Dear Tali,
Karl drew my attention to your query.
A. It would be good to do some diagnoses on the data from your 40% of patients:
First, it’s worth considering whether the data going into DCM shows a response to your task. When you say the participants have more than 30 active voxels in the ROI, is that at an uncorrected threshold of p < 0.1? This is a liberal threshold, and assuming these activations don’t survive a stricter threshold, it could be there isn’t much experimental variance in these voxels for the DCM to explain (low signal to noise).
Assuming that you’re using the eigenvariate button to extract your ROIs, you could check whether the first principal component (which is what DCM receives) is capturing much of the variance in the signal. To do this, you can look in the GUI when you extract the ROI, it’ll be shown to you as a percentage. To get this at a later stage, for each ROI (.mat file) run:
load VOI_xxx.mat;
100*xY.s(1)/sum(xY.s)
And check if each ROI is capturing a reasonable amount of signal from the timeseries.
B. Assuming your DCM has sufficient variance to be working with, is your DCM a good model for these subjects? Perhaps this sub-population are using different brain regions – have you enough subjects for a between groups SPM analysis? Alternatively, do things improve if you try more advanced DCM options (DCM.options.bilinear or DCM.options.stochastic)?
C. Regarding whether the full model should always explain more of the variance than the reduced models… The post-hoc routine will only remove a parameter when doing so won’t reduce the model evidence. Therefore, it’s perfectly possible that the full and reduced models will explain the same amount of variance – the latter will do so with fewer parameters.
Good luck!
Peter.
Peter Zeidman
Methods Group
Wellcome Trust Centre for Neuroimaging
12 Queen Square
London WC1N 3BG
[log in to unmask]
> -----Original Message-----
> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
> On Behalf Of Tali Bitan
> Sent: 18 April 2014 04:58
> To: [log in to unmask]
> Subject: [SPM] Low variance explained in DCM
>
> Dear DCM experts
>
> We have used DCM post-hoc on a group of patients and a group of controls,
> and tested the full model with spm_dcm_fmri_check. While the control
> group shows a good percentage of variance explained by the full model, for
> about 40% of the patients the variance explained by the full model is < 10%.
> This did not change by increasing the size of the VOIs, or by choosing a
> stronger peak. For these participants the variance explained in specific VOIs
> is low even for VOIs with >30 active voxels.
>
> 1. Do you have any suggestion for what we can do to avoid excluding these
> participants?
> 2. In principle - is the variance explained by the full model always greater
> than that explained by any reduced model?
>
> Thanks a lot
> Tali Bitan
> University of Haifa
|