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