Hello,
I think I realize now why there is such a significant problem with
this approach - because DR uses a multiple regression model, the
signal from your 'matching components' will also be affected by the
other components in your GLM, which likely do not match appropriately
between groups. Sorry for not picking that up earlier.
Would it be valid to run a simple regression using only the matching
component of interest as the regressor in the DR pathway? Sorry if
these are inane questions.
My issue is similar to Angela's in that I have more patients than
controls. Logically, it made sense to me to define the RSN within each
group, as spatially it may be somewhat different due to pathology, and
this would maximize sensitivity.
On Mon, Jun 20, 2011 at 9:43 AM, <[log in to unmask]> wrote:
> Hi Christian, Steve -
>
> Thanks for the reply - so it is entirely invalid then to compare z-stat images from two different dual regression streams? Beyond the sketchiness of the matching process?
>
> Graeme
> Sent from my BlackBerry device on the Rogers Wireless Network
>
> -----Original Message-----
> From: "Christian F. Beckmann" <[log in to unmask]>
> Sender: FSL - FMRIB's Software Library <[log in to unmask]>
> Date: Mon, 20 Jun 2011 09:36:02
> To: <[log in to unmask]>
> Reply-To: FSL - FMRIB's Software Library <[log in to unmask]>
> Subject: Re: [FSL] melodic, sample composition
>
> Hi Graeme,
>
> That's possible, though you;re somewhat defeating the point of a joint analysis in that you end up having to match components across runs. Once you want to do Dual Regression you will have to decide on one set of components anyways...
>
> hth
> Christian
>
>
>
> On 18 Jun 2011, at 13:00, SUBSCRIBE SPM Anonymous wrote:
>
>> I have a related question - what about running separate ICAs on the pt and ctrl groups, identifying 'matching' components, then doing cross subject stats on the z-stat outputs of the respective dual regressions?
>> Sent from my BlackBerry device on the Rogers Wireless Network
>>
>> From: Stephen Smith <[log in to unmask]>
>> Sender: FSL - FMRIB's Software Library <[log in to unmask]>
>> Date: Sat, 18 Jun 2011 08:50:30 -0400
>> To: <[log in to unmask]>
>> ReplyTo: FSL - FMRIB's Software Library <[log in to unmask]>
>> Subject: Re: [FSL] melodic, sample composition
>>
>> Hi - I think it's mostly ok.
>> Cheers.
>>
>>
>>
>> On 18 Jun 2011, at 08:36, Angela Favaro wrote:
>>
>>> Hi Stephen
>>>
>>> Thank you for clarifying that it is possible to just use the controls for
>>> the group ICA and then derive all subjects' RSNs from a "normal" set of
>>> maps. But now another question would be: can the same control subjects
>>> (already used for the group ICA) also be used in the subsequent dual
>>> regression together with all other subjects that were not included? Or
>>> would this bias the final cross-subject between-group modelling?
>>>
>>> Thank you for any help
>>>
>>> Cheers
>>> Angela
>>>
>>>
>>>
>>>
>>>> Hi -
>>>
>>>
>>>> Firstly - the set of subjects used for the group-ICA can be a different
>>> set than you then use in the dual-reg.
>>>
>>>
>>>> For the concat-mode group-ICA, you *might* find some reviewers want you
>>> to use matched >numbers - however, personally I wouldn't worry - after
>>> all in the null case (no group differences) >the group sizes would not
>>> make any difference. In the case of massive group differences you
>>>> might want to just use the controls for the group ICA, so that you can
>>> then derive all subjects' >RSNs from a "normal" set of maps.
>>>
>>>
>>>> You can use all the subjects in the dual-regression and final
>>> cross-subject between-group >modelling, as that will be unbiased by group
>>> sizes.
>>>
>>>
>>>> Cheers.
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>>
>>> On 3 Feb 2011, at 14:38, Angela Favaro wrote:
>>>
>>>
>>> Dear FSL experts,
>>> I have a methodological question about temporal-concatenated ICA. I have
>>> performed a TC-ICA in a sample of 26 subjects with a psychiatric illness,
>>> 15 subjects in remission from the same illness and 24 controls. If the
>>> hypothesis is to find some network dysfunction correlated with the risk of
>>> this illness, this sample is biased by an excess of 'cases' in comparison
>>> to 'controls'. Is there some rule in the sample composition when using
>>> TC-ICA (and then dual regression to compare groups)? do I need to use
>>> samples with the same number of cases/controls?
>>>
>>> In other words, how much the composition of the sample can affect the
>>> 'quality' of extrapolated networks? Is it correct to perform TC-ICA in
>>> homogeneous samples (only cases, only controls) to do a 'descriptive'
>>> analysis of network functioning/variation in a pathological sample?
>>>
>>> Thank you for any advice!
>>>
>>> Angela
>>>
>>>
>>>
>>>
>>>
>>> Angela Favaro, MD, PhD
>>> Psychiatric Clinic
>>> Department of Neurosciences
>>> via Giustiniani 3
>>> 35128 Padova
>>> Italy
>>>
>>>
>>>
>>>
>>>
>>> ---------------------------------------------------------------------------
>>> Stephen M. Smith, Professor of Biomedical Engineering
>>> Associate Director, Oxford University FMRIB Centre
>>>
>>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
>>> +44 (0) 1865 222726 (fax 222717)
>>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>>> ---------------------------------------------------------------------------
>>>
>>
>>
>> ---------------------------------------------------------------------------
>> Stephen M. Smith, Professor of Biomedical Engineering
>> Associate Director, Oxford University FMRIB Centre
>>
>> FMRIB, JR Hospital, Headington, Oxford OX3 9DU, UK
>> +44 (0) 1865 222726 (fax 222717)
>> [log in to unmask] http://www.fmrib.ox.ac.uk/~steve
>> ---------------------------------------------------------------------------
>>
>>
>>
>
--
Graeme C. Schwindt, HBSc
MD/PhD Student
University of Toronto
[log in to unmask]
|