Hi,

Just wanted to explain why it doesn't matter to mean center if you're looking at group differences and adjusting for a continuous covariate.  Basically, when you mean center you are moving the data to the left or right and the mean difference  contrast is simply the difference of the values where the fitted lines hit the y-axis.  With this model the lines are forced to be parallel, since there is just one regressor for the continuous covariate, and so no matter how you shift the lines to the left or right, the difference between the y-intercepts does not change.

Of course the separate means do change, but that doesn't matter if you're only looking at a difference.

Hope that helps,
Jeanette

On Thu, Dec 22, 2011 at 3:30 AM, Martha Kordulla <[log in to unmask]> wrote:
Hey Yingying,

I have the same constellation like Heather in my GLM-design and one further question to this issue.

I have two group and one nuisance variable (like the second figure on Janette's website). If I only want to look at the group differences (second figure, first contrast description on the website) demeaning is not necessary, because it has no impact. Do you agree with this?

Thank you, have a nice holiday

Martha


2011/12/21 Yingying Wang <[log in to unmask]>
Hi Heather,
 
According to the literature, FA changes with age.  Thus, you should find age and gender match controls.  You age mean is so different between patient and control group.  It is not an ideal situation to begin with.  I would prefer age- and gender- match controls.  Thus, you need to regress your age effect out (classic glm explaination).  Sometime it can still be hard to tease all the age effect out.
 
Randomise is a software package based on non-parametric test with GLM matrix setting.  From http://mumford.fmripower.org/mean_centering/ website, you should get all the information you needed.
 
Let's see a simple example.
Controls:   8  5  6  7 9   Patients: 11 9 11 14 15   
mean age for controls: 7      mean age for patients: 12
significantly different.
 
We want to know the difference between the two groups regarding measure A.  For instance: measure A usually increases with age for controls.
Measure A for controls: 30 15 20 25 35    for patients: 50 35 50 55 60
Significantly different without controlling for age.  Well, are the differences caused by group effect (patients v.s. controls) or purely age itself?  We can not know if we only use a simple two sample t-test.
 
Well, let's put age as a covariate. Bingo, group difference is not significant anymore.  From this, you can see use of age as a covariate is necessary in your case.  Detect the real differences between your conditions.
 
Demeaning age within groups (for controls and patients separately) [third figure on http://mumford.fmripower.org/mean_centering/ website ] - not right for your case. 
Demeaning all subjects (patients and controls) with the grand mean [second figure on http://mumford.fmripower.org/mean_centering/ website ] - you should use for your case.
 
Happy Holiday!
-Yingying

>>> Rajendra Morey <[log in to unmask]> 12/20/2011 4:32 PM >>>
Hi Heather :
 
If there is a relationship between age and and your dependent variable (e.g. FA) then you should definitely include age as a covariate even if age is matched between the groups. This is because age is explaining a lot of the variance associated with your dependent variable.  The best way to confirm this is to simulate this situation with a non-imaging dataset from everyday life (e.g. children's test scores and SES). Then you can run the GLM model with and without the covariate to convince yourself. It happens that TBSS is non-parametric (whereas GLM is parametric) but I think the priniciple still applies.
 
Raj
 
Duke University

From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of Heather R. Collins
Sent: Tuesday, December 20, 2011 12:50 PM
To: [log in to unmask]
Subject: Re: [FSL] TBSS - age as a covarite

Thanks, Yingying.

The two groups are controls and patients and we are interested in diffusion metrics associated with the condition and not due to developmental changes (age).

What does demeaning age within groups (for controls and patients separately) tell us and what would demeaning all subjects (patients and controls) with the grand mean tell us?

Many thanks,

--Heather

On Tue, Dec 20, 2011 at 12:00 PM, Yingying Wang <[log in to unmask]> wrote:
Hi Heather,
 
According to your message, you might want to do the two group separately.  Also it depends on your hypothesis.
 
For example: two group: controls and patients
What's your interest?
 
1) Use age as covarite (nuisance variable) if you're not interested in developmental changes. 
 
 
Best, Yingying
 
>>> Heather Collins <[log in to unmask]> 12/20/2011 9:50 AM >>>
Dear FSL experts,

I am running TBSS on 2 groups of subjects.  The mean ages are not different between the groups.  However, when any diffusion metric is plotted over age, the groups show different slopes: one increases with age whereas the other decreases with age. (It looks like an X)

I have combed through the archives and cannot find information about this kind of situation.

1) Should age be used as a covarite in this situation?

2)  What does it mean for age to be demeaned within each group versus demeaning both groups using the grand mean?  Which should I use in this situation?

Thank you in advance for your time and help!

Cheers,

--Heather