Hi JF,

Please, see below:

On 16 December 2016 at 15:50, Jean-Francois Cabana <[log in to unmask]> wrote:

Dear FSL experts,

 

I have a question concerning the design matrix and contrasts for use in PALM. Here is a short description of my problem. We are studying a group of subjects that are all affected with a particular genetic mutation, and we are comparing several MRI metrics to a group of age- and sex-matched  healthy controls. Furthermore, the subjects group can be divided into two subgroups, because the women are clinically affected differently than men (dyslexia and epilepsia respectively). I’d like to compare each of the three groups between themselves, but also compare the full subjects group (Female+Male) vs control.

 

When I first started my analysis, I did a simple ROI analysis using a two-tailed two-group t-test with additional nuisance variables added to the model (age, age², sex, age*sex, age²*sex) to compare healthy control vs all subjects with the mutation. I get a set of p-values for this test. Now I switch to the 3 groups case where I split the subjects into F and M. Let EV1, EV2 and EV3 be the variables representing healthy controls, subjects-F and subject-M respectively. I set up my two-tailed contrast matrix as this :

 

[-2 1  1 0 0 0 0 0; ... % HC vs (F+M)

 -1 1  0 0 0 0 0 0; ... % HC vs F

 -1 0  1 0 0 0 0 0; ... % HC vs M

   0 1 -1 0 0 0 0 0];    % F vs M

 

Now, my question is regarding to the first contrast here. I would (perhaps naively) expect that the first contrast of this 3 groups design would be equivalent to the earlier 2 groups design, where F and M were considered as a single group. However, the p-values I get appears to be generally higher in the 3-groups analysis. Setting p<0.05 as a significance threshold, I then get fewer significant results in the 3-groups version.

 

Here are my questions :

 

First, does this contrast here make sense for the combination of the F+M groups?


Yes.
 

 

Second, how do you interpret the difference in p-values between the two designs?


The design that has 3 groups accommodates (discounts) the possibility that M and F could have different effects and different sample sizes, whereas the other one, with 2 groups only, doesn't allow for that.

 

 

And finally, would you recommend using the 2-group design to compare the full subjects group vs controls, and then the 3-groups design to compare all subgroups, or would you rather trust the 3-group design for all statistics, including the combination of the F+M groups?


If the two groups had/have the same size and if it were known that M and F had the same effects, then lumping these two would be fine. However, that doesn't seem to be the case (e.g., different effects of epilepsy and dyslexia), so I would suggest keeping these separate.

All the best,

Anderson

 

 

 

Thank you for your help,

JF