Is poisson regression appropriate?

This is more of a general stats question regarding the best procedure to answer my question of interest. I started to address it using poisson regression in proc GLIMMIX (SAS), but now am unsure about this decision.

Experimental Design:

I sampled shrubs (stem counts by species) in quadrats in 3 forests and measured environmental variables in each of those quadrats. The environmental variable of primary interest the "wetness" of each quadrat.

In order to make more ecological sense of the data, I aggregated the shrubs by Wetland Indicator Status (e.g. a measure of the known preference of species for wetland areas).

What I want to know is how relationship between stem density and wetness differs between the different indicator groups. That is, I want to be able to compare the responses of the shrub groups to these conditions. My initial thought was to make Indicator Status a term in a poisson regression, where a significant interaction between "hydro" (the wetness measure) and "IND" the wetland indicator status of the shrub group (hydro*ind in the model statement below) would indicate that the relationship between stem count and wetness differs between shrub groups with different indicator statuses. Basically, I am trying to confirm what you would expect--that the density of shrubs with an affinity for wet soils increase as soil wetness increases, and those with a preference for drier areas decrease with increasing wetness. There is then a third group of an invasive species that does not have an indicator status, so I am trying to see it its relationship with wetness is more like the wetland or upland species.

However, I am thinking that my approach to the analysis is problematic, because indicator status (IND) is a property of the DV and not an independent variable. Is this analysis appropriate for MANOVA where my dependent variables are the counts of shrubs in the different indicator groups??

If anyone can help me come up with a way of answering this question, it would help me immensely. I do need a method that can handle non-normal data (count data, so it is has many zeroes), and random effects (forest site is a random effect).

First attempt here:

proc glimmix data=shrub maxopt=100;
class site ind;
model count= hydro ind ind*hydro / dist=poisson link=log solution;
random site;
covtest 'Global Test random effects' ZEROG / CL WALD;
output out=glimmixout predicted=pred Pearson=PearsonRes; run;

Last edited: