3DLME Group Analysis with Missing Data

AFNI Message Board
1

Hello Gang,

In my study, I collected 3 fMRI runs from each subjects (total 10 subjects). I have missing data for my regressors, therefor using GLM, I do not have 3 runs for each subject, Here is the summary of my dataset:
2 subjects only one run
6 subjects two runs
2 subjects three runs

Also my hypothesis is there is no differences between runs, we just collected 3 runs to make sure we have enough data for this study. After analyzing single subject, I have to do the group analysis. Considering the missing data, I think 3dlme is my only option, is that right(I have only one group of subjects)?
If 3dlme is the command, is this script correct for group analysis my dataset?

3dLME -prefix GroupAnalysis \
-model “1” \
-ranEff ‘Time+Subj’ \
-dataTable \
Subj Time InputFile \
s1 one s1_1+tlrc’[#0_Coef]’ \
s1 two s1_2+tlrc’[#0_Coef]’ \
s1 three s1_3+tlrc’[#0_Coef]’ \
s2 one s2_1+tlrc’[#0_Coef]’ \
s2 two s2_2+tlrc’[#0_Coef]’ \
s3 three s3_3+tlrc’[#0_Coef]’ \
…
s9 two s9_2+tlrc’[#0_Coef]’ \
s10 two s10_2+tlrc’[#0_Coef]’ \
s10 three s10_3+tlrc’[#0_Coef]’ \

Thanks,
Ahmad

2

Hi Ahmad,

Considering the missing data, I think 3dlme is my only option, is that right(I have only one group of subjects)?

Well, you have too much missing data, so I’m not so 1) whether missing-at-random is a reasonable assumption in your case, and 2) whether you have enough data to make robust inferences.

my hypothesis is there is no differences between runs

Is that your real research hypothesis or is it the straw man for the null hypothesis significance framework?

-model “1” \
-ranEff ‘Time+Subj’ \

These two lines should be

-model “Time” \
-ranEff ‘~1’ \

3

Thank you Gang for your prompt response and your help, I really appreciate it.

  • I have 4 missing data for Run1, 2 for Run2, and 4 for Run3. Do you think I just should forget about using 3 runs with 3dlme and only use Run2 and simple 3dttest++ for the group analysis?
  • Difference between runs is not the null hypothesis. The runs are similar and I do not expect any difference between them. The reason I collected 3 runs was that I could expect having several missing runs due to the nature of my study and wanted to make sure at least I have one usable run from each subject. I hope I could make it more clear.
  • It is a little bit confusing for me to assign the model and ranEff. Assigning ranEff ‘~1’ means each subject has its own intercept (please correct me if I am wrong) which is the case for my study, but not sure what -model “Time” means? Basically, I want to test the mean across input subjects datasets against 0, i.e. f I wanted to use 3dttest++ I would use the following script:
    3dttest++ -overwrite -prefix t_test_T1_${name_output[$out_count_name]} -setA
    S1/T1/GLM/stats.S1_REML+tlrc’[‘$brk’]’
    S2/T1/GLM/stats.S2_REML+tlrc’[‘$brk’]’
    S3/T1/GLM/stats.S3_REML+tlrc’[‘$brk’]’
    .
    .
    .

Thanks again for your help and your time.

Best Regards,
Ahmad

4

Ahmad,

To be clear, your goal is not to identify differences across runs at the group level, right? In that case, you can simply average each subject’s effect estimate across the available runs, and use 3dttest++, or use 3dMEMA if you average them within 3dDeconvolve/3dREMLfit

5

Thank you very much Gang for your help!