Dear Burkhard,
I think your confusion arises from misunderstanding of what DCM.R is.
It is the residual rather than the original data. So H is the model
and H+R is the original data. In my formula I divide the sum of
squares of the residual by the sum of squares of the original data and
if the fit is perfect RV will be 0 and if the model doesn't fit
anything H will be 0 and RV will be 1.
Best,
Vladimir
On Fri, Sep 30, 2011 at 10:09 AM, Burkhard Maess <[log in to unmask]> wrote:
> Hi Vladimir,
>
> I am confused by your suggestion for the computation of RV.
> Does this RV value really represent a useful Data-Model-comparison?
>
> If data and model have identical values - then RV = 1/4 = 0.25, doesn't it?
> If the model is zero - than the RV = 0.
> If the model is large compared to the data, then RV = 1
>
> However, a value RV=0.25 can not be taken as an identifier for the
> data-model-equivalence, because
> there is a infinite manifold of vector sums which has a length twice as
> large as the model.
>
> In short, I think there are certainly different expressions to compute a
> single number for the data-model-comparison, but they always should contain
> a '-'.
> Data minus Model becomes very small if they both are about the same.
> Data plus Models never reaches a special value - IMHO.
>
> all the best,
> Burkhard
>
>
>
> Vladimir Litvak wrote:
>>
>> Dear Fahimeh,
>>
>> I'm not sure your definition of RV is correct. The one I'm using is:
>>
>> RV( = sum(spm_vec(DCM.R).^2)./sum((spm_vec(DCM.H)+spm_vec(DCM.R)).^2);
>>
>> I've recently looked at a large number of models of different subjects
>> and there is no clear threshold to separate the ones with reasonable
>> fit from the ones without. If your RV is below 0.2 this is very good
>> but also there were some examples with RV of 0.5 or above where I
>> wouldn't say that DCM failed to fit the data and model comparison is
>> meaningless.
>>
>> So there is no way around of just visually looking at the 'ERPs
>> (mode)' and 'Response' displays in DCM results and seeing whether the
>> model prediction captures what you think are the main features of the
>> data or not.
>>
>> Best,
>>
>> Vladimir
>>
>> On Thu, Sep 29, 2011 at 2:02 PM, Fahimeh Mamashli <[log in to unmask]>
>> wrote:
>>
>>>
>>> Dear DCM experts,
>>>
>>> Thanks for your attention. I have two questions concerning correspondence
>>> between data and model. one well established measure is the R2 value.
>>>
>>> 1) if a model converged, does it mean that its R2 measure is high enough
>>> to explain the data? e.g. being larger than 80%, independence of SNR?
>>>
>>> R2=1-(sum(data-model).^2)/(sum(data-mean(data)).^2)
>>>
>>> 2) would you suggest other measures except R2 to do the data-model
>>> comparison?
>>>
>>> currently I have used R2 measures as a function of time, based on the
>>> modes as follows:
>>>
>>> % observed mode
>>> datamode=DCM.H{1,1};
>>>
>>> % predicted mode
>>> modepred=DCM.R{1,1};
>>>
>>>
>>> % R2 measure
>>> surat=(sum((datamode-modepred).^2));
>>> makhraj= (sum((datamode-mean(datamode)).^2));
>>>
>>> gofhigh=1-(surat./makhraj);
>>>
>>>
>>>
>>> I am looking forward to your answer.
>>>
>>> Kind regards,
>>> Fahimeh
>>> ------------------------
>>> PhD student
>>> Max Planck Institute for Human Cognitive and Brain Sciences
>>> Stephanstraße 1A
>>> 04103 Leipzig
>>> Germany
>>>
>>> Tel: +49 341 9940 - 2570
>>>
>>>
>
> --
> ------------------------------------------------------------
> Dr. Burkhard Maess
> Max Planck Institute for Human Cognitive and Brain Sciences
> Stephanstr. 1a, P.O. Box 500355, D-04303 Leipzig
> Aussenstelle Bennewitz, phone/fax: +49(3425)887525-26/-11 mail: maess 'at'
> cbs.mpg.de, http://www.cbs.mpg.de
> ------------------------------------------------------------
> Please do not attach files which may contain executable code, e.g. .rtf is
> preferred over .doc(x)
>
>
>
|