Solved – Multiple Regression – Testing for multicollinearity

I don't think it matters much. Checking it later will save you unnecessary work and agitation at needless transformations that might prove pointless if the variables won't be in the final model. That being said, checking vif(model) is not time consuming, and you can always wait with the application of solutions to potential multicollinearity until later.

The problem of multicollinearity is that it can distort the affected coefficients, change their signs and their significance. The 'good' thing (should say convenient) about multicollinearity it is that it affects only the collinear variables – yet does not affect he rest of the variables. This means that if collinear it's only exists on control variables, it often OK to disregard it.

You can check, see if it's on controls. If it is, optimize and leave it. If on main explanatory variables, deal now before optimizing (a common way is centering – which can be done using scale(var_to_scale, scale = FALSE)

Edit: the answer by @user3640761 rises a valid suggestion, that you check for high correlations in your data before doing anything else. It's easy, fast, and can give a good indication.