Contrasts in GLMM
Hello,
I am running a GLMM with a binomial distribution. I first run it with a random effect (OrderReceived), but I had a warning.
I thus ran it again without and then conducted contrasts (in the marginal mean section). But for instance, even if the model indicates significant p values, none of the contrast, be it for main effects (such as Group here), or the interaction, are systematically non-significant, even without correction. It is perhaps due to the high variability I have in one of my group, but I am not sure why the ANOVA output gives such results, but the contrasts do not follow them. What would be the option? Thanks!
Comments
I've notified our expert!
EJ
Hi Emily,
There are two reasons in an interplay here:
1) the main effect is just significant 'p = 0.02'. With more than two factor levels tested via an F-test, the F-test might be significantly while none of the pair-wise comparison is. (The reason for the F-test significance is arrangement of all of the factor levels simultaneously.)
2) the model is estimated and tested kn the logit scale, while the pairwise comparison are performed on the transformed estimates back to the probability scale (you can see the note bellow the table). You probably want to stick to interpreting the results on the probability scale as they are more natural for your interpretation.
If you probed the 2-way interaction, you would probably find significant pairwise comparisons on both probability and logistic scale.
Cheers,
Frantisek