2.3. Data analysis
We built two Linear Mixed Models (LMMs) using the nlme
(Piñeiro et al., 2009) and lme4 (Bates and Sarkar, 2007) R packages.
LMMs are particularly useful to account for nested sampling
designs (Bolker et al., 2009), where the autocorrelation of samples
(pixels) within sites (meteorological stations) is accounted for
through the introduction of random effects. We followed the steps
suggested by Zuur et al. (2009) as a procedure to build the two
models. First, using the full model for the fixed effects, we found the
optimal structure for the random component using restricted
maximum likelihood (REML). Then, to find the optimal fixed
structure, the trade-off between significance and simplicity was
evaluated by iteratively comparing more complex models to
simpler models and choosing the one with lower AIC and BIC, and
significant c2 test for the Log-likelihood (using maximum likelihood
and always maintaining the same random structure).