I am building a Linear Mixed Effects Model to investigate how activity levels (AL) change between two different types of therapies (TherapyA vs. TherapyB) and how different covariates influence the AL in TherapyA and TherapyB. Therefore I included interaction effects between by factor 'therapy' and all covariates I am interested in. For now, I am just interested in the influence of the first covariate, which is the 'center'.

I formulated my model as follows (please ignore the other interaction effects):

lmer(AL ~ center*therapy + Cov2*therapy + Cov3*therapy + (therapy|subj_ID), data = data_model)

I found a significant main effect of center and therapy, but no significant interaction effect center*therapy, when looking at the anova-output:

enter image description here

Normally, I would look at the lsmeans output for post-hoc comparisons between the single centers and would see that AL in center 2 is significantly higher than in center1: enter image description here

As far as I understood, although I get a warning message that results could be misleading, I am still allowed to interpret my main effect of 'center', because the involvement in the interaction is not significant. Am I right about this?

Next, I was curious and added the therapy in the lsmeans specification too. enter image description here

What I can see here is that center 2 has only a higher AL in therapyA, not in therapyB, which is very interesting for my analysis. It appears also to be higher than center 4, which was not revealed in my post-hoc main effect analysis. Am I allowed to interpret this at all, since the interaction wasn't significant?


1 Answer 1


The message is just saying that the model includes interactions involving center, so you need to be careful in interpreting the results. If the interaction isn't strong in any practical sense, then you may ignore the warning -- or remove the interaction from the model so that you don't get the warning at all.

In your instance, though, look at the 2 - 4 differences for the two therapies; you're not getting the same test results for those.

  • $\begingroup$ The question is rather: Am I allowed to interpret the effect of center conditional on 'therapy', although I don't have a significant interaction? $\endgroup$
    – soschnei
    Commented Nov 11, 2018 at 16:02
  • $\begingroup$ Yes. Personally I don’t think you even need to look at the anova table at all, as long as you use appropriate multiplicity adjustments in the tests of comparisons. $\endgroup$
    – Russ Lenth
    Commented Nov 11, 2018 at 17:39

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.