External Email - Use Caution
This is extremely helpful. Thank you, Kersten.
Swati
From: freesurfer-bounces@nmr.mgh.harvard.edu on behalf of "Diers, Kersten /DZNE" Kersten.Diers@dzne.de Reply-To: Freesurfer support list freesurfer@nmr.mgh.harvard.edu Date: Sunday, March 15, 2020 at 1:23 PM To: Freesurfer support list freesurfer@nmr.mgh.harvard.edu Subject: Re: [Freesurfer] longitudinal FreeSurfer help
External Email - Use Caution Hello Swati,
a model with two random effects gives greater flexibility in modeling individual variation of the slopes. This is why it is recommended; but it is not strictly necessary.
Sorry, I have to correct my response to your step #4 in my previous mail: given that you have three groups, it will not be necessary / meaningful to explicitly include the grp1 or time*grp1 effects in your model, assuming that group 1 is your reference group. In general, one of your groups has to be chosen as a reference group, and this group's effects will be modeled by the ‚intercept‘ and ’time’ effects, so there is no need to give it additional columns/regressors in the design matrix. The other columns in the design matrix will model the differences in level and slope between any other group and the reference group.
Inclusion of the reference group will also lead to the warning message that you mention in your follow-up mail, so removing the regressors should also remove the warning.
If you compare the example in the Freesurfer wiki: here, we have four groups (1: HC, 2: sMCI, 3: cMCI, 4:AD), but we only have additional regressors for groups 2-4 in the model. The HC group is chosen as the reference group.
Regarding your further questions:
The lme_mass_F function has to be run once for each contrast that you want to estimate.
Note that there can be, loosely speaking, single-row or multiple-row contrasts. Single-row contrasts test a single effect, e.g. difference in slope for group 1 vs 2. Multi-row contrasts jointly test multiple effects, e.g. differences between any pair of multiple groups, and hence are of a more global nature. Such an example is also given in the tutorial, where the ‚C‘ matrix has three rows.
Generally speaking, your contrast vector (single-row) or matrix (multi-row) needs to have as many corresponding entries (in case of vector) or columns (in case of matrix) as you have beta coefficients in your model (including the intercept). Contrast weights (often -1, 0, or 1) are then used to create linear combinations (for example, differences) of the coefficients that you want to test.
In the univariate tutorial example, which is a little bit easier to explain, we have 14 coefficients, and a three-row contrast matrix: we use it to test for any difference between the four groups. In principle, each row of this matrix could also be used as a single-row contrast, testing for a specific difference.
Looking at the contrast matrix in detail: row 1 sets coefficient #4 (which is sMCI * time) to +1,which gives the difference in slopes between the reference group (HC) and the sMCI group. Row 2 sets coefficients #4 to -1 and #6 (which is cMCI * time) to +1, which gives the difference in slopes between the sMCI and cMCI group. Row 3 sets coefficients #6 to -1 and #8 (which is AD * time) to +1, which gives the difference in slopes between the cMCI and AD group.
Best regards,
Kersten
Am 13.03.2020 um 22:40 schrieb Swati Rane <srleven@uw.edumailto:srleven@uw.edu>:
External Email - Use Caution
Also, when I run the lme_mass_fit_EMinit, lme_mass_RgGrow, lme_mass_fit_Rgw,
I get a warning: Matrix is close to singular or badle scaled. Results may be inaccurate. RCOND=3e-21/ In parallel_function>make_general_channel/channel_general (line 914)
In remoteParallelFunction (line 38).
What could cause this warning? I am afraid my results will not be correct.
Swati ____________________________________________ Swati Rane Levendovszky, PhD Interim Director, Diagnostic Imaging Sciences Center Assistant Professor, Integrated Brain Imaging Center Radiology University of Washington Box 357115 | AA035 HSB | Seattle, WA 98195 srleven@uw.edumailto:srleven@uw.edu | Ph: 206 543 6159 Web: https://sites.google.com/site/uwswatirane/ ____________________________________________
From: Swati Rane <srleven@uw.edumailto:srleven@uw.edu> Date: Thursday, March 12, 2020 at 10:18 PM To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edumailto:freesurfer@nmr.mgh.harvard.edu> Subject: Re: [Freesurfer] longitudinal FreeSurfer help
Hi Kersten,
Thanks for the reply. I think I am trying to understand the steps that follow.
Do I have to check whether my model with 2 random effects (intercept and time) is better than that with one (intercept)?
For the contrast I am interested in group wise longitudinal cortical thinning. And also, if the longitudinal thinning is different between groups. Given that my X=[intercept time grp1 time*grp1 grp2 time*grp2 grp3 time*grp3 age@timepoint1 gender]
How do I set up CM? Do I need to run the lme_mass_F multiple time per contrast I provide? Will my CM be For Longitudinal thinning in grp 1: [1 0 0 1 0 0 0 0 0 0] For Longitudinal thinning in grp 2: [1 0 0 0 0 1 0 0 0 0] For Longitudinal thinning in grp 3: [1 0 0 0 0 0 0 1 0 0] For Longitudinal thinning difference grp 2 – grp 1: [1 0 0 -1 0 1 0 0 0 0]?
Can you explain the CM set-up described in the tutorial?
Thank you.
Swati From: <freesurfer-bounces@nmr.mgh.harvard.edumailto:freesurfer-bounces@nmr.mgh.harvard.edu> on behalf of "Diers, Kersten /DZNE" <Kersten.Diers@dzne.demailto:Kersten.Diers@dzne.de> Reply-To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edumailto:freesurfer@nmr.mgh.harvard.edu> Date: Thursday, March 12, 2020 at 5:19 AM To: "freesurfer@nmr.mgh.harvard.edumailto:freesurfer@nmr.mgh.harvard.edu" <freesurfer@nmr.mgh.harvard.edumailto:freesurfer@nmr.mgh.harvard.edu> Subject: Re: [Freesurfer] longitudinal FreeSurfer help
External Email - Use Caution Hello Swati,
your points 1-4 look fine, as far as I can see.
Regarding 5, the recommended spatiotemporal mass-univariate approach consists of calling three functions:
lme_mass_fit_EMinit, lme_mass_RgGrow, lme_mass_fit_Rgw
in that order. If that does not work, it is in my eyes possible to use the more simple mass-univariate approach using
lme_mass_fit_vw
as an alternative.
Regarding 6, which specific questions do you have?
Best regards,
Kersten
On Do, 2020-03-12 at 05:58 +0000, Swati Rane wrote: External Email - Use Caution Dear FreeSurfer experts,
I simply do not understand the LME setup for longitudinal analyses.
Here is what I have: 3 groups each with variable longitudinal data (in terms of umber of time points and interval between timepoints). I am interested in detecting if the rates of longitudinal atrophy in each of the 3 groups and whether it is significantly different between the 3 groups.
1. My Qdec file contains: fsid, fsid_base, time, age, group, gender 2. I read in the relevant files with (output of surf2surf, lh.cortex, lh.sphere) 3. I read the Qdec file removed the first column, obtained the subject IDs, keep numerical data and get a sorted ‘M’ matrix 4. X=[intercept time grp1 time*grp1 grp2 time*grp2 grp3 time*grp3 age@timepoint1 gender] 5. After that, should I use voxel wise mass-univariate model or spatiotemporal mass-univariate model lme_mass_fit_vw or lme_mass_fit_EMinit? 6. After that I don’t understand the subsequent steps. Can you help me understand?
Any help is appreciated.
Thanks! Swati
_______________________________________________
Freesurfer mailing list
Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu
https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer _______________________________________________ Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edumailto:Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer