Dear Charlotte,
> I am learning to use FSL and TBSS. I have some questions about the
> regression model.
> 1. I don't really understand why we have to demean covariates? What
> clearly does that mean?
Let us say you have three data points that are [100 100 100] and that
you have an age covariate that is [50 49 51]. If you were to fit this
covariate to the data you would get a beta of 2, so that your best fit
is 2*[50 49 51] = [100 98 102]. So clearly this covariate was very
succesful in explaining your data and you would conclude that there
was a linear trend. But in reality you have only used the mean in your
age covariate to explain the mean of the data. If you de-mean your age
covariate you get [0 -1 1], which will not explain anything in the
data and you will (correctly conclude) that there is no dependence on
age.
Having said that there are all sorts of reasons why often you can get
away with not demeaning your covariates. But I think it is always
neater if you understand why, when (and even how) you should de-mean
and do that explicitly yourself.
> 2. I think I understand that demeaning is necessary for linear
> covariates, but if I had non linear covariates as viral load or
> log10 covariates, what am I suppose to do??
You are always looking for a linear relationship with the covariate
you put in, and hence it should as a general rule be de-meaned. If/
that you have performed some non-linear transform on that parameter
before you put it into GLM only affects your interpretation of the
results, it doesn't change how GLM treats it.
> 3. In my model, I have two covariates: age (demean) and sexe (0/1).
> I enter my model like that:
> EV1 (group) EV2(age demean) EV3 (sexe)
> 1 -0.2 0
> 1 1.6 1
> 1 -0.5 0
> 1 0.2 0
> If I don't enter the EV1, I can't see the age's effects. What does
> it mean?
EV1 models the mean, and if you don't do that the mean will remain as
part of the residual error (the difference between your data and your
best model fit) which means that the error gets bigger and your t-
values smaller.
Good luck Jesper
|