I have some diet data I am analyzing that was collected in different locations at different times (over 3 years). The nature of the sampling resulted in multiple predators at the same location. The locations were not randomly selected, rather the sampling area was sampled in grid-like fashion. Not all sampling stations yielded individual predators, and not all predators had prey in their stomachs. Additionally, the total number of predators with prey in their stomachs was rather low, so I am dealing with sample size issues as well. Sorry to go into so much detail, but hopefully, you can see the types of problems I am facing. Multiple predators effectively collected at the same point in space and time would not be independent from my understanding. So how would I examine, for example, whether there were differences in diet among years or predator size-classes?
The two approaches I previously considered were canonical correspondence analysis (CCA) and chi-squared. Both of these approaches seem like they would work but only if I have independent data, which I don't. I know one way to deal with dependent data is with mixed models using the variable in which the dependence is nested (sampling location in my case) as a random effect. However, as far as I know, the only way to use this in my situation would be to model each prey group individually as separate dependent variables. This is not what I want to do. For one thing, there are rare prey groups that might only have one or two occurrences. It seems very unreasonable to try and model these rare occurrences. Additionally, it still doesn't really answer the simple question of whether diets were different between groups.
So now I am stuck trying to figure out how I might be able to detect differences in diet between groups that have dependent data. Does anyone know of an ordination technique that does not have the assumption of independent data, or does a mixed multivariate linear model exist? Keep in mind I am a scientist, not a mathematician. So whatever analysis I do will most likely need to already exist in a software package (SPSS, R, Python). I'd love to hear your answers.
