Dear FSL-Randomise experts,
I have a set of binary images for 13 subjects with dyslexia and 13 controls. A voxel has value 1 if it belongs to the LGN of the subject and 0 if it doesn't. So, my independent variable IV (group membership) is categorical. And my dependent variable DV (the voxel's value) is categorical too. The question I'm trying to answer in my experiment is: Is the probability that this voxel belongs to the brain of a control higher to that of belonging to a subject with dyslexia?
In this thread (https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1302&L=fsl&P=R59199&1=fsl&9=A&I=-3&J=on&d=No+Match%3BMatch%3BMatches&z=4) it is said that there is no difference between categorical and continuous data, and that the difference between them is the interpretation we give them. That it can be modeled whatever I want in the design matrix.
1) I understand that what it is said in that thread is true for the IV, I can model it in the design matrix as a categorical variable. However, what about the values of the voxels? Is randomise assuming that these values are continuous? Does it matter if they're binary?
As I understand, a GLM is fit to each voxel's time-course, in my case a 1D vector with binary values, using OLS. However, when the DV (the voxel's time-course) is categorical, the OLS method is biased and inefficient. Does this mean what I think? That randomise can't be used on binary masks? If what I fear is true, then can you recommend other tool to be used for this same purpose?
2) If randomise doesn't assume a continuous DV, then the same thread I mentioned before says that virtually any test that can be accommodated within the GLM can have a design matrix that will be recognized by randomise. However, what if the test I need to use is not linear? This web page (http://www.ats.ucla.edu/stat/mult_pkg/whatstat/) recommends to use Fisher's exact test when there is 1 IV with 2 levels (independent groups) and the DV is categorical. According to that web page, I can't use the 2-sample t-test that seemed to be so good for my analysis. Is there something that can be done?
Thank you,
|