Hi,
The thing to realise is that GLM is not trying to calculate correlation values. It is trying to test the null hypothesis. You can, in some cases, also calculate correlation values and do a statistical test on the correlation values, and in these cases the results are exactly equivalent for the statistical testing (whether the null hypothesis is rejected or not).
The GLM is more flexible and so does not always do things that map onto correlations. It is true that we usually standardise the values in our EVs (regressors) and also in the data (e.g., grand-mean scaling), but in terms of the statistical hypothesis testing, this does not matter as you will get the same statistical results (regarding rejecting the null hypothesis) whether you standardise or not.
If you want correlations, or want to see the relationship between correlations and regression parameters, then I suggest you look at some standard statistics textbooks.
All the best,
Mark
On 3 Feb 2013, at 12:30, Ib <[log in to unmask]> wrote:
> In brain analysis generally, when talking about multiple regression/GLM, the regressors are standardized, but what about the brain data? In correlational analysis it would be automatically zscored (standardized), since it's part of the computation. But in GLM it shouldn't be. Can anyone explain me this in more detail? I know this would affect the beta coefficients, but I wonder what is the usual practice or whether there is ongoing debate about it. Thanks. Ib.
|