3dMVM vs 3dANOVA

AFNI Message Board
1

Hi experts,

The first factorial analysis I set up was a 3dMVM script due to uneven sample sizes and some missing levels (since 3dANOVA can’t handle this). As far as I understand 3dMVM would give the same results as 3dANOVA given the same (equal group sized) data input (right?).

I did not really mind the slowness of 3dMVM since we have fast machines and since I could do other things in the mean time so I have been using my 3dMVM template script for all factorial analysis out of convenience (since the script and table already was ready to go, just change some paths and variable names).

If I have even samples, would I be wise to use 3dANOVA instead (again, I don’t mind it being slow)? Do I loose sensitivity or something similar when running 3dMVM instead of 3dANOVA? For example, I now run a factorial analysis on an even amount of subjects that have done 2 MRI scans and half of them have real treatment and the rest sham. I create R-maps from seed based correlations with DMN and compare the R-maps with MVM to find out if there is an interaction (basically to make sure that the differences I find at MR 2 with 3dttest++ was not already there at baseline (MR1):


3dMVM -prefix DMN -dataTable @table_DMN.txt -jobs 10 \
-mask mask_epi_anat+tlrc \
-bsVars "Treatment" \
-wsVars "MRSession"

Is this a bad idea, results wise?

Bonus Q: I read a lot about 3dMEMA but I have never used it. I always use 3dttest++ for non 2x2 designs and 3dMVM for 2x2 designs. When would one want to use 3dMEMA and would my results be affected by e.g. 3dttest++ vs 3dMEMA or 3dMVM vs 3dMEMA?

Just making sure I’m beeing optimal (not regarding processing time).

Thanks!

2

As far as I understand 3dMVM would give the same results as 3dANOVA given the same (equal group sized) data input (right?).

Yes, 3dMVM would produce the same results as other ANOVA programs in AFNI such as 3dANOVA, 3dANOVA2, and 3dANOVA3.

If I have even samples, would I be wise to use 3dANOVA instead (again, I don’t mind it being slow)? Do I loose sensitivity or
something similar when running 3dMVM instead of 3dANOVA?

No, you are not losing anything at all with 3dMVM.

I read a lot about 3dMEMA but I have never used it. I always use 3dttest++ for non 2x2 designs and 3dMVM for 2x2 designs.
When would one want to use 3dMEMA and would my results be affected by e.g. 3dttest++ vs 3dMEMA or 3dMVM vs 3dMEMA?

I guess that you meant “3dttest++ for 2x2 designs and 3dMVM for designs that are no 2x2”.

Whenever you can perform your group analysis with 3dttest++, you could use 3dMEMA if you have also the corresponding t-statistic values available. In theory the modeling approach with 3dMEMA is always better than 3dttest++ because the former takes into consideration the reliability information of individual effect estimate. However, the extent of impact could be substantial or negligible, depending on each particular case. In other words, it is hard to know in advance when the 3dMEMA approach is guaranteed to have substantially better results unless you try out both and compare them.

3

Thanks Gang! Then I can feel safe with staying lazy regarding mvm. With MVM you also don’t have to specify number of subjects and what kind of ANOVA type it is, right? So that is also nice. But how does MVM know? If it’s “type2” or “type3” etc?

What I meant was that I use 3dttest when it’s not factorial (2x2). Like “Treatment”(real,sham) and “Stim”(cue_type1,cue_type2). I use mvm for that. And 3dttest++ for just comparing e.g. the contrast stim (cue_type1 -cue_type2) between the sets -SetA (real) -SetB(sham).
So, factorial (I called it 2x2 with MVM) and non factorial with 3dttest++. Is that not correct?

Thanks so much for the 3dMEMA explanation. I’ll try that too!

4

With MVM you also don't have to specify number of subjects and what kind of ANOVA type it is, right?

That's correct: 3dMVM can sort out those details internally.

But how does MVM know? If it's "type2" or "type3" etc?

By default, 3dMVM automatically applies type 3 for the sums of squares unless you add option "-SS_type 2".

So, factorial (I called it 2x2 with MVM) and non factorial with 3dttest++. Is that not correct?

That's fine. Even if you have a 2x2 design, you could still use 3dttest++ or 3dMEMA if you're willing to go through some tedious steps as discussed in the middle of the webpage: https://afni.nimh.nih.gov/MEMA