On Mon, Jul 27, 2009 at 6:12 PM, <[log in to unmask]> wrote:
> We need few clarifications about DCM analysis as implemented in SPM8 and
> not clarified by the manual. We would appreciate a lot your help with the
> following questions. We hope they are not too long and won't take too much
> of your time.
> 1) "Detrend" parameter on the "Data and Design" tab.
> What does it mean that "mean or low frequency drift" are modelled (how it
> is explained in the SPM manual)? Does this allow to remove effects of mean
> or slow oscillations from the analysis (the same way as we do with
> confounding effects when modelling them in GLM)?
Yes, this is exactly right. The disadvantage is that you need to do it
by eye. Alternatively you can just filter the data in the right way
during preprocessing. We don't usually use this option.
> 2) "Modes" parameter on the "Data and Design" tab.
> We understand it effectively means the number of principal components of
> the covariance matrix of the data. But we don't understand how the amount
> of explained variance affects the result (or performance) of DCM analysis.
> Perhaps we don't need some components carrying essentialy noise but how do
> we know that components are not significant? Basically the question is
> about the rule to apply when chosing the value of this parameter.
This is not about rejecting noise components but about data reduction
for computational efficiency. I'd say you need at least the number of
modes equal to the number of sources you think are present plus some
extra modes just in case. In the solution you'll see mode waveforms
and if you consistently see that the higher modes are just noise you
can reduce the number.
> 3) "Onset[s] ms" parameter on the Electromagnetic model tab.
> We do not have information when the input comes. Therefore if we put onset
> too late (too large values of this parameter) we risk to miss a moment
> when the actual input comes. That will be particularly bad in a situation
> when we test models with different input nodes. Could you please let us
> know if this way of thinking is correct or wrong?
This number is just a prior and not a very tight one. You should
specify more or less when the activity starts with respect to time
zero of your ERP. Just keep it the same for all models and it won't
> 4) After pressing "Estimate" we get question about using previous priors
> and posteriors. What do they mean? For some reasion, when using the same
> priors and not using the same posteriors we get the same log-evidence for
> all tested models. Therfore we decided not to use previous priors and
> posteriors. But we don't know if it is correct or wrong and why. Could you
> please help us explaining (may be directing to a relevant source of
Unless you have a good reason I'd answer 'No' to both questions. This
is about using a previous inversion (if you have one) to initialize
the present inversion. You can use either the same priors as
previously or you can use the posteriors of the previous inversions as
priors for the next. Lets say you did some inversions on healthy
subjects and you want to use the estimated parameters as priors for