Hi Alex,

you are not looking at a "one sample group mean" (osgd) so don't pass that flag. Your design is probably something like
1 A other_co_vars_to_regress_out
(these are column vectors).

so contrast in that case would be [ 0 1 0... ]

That should create all outputs. All of this is really cross sectional analysis where the depending variable is simply the 'change in thickness' instead of thickness itself. Take a look at the glm tutorial on the wiki, which describes the process.

Best, Martin


On 10/23/2014 02:40 PM, Alex Hanganu wrote:
Hi Martin,

thanks for confirming. I duplicated the parameter and got good results in qdec.

I also tried to repeat the analysis with mri_glmfit but I can't manage to come to an end.
In order to analyse the correlation between pc1 and parameter 'A', it seems that I have to construct an fsgd file, that is different from the .qdec file included in the "long_mris_slopes" command. Nevertheless, after doing so (presumably all "Inputs" were attibuted to subject.long.base-time1 and subject.long.base-time2) I thought that a contrast is needed, yet the "--C" and the "--osgm" flags cannot be used together.
- How can the correlation between -pc1 and parameter 'A' be performed in this case ?

Additionally, after performing the "mri_glmfit" described in the 2-stage-model page, in the tksurfer  how can I see the plot ? The y.fsgd file wasn't created. Is there another method ?

Thanks,
Alex




Le mardi 21 octobre 2014 16h40, Martin Reuter <mreuter@nmr.mgh.harvard.edu> a écrit :


Hi Alex,

you have to duplicate the parameter (it is basically fixed across time). If you put 0 for tp2, it will average the two values, which is not what you want. Otherwise I think it is the correct approach.

Best, Martin


On 10/21/2014 04:31 PM, Alex Hanganu wrote:
Dear Martin,

thank you very much for your answer ! and thanks for all the details !
- yes, we have exactly 2 time points in all subjects and the parameter is a single number.

In qdec - it seems that qdec table has to include the parameter 'A' both at time 1 and at time 2 in order for "long_qdec_table" command to create the "cross" file. I put a zero at time 2. In qdec design we analyzed parameter 'A' with -pc1 and -spc. I'm not sure that this is the correct approach.

I'll continue with LME and mri_glmfit.

Sincerely,
Alex


Le mardi 21 octobre 2014 9h19, Martin Reuter <mreuter@nmr.mgh.harvard.edu> a écrit :


Hi Alex,

the parameter is a single number that happens to be measured at time 1 right, eg baseline age? Lets call that parameter 'A' for the discussion below.  Also you have exactly 2 time points in all subjects?

There is two alternatives:

1. Simple approach (2-stage-model): You compute the atrophy rate (e.g. percent thickness change) on the cortex (long_mris_slopes) for each subject. At this point you have 1 measure per subject and work cross-sectionally. You can use qdec or mri_glmfit to correlate 'A' (independent parameter) with the thickness change (dependent variable). This is OK if you have the same number of time points and the same time distance in all subjects. Details here:
https://surfer.nmr.mgh.harvard.edu/fswiki/LongitudinalTwoStageModel

2. Better approach: use Linear Mixed Effects models (we have matlab tools for that). This model is more flexible (different manycolumn of ones, time points, different time intervals, even subjects with a single time point can be added). You'd setup a system like
Y_ij = beta_0 + b_i + beta_1 * A_i + beta_2 t_ij + beta_3 A_i * tij + error_ij
where  Y_ij is the thickness of subject i at time point j (known)
t_ij is the time from baseline of the j measurement in subject i (known),
A_i is the variable you measure at baseline in subject i (known),
the model will estimate the following:
b_i (a random effect) is the subject specific intercept (offset from the global intercept beta_0)
beta_1 another intercept offset caused by A
beta_2 the slope with respect to time (fixed effect, so it will be the same for all subjects, can also be modelled as a mixed effect)
beta_3 the interaction of A and time (<- you are interested in this)
Testing if the interaction beta_3 is different from zero will show you where A has an effect on the slope.
For the model above the X matrix would have 4 columns:
1 A T (A.*T)
where 1 is a column of 1's, A the A_ij (Ai repeated j times for each subject), T=t_ij and the coordinate wise product of A and T. Contast [ 0 0 0 1] tests the interaction. You'd tell the function that you want the intercept to be a random effect by passing [ 1] (selecting the first column). If you also want to have t_ij as a random, you can pass [1 3 ] . Details here:
https://surfer.nmr.mgh.harvard.edu/fswiki/LinearMixedEffectsModels

Best, Martin


On 10/20/2014 03:20 PM, Alexandru Hanganu wrote:
Dear FreeSurfer Experts,

How could the longitudinal analysis be performed in order to show whether a parameter at time 1 is predictive of changes in cortical thickness over time ? and can the corresponding regions be shown in FreeSurfer ?

In a statistical analysis, as we see it, we must perform the correlation between the parameter at time 1 and the cortical thickness difference (or ROI) time 2-time 1, yet in this case we cannot see it on the cortex.

Thank you,
Alex
 




-- 
Dr. Martin Reuter

Instructor in Neurology
  Harvard Medical School
Assistant in Neuroscience
  Dept. of Radiology, Massachusetts General Hospital
  Dept. of Neurology, Massachusetts General Hospital
Research Affiliate
  Computer Science and Artificial Intelligence Lab,
  Dept. of Electrical Engineering and Computer Science,
  Massachusetts Institute of Technology

A.A.Martinos Center for Biomedical Imaging
149 Thirteenth Street, Suite 2301
Charlestown, MA 02129

Phone: +1-617-724-5652
Email: 
   mreuter@nmr.mgh.harvard.edu
   reuter@mit.edu
Web  : http://reuter.mit.edu