Sorry, ambiguous language in my post. By whole volume I indeed meant
4D mean. As you point out, scaling by individual volume means is a
dangerous thing - was only really ever used for PET I think.

-Tom





On Wed, Dec 2, 2009 at 11:03 AM, Christian F. Beckmann
<[log in to unmask]> wrote:
> Hi Tom
>
> Just a quick note: by default, FEAT does _not_ scale to the whole volume mean but to the whole 4D mean (mean across all vols and across all voxels). The first type of scaling we actually discourage (it's switched off in the GUI) while the second is necessary to allow for between run/subject comparison.
> cheers
> Christian
>
> On 2 Dec 2009, at 09:39, Tom Johnstone wrote:
>
>> I might just add to Steve's comments on this, based on when we looked
>> at test-retest reliability and group analysis sensitivity in the
>> amygdala, which is relevant because that brain region is quite
>> impacted by partial volume effects. We looked at raw betas (unscaled,
>> meaning they were not scaled in the way that FSL scales to the whole
>> volume mean), % signal change calculated voxelwise, and z-scores
>> calculated voxelwise.
>>
>> We found % signal change to give the best test-retest reliability and
>> comparable sensitivity to raw betas. z-scores were about as sensitive
>> but not quite as reliable. Raw betas (which were not reported in the
>> original paper) were less reliable, but showed if anything greater
>> sensitivity than % signal change.
>>
>> Another unpublished analysis we did a bit later was to create a
>> region-based % signal change; we divided each voxel's signal change by
>> a local neighbourhood median of baseline signal (easy to do: just run
>> a median filter over the baseline signal estimate and use that to
>> perform the calculation), which we thought would be less susceptible
>> to partial volume effects than voxelwise calculation of the %. This
>> indeed turned out to be just as reliable and sensitive, but gave us
>> more confidence that had our amygdala coverage not been quite as good
>> it would have been more robust to partial volume effects.
>>
>> I suspect that FSL's PE which are scaled to whole brain mean would be
>> similar for standard group mean analysis, though if you're interested
>> in individual differences in activation in a specific region of the
>> brain, then the risk is that any measured differences might be
>> influenced by global differences in baseline signal in other areas of
>> the brain (though using a local baseline means that individual
>> differences in % signal change might be due to local differences in
>> baseline signal).
>>
>> As always then, the optimal solution depends on your specific research
>> question and data set.
>>
>> Hope this is helpful.
>>
>> -Tom
>>
>>
>> On Wed, Dec 2, 2009 at 8:31 AM, Stephen Smith <[log in to unmask]> wrote:
>>> Hi,
>>> There's arguments both ways, some theoretical and some practical.  Some of
>>> these arguments include:
>>> - PE could be more 'robust' because it doesn't involve dividing by the voxel
>>> mean intensity, which could be poorly conditioned, e.g. in partial-volumed
>>> voxels.
>>> - %change is more satisfying as it seems like a more quantitative measure
>>> - However, in the case of partial voluming, %change isn't actually any
>>> better because the result depends not just on the BOLD effect but also on
>>> the partial volume fraction, so neither measure is particularly
>>> objective/quantitative
>>> - If I wanted to choose which measure was better I probably wouldn't depend
>>> on any of these arguments but would probably test empirically which seemed
>>> more robust, sensitive and reliable.
>>> Cheers.
>>>
>>>
>>>
>>>
>>> On 30 Nov 2009, at 11:37, Hilary Watson wrote:
>>>
>>> Hiya Steve
>>>
>>> Thanks for all you're help I really appreciate it.  I was wondering, if it's
>>> not too much trouble, if you could explain to me why PE is a better measure
>>> for me to use (just in case this comes up in my viva) - I have read a few
>>> bits and pieces (attached) that suggest that PE isn't suitable.
>>>
>>> Apologies for taking up so much of your time
>>>
>>> Cheers
>>>
>>> Hilary
>>>
>>>
>>>
>>> Stephen Smith <[log in to unmask]> 30/11/2009 10:08:31 >>>
>>>
>>> Hi - I would say that there is no need to worry about converting to %
>>> signal change in this scenario, and that if anything the PE values are
>>> probably slightly more 'robust'.
>>> Cheers.
>>>
>>>
>>> On 30 Nov 2009, at 09:55, Hilary Watson wrote:
>>>
>>> Hiya Steve
>>>
>>> Thanks so much for getting back to me - I did a little extra
>>>
>>> research after
>>>
>>> emailing last night and it turns out I was looking in an older
>>>
>>> version of
>>>
>>> the FSL manual so I guess point three doesn't really apply.
>>>
>>> I just wanted to ask - in light of the fact that my experiments are
>>>
>>> event-related, the design is the same across runs and participants
>>>
>>> (in that
>>>
>>> I make the same contrasts and efficiency of these are quite similar
>>>
>>> although
>>>
>>> timings are not because presentation of items is random and coded to
>>>
>>> subsequent memory accuracy) and I get different things for % signal
>>>
>>> change
>>>
>>> and PEs - would it be incorrect to use parameter estimate values to
>>>
>>> analyse
>>>
>>> my data sets?
>>>
>>> Cheers
>>>
>>> Hilary
>>>
>>> Stephen Smith <[log in to unmask]> 30/11/2009 08:18 >>>
>>>
>>> Hi,
>>>
>>> On 29 Nov 2009, at 22:48, Hilary Watson wrote:
>>>
>>> Hi FSL users
>>>
>>> I have run a total of three fMRI studies for my PhD and I am in the
>>>
>>> process
>>>
>>> of re-analysing to write them up.
>>>
>>> A lot of my work requires extraction of effect sizes in different
>>>
>>> ROIs and
>>>
>>> then running stats on these values - generally I have run functional
>>>
>>> localisers to create functional ROIs and then queried subsequent
>>>
>>> memory
>>>
>>> effects across a variety of conditions against an active baseline
>>>
>>> condition
>>>
>>> within these.
>>>
>>> For example
>>>
>>> objects later remebered - active baseline
>>>
>>> objects later forgotten - active baseline
>>>
>>> I orginally extracted the PEs (betas) using Featquery, however I have
>>>
>>> recently come across some literature that suggests that you
>>>
>>> shouldn't use PE
>>>
>>> for your stats, instead you should use % signal change (that said I
>>>
>>> have
>>>
>>> seen plenty of recent published papers in decent journals that have
>>>
>>> used FSL
>>>
>>> PEs).
>>>
>>> First question is whether I can stick with PEs for my stats?
>>>
>>> Sure - in many cases there's little difference anyway. All data is
>>>
>>> normalised across the entire 4D dataset to have a fixed mean value, so
>>>
>>> as long as comparable designs are used for all subjects there won't be
>>>
>>> much difference.  Either choice should be acceptable, and Featquery
>>>
>>> makes it easy to do either, by turning the relevant button on or off.
>>>
>>> Secondly I have also extracted % signal change for my data and have
>>>
>>> already
>>>
>>> seen there is not a simple one to one mapping to PE.  For two of my
>>>
>>> data
>>>
>>> sets the numerical patterns are pretty much the same but for another
>>>
>>> it only
>>>
>>> appears to have an effect on one of the contrasts I am interested in
>>>
>>> looking
>>>
>>> at.  Surely if the baseline used to covert these %s is the same
>>>
>>> across
>>>
>>> conditions (within participants) why would these conversions have a
>>>
>>> greater
>>>
>>> effect on one contrast?  Obviously if it ok for me to use PE then
>>>
>>> this is so
>>>
>>> much of an issue
>>>
>>> This can happen, for example where a voxel has a fairly different mean
>>>
>>> intensity to the brain as a whole (e.g. if it's on the edge of the
>>>
>>> brain and hence partial-volumed between grey-matter and non-brain
>>>
>>> matter) then PE and %change will be more different. Also, for some
>>>
>>> contrasts the height of that contrast's 'effective regressor' (see our
>>>
>>> NeuroImage paper on design efficiency) can be different from what you
>>>
>>> might expect, particularly for more complex designs and differential
>>>
>>> contrasts.
>>>
>>> Thirdly, I have checked the FSL Feat manual and it says that you
>>>
>>> cannot use
>>>
>>> the 'covert PE into % signal change' option in Featquery for event
>>>
>>> related
>>>
>>> designs (which all of my experiments are) as this assumes the height
>>>
>>> of the
>>>
>>> waveform is 1, which is only appropriate for blocked designs.  Is
>>>
>>> there a
>>>
>>> simple way to calculate % signal change using my PEs - say a formula
>>>
>>> and
>>>
>>> somewhere I can extract a waveform height value?
>>>
>>> Are you sure the manual says that? I'm not sure it does but I may be
>>>
>>> missing something.  For most event-related designs even the simple
>>>
>>> conversion in Featquery is accurate enough; if you have a concern then
>>>
>>> have a look at J Mumford's website/tool that looks at this issue more
>>>
>>> thoroughly.
>>>
>>> Cheers.
>>>
>>>
>>> I am desperately confused and there is quite a lot riding on this so
>>>
>>> any
>>>
>>> help at all will be greatly appreciated.  Let me know if you need me
>>>
>>> to
>>>
>>> clarify anything.
>>>
>>> Thanks in advance
>>>
>>> Hilary
>>>
>>>
>>>
>>> ---------------------------------------------------------------------------
>>>
>>> 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
>>> ---------------------------------------------------------------------------
>>>
>>>
>>>
>>> ---------------------------------------------------------------------------
>>> 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
>>> ---------------------------------------------------------------------------
>>>
>>>
>>>
>>>
>