Dear Gerald,
> I have been working on a problem concerning the statistical analysis of
> a PET experiment on language. (Petersen,1988; Grabowski, 1996).
> Because of this work I struggle with some questions, that I hope you
> would be willing to discuss:
>
> 1.) For a GLM to be suitable as a regression form the following points
> appear prerequistes:
> a) The data behave approximately linearly
> b) All important factors of variance are included in the model.
>
> The question: Are you routinely checking for violations of this assumption
> (e.g. distribution of residuals, etc.) ?
Both (a) and (b) are subsumed under the requirment that the model is
appropriate (note that nonlinear relationships between various
explanatory variables and the dependent variable can be modeled with
the GLM using polynomial expansions).
Critically the assumptions are that the residuals of the model are
normally distributed. Very early versions of SPM did this but for the
past years it has been dropped because (i) the check proved unnecessary
and (ii) failing to show non-normality does not preclude violations.
In short there is no check on the distribution of residuals in SPM but
if you did check them you would generally find them to be normal unless
your model specification was incomplete or inappropriate.
I hope this helps - Karl
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|