GLM and GLMM analyses
At the lake level we analysed the determinants of
paludification using generalised linear models (GLM),
which are extensions of standard linear models and
which can accommodate various non-normal error
distributions (Nelder & Wedderburn, 1972). Because
over-dispersion (residual deviance4residual degrees of
freedom) was observed in the data we used a quasiPoisson
error distribution in all the GLM models
(McCullagh & Nelder, 1989).
However, GLMs assume that the data are independent
and so, are not suitable for all types of spatial data.
At the site level, in order to avoid biased results created
by spatially correlated results among adjacent study
transects we used generalised linear mixed models with
penalised quasi-likelihood (GLMM PQL) (Venables &
Ripley, 2002) to explain variation in lake paludification.
Like GLMs, GLMMs can accommodate various