Dear Donald, thank you very much for your helpful answer.
I don't know GLM-flex yet but I will try Some remaining problems:
1) I did one model in which I also modeled explicitly the factor
subject, but while this greatly reinforces the statistics of main
effects related to the factor condition (independently of groups), it
kills the contrasts involving a group-by-condition interaction.
Unfortunately, this is obviously where your interest usually lies
...Modeling the group-by-condition interaction only (versus including
also main effects of condition and group) doesn't change the stats at
all.
2) Why is the contrast with a single condition in a single group
invalid ? (it appears in glasher 2008, contrast 6 design2, two groups of subjects).
3) The contrast in a single group I have tried, which seems to work with script, but isn't accepted by SPM is
con = [1 1 -1 -1 0 0 0 0];
gp = [1 0]
con{1} = [gp con con.*gp(1) con.*gp(2)];
Thanks again for your kind help,
Camille
On Wed, Apr 4, 2012 at 11:30 AM, Camille Piguet <[log in to unmask]
<mailto:[log in to unmask]>> wrote:
Dear all,
I would like to have your opinion regarding a 2nd level design
(flexible factorial) I built, which includes 2 groups (patients -
controls) and 8 conditions (4 tasks x 2 emotions) measured per group.
I would recommend using GLM_flex since you have multiple within-subject
factors. This will allow you to use a partitioned variance approach to
assess task and emotion separately, as well as their interaction. In
SPM, you are pooling the variance across all 8 conditions.
The model uses 3 factors: subject, group and condition. It models
the main effect of group and conditions, as well as the
group-by-condition interaction.This is actually the design2 from the
Glasher & Gitelman tutorial, you can see what the matrix looks like
in the attached file.
You need to include the subject term in the model.
My question is the following: is it valid in this kind of model to
test for the effect of a contrast between 2 conditions (ex A1A2 >
B1B2) in *one group only*?
Yes. You are simply pooling the variance across both groups. This is
fine. However, you might want to consider my comments about GLM flex for
looking at multiple within-subject factors.
I know how to built a test for a single condition in one group
(contrast 6 from design2 of the tutorial), but for a *contrast* the
interface won't allow it. Then: is it statistically valid ?
I think you might have an error in your contrast as it should allow it.
The single condition in one group is invalid.
This question extends to other flexible factorial models used for
instance to model RM anova with a single group.
I know we can do it, but is it statistically valid to estimate
simple main effectsin such a model (and weight the non used
conditions of the design with zeros) or shall we always use all
conditions of the design in the contrasts we estimate (and build
ttest instead to estimate simple main effects or simple comparisons
between two conditions) ?
Main effect of Condition - Is any condition different than another
condition - Valid
Main effect of Group - 1 group only - is the average effect of
conditions different from 0 - invalid
Main effect of Group - 2 or more groups - are the groups different from
each other - invalid
T-test of a single condition (in a single group or multiple groups) -
does the condition differ from 0 - invalid
The reason for this is that the group effects are based on
between-subject variance, which is not computed in SPM. GLM Flex does
not have this issue.
In SPM, you would need to use a seperate model for each group effect you
want to test.
Hope this helps.
Thanks for your input !
Camille Piguet
--
MD-PhD student
Laboratory for Neurology and Imaging of Cognition
Department of Neurosciences, Faculty of Medicine
University of Geneva
Centre Médical Universitaire
1 rue Michel-Servet
1211 Geneva, Switzerland
+41.22.379.53.24 <tel:%2B41.22.379.53.24>
http://labnic.unige.ch
|