External Email - Use Caution
On Fri, Jan 24, 2020 at 10:32 AM Greve, Douglas N.,Ph.D. <
dgr...@mgh.harvard.edu> wrote:
>
>
> On 1/22/2020 2:13 PM, Graduate Imaging wrote:
>
> External Email - Use Caution
>
>
> On Tue, Jan 21, 2020 at 10:40 AM Greve, Douglas N.,Ph.D. <
> dgr...@mgh.harvard.edu> wrote:
>
>>
>>
>> On 1/17/2020 10:39 AM, Graduate Imaging wrote:
>>
>> External Email - Use Caution
>>
>>> Hello,
>>>
>>> I recently ran a vertex wise analysis on two different projects the
>>> first had three groups with three co-variates that looked at if group
>>> membership was associated with brain volume. The second project had two
>>> groups with three co-variates looking at psychiatric x BMI interaction
>>> effect on brain volume. I'm interested in running ROI analyses in matlab to
>>> ensure the same matrices were used.
>>>
>>> To run the ROI analyses I'm going to important the design matrix from
>>> the vertex wise analysis and use the fast_glmfit and fast_fratio commands
>>> as shown below:
>>> X = load('Xg.dat');
>>> C = load('C.dat');
>>> y = load('ROI.dat'); text file containing participant ROI values from
>>> the DKT atlas.
>>> [beta rvar] = fast_glmfit(y,X);
>>> [F pvalues] = fast_fratio(beta,X,rvar,C
>>>
>>> My first question is the beta values that are calculated from
>>> fast_glmfit are unstandardized is there anyway way to have it compute
>>> standardized values?
>>>
>>> I think this is how you would compute that
>>> betastddev = sqrt(rvar*diag(inv(X'*X)));
>>> betastandard = beta./betastddev;
>>>
>> When I try running this in Matlab I get the Error using *** incorrect
>> dimensions for matrix multiplication message.
>>
>> That means that the number of items in y (ie, number of subjects in
>> ROI.dat) is different than the number of rows in X (number of subjects in
>> the fsgd). What are the sizes of y and X?
>>
> That's what I suspected as well however, X = 102x12 and Y= 102x5. I ran
> [betastddev = sqrt(rvar*diag(inv(X'*X))); betastandard = beta./betastddev]
> after running successfully running [[beta rvar] = fast_glmfit(Y,X); [F
> pvalues] = fast_fratio(beta,X,rvar,C)] so its odd that the the dimensions
> wouldn't be matching.
>
> Also when using fast_glmfit and fast_fratio via Matlab for ROI analyses
> what are the appropriate papers to cite? In the Matlab files in fsfast
> there is reference to Worsley, K.J. and Friston, K.J. Analysis of fMRI
> Time-Series Revisited - Again. Neuroimage 2, 173-181, 1995. Would this be
> the correct citation?
>
> The GLM is so old I would not know what to cite. If you feel you need to
> cite something, that Worsley paper is fine.
>
Thanks. For DODS looking at an interaction term for the previously
mentioned 2 group 4 co-variate analysis it produces the beta weights for
each of the 10 regressors. Am I correct is assuming the way the analysis
works is that it creates interaction terms for the class with each of the
co-variates? If so the analysis could be summarized by Y = B0 + B1*X1 +
B2*X2 + B3*X3 + B4X4 + B5X5 + B6*X1*X2 + B7*X1*X3 + B8*X1*X4 + B9*X1*X5 +
e. Where X1= Class, X2= BMI, X3=ICV, X4= Age and X5=sex (I've since
incorporated this to be apart of class but want to continue with this
example). If that is the case to get an overall interaction term I should
be able to manually create these interaction terms and put them into a
matrix then for the contrast just regress out all of the effects except for
the BMI*class (B6*X1*X2). However, when I do this the p value is not the
same (far more significant) than using the DODS analysis with the 0 0 1 -1
0 0... contrast.
>
>>> My second question is when I ran the ROI analyses looking at the
>>> diagnosis x BMI interaction effect it outputs an array of beta values for
>>> each regressor. In addition to each groups beta value (for BMI) I'm
>>> interested in the beta value of the interaction (ie diagnosis x BMI). I was
>>> wondering how do i go about obtaining this? Would I have to create a new
>>> matrix with the interaction term included in it?
>>>
>>> If so would the analysis essentially be one group with the diagnosis x
>>> BMI interaction term plus the three co-variates?
>>>
>>> If you are using an FSGD file and have used DODS, then you can create an
>>> interaction contrast. Send me your fsgd file if you want further help
>>>
>> Yes I was using DODS approach for this analysis, I've attached my FSGD
>> file to this email.
>>
>>
>> You have 2 groups and 4 coviariates so 10 regressors. The dx BMI
>> interaction would need a contrast matrix
>> 0 0 1 -1 0 0 ... the rest 0s
>> BTW, sex should not be a covariate. You really need to have four groups,
>> P-male, P-female, HC-male, and HC-female (which would change the contrast
>> above)
>> Also, I would normalize the age and ICV
>>
>> Thanks for the correction I'll update my matrix accordingly to four
> groups and normalize the variables! The contrast matrix I ran was actually
> the same you suggested however it pastes an array of beta values for each
> 10 regressors. However, I was interested in obtaining a beta value that
> represents the interaction.
>
>>
>>> _______________________________________________
>>> Freesurfer mailing
>>> listfreesur...@nmr.mgh.harvard.eduhttps://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
>>>
>>>
>>> _______________________________________________
>>> Freesurfer mailing list
>>> Freesurfer@nmr.mgh.harvard.edu
>>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
>>
>>
>> _______________________________________________
>> Freesurfer mailing
>> listfreesur...@nmr.mgh.harvard.eduhttps://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
>>
>>
>> _______________________________________________
>> Freesurfer mailing list
>> Freesurfer@nmr.mgh.harvard.edu
>> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
>
>
> _______________________________________________
> Freesurfer mailing
> listfreesur...@nmr.mgh.harvard.eduhttps://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
>
>
> _______________________________________________
> Freesurfer mailing list
> Freesurfer@nmr.mgh.harvard.edu
> https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
_______________________________________________
Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
