Stratification criteria mess up the use of a random factor?


New Member

In my study on kelp I want to see how wave exposure and current (both continuous variables) affect kelp physical measures along the Norwegian coast. To assure a balanced design, I stratified my study area into 9 different classes, defined by all combinations of three levels of wave exposure and three levels of current and randomly selected 27 stations such that all 9 combinations of the variable levels where represented three times. Then 10 samples were taken at each station.

Analyzing these data without taking into account the dependencies between samples within each station will be wrong, due to pseudoreplication. So I wanted to perform a mixed model, including station as a random factor. But since, due to the sampling design, station is strongly confounded with the variables in question, I lose all my variation to this variable and nothing is left for the two variables in interest, i.e. wave exposure and current.

So how should I take this dependency into account? Do I have to perform my analyses on averaged values from each sample, and thereby reduce power and the possibility of treating wave exposure and current as continuous variables? I hope not! Can I possibly do some kind of nested analysis, to specify the sampling structure, without including station as a random factor?

Most likely I will perform my analyses in R, but any help/comments on this, also from non-R users, are also greatly appreciated! My best, Gunda