Predictive Modelling Strategy
We used generalized dissimilarity modelling (GDM) to predict shifts in combinations of spatial ecosystem response variables. Given this novel application of GDM, we take an exploratory modelling approach (sensu Tredennick et al. 2021) to developing three classes of models with 1) ecosystem, 2) biotic, and 3) abiotic responses. Each model is built with the same combination of predictors and the same model settings, but differing combinations of response variables. GDM formulations predict dissimilarities between pairs of sites as a function of their ecological distance in environmental and or geographic space. Since the modelled response in GDM is a pair-wise distance metric, predictor variables are also represented as pair-wise distances (i.e., difference in predictor values between site pairs). Ferrier et al. (2007) further summarize the statistical underpinnings of GDM, while Mokany et al. (2022) propose a workflow for model fitting. We adapted this workflow (see Figure 2) and employed the GDM package (Fitzpatrick et al. 2024) with default model settings (see S1 for additional detail).
We began the modelling process by selecting ecologically meaningful combinations of response variables informed by patterns summarized in the Nova Scotia Forest Ecosystem Classification (Neily et al. 2023). We built a suite of models to evaluate how varied combinations of these response variables affected model performance (see S1). Performance was assessed using deviance explained, a key GDM performance metric representing the percentage of observed variation (dissimilarity) explained by the model (Mokany et al. 2022). We retained models in all three