Print

Print


Dear Qiyuan Wu,

this paper describes the fixed-effects analysis (Group Bayes Factors)
and the new random-effects method (and its variables) now implemented
in SPM8b:
http://www.ncbi.nlm.nih.gov/pubmed/19306932?ordinalpos=1&itool=EntrezSystem2.PEntrez.Pubmed.Pubmed_ResultsPanel.Pubmed_DefaultReportPanel.Pubmed_RVDocSum

The variables stored in BMS.DCM correspond to what is plotted on the
SPM graphics window,
but it also includes some additional information (please check the
Help text in the batch editor window for further information):

In BMS.DCM.ffx (Fixed-effects analysis):
   '.F' is the Free Energy (or log-model evidence) for each model/subject
   '.P' is the Posterior Probability for each model (as plotted)
   '.SF' contains the sum of the log-evidence (F) for each model
   '.data' contains the path to the DCM files (for each
subject/session and model) that were used in this analysis.

In BMS.DCM.rfx (Random-effects analysis):
   '.F' is the Free Energy (or log-model evidence) for each model/subject
   '.SF' contains the sum of the log-evidence (F) for each model
   '.data' contains the path to the DCM files (for each
subject/session and model) that were used in this analysis.
   '.exp_r' is the expected posterior probability for each model
   '.xp' is the exceedance probability for each model
   '.alpha' are the estimated parameters of the Dirichlet distribution
for each model

Best Regards,
Maria


2009/3/28 Ԫ <[log in to unmask]>:
> Dear SPMers,
>
>
>
> I have 2 questions about DCM in SPM8b. Hope that you can help me.
>
>
>
> 1. How to report the results of comparisons of DCMs?
>
>
>
> In SPM5, it is advised to report group Bayesian factors (GBF) to show which
> model is the best. However, in SPM8b, the model comparisons use the negative
> free energy (F) as an approximation to the log evidence. In the "BMS.mat",
> there are 5 variables: BMS.DCM.rfx.SF, BMS.DCM.rfx.F, BMS.DCM.rfx.alpha,
> BMS.DCM.rfx.exp_r, BMS.DCM.rfx.xp and BMS.DCM.rfx.data. What exactly the
> first four variables represent? Or is there any paper about them?
>
>
>
> 2. The results of "average" function of DCM in SPM8b.
>
>
>
> After I got the best model after comparisons, I averaged 13 models of
> subjects and got an averaged DCM model, which contains average parameters
> and their post probabilities, but not averaged time series. Can I report the
> average parameters and their post probabilities as my final results?
>
>
>
> I also found that the average parameters and post probabilities are
> different from results of classical statistical analysis outside SPM on the
> arithmetic (non-Bayesian) means of the subject-specific connection
> parameters (n=13) with SPSS (Analysis/ Compare means to 0). For example, the
> average DCM.C is 1.3194 while the mean of DCM.Cs of 13 subjects in SPSS is
> only 0.3870. I examined all the DCM.Cs of each subject, I found that they
> were all less than 0.5. What arithmetic in the function of "Average"? And
> what result is better to be reported?
>
>
>
>
>
> Thanks a lot for advance.
>
> --
> Best Regards,
> Qi-yuan WU, [log in to unmask]
>
>