Dear Verónica,
your initial model was ill-specified and its estimation was probably
becoming sensitive to numerical precision errors that occur, eg, with
multithreading.
For the contrast you are looking at in your screenshot, you could have
used a paired t-test with data for that group alone, or, equivalently, a
one-sample t-test of the difference between the two levels of your
within-subject factor D for that group.
Also look at previous emails on this list, such as the thread with
Isadora yesterday, to see how to use models with partitioned error.
Best regards,
Guillaume.
On 06/10/16 12:15, Verónica García wrote:
> Dear Guillaume,
>
> Thank you so much for your reply. We have done your second suggestion
> and now the results are the same changing the MATLAB version. So the
> old model was not stable. I have attached the results of one of them
> (newResults.jpg). There are much less significant voxels.
> Nevertheless, there are two clusters that are located very close to
> the first two clusters of the previous results I sent you. I still
> wonder why changing the MATLAB version changed the results because I
> think the estimation is an analytical solution, doesn´t it? Maybe some
> functions in MATLAB have slightly been changed (the precision was the
> same, I checked it).
>
> We wanted to create a linear model similar to those found in some
> statistics books (see anova_model.jpg). It´s true that the model was
> very complex for the number of images we have.
>
> Thank you again.
>
> Best regards,
>
> Verónica
>
>
> On 5 October 2016 at 17:53, Guillaume Flandin <[log in to unmask]> wrote:
>>
>> Dear Verónica,
>>
>> there are many earlier posts on this mailing list about how to proceed
>> here. The recommended way would be to specify several second level
>> models, for each question you have (main effect, interaction, ...).
>>
>> Otherwise, what you can do here with the flexible factorial design is to
>> specify three factors (simplifying your two between-subject factors with
>> two levels each to a single factor with four levels):
>> * subject: equal, independent
>> * group: 4 levels, unequal, independent
>> * cond: 2 levels, equal, dependent
>> then enter:
>> * main effect: subject [1]
>> * interaction: group x cond [2 3]
>> The design matrix and covariance components will be much simpler and it
>> is likely that the results will be less dependent to the MATLAB version
>> you are using.
>>
>> Best regards,
>> Guillaume.
>>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|