not the temporal derivatives – does the inclusion of the temporal
derivatives in the GLM complicate interpretation of the gPPI results? ([url=https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1304&L=SPM&P=R74001&1=SPM&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4]Answer[/url]) - Building on #4, when
adjusting for the ‘effects of interest’ F-contrast, should that contrast
include the temporal derivative regressors or only the three task regressors?
I’m happy to provide additional detail where helpful. Thank you!
A) “Also note that PPI effects
(interaction terms in PPI model) are always relative to the baseline state (the
baseline state is defined by the zero values of the interaction term), so they
provide a relative measure of connectivity characterizing differential
task-specific effects, rather than an absolute measure of connectivity such as
that estimated using standard functional connectivity analyses (e.g. weighted
correlation measures).” (CONN toolbox manual)
Jun 8, 2018 06:06 PM | Donald McLaren
RE: gPPI: contrasting to baseline, modeling stimuli duration, and temporal derivatives
Hi Jessica,
See inline responses below.
Best Regards,
Donald McLaren, PhD
On Mon, Jun 4, 2018 at 1:11 PM, Jessica Hayes wrote:
Hello,
I'm a graduate student with experience running task GLM analyses, but my lab is now interested in performing gPPI analyses on fMRI data collected during a visual encoding task. We are interested in whether the connectivity of our regions of interest during encoding is modulated by subsequent memory (hit or miss during retrieval). During retrieval (out of scanner), participants also indicate their confidence in their decision (high confidence / low confidence).
Therefore we have three task conditions (Subsequent Hit-High Confidence [Hit-HC], Subsequent Hit-Low Confidence [Hit-LC], Subsequent Miss), with implicit baseline as a fourth condition (per listserv post). Our task GLM also includes temporal derivatives of the three task conditions, as well as six motion parameters and regressors for motion outliers.
I have a few questions about gPPI:
My understanding of PPI involves change in coupling across task conditions relative to the baseline (see footnote A). However, can this approach be used to investigate modulation of connectivity between task (i.e. encoding) and baseline? The gPPI script creates contrasts for 'none minus Hit-HC', 'Hit-HC minus none', etc. Can the results of these contrasts be reasonably interpreted as voxels where coupling with the seed region is modulated between Hit-HC and implicit baseline? This post would seem to indicate it can, but then what are the task and baseline connectivity relative to? If it's relative to the other two task conditions (Hit-LC and Miss), is it correct to assume that a general encoding (Hit-HC + Hit-LC + Miss) compared to baseline contrast is not a possibility?
>>> Yes. Comparisons against the implicit baseline are valid in gPPI.
Visual stimuli are presented for four seconds each – are there arguments to be made for modeling the stimuli as events, or giving them their 4s durations? This answer seems to suggest we should give the 4s duration, but the linked website primarily focuses on the need to give the duration when it's variable (which ours isn't).
>>> The duration provided should be the duration during which you want to compute the interaction. If you want the interaction to only be for first time bin (1/16 TR), then a 0 second duration is fine. However, I'd advise that you model the duration as you probably want to model the interaction for a longer period of time and the BOLD estimate for more than the first 1/16 TR likely has some variance with the variance increasing with more time sampled. This can make quite a difference in the estimate.
Related to #2, if we do give the duration of 4s, should this be done for both the individual first-level analysis and the PPI?
>>> Yes. This makes the modelling easier. I actually don't think its currently coded to accept different durations for the PPI model.
As noted above, our task GLM includes temporal derivatives for each of the three task conditions, which the gPPI toolbox also includes in the PPI GLM. I'm only creating interaction regressors for the three task conditions and not the temporal derivatives – does the inclusion of the temporal derivatives in the GLM complicate interpretation of the gPPI results? (Answer)
>>> I don't think this has been fully investigated. However, in theory it should not have an effect as the temporal derivative is orthogonal to the canonical HRF. The issue may arise that there shared variance between the temporal PPI terms and other terms in the model or other PPI event terms that could change the canonical estimate.
Building on #4, when adjusting for the 'effects of interest' F-contrast, should that contrast include the temporal derivative regressors or only the three task regressors?
>>> The Omnibus F-test will include the temporal derivative regressors when adjusting for the effects of interest.
See inline responses below.
Best Regards,
Donald McLaren, PhD
On Mon, Jun 4, 2018 at 1:11 PM, Jessica Hayes wrote:
Hello,
I'm a graduate student with experience running task GLM analyses, but my lab is now interested in performing gPPI analyses on fMRI data collected during a visual encoding task. We are interested in whether the connectivity of our regions of interest during encoding is modulated by subsequent memory (hit or miss during retrieval). During retrieval (out of scanner), participants also indicate their confidence in their decision (high confidence / low confidence).
Therefore we have three task conditions (Subsequent Hit-High Confidence [Hit-HC], Subsequent Hit-Low Confidence [Hit-LC], Subsequent Miss), with implicit baseline as a fourth condition (per listserv post). Our task GLM also includes temporal derivatives of the three task conditions, as well as six motion parameters and regressors for motion outliers.
I have a few questions about gPPI:
My understanding of PPI involves change in coupling across task conditions relative to the baseline (see footnote A). However, can this approach be used to investigate modulation of connectivity between task (i.e. encoding) and baseline? The gPPI script creates contrasts for 'none minus Hit-HC', 'Hit-HC minus none', etc. Can the results of these contrasts be reasonably interpreted as voxels where coupling with the seed region is modulated between Hit-HC and implicit baseline? This post would seem to indicate it can, but then what are the task and baseline connectivity relative to? If it's relative to the other two task conditions (Hit-LC and Miss), is it correct to assume that a general encoding (Hit-HC + Hit-LC + Miss) compared to baseline contrast is not a possibility?
>>> Yes. Comparisons against the implicit baseline are valid in gPPI.
Visual stimuli are presented for four seconds each – are there arguments to be made for modeling the stimuli as events, or giving them their 4s durations? This answer seems to suggest we should give the 4s duration, but the linked website primarily focuses on the need to give the duration when it's variable (which ours isn't).
>>> The duration provided should be the duration during which you want to compute the interaction. If you want the interaction to only be for first time bin (1/16 TR), then a 0 second duration is fine. However, I'd advise that you model the duration as you probably want to model the interaction for a longer period of time and the BOLD estimate for more than the first 1/16 TR likely has some variance with the variance increasing with more time sampled. This can make quite a difference in the estimate.
Related to #2, if we do give the duration of 4s, should this be done for both the individual first-level analysis and the PPI?
>>> Yes. This makes the modelling easier. I actually don't think its currently coded to accept different durations for the PPI model.
As noted above, our task GLM includes temporal derivatives for each of the three task conditions, which the gPPI toolbox also includes in the PPI GLM. I'm only creating interaction regressors for the three task conditions and not the temporal derivatives – does the inclusion of the temporal derivatives in the GLM complicate interpretation of the gPPI results? (Answer)
>>> I don't think this has been fully investigated. However, in theory it should not have an effect as the temporal derivative is orthogonal to the canonical HRF. The issue may arise that there shared variance between the temporal PPI terms and other terms in the model or other PPI event terms that could change the canonical estimate.
Building on #4, when adjusting for the 'effects of interest' F-contrast, should that contrast include the temporal derivative regressors or only the three task regressors?
>>> The Omnibus F-test will include the temporal derivative regressors when adjusting for the effects of interest.