there is two schools of thought about this, newer engines like QRPEM would be happy with a full omega and then you can reduce it if you see fit
adding a covariate can reduce the off diagonals ( explain omega correlations) and also the diagonals
see the figure 2 diagmram here (full omega before doing covariate selection)
I typically avoid doing scm (stepwise) there is tons of literature showing that it has major drawbacks:
you should look at full model that are plausible and useful to answer key clinical questions
Edited by smouksassi1, 04 October 2022 - 07:59 AM.