Dear Eleonora,
> I would like to know if, in a fixed model for a fMRI study (modelled
> with a Box-Car), the convolution with hrf function is always the better
> choice or it depends from the interscan interval. I'm right if I make a
> convolution if the interscan interval is 15 sec. and each condition has
> 5 scans?
Probably not. It is worth convolving if the interscan interval is
about 4 seconds because you partially model the hemodynamic delay (even
if there is little smoothing per se) but at 15 seconds convolution with
a HRF would be difficult to implement. I would simply specify the
onset of the epochs as occuring 6 seconds before they actually did and
omit the convolution.
> The second question is about temporal smoothing: selecting a Low-pass
> filter (hrf) the effective degrees of freedom become 80.98 (instead of
> 81). What is the meaning of a not integer value for this variable?
These are the effective number of independent samples and can be
thought of as the scan number corrected for the serial correlations
among the error terms.
> The last doubt I have (...for the moment...) is: I read somewhere in a
> mail (that I've lost) that, if I make Global Normalization in a fixed
> model for each subject, the second level statistic between subjects for
> a given contrast doesn't need again this option (also 'Grand mean
> scaling'). If I remember well what I have read, I don't understand why
> they are not necessary.
Because the data are already normalized at the first level and that
variance component has already been removed.
I hope this helps - Karl
|