Hi,
Yes - that kind of contrast is absolutely fine and very common.
All the best,Mark
On 8 Apr 2014, at 07:46, Jts67 <[log in to unmask]>wrote:
Hi Mark,
Got it. You've been a tremendous help and I can't thank you enough! I'm sorry, but I have one last question about a contrast in the glm gui. If I had the behavioral measure split into EV4 and EV5 to look for a relationship b/w GM and memory (within each group), can I also run the contrasts [0 0 0 1 -1] and [0 0 0 -1 1] to see if there are any areas for which the linear relationship b/w GM and memory is stronger for older adults than for young adults, and visa versa? I wasn't sure if that contrast made any sense, but I was trying to think through it in my head.
Best wishes,
Jon
On April 8, 2014 2:35:48 AM EDT, Mark Jenkinson <[log in to unmask]> wrote:Hi,
I meant to feed them in at the very start - as inputs to stage 1.
All the best,Mark
On 8 Apr 2014, at 07:32, Jts67 <[log in to unmask]>wrote:
Hi Mark,
Thanks for the response. I do indeed have 1's in the second column and perhaps I should've just copied part of my design matrix instead.
As for using the fsl_anat biascorr_brain images as the initial inputs for VBM, it spits out an error when I try to do so. I assumed I need to use one of those .mat files to get those images back into a space/format that fslvbm_2_template likes, or are you saying I just feed them into the fslvbm_1_bet stage and then proceed with the rest of the vbm processing pipeline?
Best,
Jon
On April 8, 2014 2:15:22 AM EDT, Mark Jenkinson <[log in to unmask]> wrote:Hi Jon,
I was assuming that the copied lines were only from the beginning of the design matrix and that there would be a set of 1's later on in the second column. If not, then Mike makes an important point here.
As for your other questions:
- You can either do two analyses, or put it all together into a single one. There can be some advantages with DOF by keeping everything together in one analysis.
- If you want to use the fsl_anat bias corrected results, then these should be included as the initial inputs into VBM.
All the best,Mark
On 5 Apr 2014, at 13:28, Michael Dwyer <[log in to unmask]> wrote:
Hi Jon,
Maybe it's just copying issue or I'm misunderstanding something, but if your design matrix is only two groups, with EV1 as Old and EV2 as Young, then where one EV is 0, the other should be 1. It looks like your EV2 is 0 for all cases, half of which are also 0 in EV1.
Best,Mike
On Sat, Apr 5, 2014 at 7:31 AM, Jon Siegel <[log in to unmask]> wrote:
Hi Mark,
Thank you so much for your response. I just pulled some arbitrary numbers, but you’re right that the ones I posted do seem like they’re centered around 1 (this isn’t the case though). It looks like I need to run two separate randomise commands, the first initially looking to see if local GM volume has a linear relationship with both groups and the second looking to see within each group (since it’s possible that GM volume predicts memory performance differently for young and older adults). I appreciate your help tremendously.
On a side note, I also used the fsl_anat script to just test it out and it spits out some biascorr_brian.nii.gz files, but is there a way to transform them back so that I could then run them through the traditional VBM pipeline (creating my study-specific template) and then into the fslvbm_3_proc stage? It seems like the fsl_anat has done an excellent job cleaning up the brains and I’m curious to see if the results are similar using that pipeline instead. Thanks again for the great advice.
Best,
Jon
On Apr 5, 2014, at 3:02 AM, Mark Jenkinson <[log in to unmask]> wrote:
> Dear Jonathan,
>
> Firstly, although you've "centered" your ICV values, they look like they are centered about 1.0, and not about 0. For the GLM, it assumes a linear relationship (not a multiplicative one) and so the ICV values need to be centred around 0. This is also true for all other covariates that you want to add to the model. With your current model there is quite a lot of shared signal between the mean of both groups and the ICV, making the results difficult to interpret cleanly.
>
> To add a new covariate to the model then you can put it as any EV (the order does not matter as they are all fitted together, not one at a time). The key is to associate the non-zero entries in the contrast with the appropriate EV. So if you keep EV3 as the ICV (but zero centred) and add your extra covariate (also zero centred) as EV4, then you can have contrasts such as [0 0 0 1] and [0 0 0 -1] that test for linear relationships (of appropriate signs) between your new covariate and the local GM volume. This test would be across both groups if you did it like this.
>
> If you wanted to see if there was a separate relationship in each group then you need to split this new covariate across two EVs - that is, EV4 and EV5 - where EV4 would have zeros for one group and the demeaned (i.e. centred) covariate values for one group and zeros for the other group, and EV5 would be the other way around (zeros for the opposite group ...). The normal recommendation for demeaning (centering) in this situation would be to take all the covariate values for this quantity and demean prior to splitting them into the two subgroups. This would then do the conservative test and account for any potential differences in the mean values of the covariate between groups, splitting any group difference in local GM volume between the group EVs (EV1 and EV2) and the new EVs. However, if you have a strong prior argument that any difference in local GM volume could not be attributed to group differences in the mean of your covariate value, then you can demean within each group separately, but be warned that this is a very strong assumption and you would need to have very good reasons and evidence to choose to do the analysis this way. The safer, and generally recommended, alternative is to demean across both groups together (as a single set of values) and then split the demeaned values into the separate EVs.
>
> For more details on demeaning and centering, see the FSL wiki page on GLM, as well as Jeanette Mumford's pages (there is a link from the FSL wiki) and also the FSL Course slides (the third fMRI lecture - Advanced Topics - is where this material is).
>
> All the best,
> Mark
>
>
>
> On 2 Apr 2014, at 20:15, Jon Siegel <[log in to unmask]> wrote:
>
>> Hi FSL Experts,
>> I've searched through the message board and the archives and it seems like there's quite a lot of questions about the design files for contrasts and matrix. I just wanted to make sure that I'm setting things up correctly for my experiment.
>> I have 2 separate groups (Older adults and Young adults; n=18 and n=21 respectively) that I'm attempting to perform VBM on, but beyond the significant volumetric differences between the two groups, I'd like to know how to create the contrasts for testing the correlation/association of these specific brain regions with memory performance measure(s). I've been able to run the whole-brain analyses with controlling for ICV by creating the following matrix:
>>
>> EV1= Old (OA), EV2= Young (YA), EV3= ICV (scaling factor centered)
>> 1 0 1.2345
>> 1 0 .8652
>> 1 0 .6587
>> 0 0 1.1365
>> 0 0 .9874
>> 0 0 1.3245
>>
>> I then created the following contrasts based off the two group difference adjusted for covariate:
>> c1 = 1 -1 0 --> Old > YA
>> c2= -1 1 0 --> YA > OA
>> c3= 0 0 1 --> Pos effect of ICV/scaling factor
>> c4= 0 0 -1 --> Neg effect of ICV/scaling factor
>> Are these setup correctly?
>>
>> If so, I'd now like to modify the analyses and investigate whether these significant differences in brain volume between OA and YA are also associated with one or more of our behavioral performance measures (ie, Memory interference). From what I've read, I can just add another EV (Proactive Interference score centered) of interest into the GLM, but how exactly do I setup the contrasts? I assume that I'd want to put the behavioral variable of interest into the GLM before controlling for any nuisance variables, which would then make EV3= memory score centered and EV4= ICV. But again, I'm confused at how to setup the contrasts.
>>
>> For your reference, I'd like to look and see whether there is a correlation/association between GM volume and memory performance overall, as well as for each group separately. Any help you can give me would be greatly appreciated. Thank you so much for your time and please let me know if you need any additional information.
>>
>> Best Regards,
>> Jonathan Siegel
>>
--
Michael G. Dwyer, Ph.D.Assistant Professor of Neurology
Director of Technical Imaging Development
Buffalo Neuroimaging Analysis Center