Dear Klaas,
thank you for your help. My last question is if I interpret the output
of the spm_BMS function correctly. In my understanding, the alpha
values are the posterior probabilities of each model, but
unnormalized, so if I normalize them with their sum, I get the
posterior distribution on my model space. Is that correct?
Thank you,
Mihály
2010/7/14 Klaas Enno Stephan <[log in to unmask]>:
> Dear Mihaly,
>
> Are you referring to a t-test on the log evidences? This is the frequentist
> random effects model we describe in the group BMS paper. As you say, it is
> much simpler than the Bayesian counterpart. However, it is less robust to
> outliers (as we show in the paper) and the inference it provides is somewhat
> different (p values instead of exceedance probabilites or posterior model
> probabilities). Both approaches can be used.
>
> Best wishes
> Klaas
>
>
> ________________________________
> Von: Bányai Mihály <[log in to unmask]>
> An: Klaas Enno Stephan <[log in to unmask]>
> Gesendet: Dienstag, den 13. Juli 2010, 15:16:10 Uhr
> Betreff: Re: [SPM] variational BMS
>
> Dear Klaas,
>
> thank you for the reference, that explains the theoretical idea behind
> the code, however it still seems overly complicated to me. What do you
> think, couldn't I directly compare the free energy values that I
> already have? Since those are approximations of the log-evidence, the
> ranking they give should be the same, and the magnitude of the
> difference between those is also informative even if I won't have a
> complete posterior distribution that way. How "worse" or less reliable
> do you think this simpler approach would be?
>
> Thank you,
> Mihály
>
> 2010/7/13 Klaas Enno Stephan <[log in to unmask]>:
>> Dear Mihaly,
>>
>> This is the paper you are looking for:
>>
>>
>> http://www.fil.ion.ucl.ac.uk/spm/doc/papers/Stephan_NeuroImage_46_1004_2009.pdf
>>
>> Best wishes
>> Klaas
>>
>>
>> ________________________________
>> Von: Bányai Mihály <[log in to unmask]>
>> An: [log in to unmask]
>> Gesendet: Dienstag, den 13. Juli 2010, 11:23:13 Uhr
>> Betreff: [SPM] variational BMS
>>
>> Dear SPM developers,
>>
>> I plan to use Bayesian model selection to rank my DCM models for a
>> group study. I read the code in spm_BMS.m, but it's not clear to me
>> why we need a variational approximation for the model probabilities if
>> the input of the algorithm is the estimate of the log-evidence. Is
>> there a documentation available that describe the theoretical
>> considerations behind this method? I use spm8.
>>
>> Thank you,
>> Mihály Bányai
>>
>>
>
>
|
