Matt Shane and I have been corresponding a bit about this off-list and,
thanks to this dialogue and Will's elucidation, I now realize that I (and
apparently others, from the looks of the archives) had been making some
improper assumptions about within-subjects or mixed designs in SPM5.
Previously, SPM5's removal of the "ANOVA (within-subjects)" option and the
addition of the independent/dependent distinction for design factors (as
well as the help blurb noting: "If you change this option to allow for
dependencies, this will violate the assumption of sphericity. It would
therefore be an example of non-sphericity. One such example would be where
you had repeated measurements from the same subjects - it may then be the
case that, over subjects, measure 1 is correlated to measure 2") had caused
me to assume that simply selecting "dependent" was sufficient to create a
within-subjects design, and some sort of lovely SPM5 magic took the place of
all those messy individual-subject columns a la SPM2.
Clearly it is time for me to learn what happens when you assume.
At any rate, if you please, just a quick clarification to make sure I and
any similarly confused individuals are on the right page. My understanding
now is that simply selecting "dependent" for a within-subjects factor (and
not explicitly modeling subject effects) is not "wrong" in the sense of
producing false inferences, but it fails to give the benefit of added power
that one expects to get from modeling subject effects and thus removing them
from the error term? I.e., by only selecting "dependent," you are
essentially still running a between-subjects ANOVA but (at least) managing
to avoid the sin of ignoring a sphericity violation where one probably occurs?
Thanks, and happy holidays to all,
On Fri, 21 Dec 2007 17:44:56 +0000, Matthew Shane <[log in to unmask]> wrote:
>Thank you Will,
>I've tried removing the Group and TrialType variables from the design
>matrix(while keeping them as factors in the model).
>Unfortunately, I'm still coming up against the same (fairly unusual)
>problem: the contrast manager continues to tell me that contrasts utilizing
>any columns other than the 30 Subject columns are invalid. Moreover, the
>contrast manager won't allow me to define a single 1 0 0 0... contrast, but
>rather requires the associated -1 to define it as valid.
>Has anyone else encountered this problem with the flexible-factorial design?
>I've never had my contrast manager act this way before, and so I'm pretty
>sure it's design-specific and not a local problem.
>I don't believe my .mat file was properly uploaded last time. I'll attach
>again, this time as a .mtt file (which will need to be renamed .mat again
>upon download). Is there a better way for me to design this model?
>On Fri, 21 Dec 2007 12:11:05 +0000, Will Penny <[log in to unmask]>
>>I would remove the 'group variables' and the 'trial type
>>You can still test for the effect of 'group' using the appropriate contrast.
>>(1) The contrast vector c=[ones(1,10),zeros(1,20)] will test for the
>>effect of group 1 (t-test for positive effect, F-test for any effect).
>>(2) The contrast matrix C=[ones(1,10),-ones(1,10),zeros(1,10);
>>zeros(1,10);ones(1,10),-ones(1,10)] will test for a main effect of
>>group. You'll need the F-test here.
>>You can also test for the main effect of trial type by using a contrast
>>that collapses over the relevant columns of the 'group x trial type
>>The main concept here is that if you've got the 'interaction terms' in
>>your design matrix you can test for the main effects using an
>>Matt Shane wrote:
>>> Dear Will (or anyone else who can help),
>>> Your reply to Michiru was very timely for me, and I have just attempted
>>> to undertake an analysis guided by your steps below. I feel like the
>>> design matrix is correct, but unfortunately the contrast manager doesn't
>>> appear to be appreciating the design I've created. And so I'm thinking
>>> that I might have gone astray from your advice in some manner.
>>> In short: I have 30 participants in a 3 (Group) x 3 (TrialType)
>>> mixed-model design. I've thus created 3 factors in the
>>> flexible-factorial model: Subject, Group and TrialType. The design
>>> matrix (which I'm attaching to this post) appears (to me) to be right: I
>>> have 30 subject columns, followed by the three group columns, followed
>>> by the three trial-type columns, and finally the group x trial type
>>> My problem arises when I try to create contrasts in the contrast
>>> manager, however: I'm able to create contrasts with the first 30
>>> 'subject' columns, but I'm told that any contrast utilizing the 'group'
>>> or 'trial type' columns is invalid. Which, obviously, is problematic
>>> since it's the group and trial type that I want to interrogate!
>>> Does anyone have any advice? Have I set up my matrix incorrectly? I'm
>>> attaching both the matrix and the .mat file, and would be ever thankful
>>> for anyone willing to take the time to look it over.
>>> Matthew S. Shane, Ph.D.
>>> Research Scientist
>>> The MIND Institute
>>> 1101 Yale Blvd NE
>>> Albuquerque, NM, 87131
>>> (505) 272-4374
>>> [log in to unmask]
>>> -----Original Message-----
>>> From: SPM (Statistical Parametric Mapping) on behalf of Will Penny
>>> Sent: Thu 12/20/2007 9:20 AM
>>> To: [log in to unmask]
>>> Subject: Re: [SPM] questions on perfroming 2 x 2 within-subjects ANOVA
>>> in SPM5
>>> Dear Michiru,
>>> This is most easily done using the 'Flexible Factorial'
>>> 1. Create two factors.
>>> 2. Call the the first one Subject. Independence Yes, Variance Equal.
>>> 3. Call the second one 'Condition'. Independence Yes, Variance Unequal.
>>> 4. Under, Specify Subjects or all Scans, Choose Subjects
>>> 5. Under Subjects, create a new 'Subject' for each subject that you have
>>> eg. 5.
>>> 6. Then, for each Subject, under 'Scans'. Enter the 4 scans you have for
>>> each subject.
>>> 7. Also, for each Subject, under 'Conditions' enter the vector [1:4]
>>> 8. Under Main effects and Interactions create 2 main effects; factor 1
>>> and factor 2.
>>> 9. Specify other covariates as necessary and your o/p directory.
>>> 10. Then save your design job as 'within_subject_design' and press run.
>>> I have attached my saved job file 'within_subject_design.mat' as a
>>> template for you. When you run it, SPM should create the design matrix
>>> shown in 'design-matrix.png'.
>>> Note the 5 subject columns on the left. Without these 5 columns
>>> you do not have a 'within-subject' design.
>>> Also I have treated your 2 x 2 design as a 1 x 4. So you'll need to bear
>>> this in mind when doing your contrasts eg. 1 1 -1 -1 and 1 -1 1 -1 to
>>> test for main effects and 1 -1 -1 1 for the interaction (of course,
>>> pre-pad these with 5 0's).
>>> Best wishes,
>>> Michiru Makuuchi wrote:
>>> > Hi,
>>> > I have tired to perfrom 2 x 2 within-subjects ANOVA in SPM5, but I
>>> > couldn't find
>>> > how I could do that in 'Full factorial' dsign. Therefore I designed the
>>> > design matrix
>>> > via 'Multiple regression' option. The resulted design matrix was similar
>>> > to Fig 7 of
>>> > Henson and Penny's online document (ANOVA and SPM). The difference was
>>> > only the position
>>> > of constant term. In Fig 7, it was the 4th column, but it was on the
>>> > last column in my design matirix.
>>> > Here are my questions.
>>> > Is my approach acceptible for the purpose?
>>> > Can someone point out the exact procedure to build the model for 2 x 2
>>> > within-subject ANOVA?
>>> > Best regards,
>>> > Michiru
>>> > Michiru Makuuchi
>>> > Max Planck Institute
>>> > for Human Cognitive and Brain Sciences
>>> > Stephanstrasse 1a, 04103 Leipzig, Germany
>>> William D. Penny
>>> Wellcome Trust Centre for Neuroimaging
>>> University College London
>>> 12 Queen Square
>>> London WC1N 3BG
>>> Tel: 020 7833 7475
>>> FAX: 020 7813 1420
>>> Email: [log in to unmask]
>>> URL: http://www.fil.ion.ucl.ac.uk/~wpenny/
>>William D. Penny
>>Wellcome Trust Centre for Neuroimaging
>>University College London
>>12 Queen Square
>>London WC1N 3BG
>>Tel: 020 7833 7475
>>FAX: 020 7813 1420
>>Email: [log in to unmask]