3dLMEr between-subjects contrast clarification

Dear AFNI team (Gang!)

I hope you’re all keeping safe and well.

I wanted to ask for your advice on modelling between-subjects effects in 3dLMEr.

I have a dataset where participants listen to a story twice. On the second listening, the story is either temporally scrambled at the sentence or word level.

I’d like to contrast the within-subject effects of the intact vs scrambled versions of the story. Further, I’d also like to contrast the between-subjects effects of listening to the sentence (long) vs word (short) scrambled versions of the story.

I’ve specified the model and post-hoc GLTs as follows, but I’m concerned that 3dLMEr may not account for differences in group variance between groups that listened to the short vs long scrambled versions of the story. Please could you advise as to whether my model specification here is sufficient?


-model 'Age+Gender+scram+(1|Subj)' \
-qVars "Age" \
-qVarsCenters 20.66 \
-gltCode intact 'scram : 1*intact' \
-gltCode longscram 'scram : 1*long' \
-gltCode shortscram 'scram : 1*short' \
-gltCode intactVscram 'scram : 1*intact -0.5*long -0.5*short' \
-gltCode intactVlong 'scram : 1*intact -1*long' \
-gltCode intactVshort 'scram : 1*intact -1*short' \
-gltCode longVshort 'scram : 1*long -1*short' \
-dataTable \
Subj Age Gender scram InputFile \

I have a dataset where participants listen to a story twice. On the second listening, the story is either
temporally scrambled at the sentence or word level.

What do the subjects listen to during the first session? During the second session does each subject listen to only one of the two scrambling cases or both?

I’d like to contrast the within-subject effects of the intact vs scrambled versions of the story. Further, I’d
also like to contrast the between-subjects effects of listening to the sentence (long) vs word (short) scrambled
versions of the story.

I suggest that you build two separate models: one for the contrast of the intact vs scrambled versions using 3dMVM, 3dttest++, or 3dLMEr, and the other for the contrast between the long and short scrambled versions using 3dMVM or 3dttest++.

Thank you very much for your prompt response!

What do the subjects listen to during the first session? During the second session does each subject listen to only one of the two scrambling cases or both?

All subjects listen to the temporally intact story during the first session. Subjects then listen to only one of the scrambling cases. Slightly questionable study design, I know. It’s from an open-source dataset!

I suggest that you build two separate models: one for the contrast of the intact vs scrambled versions using 3dMVM, 3dttest++, or 3dLMEr

This makes some sense. Just to clarify, could I use the same model that I’ve applied previously for this purpose, and just change my post-hoc GLT codings slightly? e.g.


-model 'Age+Gender+scram+(1|Subj)' \
-qVars "Age" \
-qVarsCenters 20.66 \
-gltCode intact 'scram : 1*intact' \
-gltCode longscram 'scram : 1*long' \
-gltCode shortscram 'scram : 1*short' \
-gltCode intactVlong 'scram : 1*intact -1*long' \
-gltCode intactVshort 'scram : 1*intact -1*short' \

and the other for the contrast between the long and short scrambled versions using 3dMVM or 3dttest++.

Again, just to clarify, if I were to run this in 3dLMEr, could I achieve this between-subjects contrast by removing all subject-level analyses of the ‘intact story’ from my input files table, only retaining the long and short scrambled versions of the story using the following model?


-model 'Age+Gender+scram+(1|Subj)' \
-qVars "Age" \
-qVarsCenters 20.66 \
-gltCode intactVlong 'scram : 1*intact -1*long' \
-gltCode intactVshort 'scram : 1*intact -1*short' \

If so, I’m a little confused. Why should these have to be run separately?

Thanks again for your help, Gang! I really appreciate it.

could I use the same model that I’ve applied previously for this purpose, and just change my post-hoc GLT codings slightly?

In that case, you code the factor “scram” with a mixture of between-subject and within-subject types among the 3 levels. I’m not so sure if 3dLMEr would work that way. Does the program complain about such a specification? Even if it does not complain, try the approach I suggested earlier, and make sure the two approaches converge.