[R] general question about dropping terms of glm model fits

[Sacha Viquerat]

hello dear list!
as I am currently helping someone with their statistical analysis of a 
count survey, I stumbled upon a very basic question upon model optimization:

when fitting a model like:

model<-lmer(abundance~(x+y+z)^3,family=poisson,...)

in which x,y,z are continuous abiotic parameters such as po4 
concentration, no2-concentration, which terms / interaction terms would 
you recommend removing FIRST?

the ones of lowest significance (i.e. the ones with highest p-value) OR

the ones with the most complex interaction structure (even though 
p-values may be low-ish)?

another question just popped in my mind:

let's say I've reduced my model to significant terms:

y ~ temperature + po4 + po4:temperature

and I know that correlation between po4 and temperature is high. would 
you say that this is reason enough to remove the interaction term?

any opinion is a welcome opinion!