I am trying to compare 2 models (a GLMER with a random effect and a GLM with the random effect removed). However, I was told you can't use an AIC for GLM's but I thought you could!?


1 Answer 1


I don't know exactly what models you are working on, however I don't think this is strictly necessary for a general answer. How are you estimating the GLM? Are you using Maximum likelihood? If so, why you should not use Akaike's (same story for LR test)? A couple of links pointing to this may be of some help link 1 link 2

Edited: If your aim is to test the joint significance of the two independent variables (i.e. to test whether your model as a whole is significant or not), then I would use a Wald test link wald or a LR test where you restrict both the variables to be 0 and compare the restricted model to the model with both variablkes left inside. If instead your ambition is to define whether to use 1 or 2 variables for the model and which variables then use a specification crierion or info criteria (different IC may give different answers as they impose a different penalty function on the number of parameters Akaike's link, Bayesian link, Hannan-Quinn link)

  • $\begingroup$ My model is looking at how 2 categorical variables affected a count variable (e.g. number of days). e.g. x~y*z. However, when I run my AIC and LRT on the model they give contrasting results as to what model is best. My overall aim of the model is to determine if these 2 categorical variables significantly impact the count variable. My models are nested too $\endgroup$
    – user255144
    Commented Aug 2, 2019 at 11:23
  • $\begingroup$ @user255144 Edited my answer in light of your comment $\endgroup$
    – Fr1
    Commented Aug 2, 2019 at 11:40

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.