3dLME, F tests and GLts

AFNI Message Board
1

Dear colleagues,

I have a few questions about the 3dLME.
My design is originally a 2x2 with two within-subject factors (Session: T1, T2) and Velocity (Slow, Fast).
I realized that one of the two sessions make subjects moving significantly more than the other (this was quantified with average movement per TR).
Therefore, in order to exclude that the motion, even after motion correction, could explain the variance between the levels of the factor Session, I decided to add a covariate of no interest “Movement”. Since this would mean that every subject would have a different value for each session, but the same value for each velocity in each session, I decided to use 3dLME. The script looks as follow (data for one subject is shown as example in -dataTable):


3dLME -prefix lme_qvars0 -jobs 30 \
-model "Session*Velocity*Movement" \
-mask simulation_mask+tlrc[0] \
-qVars "Movement" \
-qVarsCenters 0 \
-ranEff "~1+Movement" \
-SS_type 3 \
-num_glt 4 \
-gltLabel 1 'S_F_mov' \
-gltCode 1 'Velocity : 1*Slow -1*Fast Movement :' \
-gltLabel 2 'T1_T2_mov' \
-gltCode 2 'Session : 1*T1 -1*T2 Movement :' \
-gltLabel 3 'S_F' \
-gltCode 3 'Velocity : 1*Slow -1*Fast' \
-gltLabel 4 'T1_T2' \
-gltCode 4 'Session : 1*T1 -1*T2' \
-dataTable 
Subj Session Velocity Movement InputFile
501 T1 Fast 0.011  path/S2_sub501.results/stats.S2_sub501_REML+tlrc[1] \
501 T1 Slow 0.011 path/S2_sub501.results/stats.S2_sub501_REML+tlrc[4] \
501 T2 Fast -0.008 path/S1_sub501.results/stats.S1_sub501_REML+tlrc[1] \
501 T2 Slow -0.008 path/S1_sub501.results/stats.S1_sub501_REML+tlrc[4] \
...continues...

Here are my questions:

[ol]
[li] Since a t test of Movement values between T1 and T2 comes significant, I first centered around the mean of each session. Then in the 3dLME script I am testing what the effects would be if Movement was not happening, i.e. =0 (which I guess would be the default anyway). Is this interpretation correct?
[/li][li] The output of the 3dLME shows the F tests for main effects and interactions. If I want to see the direction of the main effects, should I code GLTs as shown in the script? And why in that case the clusters I amobtaining with the Z tests (corresponding to glt) do not correspond 100% to those of the F tests (e.g., in terms of cluster size)? As you can see I have both tried with Movement effect specified (GLT 2) and not (GLT 4, which I guess would be the way to go), but none of these match 100% the results of the Main Effect.
[/li][/ol]

Thank you in advance for your help!

2

I first centered around the mean of each session. Then in the 3dLME script I am testing what the effects would be
if Movement was not happening, i.e. =0 (which I guess would be the default anyway). Is this interpretation correct?

Yes, your interpretation seems correct. If you meant the default of 3dLME for centering, the center value is actually the overall average, not 0.

If I want to see the direction of the main effects, should I code GLTs as shown in the script?

They look fine.

why in that case the clusters I amobtaining with the Z tests (corresponding to glt) do not correspond 100% to those
of the F tests (e.g., in terms of cluster size)?

Under the conventional ANOVA framework, they would usually end up the same. Under LME, these two types of statistics (F- and Z-statistic) are obtained through a slightly different formulation. So it’s expected the two do not exactly match. As long as they are roughly similar, it should be fine. The exact number of voxels in a cluster is defined arbitrarily anyway because of the artificial dichotomization.

3

Hi again,

I have a follow up to my previous questions.
In the end I had a design with two within factors (2 levels each) and 3 covariates.
Let’s say I observe clusters for the F test of one covariate. I would interpret that as: those voxels covary with the covariate. How do I investigate if it is a positive or negative relation though? If I extract the betas and correlate those with the covariate they do not always correlate significantly (sometimes trendlines are even flat and correlation coefficients are close to zero!). Now I am aware this is there as a result of a correlation… but how do I investigate it further?

4

How do I investigate if it is a positive or negative relation though? If I extract the betas and correlate those with the covariate
they do not always correlate significantly (sometimes trendlines are even flat and correlation coefficients are close to zero!).
Now I am aware this is there as a result of a correlation… but how do I investigate it further?

Visualizing data with hierarchical structures is as simple and straightforward as it sounds like. It would be more revealing to plot the data per subject and per level of each within-subject factor, which can be very challenging.

To assess the directionality for a quantitative variable, add something like the following to your script:

-gltCode … ‘Session : Movement :’
-gltCode … ‘Session : 1T1 Movement :’
-gltCode … 'Session : 1T2 Movement :’ \