I have a doubt about nesting random effects. I'm using R with the lme4 package and, in particular, the glmer function with binary family. I will describe the data first --
In this experiment, people answer sentences related to some verbs with a binary
answer. They are given a
prompt to formulate their answer, and this prompt is one of the variables. Each sentence appears with one different prompts (of two) to different participants in a randomised manner.
Two more binary variables are present and they're bound to the verb: verbs can be divided in two different ways: two
classes, A and B, and two
types, Y and X. But a verb cannot appear in both A and B forms depending on the context, so each item/verb in the experiment only appears in one of the two variants. That means that verb1 is always A and X, verb2 is always B and X and so on.
My doubt is how to define the random effect for verb/item, knowing that the data does not allow for each item to appear in both conditions of
type. I thought nesting would solve it in this way:
Answer ~ Prompt + Class + Type + (Prompt + Class + Type | participant) + (Prompt | verb:class) + (Prompt | verb:type)
But I'm not totally sure if it makes sense. My questions are:
- Is it okay to include verb as a random effect nesting it twice (within class and withing type), as in the formula above?
- Is it possible that this definition of the random effects would "absorb" the fixed effects, reducing their estimates or significance?
- Does it make sense to include verb as a random effect at all, given that each verb can only appear in one class and one type?