Random effects with 3dLME

Hi there,

I am trying to implement two analyses with the following longitudinal designs:

In case A, we have three measurements per subject (with some missing data) and the explanatory variables:
(1) ISI: time between measurements, W/S
(2) Age: age at baseline, B/S
(3) Cond: control or cond1, B/S

In case B, two measurements per subject and same explanatory variables (1) and (2) but (3) is different:
(3) Cond: control, cond1 or cond2

In both cases A and B, scanner is a random effect. For most, the subject’s complete data set comes solely from one of three scanners, but in a few subjects, scanner source differed across time points. (We use Fortin’s COMBAT algorithm prior to this analysis to remove most of the scanner-contributed noise, though some appears to remain.)

For case A, I used 3dLME to model the data with the following model specification which seems to run fine:

3dLME -prefix ${seed}.nii
-jobs 32
-model “Cond+Age+ISI+Cond:Age+Cond:ISI+Cond:ISI:Age”
-qVars “Age,ISI”
-qVarCenters “46.26,7.18”
-ranEff ‘~1+Scanner’
-SS_type 3
-num_glt 1
-gltLabel 1 ‘carrier-control’ -gltCode 1 ‘Cond : 1carrier -1control’
-dataTable @${seed}.table.txt

However, in case B with the following, the model hangs and eventually fails with this:

3dLME -prefix ${seed}.all.nii
-jobs 32
-model “Cond+Age+ISI+Cond:Age+Cond:ISI+Cond:ISI:Age”
-qVars “Age,ISI”
-qVarCenters “51.73,2.78”
-ranEff “~1+Scanner”
-SS_type 3
-dataTable @${seed}.table.all.txt

This model will not run (hangs at the “If the program hangs here…” step, unless I remove the Scanner from
-ranEff “~1+Scanner”
(it will run with: -ranEff “~1” )

My main question is, why does adding “Scanner” as a random effect in case B cause it to hang and fail? What is it about the differences between A and B that is causing the problem? And does this seem like an appropriate way to model and conceptualize this?

I would appreciate any advice on this. Thanks!

For most, the subject’s complete data set comes solely from one of three scanners, but in
a few subjects, scanner source differed across time points.

Scanner is largely a between-subject factor? In that case, it would make more sense to try the following for both of your cases:

-model “Cond+Age+ISI+Cond:Age+Cond:ISI+Cond:ISI:Age+Scanner” \

-ranEff ‘~1+ISI’ \

Thanks, that makes sense. I took a look at the ratio of instances of mixed scanner time points within subjects for both cases and in case B (hanging failure) ~13% of the subjects contributed mixed data while in case A, 25% of the subjects had mixed scanner within-subject, so perhaps that is the difference.