Dear All,
Is there a most relevant paper for model selection in the context of
DCM as summarized so elegantly in this thread? Thanks in advance.
Eric
Quoting Klaas Enno Stephan <[log in to unmask]>:
> Dear Darren,
>
> The choice between fixed effects (FFX) and random effects (RFX)
> analyses in the context of BMS is no different from choosing between
> FFX and RFX in the context of any other statistical analysis (like
> SPM): if one believes that the effect of interest (here: model
> structure) is a fixed property of the population studied, one should
> use a FFX analysis. If, however, if one believes that the effect
> of interest is a random variable in the population studied, a RFX
> analysis is preferable.
>
> FFX analyses are appropriate, for example, when studying low-level
> physiological phenomena where it can (relatively safely) be assumed
> that these phenomena exist as fixed properties of the population and
> that variability across subjects is due to measurement noise alone.
> FFX BMS requires summing of the log evidences across subjects and
> then comparing this across models (equivalently: multiplication of
> Bayes factors).
>
> RFX analyses should be preferred when studying cognitive processes
> (due to potential inter-subject variability in strategy and, for
> systems with degeneracy, the possibility that networks are used
> differently across subjects to implement task demands) or patients
> (due to potential variability in pathophysiology or in the degree in
> which brain function has been compromised by the disease). RFX BMS
> uses the new Variational Bayes method in SPM8.
>
> Best wishes,
> Klaas
>
>
>
>
>
> ________________________________
> Von: Darren Gitelman <[log in to unmask]>
> An: [log in to unmask]
> Gesendet: Freitag, den 22. Mai 2009, 16:40:54 Uhr
> Betreff: Re: [SPM] DCM: fixed vs. random effects BMS; direction of
> connectivity change
>
> Narender
>
> Thanks for this update. I have some questions about it.
>
>
> On Fri, May 22, 2009 at 3:23 AM, Narender Ramnani
> <[log in to unmask]> wrote:
>> -------------------------------------------
>> Dear Narender,
>>
>>> We are investigating connectivity between two areas using DCM. Our
>>> anatomical model simply consisted of
>>> forward and backward connections between them.
>> ...[Details deleted for brevity]...
>>
>>> We used random effects Bayesian model selection to distinguish between a
>>> number of
>>> models (varying the modulating influence of our experimental effect on
>>> each connection, and varying the location of the input).
>> ...[Details deleted for brevity]...
>>
>>>
>>> (1) Is our model comparison approach appropriate?
>>
>> Yes - it is compelling and the model space is conceptually nice.
>> The only point I would make is that you appear to have used a random-effects
>> inference over subjects. This means that a priori, you expect each subject
>> could have a different architecture. Usually, in straightforward systems
>> neuroscience studies, one assumes that all subjects have the same
>> basic architecture (but different parameters). This means the fixed-effect
>> pooling of log evidence is more appropriate and usually gives more
>> significant results.
>
> If I understand correctly, this suggests it would be reasonable to do
> a fixed effects analysis to compare models across subjects, as long as
> one thinks that "all subjects have the same basic architecture". This
> seems to apply best to groups of normal subjects (or homogeneous
> groups). However, in the case of abnormal subjects (e.g.,
> neurodegenerative disease) the assumption of having the same basic
> architecture might no longer apply, and in that case a random effects
> model would more properly apply?
>
>>
>>> (2) We are interested in whether or not we can say that modulations
>>> represent increased or decreased connectivity. Are we able to make
>>> inferences about this from the contents of matrix B (e.g. some values in
>>> some subjects are negative)?
>>
>> Yes, exactly. The best way to report these is to report the % increase or
>> decrease in the fixed connectivity (A) implied by the condition specific
>> effect (B). This means commuting (100*DCM.Ep.B{i}(j,k)/DCM.EpA(j,k))
>> for each connection (j,k) and condition (i). (this is for fMRI, in M/EEG the
>> parameters are already gain parameters). For group results you can
>> use the group average,where each subject's estimate is weighted
>> by its precision
>
> Does this take the place of doing direct t-tests on the B matrix
> parameters as has been done in the past? Would it be best to analyze
> the "normalized" (B/A) matrix parameters for each condition instead?
>
> Thanks
> Darren
>
>
>> I hope this helps,
>>
>> Karl
>>
>>
>>
>>
>>
>> --
>> Narender Ramnani
>> Reader in Cognitive Neuroscience
>>
>> Cognitive Neuroscience Laboratory
>> Department of Psychology
>> Royal Holloway University of London
>> Egham, Surrey TW20 0EX
>>
>> Tel: 01784 443519 (Direct)
>> Fax: 01784 434347 (Departmental)
>> email: [log in to unmask]
>>
>> www.pc.rhul.ac.uk/staff/n.ramnani
>>
>
>
>
>
|