BayesianGroupAna and gPPI question

AFNI Message Board
1

Hi Gang - quick (hah!) question:

I have a complex task that I am calculating gPPI on 2 conditions, with the following event of interest structure:

ConditionA
initiation
outcome1
outcome2

ConditionB
initiation
outcome1
outcome2

At the individual level, I am calculating seed to whole brain gPPI and I have 8 a priori regions of interest that I’d like to use your new Bayesian ROI group method. My question is: Which stats should I enter into the group analysis?

Should I use the PPIinteraction>baseline effects, or should I do post hoc t-tests (say, ‘SYM: +initiationA -initiationB’)? These are the tests I’m interested in running for each ROI:

+initiationA -initiationB
+outcome1A +outcome2A -outcome1B -outcome2B (main effect of condition)
+outcome1A -outcome2A +outcome1B -outcome2B (main effect of reply type)
+outcome1A -outcome2A -outcome1B +outcome2B (interaction)
+outcome1A -outcome2A
+outcome1A -outcome1B
+outcome2A -outcome2B

So, for each participant, I would like to have each of those contrasts within each ROI. For each event, I could extract the mean Z value for eventPPI>baseline or I can extract the mean Z from the post hoc t-tests. Do you know which would be better?

Also, just a heads up, I have two groups. So I’m also interested in the effect of group on the above tests as well and, to make it even more complicated, I’d like to add a continuous (e.g., age) or interval (e.g., likert-type scale) variable as well.

This seems like a good fit for your bayes multilevel model, but in that paper it only looks like you use one group and one effect. I always seem to be making your new analyses much more complicated :slight_smile:

Also, related question: for a voxelwise group PPI analysis would you recommend I take the PPI>baseline or individual level contrasts to the group level. With the PPI>baseline then I’d have a lot of repeated measures and I should probably use 3dLME, whereas with the contrasts I could probably get away with 3dMEMA.

Thanks!!

Dustin

2

Hi Dustin,

For each event, I could extract the mean Z value for eventPPI>baseline or I can extract the mean Z
from the post hoc t-tests. Do you know which would be better?

It’s more straightforward to extract each effect of interest from the post hoc test, and use that as input for BayesianGroupAna.

So I’m also interested in the effect of group on the above tests as well and, to make it even more
complicated, I’d like to add a continuous (e.g., age) or interval (e.g., likert-type scale) variable as well.

Just dummy-code your groups with 0s and 1s (or -1s and 1s), and use the coding, together with other quantitative variables for BayesianGroupAna.

for a voxelwise group PPI analysis would you recommend I take the PPI>baseline or individual level
contrasts to the group level. With the PPI>baseline then I’d have a lot of repeated measures and I
should probably use 3dLME, whereas with the contrasts I could probably get away with 3dMEMA.

Either way is fine, but with slightly different assumptions.

3

Thanks for the info, Gang. I am now turning to this project.

BayesianGroupAna.py
-dataTable $data
-prefix gPPI.Bayesout
-y beta -x part con
-chains 4 -iterations 1000
-plot -more_plots rhat violin

What if I want to include some interactions?

My formula is:

beta ~ 1 + part + con + (1 | Subj) + (1 + part + con | ROI)

4

What kind of variable are “part” and “con”? Are they between- or within-subject factors?

5

They are within-subject factors.

Our event-related design is a 2x2 factorial. There is a factor of social partner (peer or computer) and type of reply to a questions (agree or away). So there are 4 conditions nested within two factors.

Currently I am taking your previously-suggested approach and extracting the 1st level beta values from the post-hoc contrasts peer-comp.agree and peer-comp.away and then looking at the effect of agree/away at the 2nd level.

My research question primarily revolves around the peer-computer contrast, but I want to account for the other factor in the analysis because I would expect an interaction between partner and reply type.

6

My research question primarily revolves around the peer-computer contrast, but I want to account for the
other factor in the analysis because I would expect an interaction between partner and reply type.

The most straightforward approach is to obtain the contrast of peer-computer from each subject, and then directly run BayesianGroupAna.py with the contrast from each subject as input:

beta ~ 1 + (1 | Subj) + (1 | ROI)

For the contrast of peer-computer, you need to average across the two questions (agree and away).