Dear all

Following for this discussion, i have some following question and help you could give me some answer.

After reading Jeanette A Mumford's website, the situation in this discussion thread is likely the example 1.

If i model the intercept in the design matrix, and interest in the contrast of [0 1] or [0 -1] (Correlation between FA and behavior measurements), do i still need to demean the behavior measurements and FA ?  It seems like i don't need to demean the behavior measurement and FA data when i have model the intercept in the design matrix, am i right ?

Best

Paul

---

Date: Fri, 19 Apr 2013 22:15:36 +0200
From: [log in to unmask]
Subject: Re: [FSL] Demean Covariate
To: [log in to unmask]

Hi Mike,

I guess, the issue of demeaning EVs in FSL is simply because of the lack of demeaning routine in FSL, which may be necessary for a GLM or a linear regression, and actually implemented to deal with automatically in other statistical programs such as SPSS or other matlab functions.

As you can see on http://mumford.fmripower.org/mean_centering/, as Dr. Smith directed, 

I guess you used other co-variates as well, since you found different results with/without demeaning. If not, I suggest you to check if your contrast vector is in accordance with your intention carefully. Or maybe you didn't include the interceptor... I don't know.

Re1: Because the model that you fit was different. 

Re2: It depends on your question of interest, your model and contrast you tested. When you model the response (Y) as Y=beta_0 + beta_1*X, and you test a contrast of [0 1] or [0 -1] to know if the beta_1 is non-zero, then it is reasonable to demean X, because you already have the interceptor beta_0.

Re3: Actually that's what a linear regression does if you include an interceptor (beta_0) in your model. Demeaning the response variable Y (here FA values) only affects beta_0 (if you used a model: Y=beta_0 + beta_1*X), which might be not of your interest. What you really want to know is the effect of variable X (significance of beta_1), but not the mean FA value of all subjects, right?

While I'm writhing the answers, the feeling that you didn't include an interceptor in your model has grown... did you?

---

Seung-Goo KIM

On 19, Apr, 2013, at 3:02 PM, Michael Skeide wrote:

Hi,

similar questions have been asked before, I know, but like many others I got confused studying the answers, so I have to ask again.

I have created skeletonized FA maps of 40 participants using TBSS and want to find out if 1 behavioral measure (range 19-71, mean 47.7674) is associated with these FA maps. In the FEAT user guide you recommend to enter the behavioral measure as an extra EV, which should be orthogonal wrt the group mean EV, i.e. demeaned.

I have run the analysis with both  raw and demeaned behavioral scores and the results are very different.

Now my questions are:

1. Why are the results different given that the differences between the individual behavioral scores are the same?  

2. Is the analysis really more valid when I use demeaned behavioral covariates? If yes, why? If not, why?

3. If I demean the behavioral covariate, don't I have to demean the FA maps as well?

Thanks for your patient answers,

Mike