Dear Soha,

>For example in my models I am getting big negative DCM.F values (up to -1200),

The absolute value of F is unimportant. What matters is how different F’s are for different models.

>DCM. Ce values are above 0.1

This means the error SD is sqrt(0.1)=0.316. This can only be interpreted in relation to the size of the signal.

Inferences about model parameters are based on the posterior density – these take into account both signal and noise.

> In DCM10 version, BMS algorithm compare the models based on AIC or BIC criteria or it started to take free energy difference into account?

We recommend using the Free Energy. See our recent paper: Neuroimage. 2012 January 2; 59(1): 319–330.

This paper also looks at model selection as a function of SNR.

Best wishes,

Will.

Sent: 06 February 2012 17:19
Subject: Re: [SPM] Signal to Noise Ratio in DCM

Hello :

I have few questions about free energy and error variance in the data.
How big  the DCM.F value can be, the same for the error variance. For example in my models I am getting big negative DCM.F values (up to -1200), also DCM.Ce values are above 0.1.  Does that mean the data is very noisy?  Also is the difference in log evidence should be 3 minimum or the difference in DCM.F values?
In DCM10 version, BMS algorithm compare the models based on AIC or BIC criteria or it started to take free energy difference into account?

Thanks, I appreciate your help, and any hints you can give me regarding model selection.

Best regards,

Soha

Date: Fri, 3 Feb 2012 20:26:40 +0000
Subject: Re: [SPM] Signal to Noise Ratio in DCM

Dear Mobin,

DCM can handle any amount of signal to noise. To get an idea of how much noise can be tolerated for any particular question, one would do simulations with increasing levels of noise and simulate the model comparison of interest.  The noise level that renders the difference in free energy (DCM.F) less than 3 is the level of noise at which your data are too noisy to provide evidence for one model over another. I think Klaas wrote an SPM routine to generate simulated data, given a model specification (a DCM structure).

I hope this helps.

With very best wishes,

Karl

Sent: 03 February 2012 19:54
To: Friston, Karl
Cc: Stephan, Klaas; Penny, William
Subject: Signal to Noise Ratio in DCM

Dr. Friston Dr. Stephan and Dr. Penny

I asked the SPM list  about how much Signal to noise ratio

can DCM handle when we have cognitive emotional tasks and nobody