Comparison of distance-based and model-based ordinations.

Area of Science:

• Community ecology
• Ecological statistics
• Ordination methods

Background:

• Distance-based ordinations are crucial in community ecology for analyzing sample unit relationships in high-dimensional spaces.
• Recent advancements include model-based ordinations utilizing latent vectors and individual species response models.

Purpose of the Study:

• Compare the performance of two distance-based methods (NMDS, t-SNE) against two model-based methods (BORAL, REO).
• Evaluate their superiority in identifying ecological drivers of community variability and estimating species distributions.
• Assess the goodness-of-fit for species response models in low-dimensional ordination space.

Main Methods:

• Applied Non-metric Multidimensional Scaling (NMDS), t-distributed Stochastic Neighbor Embedding (t-SNE), Bayesian Analysis of Community Ecology (BORAL), and Regression with Environmental Overlays (REO) to five ecological datasets.
• Fitted environmental variables and species data using Generalized Additive Models (GAMs).
• Replicated analyses three times, averaging results to mitigate stochasticity and evaluated using deviance explained and AIC.

Main Results:

• Distance-based methods consistently outperformed model-based methods across all assessments.
• t-SNE demonstrated superior performance, particularly on complex datasets, surpassing NMDS.
• Gradient-based theory and data sufficiency of distance-based methods contributed to their superior performance.

Conclusions:

• Distance-based ordination methods, especially t-SNE, are more effective for ecological community analysis than model-based approaches.
• These findings highlight the utility of distance-based methods for uncovering ecological drivers and enhancing species distribution modeling.