Dear Sahil,
Yes, you are right - the results of BMS and BMA here will not always concord.
The reason is that BMS considers the contribution of all these connections together - it is like a multivariate test - whereas the t-tests on individual connections consider the contributions individually - these are univariate tests.
I think you can report both and explain them using the above paragraph.
Best wishes, Will
________________________________
From: SPM (Statistical Parametric Mapping) [[log in to unmask]] on behalf of Sahil Bajaj [[log in to unmask]]
Sent: 20 August 2012 15:15
To: [log in to unmask]
Subject: [SPM] DCM results
Dear DCM experts,
I posted a doubt regarding BMS vs BMA results few days back, I am reposting the same with some changes. I would greatly appreciate any kind of help/suggestion.
I have analyzed 35 subjects perceptual decision making data using DCM. I used BMS to find the optimal model (model number 5) out of 10 defined models and also used BMA to interpret about the parameters by averaging over all the models over all the subjects. Each of the models contains 4 areas e.g. a,b,c & d.
My questions are:
1). My winning model, model number 5 e.g. contains facial affect modulating from region a to b, a to c & a to d- using BMS.
whereas BMA results show there is no significant modulation by that facial affect from region a to b, a to c & a to d - using one way t-test on the parameters.
And similarly for significant intrinsic connections as well.
So is it always true that the connections/modulations on the winning model should match with the significant endogenous/modulatory parameters obtained from BMA ?
If not (this is what I understood from Dr. Friston & Dr. Stephan's paper- 'Ten simple rules for DCM' that we use BMS to infer about model structure & BMA to infer the parameters and BMA abandons the dependence of parameter inference on particular model chosen and so it may not be matched.)
then whether to believe BMS results or BMA results and which one to report or both results can be reported in writing part with different inferences ?
2). What the minimum exceedance probability value is enough for a model to be selected as optimal (of course it should have maximum compared to other models but what value should be OK ?) ?
Thanks a lot !!
Regards,
Sahil Bajaj
