Climatic associations

We used Partial Least-Squares Regression (PLSR) models to evaluate the importance of ECV’s trends in describing the ΔCF variability. We first estimated the optimal number of components required for the model using 10-fold cross-validation models repeated 100 times following (Kuhn and Johnson, 2016). The optimal number of components was selected as the lowest Root Mean Squared Error of Prediction (RMSEP). Knowing the optimal number of components, we then developed a final iterative PLSR model. This last model was the average of 5000 iterations, each using 50% of the data randomly selected to build the model and test it using all the samples. We evaluated the model performance for each iteration by examining the coefficient of determination (R2) and the Root Mean Square Error (RMSE). We also assessed the importance of each ECV in describing the ΔCF variability by estimating the Variable of Importance of Prediction (VIP). The PLSR models were performed in R using the pls package (Liland et al., 2021), whereas the VIP was estimated using the plsVarSel package (Mehmood et al., 2012). We did not split our data into training and testing datasets because our goal was to disentangle the association between ECVs and ΔCF more than creating a prediction model.