Hi Gang,
Thank you for your message, very interesting and helpful.
Regarding modeling our data using 3dMVM with a covariate: our groups have a difference in age (t-test p = 0.0457; M/SD/min/max, GRP1 = 30.2/11.2/18.9/52.1, GRP2 = 38.1/9.4/22.8/51.1). Based on your webpage on covariates and centering it seems that best would have been to recruit groups that did not differ by age. Since I cannot go back to recruitment, it seems that at a minimum I can try modeling the interaction of age with the variables of interest (group). If I see no interaction effects then I can go back to not modeling the interaction. It may also be helpful that my groups are not extremely different in age, and also that a similar range was sampled.
Thus, I think this may be the best option:
3dMVM -prefix 3dmvm_output -jobs 12 -bsVars "group*age" -wsVars "condition" -qVars "age" @mvm_table.txt > 3dmvm_script_output.txt
This is the same model I was running before, but including interaction term between age and group. I can check the results of this and see if the interaction is picking up variance, and if it is, keep it, if not, then remove it from the model.
With that said, your covariate/centering website I see that it is actually not a good idea to model covariates when they differ across groups, as this violates the assumption of ANCOVA that the covariate is independent of the subject-grouping variable. With this in mind I am no longer sure how to proceed, as it seems like a fundamental violation that could lead to misinterpretation or misleading conclusions from the resultant model.
What do you advise in my situation?
Thank you,
Matthew