Hi Anderson,

this was very helpful. Thank you very much so far. I would like to ask a first, final question in all this FA-covariate interaction-stuff which I am dealing with at the moment, referring to my very initial question (see also below).

For summary:
Question of interest: I am looking for voxels with significant interactions between FA and one covariate COV in explaining behavior BEH, correcting for AGE and GENDER.

This is how I actually set up my model:

In principal: DV ~ [ GROUP + AGE + GENDER + COV + FA + COV*FA ]   [parameters models in FSL GLM]

DV: 4D skeletonised BEH, demeaned (i.e. all skeleton voxels filled with BEH for each subject)
GROUP: all 1

AGE, GENDER, COV: demeaned
FA: 4D skeletonised TBSS data (as usually used), but demeaned (!) (i.e. each voxel on the skeleton is demeaned taking the group average of this specific voxel into account)

COV*FA: product of COV and FA 4D ==> giving another 4D image

Contrast: 0 0 0 0 0 1 ==> Voxels with sign. pos. interaction between COV and FA.

My randomise command is:
randomise -i all_BEH_demeaned_skel -o results -m mean_FA_skeleton_mask -d model.mat t model.con -n 500 --T2 --vxl=5,6 --vxf= all_FA_demeaned_skel,all_product_skel

Actually, the calculations seems to take a while.... Eyerything correct at this stage?

Thank you so much for your help.

Kind regards,
Robert


###########################################

Hi Robert,

The F-test it seems uses C3 and C4, but C4 is the same as C3 multiplied by -1. Instead, define the F-test as just C3 or as just C4 (either will give the same result) but not both at the same time.

That said, you also need to make sure the option -f <design.fts> is included when invoking randomise.

All the best,

Anderson


On 6 January 2016 at 17:12, Robert Schulz <[log in to unmask]> wrote:

    Dear Matthew,

    I see. However, it says, the F test would not valid as each included contrast cannot be a linear combination of the others.
    I guess it has not been caculated. This message also pops up if I try to view the design - which I cannot; it says child process exited abnormally. Or how do I get to the F test results?

    Otherwise, would this approach including its GLM and contrasts be correct?
    Thank you for your help.
    Kind regards, Robert


    ###############################


    Hello Robert,
                            Depending on the values used for the "dummy" EVs used as placeholders for the voxelwise EVs, the GUI may warn about linear combinations. In this case, the message can be safely ignored.

    Kind regards
    Matthew

    > Dear Anderson,
    >
    > thank you very much for your quick reply.
    > According to your suggestion, we changed our randomise setup in the following way as a first approach. The aim of this analysis is to uncover an interaction between regional FA and GROUP in explaining a dependent behavioural variable.
    >
    > We calculated a new stock of skeletonised images inlcuding our dependent variable, DV (Behavioural), one for each subject and the whole skeleton. ==> new 4D file all_DV.nii.gz
    >
    > We set up a new GLM including
    >
    > EV1 = GROUP1 (coded as 0,1)
    > EV2 = GROUP2 (coded as 0,1)
    > COV1 = (demeaned across both groups, age)
    > COV2 = (demeaned across both groups, gender)
    >
    > Additionally, we inlcuded  - as voxelwise EVs - two 4D files with subjects specific FA values on the skeleton, one for GROUP1 (EV5, all_FA_G1.nii.gz), another for GROUP2 (EV6, all_FA_G2.nii.gz).
    >
    > For contrast, we set up 4 contrasts:
    >
    > 1. T 0 0 0 0 1 0 (pos. correlation between FA and DV for Group1)
    > 2. T 0 0 0 0 0 1 (pos. correlation between FA and DV for Group2)
    > 3. 0 0 0 0 1 -1 + 4. 0 0 0 0 -1 1
    > Contrasts 3 and 4 are summarized to one F contrast to investigate an INTERACTION between GROUP and FA in explaining DV.
    >
    > Firstly, when saving this GLM, I got a warning that some EVs might be linear combinations of others. Why this? Where did I make a mistake? Do I have to demean the voxelwise EVs (4D FA files) across both groups as well? Could I use -D flag instead?
    >
    > My actual randomise command would be
    >
    > randomise -i all_DV -o tbss -d randomise.mat -t randomise.con --vxl=5,6 --vxf=all_FA_G1,all_FA_G2 -n 500 -T
    >
    > Does this approach sound valid to you? For C1, I find some clusters, nothing for C2. For C3 I wonder that I don't find anything, for C4 I get some signals. Nothing after correction. F statistics are not calculated. I am confused.
    >
    > I am looking forward to your answer.
    >
    > Thank you so much.
    >
    > Kind regards,
    > Robert
    >
    >
    >
    >
    >
    >
    >
    >
    >
    > ###############################################
    >
    >
    >
    > Hi Robert,
    >
    > Please, see below:
    >
    >
    > On 23 December 2015 at 09:03, Robert Schulz <[log in to unmask]> wrote:
    >
    >    Dear all,
    >
    >    I wonder whether it would be possible to use TBSS / randomise on FA data to answer the following 2 questions:
    >
    >    1. I would like to visualize voxels with a significant interaction GROUP[2 levels, 1,0]*FA in correlating with a continuous COV, adjusting for AGE and GENDER. I would set up the following GLM with the EVs:
    >
    >    Intercept  GROUP1   GROUP0   COV_G1(cont., demean)   COV_G0(cont., demean)    AGE(cont., demean)   GENDER (0,1, demean)
    >
    >    Contrasts (randomise without -D option as intercept is included):
    >    Pos. Correlation FA-COV for Group1: 0 0 0 1 0 0 0
    >    Pos. Correlation FA-COV for Group0: 0 0 0 0 1 0 0
    >    Sign. Interaction FA-GROUP in the explanation of COV: F contrast 0 0 0 1 -1 0 0 / 0 0 0 -1 1 0 0
    >
    >    Correct?
    >
    >
    > Almost. Use:
    > EV1: Group 0 (coded as 0 and 1)
    > EV2: Group 1 (coded as 0 and 1)
    > EV3: Cov, group 0
    > EV4: Cov, group 1
    > EV5: age
    > EV6: sex.
    >
    > The contrasts are nearly identical, except that you will have dropped EV1.
    >
    > The reason for the change is that Groups 0 and 1 added together correspond to the Intercept, which would introduce redundancies to the design.
    >
    >
    >
    >
    >    2. Now, if I would be interested in an interaction between FA and a continuous variable BEH instead of a categorial factor GROUP:
    >    In R statistical package, the model would be - e.g. think of one single FA value in one voxel: lm(COV ~ AGE + GENDER + BEH + FA + BEH*FA, data=dat). How should I set up the model - if possible - for randomise? I could imagine
    >
    >    Intercept  BEH(cont., demean) COV(cont., demean) BEH*COV(demean)  AGE(cont., demean)   GENDER (0,1, demean) with
    >    contrast: 0 0 1 0 0.
    >
    >    However, here I would be modeling the interaction between BEH and COV in explaining FA, but this is actually not really what I want. I want to model the interaction BEH*FA in explaining COV (see R model). I would really appreciate any help.
    >
    >
    > For this you'd need to convert the COV to a 4D image, then run randomise with voxelwise EVs. To convert COV to an image, use fslmaths, multiplying a 3D mask of ones by the value of each subject, then merging as a 4D.
    >
    > I think PALM expands an input single column .csv file to the size of an image if there are voxelwise EVs (I can't recall if this is just for NPC or any voxelwise EV) so it could be an option.
    >
    > All the best,
    >
    > Anderson
    >
    >
    >
    >
    >
    >    Thank you very much for your help.
    >    Kind regards, Robert