3dLMEr interaction chi-square versus gltcode

AFNI Message Board
, ,
1

AFNI version info (afni -ver): Latest

Hi @Gang and all,

I am doing a 3dLMEr model below

3dLMEr -prefix results \
-jobs 20 \
-mask mask.nii.gz \
-model 'Drug*Time+Age+Sex+(1|Subj)' \
-qVars Age \
-gltCode V5_V0 'Time : 1*Post -1*Pre' \
-gltCode drug_by_time 'Drug : 1*DrugA -1*DrugB Time : 1*Post -1*Pre' \
-dataTable @model_variables_${roi}.txt

My question: if my Drug*Time chi-square is not significant (i.e., does not survive cluster thresholding), but the -gltCode drug_by_time Drug : 1*DrugA -1*DrugB Time : 1*Post -1*Pre shows significant clusters, would it be erroneous to say that there is a significant Drug*Time interaction? Is it the case that they are both similar tests but using different stats? Or is the latter considered a "post-hoc" test and can only be done if the chi-square is significant?

2

Do both factors Drug and Time have two levels? If so, the two tests should yield approximately similar statistical evidence. Additionally, do they exhibit comparable patterns when a more lenient threshold is applied? The conventional cluster-based adjustment, commonly employed, tends to be overly conservative. Consider adopting a “highlight, but don’t hide” approach in your result reporting.

Gang Chen

3

Thank you @Gang . Yes, both factors only have two levels. However the chi-square results do not survive the minimum cluster size (but yes, it does survive under a more lenient cluster-size threshold), whereas the -gltcode-based test does. What is the source of the difference between these results?

I also noticed this post with the same question:

and @Gang your reply seemed to suggest one can report either the chi-square or the z-test. However, in most papers, we noticed that people using AFNI 3dlmer report the chi-squared results to report interactions and main effects - hence we are a bit confused.

4

Both of these results are conceptually equivalent, but they are evaluated slightly differently. You may choose either one to report as an interaction effect. If you’d like me to examine them closely, feel free to reach out to me via email.

Gang Chen