2
$\begingroup$

I'm new to checking the VIF value for a glm model so I just want to make sure i"m understanding this correctly.

I have 4 predictors for my count model and the model looks like this:

model1<-glm(Number~dts+dss+dtn+dsn, family=poisson, data=birds)

I then checked the collinearity on the model using the car::vif function and got this output;

dts         dss         dtn       dsn
2.261840    2.281326    2.016644  2.073556 

so from my understanding and reading online, due to all 4 being below vif of 3, then there is no multicollinearity and I can now proceed with finding the "simplest" model.

After I did this, i then performed the "summary" function, and found 3 out of the 4 to be significant, the model would then become;

dts, dss, dsn

Would that therefore mean that my best model would be;

glm(Numer~dts+dss+dsn, family=poisson, data=bird)

or am i completely misunderstanding the car:vif function

$\endgroup$
3
  • $\begingroup$ What is your interest in multicollinearity or even in building this model? $\endgroup$ – Dave Jul 31 '20 at 12:23
  • $\begingroup$ I'm trying to see if any of those 4 predictors influence the number of birds being present. So i wanted to make sure those 4 predictors did not have any collinearity $\endgroup$ – andy Jul 31 '20 at 12:25
  • $\begingroup$ Well then why not test the full model against a null model with just an intercept? $\endgroup$ – Dave Jul 31 '20 at 12:40
2
$\begingroup$

From what you describe, there's no reason to abandon the full four-predictor model.

Unless it's extreme, VIF isn't much of a criterion for evaluating a model. It describes the correlations among the coefficient estimates. A high VIF just means some imprecision in those estimates. See this page among many other for discussion on implications of VIFs. Things like goodness of fit and predictive performance are much more important.

There's little to be gained and potentially a lot to be lost if you remove the nominally "insignificant" predictor. There's a problem even in ordinary least-squares regression with omitted-variable bias: if you omit from a model any predictor that is correlated both with outcome and with another predictor, then the estimates of the coefficients for the retained predictors will be biased. In other types of regressions like logistic regression, you can have this problem even if the omitted predictor isn't correlated with the retained predictors.

So I don't see much reason to omit dtn from your model.

$\endgroup$

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.