Bas Neggers wrote:
> But still it is hard to grasp why GLM estimation with both spm2 and spm5
> having maxmem at 2^20 yields an 8x speed difference in estimating the
> model.
That was what I was wondering...
One thought I've had is that if you've specified inter-level
dependencies or unequal variances in the model (this might not have
been an explicit decision on your part, e.g. two-sample t-test has
unequal variance set by default) then because of the resulting
non-sphericity SPM5 will perform ReML to estimate the variance
components, which would slow things down a lot compared to just
ordinary least squares.
> But who cares, it is fast now ;-)
Ah, the spirit of scientific enquiry ;-)
Best,
Ged.
|