I am looking at the effect of land cover (tree species, grass, woodland) on soil carbon at 3 depths. I have site as a random factor and biomass a covariate.
I ran a ranova which revealed there was no significant effect of site - therefore I dropped this from the model and have been trying to run a linear model.
However, my covariate, biomass, is dependent on the predictor variable, land cover, which violates the assumptions of ancova.
The linear model C ~ Land_Cover*biomass + Depth violates the assumption of homogeneity of variance and linearity. Transforming the response to the reciprocal has improved the linearity and although it has also improved variance there is still increasing variance- transforming the covariate doesn't seem to have any effect.
Am I doing the right thing using a linear model? Will I always have variance due to the dependence of the covariate on the predictor and does this mean I should use some sort of mixed model even though my random effect isn't significant?