Dear Ilyan,
Generally, the difference lies in something called the pooled variance
assumption, which means that a model that includes both groups will be
more powerful than an analysis of each group separately (because the
combined model has more degrees of freedom to estimate the error
variance). This means that the separate analysis of groups is more
conservative (and more robust to violation of the pooled variance
assumption). The pooled variance assumption in SPM is quite subtle
because it operates over voxels (SPSS and SAS do not have to worry
about this because they do not handle mass-univariate data). This
means that one can pool the relative amounts of error variance (and
their correlations) between groups without assuming the variances are
equal.
There is no right or wrong way of doing this; the only thing to
remember is to qualify your inference by saying that you used a pooled
error variance assumption implicit in the non-sphericity estimation
used by SPM.
Having said this, it is irrelevant for you because you want to compare
groups. This means all the contrasts from subjects in all groups have
to be modelled in the same ANOVA model. In other words, you cannot
compare groups if you have already reduced the summary of activations
in each group to a single contrast (image) per group.
I hope this helps :)
With very best wishes,
Karl
From: iliyan ivanov
<[log in to unmask]>
To: Karl Friston <[log in to unmask]>
Sent: Wed, 10 February, 2010 16:45:22
Subject:
Dear Karl,
As an educational point for myself - using a single model for the
the 3 subgroups of children offers the benefit of increasing the power of
the analyses via increasing the degrees of freedom (one group of 30+
subjects vs. 3 groups of about 10 subjects each) - is that correct?
However, are there any benefits for analysing the groups separately and
then comparing them to one another to ascertain the effects of ADHD and
Fam HX - some may say that is the most robust way of doing it
Thanks Iliyan
