Dear Carlos,
On Fri, 4 Nov 2022 13:31:43 +0000, Carlos Murillo Ezcurra <[log in to unmask]> wrote:
>Dear CAT12 developers/users,
>
>I am currently using CAT for my VBM analysis, and I have some doubts about the integration of TIV as covariate based on the statistical analysis. I have already read some posts in the mailing list, but I could not really find the information:
>
>(1) Adjustment by TIV in longitudinal analysis with Group*time interaction. The CAT documentation states that using TIV as covariate in longitudinal analysis is not necessary because we are analyzing intra-individual effects. However, I understand the would be indeed the case if we are interested in the main effect for Time. However, if the main interest is the Group*time, shouldn’t we include the baseline value as we did with the cross-sectional analysis? (Same applies to gender and age). Then if we do want to focus on the Main Effect of Time (e.g., no Group*time significant) another model with non-time varying covariates (TIV, age and gender) can be fitted?
>
You are right that for the main effect it should be helpful to include TIV as covariate. However, the contrast for main effect in long. designs may artificially inflate your results. More information about that issue can be found here:
https://doi.org/10.3389/fnins.2019.00352
Therefore, I would not really recommend investigating main effects in long. designs because of these issues.
>(2) Adjustment by TIV in partial correlation/multiple regressions (Y ~ bo + B + TIV + age + Gender) in only one group (cases). We Y denotes the scores of a questionnaire and B the extracted betas from a significantly different (cases vs controls) cluster. There is multicollinearity to some extent between TIV, gender and B. The addition of TIV alters the pattern of findings (significant --> non-significant) in the same way as described in a recent seminar paper https://www.sciencedirect.com/science/article/pii/S105381191930816X
>Which is suggested to be due to removal of meaningful variance from the other predictor(s).
>
>Since we can expect that TIV cofounds the gray matter betas but NOT the questionnaire scores and GM is not being compare between subjects, it could be argued (if usual confounding adjustment rationale applies) that adjusting for TIV is not required since it is not common cause of both. Am I right or should I adjust for TIV anyway because another reason that I am missing?.
The correlation between TIV and you effect of interest is important and if questionnaire scores are not correlating with TIV this should be fine. You could also try using TIV as global scaling. However, this will not consider local (voxel-wise) effects and is less efficient in removing non-linear effects.
Best,
Christian
>
>Thank so much for your help in advance. I really appreciate it!
>Kind regards
>Carlos
|