Hello,
The method you describe sounds fine to me from a mathematical point of view, as long as you have demeaned all of your covariates (including the salthouse test score).
Whether you need to separate age and gender covariates in the two groups is another question, and it really depends on whether you want to model potentially different levels of correlation between your FA values and the age/gender/education values.
The fact that your result disappears when adding extra covariates is quite common. It can either mean that these covariates are the main cause of the effect, or that there is a lot of ambiguous and shared variance so that the GLM cannot distinguish whether the effect is *uniquely* due to a relationship with the salthouse test or a relationship due to age/gender/education. The key word here is uniquely. The GLM always takes the conservative approach so that if it is potentially due to one or the other effect then it will not be found significant in any single contrast. To find out whether it is due to a combination of factors (e.g. salthouse and/or age and/or education) then you can use an F-contrast over a number of individual t-contrasts (e.g. salthouse contrast & age contrast & education contrast). This will then show you results whenever FA changes could be attributed to *any combination* of these factors. This can often show effects that cannot be seen in individual t-contrasts, since for the F-contrast it is not necessary to separate the effects. Note that in my example I have excluded gender (just as an example) but if there was a lot of shared variance between gender and any of the other factors, then this particular F-contrast might also show nothing.
Anyway, I hope this helps and gives you some things to explore.
All the best,
Mark
On 22 Nov 2012, at 09:07, SUBSCRIBE FSL Anonymous <[log in to unmask]> wrote:
> Hi FSLrs,
>
>
> We are doing analyses with TBSS and performed a regression analysis between cognitive functions and FA maps in patients and controls. We did a matrix with the next EVīs: one for the group of patients, other for the group of normal controls, other for the neuropsychological task (score in salthouse test) in patients and another for the neuropsychological task in normal controls. The correlation we obtained between patients FA maps and salthouse scores were significant. However, when we introduce age, gender, and years of education as covariates in the regression analyses this significant results disappear. The way we add the covariates to the design matrix was the following: We added two EVīs for gender (one for patients group and other for normal control), two EVīs for age (one for patients group and other for normal control) and another two EVs for years of education (one for patients group and other for normal control). In the contrasts we just put 0s in the covariates.
> Is this the correct way of introducing confounding covariates in the matrix?
> Is it possible to do include many covariates in a correlation analysis?
>
> Thank you for your help!
>
|