Non-independence in data and GLMMs I am working with a dataset that consists of categorical variables and count data. My response variable is DPM SpeciesA (detection positive minutes of Species A, where for each hour I have a count of how many minutes species A was vocalizing), and my predictors are Season, Time of day (classified as "Morning", "Evening", "Night" and "Day"), Tidal cycle ("High tide", "Ebbing tide"...) and DPM SpeciesB.
I am wondering how to account for the non-independence in my data. Because the vocalizations of Species A are being recorded, but we have no way of knowing which individual is vocalizing, there is non-independence since the same individual may be vocalizing for multiple minutes, or different individuals, and each hour it could be the same individuals, or different individuals.
So my question is, is it not possible to use GLMs in this case because of the non-independence? And if it is not possible, then how could I account for the non-independence in a mixed model? I don't understand how to incorporate any dependency into the model because I can't add the ID of the individual as a random factor, since we never know which individual is being counted. 
Thanks for your help everyone!
Edit: I have now included some plots that I have made to try to determine whether violating independence would be a problem (at least in terms of temporal autocorrelation) in a quasipoisson GLM for this data. These are Acf, Residuals vs Index and a residuals lag plot. But I am not very experienced interpreting these plots! To me it looks like there is no particular problem with the residuals vs index (except that everything is above zero, not sure if there should also be negatives) but the Acf and Residuals lag both seem to have a pattern. 



 A: Welcome to the site, roc. It may be possible to employ a mixed modeling framework to your data. How many different species did you record? If there are 10-15 or more, then you could have the grouping variable be species. This would indicate that you believe that vocalization frequency is more correlated within species than it is across different species. That being said, this doesn't really solve the problem you identified about individual members of the species. That is not something you can deal with given your data. You would have to use some sort of machine learning algorithm to try to identify individuals within species, and I'm not sure that would work (it might, I just don't know much about that).
However, if you only have two species (you mention SpeciesA and SpeciesB), then you are in a more difficult position. Since you cannot identify individuals within SpeciesA and SpeciesB, you cannot account for the likely fact that some members of the Species will be more likely to repeatedly call. Is it correct to assume that you want to have a predictor for SpeciesB in the model for SpeciesA vocalizations? Or were you just saying you wanted to generate the same model for each species?  
If you have multiple species, there are many different R packages for doing this type of analysis with your count outcome, including glmer() in lme4, glmmTMB() and GLMMadaptive. 
