3dLME vs. 3dMVM

AFNI Message Board
1

Dear AFNI experts,

I am trying to understand the difference between 3dLME vs. 3dMVM GLT contrasts.

In brief, I have a model with all within-subjects variables of interest: 3 time (TRs), 2 scan sessions, and 2 attention conditions. I also have some nuisance factors that I am not interested (e.g., participants’ scan sequence orders, body weights, scan time onsets; all centered at 0 across subjects). Because of these nuisance factors, I have decided to run 3dLME framework. I am defining many different contrasts using GLT set up (attached below). Without the nuisance factors, I also have run the very similar GLT’s under 3dMVM.

The big discrepancy I see is that the effect seems to be very big under 3dLME compared to 3dMVM. For instance, for the 1st GLT, I see a very large region showing a significant contrast under 3dLME, but greatly reduced # of voxels under 3dMVM at the same uncorrected thresholds. It almost feels like 3dLME inflates this contrast.

Could you give me a pointer whether my model specification is correct for the conditions that I have? In essence, I am not so interested in time (TR) factor, but rather interested in 2 scan session (DO and PL) and attType (such as GLT # 1, 2, 21, 25, 26, 27).
Would it be more appropriate to run it under 3dMVM?

Thank you very much for your help!
Michelle.

3dLME
-prefix …/ANOVA_results/3dLME_VvsN_CueRetProbeTENT-2-4TRs_x_attType_x_scanType-allCOVs_final.nii
-jobs 1
-model ‘scanTypeattTypetimeTR+sequence+weight+scanTime’
-qVars ‘sequence,weight,scanTime’
-qVarCenters ‘0,0,0’
-ranEff ‘~1’
-num_glt 27
-gltLabel 1 ‘DO-PL’ -gltCode 1 ‘scanType : 1D -1P’
-gltLabel 2 ‘Valid-Neutral’ -gltCode 2 ‘attType : 1V -1N’
-gltLabel 3 ‘Valid-Neutral_4s’ -gltCode 3 ‘attType : 1V -1N timeTR : 1t2’
-gltLabel 4 ‘Valid-Neutral_6s’ -gltCode 4 'attType : 1V -1N timeTR : 1t3’
-gltLabel 5 ‘Valid-Neutral_8s’ -gltCode 5 ‘attType : 1V -1N timeTR : 1t4’
-gltLabel 6 ‘attType_x_scanType_4s’ -gltCode 6 'attType : 1V -1N scanType : 1D -1P timeTR : 1t2’
-gltLabel 7 ‘attType_x_scanType_6s’ -gltCode 7 ‘attType : 1V -1N scanType : 1D -1P timeTR : 1t3’
-gltLabel 8 ‘attType_x_scanType_8s’ -gltCode 8 'attType : 1V -1N scanType : 1D -1P timeTR : 1t4’
-gltLabel 9 ‘DO-PL_Neutral_4s’ -gltCode 9 ‘attType : 1N scanType : 1D -1P timeTR : 1t2’
-gltLabel 10 ‘DO-PL_Neutral_6s’ -gltCode 10 ‘attType : 1N scanType : 1D -1P timeTR : 1t3’
-gltLabel 11 ‘DO-PL_Neutral_8s’ -gltCode 11 ‘attType : 1N scanType : 1D -1P timeTR : 1t4’
-gltLabel 12 ‘DO-PL_Valid_4s’ -gltCode 12 ‘attType : 1V scanType : 1D -1P timeTR : 1t2’
-gltLabel 13 ‘DO-PL_Valid_6s’ -gltCode 13 ‘attType : 1V scanType : 1D -1P timeTR : 1t3’
-gltLabel 14 ‘DO-PL_Valid_8s’ -gltCode 14 ‘attType : 1V scanType : 1D -1P timeTR : 1t4’
-gltLabel 15 ‘DO-PL_at_4s’ -gltCode 15 ‘scanType : 1D -1P timeTR : 1t2’
-gltLabel 16 ‘DO-PL_at_6s’ -gltCode 16 'scanType : 1D -1P timeTR : 1t3’
-gltLabel 17 ‘DO-PL_at_8s’ -gltCode 17 ‘scanType : 1D -1P timeTR : 1t4’
-gltLabel 18 ‘DO_Valid-Neutral_4s’ -gltCode 18 'scanType : 1D attType : 1V -1N timeTR : 1t2’
-gltLabel 19 ‘DO_Valid-Neutral_6s’ -gltCode 19 'scanType : 1D attType : 1V -1N timeTR : 1t3’
-gltLabel 20 ‘DO_Valid-Neutral_8s’ -gltCode 20 'scanType : 1D attType : 1V -1N timeTR : 1t4’
-gltLabel 21 ‘DO_Valid-Neutral_allTRs’ -gltCode 21 'scanType : 1D attType : 1V -1N’
-gltLabel 22 ‘PL_Valid-Neutral_4s’ -gltCode 22 ‘scanType : 1P attType : 1V -1N timeTR : 1t2’
-gltLabel 23 ‘PL_Valid-Neutral_6s’ -gltCode 23 ‘scanType : 1P attType : 1V -1N timeTR : 1t3’
-gltLabel 24 ‘PL_Valid-Neutral_8s’ -gltCode 24 ‘scanType : 1P attType : 1V -1N timeTR : 1t4’
-gltLabel 25 ‘PL_Valid-Neutral_allTRs’ -gltCode 25 ‘scanType : 1P attType : 1V -1N’
-gltLabel 26 ‘DO-PL_Neutral_allTRs’ -gltCode 26 'attType : 1N scanType : 1D -1P’
-gltLabel 27 ‘DO-PL_Valid_allTRs’ -gltCode 27 ‘attType : 1V scanType : 1D -1*P’
-mask /Users/sungjoo/DATA/RETROCUE_fMRI/GroupMask/mask_group_N=22_fromNative+tlrc.
-resid …/ANOVA_results/resid_3dLME_VvsN_CueRetProbeTENT-2-4TRs_x_attType_x_scanType-allCOVs_final_corStr.nii
-dataTable
Subj sequence weight scanTime scanType attType timeTR InputFile
Ss_210 1 -0.63657 -1.36364 D V t2 …/…/Ss_210/session2/allRuns_tlrc_e2a_censor_noss_smooth_native/stats_test_cueContrast_allfields_CueRetProbeTENT_tlrc.Ss_210+tlrc’[79]’
Ss_212 -1 0.95703 -1.36364 D V t2 …/…/Ss_212/session1/allRuns_tlrc_e2a_censor_noss_smooth_native/stats_test_cueContrast_allfields_CueRetProbeTENT_tlrc.Ss_212+tlrc’[79]’
…

2

There are a couple of possibilities for the discrepancies between 3dMVM and 3dLME.

  1. If one quantitative is quite skewed among the subjects, that may have a large impact,

  2. The way 3dLME counts the degrees of freedom is slightly different from 3dMVM, and the impact could be substantial if the number of subjects is on the lower end.

Are all the quantitative variable between- or within-subject? With the former, you could still run 3dMVM with those covariates.

3

Dear Gang,

Thank you very much for your reply. The covariates of no interest are between-subjects quantitative variables – for this, then I would use 3dMVM.

But in the further analysis, I might want to pursue to add quantitative variables that are within-subjects (associated with each of w/s-condition). This would require 3dLME in the end? If it is within-subjects variable, the qVarCenter should specify the centering value considering all conditions across all subjects? So when the model is being tested, data will be all scaled (subtracted by qVarcenter value)?

Thanks a lot for your help!

Michelle.

4

But in the further analysis, I might want to pursue to add quantitative variables that are within-subjects (associated
with each of w/s-condition). This would require 3dLME in the end?

That’s right.

If it is within-subjects variable, the qVarCenter should specify the centering value considering all conditions across
all subjects?

Correct.

So when the model is being tested, data will be all scaled (subtracted by qVarcenter value)?

Centered, but no scaling involved.

5

Gang Wrote:

There are a couple of possibilities for the
discrepancies between 3dMVM and 3dLME.

  1. If one quantitative is quite skewed among the
    subjects, that may have a large impact,

  2. The way 3dLME counts the degrees of freedom is
    slightly different from 3dMVM, and the impact
    could be substantial if the number of subjects is
    on the lower end.

Are all the quantitative variable between- or
within-subject? With the former, you could still
run 3dMVM with those covariates.

I have a similar question regarding 3dLME vs 3dMVM

In our setup, we are running a 2x4 repeated measures design. Both factors are within subjects. We’ve run both analyses using 3dMVM and 3dLME (with a random intercept). We get similar results in terms of location activated voxels, but 3dLME does have larger clusters than 3dMVM at the same p-value level. What is the cause of the difference in spatial extent? I’ve attached a picture showing the differences.

Also, should we ONLY use 3dLME if we have missing values. I thought 3dLME with a random intercept and 3dMVM should yield similar results. Does including a random intercept make that much of a difference?

When should we use 3dMVM vs 3dLME in the above example? Which result should we trust most?

Your help would greatly be appreciated.

Thanks.

Michael

3dMVM vs 3dLME.JPG
3dMVM vs 3dLME.JPG946×534 67.5 KB

6

The differences are most likely caused by the fact that the LME approach may count a higher number of DFs under some scenarios. Under those scenarios 3dMVM is more accurate.

7

Thanks, we will keep that in mind.

Thank again.

Michael