I am running a glmm with three fixed effects:

opponent 1 size ("1")
opponent 2 size ("2")
opponent 1 size - opponent 2 size ("diff")

I am unable to run all three variables in the model at once because of the "diff" variable being correlated with the "1" and "2" variables. How, then, do I decide which is the best model, when all the different combinations I can test are not nested within one another? Or can I consider the first two variables to be nested within the third one, since they are both used to calculate "diff"? I have these combinations of variables and their corresponding AICs (variable(s) followed by AIC in parentheses):

variable(s) (AIC):
diff (223)
1 (231)
2 (262)
1,2 (265)
1,diff (265)
2,diff (265)


1 Answer 1


Run the model with the first two predictors. It is not necessary to include the third, since this is merely a linear combination of the first two.

  • $\begingroup$ But the predictive power of the third variable is the one I (and my reviewers) are most interested in! :) $\endgroup$ Sep 10, 2014 at 18:58
  • $\begingroup$ Can you leave one of the first predictors out, then? $\endgroup$
    – coanil
    Sep 10, 2014 at 20:28
  • $\begingroup$ No, unfortunately not. The goal is to determine which of those three factors best explains the outcome of a contest. $\endgroup$ Sep 10, 2014 at 20:35

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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