I'm working on a project where I have two response variables of animal behaviour: one is count data (Poisson distribution), and the other is proportion data (Binomial distribution). I constructed GLMM model for both response variables using the same fixed effects but encountered a situation where different random effects structures seem to be more appropriate for each response variable.
- For the count data (Poisson model), the model with NestID as a random effect has a lower AIC value.
- For the proportion data (Binomial model), the model with IndividualID as a random effect has a lower AIC value.
I understand that using different random effects structures might be unconventional, but I have a scientific justification for each choice based on the nature of the data. However, I'm uncertain about how to approach this situation and whether it might raise concerns during the manuscript review process.
Code: Here is an example of how I specified the mixed-effects models in R:
proportionmodel <- glmer(Wing~HourTemp+HourHum+HourWin+HourPre+BreeStage +
(1|Nest.ID),
data = thermo, family = binomial(link = "logit"))
countmodel <- glmer(UroEvents~HourTemp+HourHum+HourWin+HourPre+BreeStage +
(1|Individual.ID),
data = thermo, family = poisson(link = "log"))
I tried using the same random effects, for example Nest.ID, for both the responses creating separate models and individual id for both response creating separate models; however, the results are not according to theory and not justifying the observations.
