Dear Luca,
If you want to test for a group x score interaction, you can still use a
"one-way ANOVA" but you need to set "interaction with factor 1" for the
covariate. You would then use an F-test with contrast:
[0 0 0 1 -1 0
0 0 0 0 1 -1]
Note that a "one-way ANOVA" is the same thing than a "full factorial"
design with a single factor - this is just an interface shortcut.
Best regards,
Guillaume.
On 16/10/17 09:33, PRESOTTO LUCA wrote:
> Dear experts,
>
>
>
> I know these questions pop up regularly on this list but, even after
> googling quite a lot, I’m still confused.
>
> Here’s my setup. I’m doing a second level analysis were I want to study
> the correlation between the patient score in a test and the imaging
> data. I have patients clearly divided in three groups (they suffer from
> 3 completely different pathologies).
>
> The three groups have (very) different numerosities (e.g.: 35, 7, 13).
>
> What I’d like to do is a general linear model where the effect of
> differences due to the pathology are correct for before testing for the
> correlation. E.g., for a single pixel,
>
> y_i = b0 + b1* X_1_i + b2 * X_2_i + b3*X_3 _i+ b4 * S_i + eps_i
>
> where i is the patient index, b1,2 and 3 are the average deviations from
> the means for the pathologies 1, 2 and 3, X1-3 are the index that are 0
> and 1 if patient i has pathology 1 2 or 3, S_i is the score for patient
> i in the test I’m interested and b4 is my regressor, the parameter I
> want to estimate.
>
>
>
> So, if I understood SPM correctly I can use “one way ANOVA” to do this,
> if I’m not interested in studying interactions. Is this correct? Then
> I’d create 3 cells, each with the scans for that specific pathology, and
> then a single covariate vector with all the scores for all the subjects.
> Finally, to see the results, I’d input as a contrast 0 0 0 -1 (I’m
> interested only in negative correlation). Is everything correct until now?
>
>
>
> What if I want to test interactions? I should use a full factorial
> design I’d say. And I should put in 1 factor with 3 levels. Then 3
> cells, one for each pathology, indexed with levels 1, 2 and 3. Right?
>
> What about the contrasts? Is it still going to be 0 0 0 -1? Or should I
> do one of the weird things with like dozens of numbers in it? And what
> about the interactions? I’ve read a dozen tutorials about this, but I’ve
> still got no clue!
>
>
>
> Thanks in advance!
>
> Luca
>
>
>
> /Rispetta l’ambiente: non stampare questa mail se non è necessario.
> Respect the environment: if it's not necessary, don't print this mail./
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|