Dear FSL Users & Experts,
I came across this message & reply in the thread.
On Thu, Feb 10, 2011 at 7:48 AM, Elif Alkan <[log in to unmask]> wrote:
Hello!
I have an experiment with two groups (patients and controlls). Unfortunately patients have sig. higher scores on scales for depression and trait-anxiety. I want to control for this two variables in my group analyses. I have demeaned the depression and anxiety scores and included them in my GLM as EV3 EV4. Now I have two questions about the GLM.
1) Feat finds it rank deficient, when I name two groups (Group (111.....11112222.....22222) the very first column).I have always done it this way for a normal 3rd level analyses without covariates. I really don't get, why it is okay to name two seperate groups, when I make the analyses without additional covariates and why it is not, when I have the additional covariates. Is it allright, how I did it? (Group 1111.....1111...11111)
Group EV1 EV2 EV3 EV4
Patients Cont. Dep. Anxiety
Input1 1 1 0 10.5 12.3
Input2 1 1 0 3.4 7.5
....
Input30 1 0 1 -1.5 2.1
..
Input57 1 0 1 2.3 -3.4
Contrasts
title EV1 EV2 EV3 EV4
C1 Patients mean 1 0 0 0
C2 Cont. mean 0 1 0 0
C3 Pat.> Cont 1 -1 0 0
C4 Cont.> Pat. -1 1 0 0
C5 whole sample 1 1 0 0
Jeanette's reply: Are you sure it was rank deficient? I think the error message you received was probably something along the lines of having a non-separable design. Basically, if you want to use the "Group" column to identify different groups in the variance estimate the design must be "separable". This means each EV can only have nonzero entries within a single group.
Running with all 1's in the "group" column will only make a difference if your between subject variability greatly differs between your two groups.
The only way I can see to run with 1's & 2's in the "group" column is to run the interaction model, but that's a whole other story and interpretation isn't as straightforward if the interaction between your continuous covariate and group is significant. In this model you'd split each of your continuous covariates into 2 regressors by group (but still demean over ALL subjects).
***
Essentially, we are doing a similar and we thought we would run the following model, based on the above suggestion to run an interaction model:
Group EV1 EV2 EV3 EV4
Patients Cont. Age_Pat Age_Control
Input1 1 1 0 10.5 0
Input2 1 1 0 3.4 0
....
Input30 2 0 1 0 -1.5
..
Input57 2 0 1 0 2.3
(the ages are demeaned across both groups)
We were wondering the following things:
- What contrasts would we do? Is the following ok?
Title EV1 EV2 EV3 EV4
C1 Pat>Cont 1 -1 0 0
C2 Cont>Pat -1 1 0 0
C3 Pat mean 1 0 0 0
C4 Cont mean 0 1 0 0
C5 Age effect (pos) 0 0 1 1
C6 Age effect (neg) 0 0 -1 -1
Many thanks in advance for your help.
Kind regards,
Lucia
|