Dear Vladimir,
Yes, I was wrong about inputs because I was using the drop-down menu at the bottom of DCM GUI and I thought that it is the estimated input. Actually I checked the two models with BMS again and I am not sure why I thought they were the same. Actually it shows that 40ms is better than 240 which I think is because I am using EEG signals in [0 700]ms range and the algorithm just predicts the signal after onset and predicts the signal before onset as zero which increases the error as the real signal before onset is not zero.
But about the source reconstruction, as my actual experiment is with other stimuli than just tones, I need to find out if any other source (other than just the ones which usually get active for tones) is also involved in the response. I was just trying to test MSP with tones and if it passed I could be more confident with my own stimuli.
And for other users reference, DCM.Ep is posteriors and DCM.M.pE is priors.
Best,
Pegah
-----Original Message-----
From: Vladimir Litvak [mailto:[log in to unmask]]
Sent: 15 August 2012 12:03
To: Tayaranian Hosseini P.
Cc: [log in to unmask]
Subject: Re: [SPM] DCM and source reconstruction results
Dear Pegah,
Since you are using a well-established paradigm I think it is perfectly OK to start with the established DCM model and not do source reconstruction for your data. Regarding the onset time, it is optimised but there are some constraints and also there can be different local maxima depending on the prior. Also perhaps you didn't look in the right place as the prior and posterior values are stored separately (look in DCM.M.pE). I find it hard to believe that inputs at 40 and 240 ms would yield the same fit without any optimisation.
Best,
Vladimir
On Wed, Aug 15, 2012 at 11:51 AM, Tayaranian Hosseini P.
<[log in to unmask]> wrote:
> Dear Vladimir,
>
>
>
> Thank you very much for your last email. It helped me a lot. Now, I
> have these new concerns.
>
>
>
> I have recorded one data similar to the one in DCM papers (480 trials
> of 2000Hz and 120 trials of 1000Hz tones) and applied source
> reconstruction on it. The MSP source reconstruction gives me some
> source locations as the most active ones (which are by the way not
> very reasonable). I applied DCM on the data using these source
> positions and the predicted modes are very similar to the real ones.
> Then I applied DCM on the same data but with the source positions that
> are used in DCM papers and SPM manual and again it gave me very good
> prediction. But when I compare the two models using BMS, the second one gives me much higher log-evidence than the first one.
>
> So, I decided that I cannot trust MSP much and have to find my source
> positions using some other method.
>
>
>
> Also, I had read in DCM papers that the input delay is being optimised
> too but when I tested this with different input times, it did not optimise it.
> My stimulus onset is at 200ms so I set the DCM input around 240ms.
> After estimation, the input remained around 240ms. Again, I changed
> the input time to 40ms (which means even before the onset) and after
> estimation, the input was around 40ms and the evidence of both models
> (for 40ms and 240ms) where very close.
>
> How should I interpret these results?
>
>
>
> Best,
>
> Pegah
>
>
>
> --------------------------------------------------------
>
> Pegah Tayaranian Hosseini
>
> PhD Student
>
> Room 4077, Tizard building (13)
>
> Institute of Sound and Vibration Research
>
> University of Southampton, SO17 1BJ, UK
>
>
>
> Tel: 023 8059 2850
>
> email: [log in to unmask]
>
>
|