Hi Mark,
is there a lower thresh you'd recommend for the tool? I.e. would 95%, for example, still be ok?
All the best-
Andreas
________________________________________
Von: FSL - FMRIB's Software Library [[log in to unmask]] im Auftrag von Mark Jenkinson [[log in to unmask]]
Gesendet: Dienstag, 6. Oktober 2009 13:24
An: [log in to unmask]
Betreff: Re: [FSL] cutoffcalc
Hi,
It is a simple method and does not take account of autocorrelation
except for the lower limit which is put there precisely for reasons
of prewhitening.
I imagine that there may be ways of incorporating autocorrelation
estimation as well, but I suspect that the differences would be small
and that it is probably not worth the effort to do.
All the best,
Mark
On 6 Oct 2009, at 10:41, wolf zinke wrote:
> Hi,
>
> Thanks for the explanations. Just a follow up. It is just based on
> the temporal structure of the model, but does not take the noise and
> autocorrelation of the data into account. The paper by Smith et al.
> ('Meaningful design and contrast estimability in FMRI, NeuroImage
> 2007) did comment on the interaction between high-pass filtering and
> the autocorrelation. Is there a simple method to use both
> informations (retained power and sufficient data to do a valid
> prewhitening) for the determination of an optimal HP cutoff.
> Otherwise it is most likely that for event related designs the lower
> limit is always given by the shortest period which one dares to
> apply without affecting the AR modelling to much. Is there a
> comparable method available to obtain a cutoff values based on AR
> modelling?
>
> thanks,
> wolf
>
> On 10/06/2009 09:55 AM, Mark Jenkinson wrote:
>> Hi,
>>
>> It is quite a simple tool so doesn't really have documentation yet.
>> Essentially it tried different filter cutoffs and calculates the
>> percentage
>> of retained power in the EVs of the design matrix. Once this goes
>> below
>> a threshold (default is 99%) then it chooses the last value for
>> which it
>> passed the threshold. However, it also has a lower limit on the
>> period
>> of the cutoff (default 90s) to avoid it being too aggressive and
>> removing
>> information that the pre-whitening needs to work.
>> Note that the output is in seconds - just paste it into the FEAT GUI.
>>
>> You are right about running it on design matrices with confound
>> regressors
>> in. It does not know about this and so doesn't do any adjustment.
>> You need
>> to save a design without these (no need to run anything in FEAT -
>> just save
>> the design from the GUI and use the *.mat file as input). Running
>> it on such
>> a subset (no confounds) is the best approach to use.
>>
>> All the best,
>> Mark
>>
>>
>> On 5 Oct 2009, at 16:46, wolf zinke wrote:
>>
>>> Hi,
>>>
>>> I came across the recently added and quiet handy tool cutoffcalc.
>>> Is there any documentation/reference how it determines the minimal
>>> period for the high pass filter? If I understand it correctly, it
>>> uses all regressors of the design matrix for the calculation.
>>> Hence, if I run it on the 'pure' design matrix I'll get different
>>> results than running it on the feat design matrix with motion and
>>> confound regressors added. I guess it is reasonable to apply
>>> cutoffcalc only on the regressors of interest. Do I need to
>>> generate a reduced model for this purpose or can I tell cutoffcalc
>>> to use only specified regressors for the calculation (or is it
>>> completely stupid to use a regressor subset)
>>>
>>> Thanks for all the help,
>>> wolf?
>>>
>
|