Dear David

 

1.      In case I'm using differently sized ROIs (e.g. one with 20 voxels, one with 500), does this strongly affect the estimation of the DCM parameters due to the difference in signal-to-noise ratio? 

 

 

I don’t think so – the precision of the observation noise is estimated on a per-region basis (DCM.Ce). However, there may be other effects of ROI size. For instance, if your effect is small (a few voxels) and you have a 500 voxel ROI, then the principle mode of variance in that ROI is unlikely due to be related to your task. That means that the representative timeseries used in DCM for that region may not reflect your effect of interest. So try to pick ROIs where most voxels express the effect of interest.

 

I understand. I'm just wondering if it will be a problem in a resting state (spectral) DCM? 

 

 

I’m afraid I don’t know. Perhaps you could try different sizes and report back. However, to simplify your analysis and avoid post-hoc inferences, I would recommend not getting bogged down by this. Come up with a principled way of selecting ROI size – e.g.  by convention (e.g. the size of your smoothing your kernel) or by using an ICA component as a mask or by an anatomical mask. Then stick to it.

 

 

2.      Is there a way to calculate contrasts between connectivity parameters of a PEB? That is, is it possible to compare the connectivity estimates between conditions, e.g. if there is a difference in the strength of modulation of some between-region connection between an emotional or neutral condition?

 

This feature is on my to-do list! At the moment, the most straight-forward way is to form regressors which capture the contrast or difference of interest. Can you give more details as to the regressors in your PEB and your experimental design?

 

Do you mean adding regressors within the PEB design matrix? I did not have a specific design in mind, but an easy one would be a single group, 3 ROIs and two conditions. So there will be only the regressor with all ones for the group mean. What regressor to include in order to, for example, contrast the modulation between the connectivity of ROI 1 to ROI 2 of the two conditions? In other words is the modulation of connectivity of R1-R2 stronger in the emotional than in the neutral condition. 

 

Yes you can do this without additional regressors – we call this a Bayesian contrast. You get the expected value of the parameters from BMA.Ep and the covariance from BMA.Cp and then compute the probability of a difference between the two parameters of interest. I’m happy to write a script to do this – which I’d adapt from spm_dcm_review lines 253-273 – but I don’t have time today. Perhaps you could get back in touch at such a time that you want to perform this analysis?

 

Best

Peter