On May 11 2005, Stephen J. Fromm wrote:
> Are parameter estimates (the beta images) going to be the same (or
> essentially the same) in SPM2 and SPM99? This excerpt from the intro
> comments to SPM2's spm_spm.m indicates they *won't* be, but I wanted
> someone in the know to verify for me.
>
> Excerpt: "The ReML estimates can then used to whiten the data giving
> maximum likelihood (ML) or Gauss-Markov estimators. This entails a second
> pass of the data with an augmented model K*W*Y = K*W*X*B + K*W*e where
> W*W' = inv(xVi.V)."
>
If you select the autocorrelation option and specify AR(1), you will get ML
estimators as in your excerpt. In this case your betas will be different
from those obtained using SPM99.
However, if you select 'none' for autocorrelation, then estimation will
proceed using OLS, as in SPM99, assuming your errors are iid. This will
give betas similar to those from SPM99, assuming other things are equal. It
will not estimate the temporal nonsphericity at all (and so will be
quicker). If the assumptions of the autocorrelation estimation are met then
your betas using the SPM2 method will, however, be the best ones. Using the
pure OLS method, however, although the betas will be similar to those you
would get from SPM99, the effective degrees of freedom will be too great if
you want to do a first level inference. The right corrected degrees of
freedom can (if I'm reading spm_spm help right) be obtained along with OLS
estimators using the following 'third way' method:
Excerpt: If you do not want ML estimates but want to use ordinary least
squares (OLS) then simply set SPM.xX.W to the identity matrix. Any
non-sphericity V will still be estimated but will be used to adjust the
degrees of freedom of the ensuing statistics using the Satterthwaite
approximation (c.f. the Greenhouse-Giesser corrections)
I think this means that setting SPM.xX.W, an optional field, from the
command line gives you this combination, which is what SPM99 gives you. But
there's no reason to do this if you are using a 2-stage random effects
approach as the 1st level df are irrelevant.
Hope this is helpful
Alexa
|