Dear SPMrs,
we read the mail reported below and tried to do the suggested multiple
regression analysis.
We entered at the second level 22 contrast images and 2 covariates [ 1 1
1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 0 0 0 0 ]
and [0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 ] (in basic models,
multiple regression).
But then, it is impossible to do the suggested contrast (1 0) or (0 1).
It says it is an invalid contrast.
Valid contrasts are [1 0 1] and [0 1 1], that aren't orthogonal. The
last column is the "block" column.
Do these contrasts have any sense?
We also made the conjunction between these two contrasts. Is it right?
So, we are still interested in the Gaspare problem.
Can somebody suggest a way to do the conjunction analysis in random
effect model?
As for conjunction analysis, how should we set the significance
threshold?
Usually in random effect model we set 0.01 uncorrected threshold at
voxel level
and 0.05 threshold at cluster level.
In the conjunction there is not cluster level. How can we establish the
significance?
We attached the design description.
Thanks in advance.
Date: Thu, 2 Sep 1999 13:37:21 +0100 (BST)
Reply-To: Karl Friston <[log in to unmask]>
Sender: spm
From: Karl Friston <[log in to unmask]>
Subject: conjunctions and random-effect analysis - SPM99b
Dear Gaspare,
> I need your help to select the appropriate analysis model for a 2x2
> factorial fMRI experiment, where 10 subjects performed an Activation
> (A) and a Baseline (B) task, with stimuli delivered either in the
> visual (v) or in the tactile (t) modality.
>
> I've specified and estimated separate subject-specific models,
entering
> 4 conditions (Av Bv At Bt), modeled as box-cars (convolved with hrf),
> and tested for the main effect of task (1 -1 1 -1) and the
interaction
> task x modality (1 -1 -1 1 and -1 1 1 -1). I then entered the
contrast
> images into a 2nd level random-effect analysis (one sample t-test).
>
> The problem is the following: since I am interested in finding
regions
> which are activated by task A vs. B to the same degree in the two
> modalities, this should be more properly tested by a conjunction
> analysis (i.e. the conjunction between the "simple effects" of task
in
> the two modalities: 1 -1 0 0 and 0 0 1 -1), rather than by the "main
> effect" contrast (1 -1 1 -1). With the conjunction approach, regions
> where interactions occur should be discarded from the results.
>
> But: - If I do the conjunction analysis at the 1st
(subject-specific)
> stage, I do not obtain any contrast image to enter at the 2nd
> (random-effects) stage. - On the other hand, I can't imagine an easy
> way to do the conjunction analysis at the 2nd level, for example
using
> "simple effects" contrast images. Is there any? - If not, I could
> build a big multi-subject fixed-effects model and test for the
> conjunction between 1 -1 0 0 1 -1 0 0 ... (repeated for all subjects)
> and 0 0 1 -1 0 0 1 -1 ... (again repeated for all subjects). In this
> case, however, I understand that my inference would be limited to the
> subjects studied, because this is a fixed-effects approach.
This is an interesting issue which we have been wrestling with as
well. One approach would be to take estimates of the simple main
effects to the second level (e.g with contrasts 1 -1 0 0, 0 0 1 -1) and
model these seperately (using 'multiple regression' in 'Basic models'
and [0,...0, 1,...1] for the first regressor and [1,...1 0,...0,] for
the second). A conjunction of second level contrasts [0 1 and 1 0]
should give you what you are after. The assumption you are making here
is that the simple main effects have the same error variance and are
independent of subject (i.e. show sphericity).
Perhaps others would like to comment on, or qualify, this?
I hope this helps - Karl
--
Iole Indovina.
Laboratorio di Neuroimmagini funzionali
Fondazione Santa Lucia
Via Ardeatina 306
Roma 00179 (Italy)
Tel (+39) 0651501376
Fax (+39) 0651501213
|