Hello All,
Any updates on how to get around this problem? Maybe passing flameo or randomise a 4d voxelwise ev that's 1 on all the missing voxels in each subject would get you the right parameter estimates. But what can you do to get accurate z statistics (with the dof varying across voxels)?

Thank you in advance,
Connor

On Thu, 7 Apr 2011 14:33:56 -0700, Kirstie Whitaker <[log in to unmask]> wrote:

>So helpful!! Thank you for all the links! And for your comments.
>
>Kx
>
>On 7 April 2011 13:52, Michael Harms <[log in to unmask]> wrote:
>
>>
>> Hi Kirstie,
>>
>> Unfortunately, there isn't an easy solution to your problem.
>>
>> See this thread, and its follow-up posts, in which I asked the same basic
>> question:
>>
>>
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1006&L=FSL&P=R41688&1=FSL&9=A&J=on&d=No+Match%3BMatch%3BMatches&z=4
>>
>> (or search for "missing voxel data going to 2nd level" thread if that link
>> doesn't work).
>>
>> We have some local software here at WU for doing GLM's that allows for
>> missing voxels by letting the dof vary across voxels, but neither FEAT nor
>> SPM currently support that as an alternative analysis approach -- with both
>> packages it is either all or none. So, when using higher-level FEAT we have
>> been looking at how much spatial coverage we lose, and then making an
>> empirical decision about who to exclude, so as to achieve a balance between
>> maximal spatial coverage and maximal number of subjects.
>>
>> Note that if you go with an ROI approach, you still need to keep track of
>> the number of voxels with data within the ROI, and you'll have to make a
>> decision regarding the minimal number of voxels that must be present within
>> the ROI such that you're willing to include that subject's ROI value in the
>> analysis -- e.g., are you going to include an ROI if it has only 10% of the
>> voxels of the "full" ROI? 20%? 30%? (Such is the bane of missing data --
>> there is no simple, universal work-around).
>>
>> Also note that 'featquery' includes any zeros within your ROI (mask) as
>> part of its mean/median/percentile/std calculations, but the reported
>> "number of voxels" number represents just the non-zero voxels within the
>> mask. I consider the former to be a "bug", but the FSL folks didn't agree
>> with this characterization. Regardless, if you want ROI values for ROIs
>> that cover voxels with missing data, you'll want to exclude the "zero"
>> voxels, which means that you can't use 'featquery' as is, but will need to
>> do a little extra work to get what you want. For comments on that see these
>> threads:
>>
>>
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1006&L=FSL&P=R29647&1=FSL&9=A&I=-3&J=on&X=3C7ED4135F095CC0DC&Y=mharms%40conte.wustl.edu&d=No+Match%3BMatch%3BMatches&z=4
>>
>>
>> https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=ind1006&L=FSL&P=R36078&1=FSL&9=A&I=-3&J=on&X=23BB2370CAA058FCA1&Y=mharms%40conte.wustl.edu&d=No+Match%3BMatch%3BMatches&z=4
>>
>> cheers,
>> -MH
>>
>>
>>
>>
>> On Thu, 2011-04-07 at 12:58 -0700, Kirstie Whitaker wrote:
>>
>> Hi FSL Community,
>>
>> I have a problem that in my huge data set of children aged 6-18 some of
>> them move at some point between runs and I end up not collecting the whole
>> brain. For one or two subjects that's a total disaster and I'm missing half
>> their brain, and they're the ones I will exclude completely, but for many I
>> still have large amounts of brain, maybe they've just lost a little bit of
>> inferior temporal lobe, posterior occipital or primary sensory-motor cortex
>> in their last (of 4) runs. Since many of my questions are focused on
>> prefrontal and parietal regions (which I *do* still have data for) I would
>> really like to salvage these subjects' runs.
>>
>> When I run a higher-level FEAT analysis including these subjects the mask
>> that is created is only for the regions which ALL subjects have in all
>> runs. This makes my results look really strange because when all the
>> "little bits" of missing data are added together I end up losing a lot of
>> brain in total. To give you some sense of numbers, around 100 subjects have
>> complete brain coverage, while around 20 are missing some tiny parts.
>> Almost every voxel which is real brain but is excluded due to missing data
>> is missing less than 4 subjects.
>>
>> One possible solution is to run the whole brain analysis with the subjects
>> with whole brain coverage but then include the other subjects in ROI
>> analyses. I'm concerned that if I use functional ROIs I will have biased my
>> results towards the subjects who have whole brain coverage. Do you have any
>> thoughts on that approach?
>>
>> Are there any other suggestions for including subjects without data in the
>> whole brain voxel based analyses? Are there any FEAT options I haven't
>> caught on to?
>>
>> Thank you so much for your help!!
>>
>> Kx
>>
>> --
>> I'm riding to LA, again, again! (3rd time)
>> Every dollar you donate not only pushes me along an incredible journey but
>> also supports the treatment and prevention of AIDS for those living in the
>> San Francisco Bay Area. Please consider sponsoring me at
>> www.tofighthiv.org/goto/kirstie
>>
>> Kirstie Whitaker
>> Doctoral Candidate
>> Cognitive Control and Development Laboratory
>> Helen Wills Neuroscience Institute
>> University of California at Berkeley
>> tel: 510 684 2456
>> web: bungelab.berkeley.edu
>>
>>
>
>
>--
>*I'm riding to LA, again, again! (3rd time)*
>Every dollar you donate not only pushes me along an incredible journey but
>also supports the treatment and prevention of AIDS for those living in the
>San Francisco Bay Area. Please consider sponsoring me at
>www.tofighthiv.org/goto/kirstie
>*
>*Kirstie Whitaker
>Doctoral Candidate
>Cognitive Control and Development Laboratory
>Helen Wills Neuroscience Institute
>University of California at Berkeley
>tel: 510 684 2456
>web: bungelab.berkeley.edu
>