Yes, it was arbitrary. A linear trajectory was a sufficiently good model for the mean response over time for most of the cortex points in my analyses. Also if you only have at most three time points per user then you can only have at most two random effects in your mixed-model (intercept and time). Time squared, if considered, would be just a fixed effect. Best-Jorge
De: Jordi Pegueroles jordi@pegueroles.cat Para: jorge luis jbernal0019@yahoo.es CC: Freesurfer freesurfer@nmr.mgh.harvard.edu Enviado: Lunes 28 de noviembre de 2016 11:34 Asunto: Re: LME: contrast questions
Thank you for your quickly answer Jorge.
What is the rule for decide if we have to use a quadratic term in the model? That is, in your example you use the 80% of the length of the lh.cortex as a threshold. Is this value an arbitrary one?
Thank you again,
A 2016-11-28 16:05, jorge luis escrigué:
Hi Jordi
You have implicitly included the reference group in your statistical model by including an intercept term and a time variable:
group0 intercept: ß1 group0 time: ß2 group1 intercept: ß1+ß3 group1 time: ß2+ß4 group2 intercept: ß1+ß5 group2 time: ß2+ß6
From here you can quickly see that your proposed contrast for the differences in the mean response over time between group1 and group2 is correct.
Finally the region growing estimation method should yield more precise (lower variance) estimates for the ßs but I think it’s (might be) too complicated to setup with the current scripts. The vw method will produce good estimates too. So I recommend the vw estimation method if things get too complicated for you with the region-wise method.
Cheers -Jorge
DE: Jordi Pegueroles jordi@pegueroles.cat PARA: Freesurfer freesurfer@nmr.mgh.harvard.edu; jbernal0019@yahoo.es ENVIADO: Lunes 28 de noviembre de 2016 5:11 ASUNTO: LME: contrast questions
Dear FS experts,
I am trying to run a LME model for three groups (group0, group1 and group2) and (almost all) with three time points, all processed using the Longitudinal FS stream.
Our intention is to compare the differences between group0 against group1 and group0 against group2. I have followed the online LME tutorial and I have adapted your code to my data. However, I think I have not well understood the steps and I have few questions that you will be able to answer me...
i) In the X matrix I put the following columns:
- intercept
- time (years from baseline)
- group1 (binary variable: 1 if belongs to group1)
- group1.*time
- group2 (binary variable: 1 if belongs to group2)
- group2.*time
I don't understand why I do not have to include the reference group (in my case group0 and group0.*time) in the design matrix. How I can see the progresion of the reference group if I did not include it in the design matrix? How I can compare both progressions if the time of the group0 subjects is not taking into account?
ii) Linked to the previous question, if I want to compare the differences in time between group1 and group2... Should the contrast be C = [0 0 0 1 0 -1]?
iii) Finally, between the vw method and the region growing method which one is more accurate? In other words, which one do you recommend to use?
Thank you,
-- Jordi.