Help with 3dMVM

AFNI Message Board
1

Hello,

I cannot make this 3dMVM work.

3dMVM -prefix mvm -jobs 30 \
-SS_type 3 \
-bsVars '1*Int_Cov+NoInt_Cov' \
-wsVars Session \
-mask simulation_mask+tlrc[0] \
-qVars "Int_Cov, NoInt_Cov" \
-num_glt 1 \
-gltLabel 1 Session*IntCov \
-gltCode 1 'Session : 1*A -1*B IntCov: ' \
-dataTable \
Subj    Session   NoIntCov   IntCov     InputFile    \
S1       A              1                7             file_S1_A   \
S2       A              2                8             file_S2_A   \
S3       A              3                9             file_S3_A   \
...
S1      B              4                10           file_S1_B   \
S2      B              5                11           file_S2_B   \
S3      B              6                12           file_S3_B   \

I have tried many ways, even removing the glt.
I am interested in testing the difference between the covariation of the two levels of the within-subject factor (A,B) and the Covariate of interest (IntCov) while controlling for a covariate of no interest (NoIntCov).
Any idea?

Thanks,
Giovanni

...

2

Giovanni,

Try the following:

3dMVM ... \
-bsVars 'Int_Cov+NoInt_Cov' \
-wsVars Session \
-mask simulation_mask+tlrc'[0]' \
-qVars "Int_Cov,NoInt_Cov" \
-num_glt 1 \
-gltLabel 1 SessionIntCov \
-gltCode 1 'Session : 1*A -1*B IntCov : ' \
...

Gang Chen

3

Dear Gang,

When I try:
-bsVars 'Int_Cov+NoInt_Cov' \
or
-bsVars 'NoInt_Cov+Int_Cov' \

the model fails and it takes the first listed Covariate in bsVars as a categorical variable for which it shows the contingency table.
I tried:

-bsVars '1' \

and that worked, but the glt still fails.

4

Giovanni,

-bsVars 'Int_Cov+NoInt_Cov' \

What kind of variables are Int_Cov and NoInt_Cov? Are they truly between-individual quantitative variables are declared above? If they vary within individuals, they should be treated as within-individual variables. In that case, you should use 3dLMEr:

3dLMEr ... \
-model 'Session*Int_Cov+Session*NoInt_Cov+(1+Int_Cov+NoInt_Cov|Subj)' \
-mask simulation_mask+tlrc'[0]' \
-qVars "Int_Cov,NoInt_Cov" \
-gltCode SessionIntCov 'Session : 1*A -1*B IntCov: ' \
...

Check out the discussion here for the general usage of 3dLMEr.

Gang Chen

5

Dear Gang,

Let me take one step back. I have started with a model that had an additional within subject factor (Velocity, two levels: fast, slow).
So I ran a 3dLMEr with

-model "Session*Velocity*IntCov+NoIntCov+(1+IntCov|Subj)" \

This worked, glt included. I am still not sure though why I should choose this model over, let's say, (1+IntCov+NoIntCov|Subj) or (IntCov|Subj).
Then I have a regressor in the first level which is not varying across Velocity, for which I would like to run a similar analysis, therefore the original question and model. Just removing Velocity from the model and running:

-model "Session*IntCov+NoIntCov+(1+IntCov|Subj)" \

does not work.
Also, in your suggestion you specify the interaction between Session and NoIntCov, which I am not interested in. Why is that?

6

in your suggestion you specify the interaction between Session and NoIntCov, which I am not interested in. Why is that?

A statistical model should be constructed to closely reflect the underlying data-generating process, rather than the investigator's interests. As the data owner, you possess prior knowledge that allows you to determine the presence of any interaction effects.

I have started with a model that had an additional within subject factor (Velocity, two levels: fast, slow).

One possibility is to specify the individual-varying (random) effects as the following:

"...+(1+IntCov|Subj)+(1+IntCov|Subj:Session)+(1+IntCov|Subj: Velocity)"

Depending on whether NoIntCov varies across the levels of Session/Velocity, you may want to include NoIntCov as a varying slope in some of the random-effects terms.

Gang Chen