Print

Print
Hi Matthieu,

Please see responses below:

On 8 October 2015 at 09:24, Matthieu Vanhoutte <[log in to unmask]> wrote:
Dear Anderson,

Thank you for your detailed answer.

Please see below my questions.

Best regards,

Matthieu

Le 08/10/2015 09:48, Anderson M. Winkler a écrit :
Hi Matthieu,

There are multiple ways to analyse repeated measurements data, but if some timepoints are missing, it can be difficult. You mention "for each single subject", so this means dropping a group analysis. Even that is difficult: permutations within subject with many timepoints aren't allowed (same reason we don't do for first level fMRI). Parametric tests could be considered, but FA data shows skewness (you could address with a data transformation, like probit), but the random field theory won't work easily with TBSS data.

It is possible, however, to instead run fsl_glm and compute, for each subject, an image with the COPEs (in this case the same as the PEs, i.e., the "betas" of the GLM) for, e.g., a linear trend. With multiple subjects, say A, B, C, D, etc. you'll have COPE_A, COPE_B, COPE_C, etc. which can be tested in a 1-sample t-test. Some of these will have 4 timepoints, some 5, some 6, etc, which means they cannot simply be shuffled in randomise, as their variances aren't the same (more visits, lower variances around these estimates).
 I really want to do separate longitudinal analysis on each subject independantly. Is this way you explained above allowing some kind of statistical analysis on one subject with several timepoints ?

As in the previous paragraph (the one that starts with "There are..." parametric tests could be considered, but it is difficult to correct for multiple testing using the RFT in TBSS data. Permutation tests cannot be used for the same reason they aren't used for 1st level FMRI, i.e., permuting would destroy the temporal correlation.

 
If this is the case, I don't see well how with fsl_glm to compute this COPE image for a subject ?  Is this linear trend gives some information of increase/decrease of FA with increasing time for one subject ?

The suggestion is to fit a linear model with a linear trend over time, i.e., that FA grows or reduces linearly over time within subject. The design would have timings of the scans in a single, mean-centered EV. Data would also be mean-centered. Then take the slope of this linear fit (the COPE) and see if, across subjects (group effect), it is significantly different than zero. It can be done even with a different number of timepoints (next paragraph).

 


However, you can define one variance group (VG) for each set of subjects that have the same number of timepoints (i.e., one VG for subjects with 4 visits, another for subjects with 5 visits, and so on), and use PALM with the options -vg and -ise (for independent and symmetric errors). This will compute a statistic that is robust to heteroscedasticity, in this case, the well known Aspin-Welch statistic.
 Do you mean that beyond the longitudinal single subject, I could make a group analysis with the few subjects (7 with 2 up to 4 timepoints) I have with PALM ?

Yes. The above approach is equivalent to whole-block sign-flipping if you were to construct a single, large, univariate model encompassing all subjects and all timepoints. But computing the COPEs with fsl_glm for each subject is much simpler, and requires no explicit definition of exchangeability blocks.
 
If so, how to define each VG before launching PALM ?

Use a text editor. One line per subject in a single-column csv file. The value for each subject can be simply the number of visits (the VG numbers don't have to be sequential integers). So, if subject 1 had 5 visits, put 5 as his VG number. If subject 2 had 3 visits, put 3 as his VG number.

When running PALM, don't use the option -eb, but use the options -vg and -ise.

All the best,

Anderson


 


This is just one possibility, but we have seen that ultimately the power is determined mostly by the size of the smallest VG.

All the best,

Anderson



On 7 October 2015 at 12:03, Matthieu Vanhoutte <[log in to unmask]> wrote:
Dear FSl experts,

I have a few subjects with different numbers of timepoints that are not equally time-spaced.

I would like to longitudinally analyze the FA (MD, ADC,...) evolution for each single subject. Is there a method I could apply as TBSS is designed for longitudinal group analysis ?

Best regards,

Matthieu