Dear Helmut,
>Dear Klaas,
>
>>Will might be answering this at the same time as I do, but I have been
>>too lazy to walk down the stairs to find out... let's see whether we have
>>e-mail crossings again ;-)
>
>Any one answer will be fine...
>
>>Q1: In principle, there is no need to worry about large Bayes
>>factors. It is perfectly reasonable that some models should be much
>>better in explaining the empirical data than others. I am more puzzled
>>about two of your models (mode 1 and 2) showing perfectly identical model
>>fit. This does seem unlikely to me, unless the models have identical
>>structure.
>
>This is so because I selected two different regions (one in either model)
>which are anatomically distant but functionally connected.
Models with different VOIs cannot be compared by BMS within DCM - see my
previous e-mail.
>This should explain the result. Which at the same time brings up my next
>question: is there any way in distinguishing which one of different
>regions for which functional connectivity has been established (SVD) is
>the driving region? Or are they all equivalent? So far, I have not
>included the within one model. Would doing this be the key?
This question is exactly what you can address with BMS. For your case (as
I remember it) you could define two models, each of which consists of two
areas. In model 1, area 1 is driven by the input and in model 2, area 2 is
driven by the input. You can then cross this with the additional question
whether you have a unidirectional connection (from the "driven" area to the
other) or reciprocal connections between the two areas. Altogether this
gives you four models which you can compare by BMS:
Model 1: INPUT --> A1 --> A2
Model 2: INPUT --> A2 --> A1
Model 3: INPUT --> A1 <--> A2
Model 4: INPUT --> A2 <--> A1
>>Q2: If I understand you correctly you are asking for measures of
>>"absolute" model fit. You could, in principle, compute measures like
>>"percent variance accounted for", but such measures are purely
>>descriptive. How to derive a measure of absolute fit that has
>>inferential meaning is something that has been bugging me as well, and I
>>do not have a good answer, I'm afraid. One thing you could do within the
>>Bayesian Model Comparison framework is to compare your DCM against its
>>GLM-equivalent, i.e. a model in which all inputs affect all regions, but
>>no connections exist between regions (see Fig. 2 in Stephan 2004, J.
>>Anat. 205:443-470).
>
>Thank you for this hint which sounds promising. I will have a look. The
>basic question is: if I was ever to report this in a paper, I could - as
>others have done - just not talk about the fit ("don't wake a sleeping
>dog" - gibt's das?), but if I were - how could I make the point that in
>our subjects this is similar to the - equally bad :-( - fit of the GLM for
>those regions.
A simple way of doing this is to show the modelled time series in
comparison to the measured one, both for DCM and GLM. For DCM, you get
this plot through the GUI when choosing "review" and then "outputs".
Best wishes,
Klaas
_____________________________________
Dr Klaas Enno Stephan
Wellcome Dept. of Imaging Neuroscience
12 Queen Square, WC1N 3BG, London, UK
phone: +44-207-8337485
fax: +44-207-8131420
web: http://www.fil.ion.ucl.ac.uk/~kstephan/
|