Dear SPMers,
I'm currently analyzing fMRI-data using the SPM99-approach and the random
effect model.
AFAIK, in this model the statistical test regarding contrasts is based on
individual differences of the parameter estimations. Therefore, no
individual variance is considered in the group analysis. Now I am wondering,
how the temporal autocorrelation of the time-series is accounted for.
Typically, design-matrix and functional data are smoothed by a Gaussian
kernel to increase the auto-correlation to a known extend, so that one can
account for it by an adjustment of the degrees of freedom (df). But this
seems to make sense only in SPM96 (fixed-effect model), since in this model
individual variance is considered in the statistical test and the group
statistic is based on individual t- or z-values. This calculation of
individual t-/z-values enables the df-correction. Because individual
variance is not considered in the random effect-model, a correction of the
df should be impossible. In this model, group analysis is based on
"raw"-differences in parameter estimates, and not t-/z-values. But
differences in parameter estimates are not adjustable by means of df.
However, when a correction for temporal autocorrelation does not take place,
does that mean that I can use any filter (e.g. lowpass filter, which was not
recommended in the fixed-effect model, since it increases autocorrelation to
an unknown extend)?
Maybe, there is a flaw in my argumentation. If so, please correct me!
Thanks for your help,
Sincerely,
Andre Szameitat
-------
Andre Szameitat
Ph.D. student
Max-Planck-Institute of Cognitive Neuroscience
Leipzig, Germany
[log in to unmask]
|