Print

Print
Hi Michiko,

It's ok to test the nuisance that is present only in patients using this model. Although a separate model could be considered only with patients, then it wouldn't allow for multiple testing correction across contrasts.

In any case, if you use randomise such correction isn't present anyway so if you want, you might as well run a separate model just with those subjects. Both ways are correct in their own way. Running a separate model has fewer assumptions about the data of patients being similar to the data of controls, but the more such tests, more exposed you are to false positives.

All the best,

Anderson


On 22 May 2017 at 02:49, Michiko HZ <[log in to unmask]> wrote:
Hi Anderson,

Thanks very much for the quick reply and advice.

In the event that I wish to look beyond comparing groups, e.g. looking at the main effects of say, patient specific variables like disease duration or amount of medication, would you suggest to proceed with the 2-model design as I proposed? If so, how would I do mean-centering and what should be done about the template?

Thanks again.
Michiko


On Mon, May 22, 2017 at 12:37 AM, Anderson M. Winkler <[log in to unmask]> wrote:
Hi Michiko,

How about this model:

EV1: Patient group 1
EV2: Patient group 2
EV3: Controls
EV4: age
EV5: sex
EV6: Patient-specific nuisance variable.

About mean-centering: if the contrasts will only compare groups (EV1, EV2 and EV3), such that all other variables are never tested on their own, then there is no need for any mean-centering.

This should also address the template issue: use all subjects, or a balanced random composition of members from all three groups.

All the best,

Anderson


On 20 May 2017 at 15:28, Michiko <[log in to unmask]> wrote:
Dear FSL experts,

I am planning to run a 3 group analysis with TBSS on DTI data and have some specific questions I hope to get some advice on:

The aim is to compare if diffusion indices (e.g. FA) significantly differ between the 3 groups - particularly between patient groups. Running a quick t-test shows there is significant difference in disease duration, one of the variables specific to the patient groups and which we have to account for in the model. What is the best way to conduct the analysis?

I am thinking of running 2 models as follows:
Model 1 -
subjects: everyone
regressors: mean ptgrp 1, mean ptgrp 2, mean HC, confounders available for everyone (e.g. age, demeaned across everyone)
contrasts: ptgrp 1 vs HC, ptgrp 2 vs HC, ptgrp 1 vs ptgrp 2

Should this model test only ptgrp 1 vs HC and ptgrp 2 vs HC as opposed to also including testing ptgrp 1 vs ptgrp 2?
With this model, is it OK to leave it as such and not include the other pt-specific confounders since I am conducting model 2? Or is it better to regress out within-pt group duration differences (by demeaning within groups)?

Model 2 -
subjects: ptgrp 1 and ptgrp 2
regressors: mean ptgrp 1, mean ptgrp 2, all possible confounders (demeaned across all pts, also including previous covariates in Model 1, but re-demeaned to new mean calculated from pt-only groups???)
contrasts: ptgrp 1 vs ptgrp 2

I am also using the -n flag for tbss_2_reg (study-specific template). In this case, because model 2 only includes the patient groups, would I have to redo the step involving creating the study-specific template in order to run model 2? i.e. having to generate a new template for this purpose.

Thanks very much in advance!