Dear FS experts, I want to add some in my early mail. If the contrast CM.C= [0 0 0 1] is designed for the interaction effect(time and group). How to design contrasts to do the main effect about the time and group? CM.C=[0 1 0 1] and CM.C=[0 0 1 1]? Or other? I wonder I may be wrong in designing matrix X in early mail. Please help me.Looking forward to reply, and thanks very much. Kind regards, Livia
Hi Livia,
yes, 0 0 0 1 is the interaction (if longitudinal slopes = atrophy rates, differ across groups).
When interpreting these, it makes sense to look at your model:
Y_ij = b0 + b1 t_ij + b2 g_i + b3 t_ij g_i where t is time, g is group. For group=0 you have Y_ij = b0 + b1 t_ij so b0 is the intercept of group_0, and b1 the slope
For group=1 you have Y_ij= ( b0 + b2 ) + (b1 + b3) t_ij so (b0 + b2) is the intercept for group_1 and (b1 + b3) the slope.
So slope difference across groups is (b1 + b3) - b1 = b3 so c=( 0 0 0 1 ) (this is group_1 minus group_0) and intercept difference is (b0 + b2 ) - b0 = b2, so c = ( 0 0 1 0 )
Average slope across both groups is:
0.5 * (b1 + b3 + b1) = b1 + 0.5 b3 so c =( 0 1 0 0.5 )
Best, Martin
On 6. Jun 2017, at 07:44, Livia Liu livialiu333@gmail.com wrote:
Dear FS experts, I want to add some in my early mail. If the contrast CM.C= [0 0 0 1] is designed for the interaction effect(time and group). How to design contrasts to do the main effect about the time and group? CM.C=[0 1 0 1] and CM.C=[0 0 1 1]? Or other? I wonder I may be wrong in designing matrix X in early mail. Please help me.Looking forward to reply, and thanks very much. Kind regards, Livia
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
Dear Martin, Thank you very much for your quick reply.
I use voxel-wise mixed model analysis, and I see different design:
lhstats = lme_mass_fit_vw(X,[1],Y,ni,lhcortex);----Summary: Algorithm did not converge at 0 percent of the total number of locations.
lhstats = lme_mass_fit_vw(X,[1 2],Y,ni,lhcortex);----Summary: Algorithm did not converge at 9.231.. percent of...
What the meaning and difference between [1] [1 2] or other?
In the end of the page https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels
It use Region-wise linear mixed-effects estimation:lme_mass_fit_Rgw , then do FDR.
Sorry, I can not make sure how to do FDR in my result.Please tell me some details.
And should I try this Region-wise model in my analysis?
My purpose is to compare cortical thickness changes in patient group after treatment and placebo group in 2 time points.
Then You say in the second mail:(if longitudinal slopes = atrophy rates, differ across groups),I am so sorry that I can't understand.
If I get the interaction result, how to explain?
I wonder I must get the main effect about the time or group to measure what cause the interaction.
Could you please give me some direction? Looking forward to reply, and thanks very much.
Kind regards, Livia
Hi Livia,
[1] only the intercept is used as random effect in the model
[1 2 ] both intercept and time (slopes) are used as random effect in the model. Here many more parameters get estimated so you need to have sufficient data. Often the more simple approach is better (but that can be tested by comparing the models, described on the wiki).
FDR correction can be done independent of the vertex wise or region wise approach.
The region-wise approach may be better (it definitely should be faster, but also benefits from some more statistical power).
this is a LME model, which is different from standard GLM models. The interaction of group and time is slope (e.g. atroph rate) difference across groups ( C = [ 0 0 0 1]). I explained it in detail in my last email.
Best, Martin
On 06/06/2017 03:50 PM, Livia Liu wrote:
Dear Martin, Thank you very much for your quick reply. I use voxel-wise mixed model analysis, and I see different design: lhstats = lme_mass_fit_vw(X,[1],Y,ni,lhcortex);----Summary: Algorithm did not converge at 0 percent of the total number of locations. lhstats = lme_mass_fit_vw(X,[1 2],Y,ni,lhcortex);----Summary: Algorithm did not converge at 9.231.. percent of... What the meaning and difference between [1] [1 2] or other? In the end of the pagehttps://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels It use Region-wise linear mixed-effects estimation:lme_mass_fit_Rgw , then do FDR. Sorry, I can not make sure how to do FDR in my result.Please tell me some details. And should I try this Region-wise model in my analysis? My purpose is to compare cortical thickness changes in patient group after treatment and placebo group in 2 time points. Then You say in the second mail:(if longitudinal slopes = atrophy rates, differ across groups),I am so sorry that I can't understand. If I get the interaction result, how to explain? I wonder I must get the main effect about the time or group to measure what cause the interaction.
Could you please give me some direction? Looking forward to reply, and thanks very much. Kind regards, Livia
Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer
freesurfer@nmr.mgh.harvard.edu
-
Livia Liu -
Martin Reuter