Hi David,
Yes, absolutely, it should have been:
beta = mean_random_effects_var1
beta_outlier = mean_outlier_random_effects_var1
Cheers, Mark.
----
Dr Mark Woolrich
EPSRC Advanced Research Fellow University Research Lecturer
Oxford University Centre for Functional MRI of the Brain (FMRIB),
John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
Tel: (+44)1865-222782 Homepage: http://www.fmrib.ox.ac.uk/~woolrich
On 11 Jun 2009, at 12:58, David Paulsen wrote:
> Mark,
>
> To be clear, should
> beta = mean_random_effects_var1
> beta_outlier = mean_outlier_random_effects_var1
> ?
>
> The previous post has it the other way around and it seemed odd that
> the
> beta_outlier file did not have 'outlier' in the title whereas the
> non-outlier file did.
>
> Thanks,
> david
>
>
> On Sat, 6 Jun 2009 15:45:44 +0100, Mark Woolrich <[log in to unmask]
> >
> wrote:
>
>> David,
>>
>> Just to supplement my other reply which was:
>>
>>
>>> Hi David,
>>>
>>> Apologies for the delay in replying.
>>>
>>> The weighting that is effectively applied to each subject is not
>>> currently stored in the feat stats directory. It is a function of:
>>>
>>> S: the lower-level variance (var_filtered_func_data in the feat dir)
>>> beta: the non-outlier random-effects variance
>>> (mean_outlier_random_effects_var1 in the stats dir)
>>> beta_outlier: the additional random-effects variance for outliers
>>> (mean_random_effects_var1 in the stats dir),
>>> probout: the probability of being an outlier (prob_outlier1 in the
>>> stats dir)
>>>
>>> Note that the 1's in the file names indexes the variance group.
>>> Then:
>>>
>>> weighting(i) = 1/(sqrt(A(i))
>>> where A(i)=(beta(i)+S(i))*(1-probout)^2 +(beta(i)+beta_outlier(i)
>>> +S(i))*probout.^2;
>>> and i indexes the subject.
>>>
>>> We'll have a think about perhaps putting this in as an output in a
>>> future release.
>>
>> Strictly, the relevant eqn should have been:
>> A(i)=(beta+S(i))*(1-probout(i))^2 +(beta+beta_outlier
>> +S(i))*probout(i).^2;
>> apologies for that.
>>
>> So this equation applies at any level in the hierarchy (e.g. 2nd or
>> 3rd) above the first level when random effects with outlier inference
>> is being used.
>> If you want the weighting that is effectively being used without
>> outlier inference then naturally this corresponds to probout(i)=0,
>> i.e:
>> A(i)=beta+S(i);
>>
>> Note that these weightings correpond to flame1. Weightings for ols
>> would be the same but with S(i)=0. If flame2 is being used then the
>> weighting is integrated over the uncertainty in the random effects
>> variance and so can not be summarised with a single equation.
>>
>> Cheers, Mark.
>>
>> ----
>> Dr Mark Woolrich
>> EPSRC Advanced Research Fellow University Research Lecturer
>>
>> Oxford University Centre for Functional MRI of the Brain (FMRIB),
>> John Radcliffe Hospital, Headington, Oxford OX3 9DU, UK.
>>
>> Tel: (+44)1865-222782 Homepage: http://www.fmrib.ox.ac.uk/~woolrich
>>
>>
>>
>>
>> On 4 Jun 2009, at 20:15, David Paulsen wrote:
>>
>>> Dear List,
>>>
>>> As I understand it, there are two ways in which individual's second
>>> level
>>> parameter estimates are differentially weighted in third level
>>> statistics:
>>> based on variance of second and first level PEs and if outlier
>>> deweighting
>>> is used.
>>>
>>> I would like to know the weights that were used for each subject at
>>> the
>>> third level, in both cases. Is there anyway to retrieve this
>>> information?
>>>
>>> Please & Thank You,
>>> David
>>>
>
|