Understanding interaction in 3dLMEr between ws factor and quantitative var

[sinandrei]

Hello,

I’ve read the other related past post, but they weren’t really related to my case so I thought its best to ask separately.

I have a within-subject factor (“hierarchy”) with two levels and a continuous mean performance measure (“infPerf”) that is centered outside of AFNI according to the hierarchy levels - so it is also within-subject hence me getting into 3dLMEr. There is also a nuisance covariate “encStr”

I have the following command prepared:
3dLMEr -model ‘hierarchy*infPerf+encStr+(1|Subj)’ -prefix XXX -qVars ‘infPerf,encStr’ -resid lmerResid […] -dataTable @covariates.txt.

I’m primarily interested in the “hierarchy:infPerf” which if I understand correctly from the documentation 3dLMEr will give me automatically in the form of a chi-square map

1. Were I to look for group level cluster for “hierarchy:infPerf” would I go about this via the 3dFWHMx + 3dClustsim route?
2. If so - and assuming something survives what would be the best way to unpackage/visualize this interaction? Do I extract and average the individual condition level signal masked by the surviving cluster and put it on scatterplot? But how would I extract the fitted model estimates to plot the model predicted relationship?
3. Is there a way to extract the R “lmer” object and work with that? I have more experience plotting lmers…

[Gang]

@sinandrei

I’m primarily interested in the “hierarchy:infPerf” which if I understand correctly from the documentation 3dLMEr will give me automatically in the form of a chi-square map

You may also use -gltCode in 3dLMEr to obtain the slope effect of infPerf for each level of hierarchy to help you parse the interaction effect.

Were I to look for group level cluster for “hierarchy:infPerf” would I go about this via the 3dFWHMx + 3dClustsim route?

You could. Alternatively, you may simply adopt a soft cluster threshold (e.g., 20 voxels at the p-value of 0.01), and use a highlight-but-not-hide approach as suggested in this recent paper.

Do I extract and average the individual condition level signal masked by the surviving cluster and put it on scatterplot

Some people think such an extraction approach may lead to selection bias or what is popularly called "double dipping". Alternatively, you can select a region defined independently of your own data such as atlas or parcellation.

Is there a way to extract the R “lmer” object and work with that? I have more experience plotting lmers

With the extracted data, you can directly adopt the same model as the one you specified with 3dLMEr.

[sinandrei]

Thank you for the quick reply.

@Gang

@sinandrei

I’m primarily interested in the “hierarchy:infPerf” which if I understand correctly from the documentation 3dLMEr will give me automatically in the form of a chi-square map

You may also use -gltCode in 3dLMEr to obtain the slope effect of infPerf for each level of hierarchy to help you parse the interaction effect.

So this would mean one gltcode for hierarchy: 1H1 infPerf:'* , one for gltcode for hierarchy: 1H2 infPerf:'* and then if there is a “hierarchy:infPerf” cluster one could look at the estimates/z-score for in those maps?

@Gang

Were I to look for group level cluster for “hierarchy:infPerf” would I go about this via the 3dFWHMx + 3dClustsim route?

You could. Alternatively, you may simply adopt a soft cluster threshold (e.g., 20 voxels at the p-value of 0.01), and use a highlight-but-not-hide approach as suggested in this recent paper.

I'm big fan of that approach - am I correct in thinking that there isn't a terminal command to generate those plots? Only through the GUI?

@Gang

Do I extract and average the individual condition level signal masked by the surviving cluster and put it on scatterplot

Some people think such an extraction approach may lead to selection bias or what is popularly called “double dipping”. Alternatively, you can select a region defined independently of your own data such as atlas or parcellation.

I see. So to avoid double dipping would the recommendation be to use the whole roi/atlas/parcellation mask that corresponds/overlaps with where the cluster was found?

[Gang]

@sinandrei

So this would mean one gltcode for hierarchy: 1H1 infPerf:‘* , one for gltcode for hierarchy: 1H2 infPerf:’* and then if there is a “hierarchy:infPerf” cluster one could look at the estimates/z-score for in those maps?

Yes, but be careful with the space:

hierarchy : 1*H1 infPerf :
hierarchy : 1*H2 infPerf :

I’m big fan of that approach - am I correct in thinking that there isn’t a terminal command to generate those plots? Only through the GUI?

Yes, I do believe it's a more transparent and healthy approach than artificially dichotomizing the evidence. You can do it

1. through the GUI as shown in this video, or

2. using command lines.

So to avoid double dipping would the recommendation be to use the whole roi/atlas/parcellation mask that corresponds/overlaps with where the cluster was found?

You can even analyze the whole dataset directly on a list of regions (instead of the voxelwise approach) as demonstrated in the recent paper and dissolve the multiple testing issue.