Alberto,
as you say this is an important and general question, but unfortunately there is no one general answer to it.
An example of what you describe might be where you attempt to correlate your MR-metric with disease duration. Let us say that you find an area which correlates nicely with disease duration. Let us further say that someone suggests you add age as a "covariate of no interest" in your model, and that when you do that your correlation disappears. The reason for that being that age and disease duration were highly correlated so that anything in the MR data that could be explained by disease duration could equally well be explained by age.
In this case the good solution would be to make sure already in the recruitment phase to try and get subjects that break this correlation, i.e. to find young subjects with a long disease duration and old subjects with a short disease duration (if at all possible).
Another example is when you have performed a whole battery of psychometric tests on your subjects and it turns out you find no correlations between your MR-metric and any of the tests. In this case it is quite possible that the problem is that many of those tests asses the same or similar functions and that they are therefore highly correlated.
In that case one solution is to try and reduce the number of covariates by performing a factor analysis of your psychometric scores, thereby reducing N tests to M (where M<N) underlying factors. You can then either enter your factor scores directly, or you can keep the original scores and use the factor weights to guide you in defining M F-tests.
There are other scenarios as well, but as I said there is no good general answer and each case need to be approached with care and common sense.
Good luck Jesper
On 20 Apr 2012, at 11:35, Alberto Inuggi wrote:
> dear fsl experts
>
> I have a very general and important question, which i think disturb whichever researcher dreams.
> are there any rules to define the GLM model when i have several variables and i want to correlate my MRI images to (FA, MD, VBM, RS, Tracto..or whichever measure). how do i have to arrange my parameters ??
>
> consider for example a disease investigation, just the patient group correlation with several scores.
>
> demographic values, like age,gender, education...etc (D1, D2,.., Dy)
> clinical scores (C1, C2,..Cn),
> neuronal integrity marker like NNA/Cr or NAA/Cho..etc.....(S1, S2,..Sm)
> Neurophysiology scores (N1, N2, ..., Nx).
>
> these values may few (6,7) but also many (20,30,40).
> assuming Z the number of mandatory covariates to correct for (like age for example), and W the number of scores of interest
>
> we all know that there is a dramatic difference in performing W analyses with Z+2 columns (mean, the score and Z covariates) or one analysis with W+Z+1 columns (mean, Z cov, W variables of interest)
>
> we also know that we are free to declare how many scores we calculated and/or investigated. for example for spectroscopy parameters, we can calculate whichever score we want NNA, Cr, Cho, mI, H2O, NNA/Cr, NAA/Cho, NNA/Cho+Cr, NAA/mI, etc....
>
> so we can play with this....but we don't like to play with this.
>
> of course there is a plenty of physiological reason to couple variables, for example, when you suspect a relationship between two variables, in order to separate their contribute to the investigated MRI voxel property, you should put them in a same GLM.
>
> but if you find, when making single score analyses, that some scores all correlates with a voxel, untill you are not interested in assess to which score your voxel correlates more/less, can you also report the results of the separated analyses, reporting only those voxel with a p value corrected by the number of scores tested (e.g. you make five 1-score analyses, you report only voxels < 0.01, or > 0.99 if tfce_corrp) ??is it ok ??.
> when you want to compare their relation,
> if you declare that you investigated 5 scores, 3 do not correlate at all, is it ok if you declare that, and then make a 2-scores GLM with only the two scores correlating.
> and if, finally, you find that only one (A) of the two scores correlated, you may declare that, although both A and B singularly correlated, a more precise analysis revealed that actually such correlation was due to the A score.
> but, importantly, also B correlated, before considering A also.
>
> to conclude, can i preliminary correlation analyses of the scores itself (without GLM with MRI voxel data) guides us in coupling variables ?
> is it correct to perform several 1-score GLMs to reduce the dimensions of the final GLM matrix, that is exclude scores from further analyses?
> should all these steps carefully reported in a paper ??
>
> well I could go on with several other examples and questions but, surely, some of you, much more expert, can formalize my problem in a more concise and better form.
> thanks in advance
> regards
> Alberto
>
|