Non uniform sampling evaluation

[robbb]

Hi everyone!

I am working on an animal epidemiology project; dead animals (samples) are opportunistically collected, then checked for the disease (positive/negative). Positive samples are restricted to one specific area but, as some areas provide more samples than others, I would like to check for biases in the sampling process (i.e. do I get more positives just because sampling was more intense on that area?)...Does anyone have any idea of which statistical analysis should I apply on my dataset to evaluate that? I suspect it might be trivial for a statistician, so I apologize up front if I'm lowering the bar of the discussion on you guys, I'm just an ecologist =) still, I'm not finding any clue on this!

Thanx to any help, and cheers for the good work done up to now!

Roberto

 

[obh]

Hi everyone!
Positive samples are restricted to one specific area but, as some areas provide more samples than others
Roberto

What do you mean?

 

[robbb]

Hi and thanx! For example, let's say that every State in Europe provides some samples, but some State provide more, some other less samples. Negative samples (i.e. free from the disease I am studying) are overall, but positives are found only in one specific State. I'd like to know if the different sampling intensity between States influences the location of my positive results.

 

[hlsmith]

Not sure your comfort with stats, but this seems like it may be addressed using multilevel modeling with a Poisson (count) outcome. This model would also allow you to control for secondary variable at the observation (animal) or state level. This model account for variability between and within groups. In your case a group is a state.

 

[robbb]

My comfort with stats is indeed quite uncomfortable! What you suggest is the final object of the research (finding "why" the epidemiological focus is where it is, including environmental variables, animal ecology etc.), for now I just would like to exclude the possibility that positives are there just because a lot of samples come from there. I think something like a Pearson's correlation test may work, but as I say I'm not a stats-grinder =)