Dear SPMers,
Sorry to come back on this again but I would be very glad to get some help
on this issue. I'll try to rephrase in a few words :
I have an experiment involving 2 groups of subjects (Patient, Control: 10
subjects per group) and 4 sessions (1 condition versus implicit baseline per
session). I wonder if I can perform a within-subject ANOVA with two factors:
subject + condition (instead of classical one-sample t-tests) including all
data available in one group, and test for each condition separately (design
matrix attached).
For instance, this contrast would test the effect of condition 1 :
[ones(1,10)/10 1 0 0 0]
This approach could also be extended to a mixed ANOVA using data from both
groups. On my point of view, the big advantage of within-subject ANOVA
versus multiple one-sample t-tests is to estimate the model based on all
data available instead on focusing on sessions one after another and as we
discussed earlier it's also the way to model within-subject variance.
This approach is therefore coherent with contrast n°4 given in p.6 of
"Contrast weights in flexible factorial design with multiple groups of
subjects" (conweights.pdf) by J. Glasher and D. Gitelman, quoted numerous
times on this list.
However I am not sure if this design is applicable in my case.
Any help would be greatly appreciated,
Cheers,
Camille
On Wed, 28 Jan 2009 10:51:25 +0000, Camille Maumet <[log in to unmask]>
wrote:
>Dear Volkmar,
>
>First of all, thanks for your quick answer and these explanations.
>
>>there is one likely explanation for the increase of the number of
>>significant voxels in an ANOVA with multiple conditions compared to a
>>series of one-sample t-test: the ANOVA model has much higher degrees of
>>freedom (if I'm right there should be 27 for 10 subjects and 4
>>conditions) than the one-sample t-test (9).
>>There are other differences between the two models as well: in a
>>one-sample t-test the only source of variance is between-subject,
>>whereas in the ANOVA model there is both between-subject and
>>within-subject variance.
>>To compare 2 conditions, one could even model differences between
>>conditions on the 1st level and take the resulting con images to the 2nd
>>level. (This holds for any t-contrast that can be computed on 1st
>>level.)
>
>I still have a doubt on whether or not this model (within-subject ANOVA with
>2 factors : subject + condition) is valid to assess the effect of each
>single condition. As you said :
>
>>There is a lot more to say about the implications of different models on
>>the results, but as a rule of thumb you should only include those
>>conditions in your 2nd level that you are using in your contrasts.
>
>Here is my question, what if I want to take into account within-subject
>variance in the study of each single condition ? Maybe it's a silly thing to
>do... However, if I understand well, this would allow to model the
>variability held by each subject which seems sensible to me in the case of
>fMRI studies since the answer to MRI scanner can be quite different from a
>subject to another. This led me to a within-subject ANOVA with 4 different
>contrasts (as specified in my previous message) to assess the effect of each
>condition separately. To follow your point, I am actually using all
>conditions specified in my 2nd level but in 4 different contrasts (I could
>also test for comparison between conditions but that's another matter).
>Could you tell me if this make sense ?
>
>Thank you,
>
>Camille
>
>>
>>Am Dienstag, den 27.01.2009, 17:22 +0000 schrieb Camille Maumet:
>>> Dear SPMers,
>>>
>>> I have 10 subjects and 4 different conditions. First level stats give me one
>>> contrast image per subject per condition. My question arises with the
>>> second-level statistical analysis. I want to find the effect of each
>>> condition on my group of subjects.
>>>
>>> I first performed four different one-sample t-test to test each condition on
>>> my group of subjects. This worked fine, however I thought that it would be
>>> better to use a one-way ANOVA (flexible factorial) in order to take into
>>> consideration the effect of each subject.
>>>
>>> To this aim, I used the "flexible factorial" design with two factors :
>>> subject and condition and two main effects subject and condition. To isolate
>>> the effect of one condition I used the following T-contrasts :
>>> Condition 1 : [ones(1,10)/10 1 0 0 0]
>>> Condition 2 : [ones(1,10)/10 0 1 0 0]
>>> Condition 3 : [ones(1,10)/10 0 0 1 0]
>>> Condition 4 : [ones(1,10)/10 0 0 0 1]
>>>
>>> The ANOVA leads to much more voxels activated with a FWE-corrected threshold
>>> than the solution using 4 different one-sample t-test. I wonder if that can
>>> be explained by the fact that subjects variability have been identified in
>>> the model ?
>>>
>>> Furthermore, I wonder if this approach is right on a statistical point of
>>> view ? I searched the mailing list and only found examples in which one-way
>>> within-subject ANOVA was used for conjunction analysis or difference
>>> between-conditions.
>>>
>>> Any thought on this would be highly appreciated,
>>>
>>> Camille
>>>
>>--
>>Volkmar Glauche
>>-
>>Department of Neurology [log in to unmask]
>>Universitaetsklinikum Freiburg Phone 49(0)761-270-5331
>>Breisacher Str. 64 Fax 49(0)761-270-5416
>>79106 Freiburg http://fbi.uniklinik-freiburg.de/
|