When I build my most parsimonious model using a backward selection approach, do I have to worry about multicollinearity. I mean, do I first check for multicollnearity and drop the variables which has highly collinear and then put only those independent variables in my model that are not correlated and then run a backward selection approach to build the most parsimonious model.

Or should I just put all variables in the model irrespective of whether there is multicollnearity issue since anyway I will end up with the most parsimonious model using a backward selection appraoch.

mdl<-lm(y ~ x1 + x2 + x3 + x4 + x5 + x6 + x7)
drop1(mdl, test="F") 

Please let me know if the question is not clear and I will try to change it.


  • $\begingroup$ backward elimination is really terrible $\endgroup$ – bdeonovic Feb 2 '15 at 13:32
  • $\begingroup$ Ya, but this question is just to teach someone the backward elimination (not telling them it is the best) $\endgroup$ – user53020 Feb 2 '15 at 14:07
  • $\begingroup$ Its like asking whats the best way to dig a hole with a feather; no matter how you do it, you won't get the result you want. $\endgroup$ – bdeonovic Feb 2 '15 at 22:34

I might avoid absolute statements. Every methods has its limitations.

(1) Often multicolinear terms are included in a model. Or alternatives to dropping one of the variables are found. The multicolinear term results in unstable parameter/coefficient estimates.
(2) Backward elimination has many limitation. One of those is unstable variable/model selection. Collinear terms may well increase this instability. Usually backward elimination is coupled with some form of bootstrap to evaluate model stability. I thought this might be built into the rms package, but couldn't find it on a quick look.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.