Several months later, I can actually my own question.
What I was actually asking was if colinearity in my cluster/grouping variables was acceptable. It looked something like this:
glmer(ret15i~AARC_Ret+AARC_NonRet|METRO)+ret02i,data=regressing9,family="poisson",control=glmerControl(optimizer="bobyqa"), nAGQ=0)
What I was asking was if co-linearity between AARC_Ret and AARC_NonRet matters. Turn out, it does, for the same reason it matters everywhere: it biases the estimates for both variables.
Judging by the number of replies my question received, that was a pile of 'duh'. But I post here to clarify for others what I was (wrongly) thinking.
In a nutshell, I didn't care about the significance of my grouping variables, because I only cared about the significance of my lower level variables.
In a nutshell, it's a question of variance pooling. Three ways to do that: All variance pooled, no variance pooled, some variance pooled. All variance is when you run the regression, ignoring the sub-pools of your grouping variable. No variance pooled is when you run a regression for each grouping variable. Some variance pooled is what HLM does: It allocates some of the variance to your group variable.
So when I added more co-linear variables to my grouping level, it sucked variance out of my lower pool, which meant there was less to explain there...which meant the lower level variables I cared about were not showing up as significant.
