On Sat, 26 Aug 2006, Ray Fix wrote:

Hello,

I have some questions about doing group analysis. I know these are really basic... but everyone loves easy softballs, right?

Question 1. How should I define multiple classes?

The tutorial examples show two classes, male and female. If I have patient and controls, in addition to male and female, should I define it :

class Patient class Control class Male class Female Variable Age Input subject1 Patient Male 24 Input subject2 Control Female 39 :

or do I have to boil it down to one column??

class PatientFemale class PatientMale class ControlFemale class ControlMale Variable Age Input subject1 PatientMale 24 Input subject2 ControlFemale 39 :

or should I separate these into different FSGD files?

or some other way?

Your 2nd method is correct. You should not confuse classes with factors/levels. You need a class for every combination of factor levels. The number of classes equals the product of the number of levels of each factor.

Question 2: Best practice for a small study

If I have lots of slopes, but not so many samples (<50), is it advisable to make one big FSFD file or make several small FSGD files with all those slopes?

Question 3: DOSS versus DODS

By staring at Doug Greve's excellent slides http:// surfer.nmr.mgh.harvard.edu/docs/ftp/pub/docs/fsgroupana.pdf I was able to figure out whad doss and dods do the design matrix. Can someone make a simple statement as when to use one instead of the other?

Questions 3 and 4 relate to how you are going to model the effects. But sounds like you are on a fishing expedition? If so, then you'll probably want to create separate models (fsgd files) for each effect, but remember you'll need to correct your p values for all the effects you test (one of the problems with fishing trip). The DOSS and DODS are just another aspect of the model. The one you choose will depend on what you believe to be true. DOSS will create one slope across all classes. Do you believe this to be true? Eg, if you variable is Age, do you believe that the change in thickenss wrt to age is the same across all classes? If not, then you'll need DODS.

Question 4: Multiple Comparison

Section 4 of the tutorial gets into multiple comparison. I started to run the sample and realized it will probably take my computer 3 years to finish executing ;) Can someone tell me what problem this is solving for me? What is multiple comparison and where do you run into trouble by not doing it?

The "Multiple comparsions" problem is the problem that if you simply threshold each voxel/vertex of your map to achieve some false positive rate (eg, p < .01), that FPR only applies to one voxel/vertex. The meaning of the threshold is that 1/100 vertices will be active purely by chance (ie, in the case where you have pure noise). Thus, with 1000s of vertices/voxels you will have many false positives. One correction is bonferroni in which you divide your threshold by the number of voxels/vertices, but this will lower the threshold to something like 10^-6 (or lower), and not much will survive. Another method is to perform clustering with the assumption that two or more continuous vertices/voxels would not all be active by chance. This is what the simulation is doing. Which sim are you running? Probably mc-full, which I think is what is recommended on the wiki. You can try running mc-z, which will be much faster but more conservative.

doug

That's all I can think of for now. Thank you.

Ray

PS: I did notice there is a broken link in the wiki tutorial. On the page https://surfer.nmr.mgh.harvard.edu/fswiki/FsTutorial/GroupAnalysis there is a link in section 3.0 that points to information about creating contrast vectors. It is broken. I think it should point to: https://surfer.nmr.mgh.harvard.edu/fswiki/FsTutorial_2fCreateContrastVectors

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer