Print

Print
Dear Peter,

Thanks for your reply. 

I first thought using a log transformation to address the outliers problem, but realized that I have a lot of negative values;)

Yes, it seems I have to go some steps back (although I think the evidence in one group for model and family inferences is quite clear). I think the critical step is the time series extraction. I extract them from the global maxima within a anatomical mask using the effects of interest F-contrast (at p=0.1). I adjusted the extraction with the same F contrast and then used the same mask as before. Is it probably better to use a sphere instead of the mask for the last step (VOI definition)? All of my regions are quite small (anatomically) and I'm not entirely sure a meaningful DCM is possible.

Best wishes
André 
Von: Zeidman, Peter [[log in to unmask]]
Gesendet: Mittwoch, 20. Mai 2015 17:59
An: André Schmidt; [log in to unmask]
Betreff: RE: outliers in B parameters after BMA

Dear André,

It is reassuring that you get a similar pattern of families in each group and model 2 has the highest log model evidence in each group. However, I don’t think the evidence is strong enough to call this a “winning model” or “winning family”. (A log evidence difference of 3 between models or families is equivalent to 95% probability of a difference – yours is much less than this.) It is not surprising, therefore, that your subjects have varying parameter estimates.

 

I didn’t understand your sentence regarding negative values and log transformation.

 

Perhaps you should revisit your ROI selection, task modelling and DCM model setup to work out why your families didn’t have a strong difference in evidence?

 

Best,

P

 

From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of André Schmidt
Sent: 19 May 2015 16:50
To: [log in to unmask]
Subject: [SPM] outliers in B parameters after BMA

 

Dear list,

 

I've performed a DCM analysis with 21 models per subject (32 in one and 28 in the other group). BMS analysis revealed the same winning family and model for both groups (see attached). I then decided to use BMA over the models from the winning family (first seven) to compare the values across groups. However, I think I have some outliers in the B parameters (obtained from BMS.DCM.rfx.bma.mEps.B) and would like to ask how I can deal with them? I have a lot of negative values, log transformation will not work.

 

Many thanks for your help and suggestions.

André