Analog impact models: a comparison with commonly used empirical approaches to model climate impacts

Forecasts of climate change impacts are crucial to respond and adapt to ongoing and future environmental changes. Analog impact models (AIMs) are increasingly used to project climate impacts to biological systems. However, despite their growing use, the skill of AIMs is underexplored, and validation of AIMs is not yet a standard practice. More importantly, there is a limited understanding of how AIMs compare to other modeling approaches. In this study we assess the skill of AIMs using contemporary validation, and compare their performance to commonly used empirical models: generalized linear model (GLM) and random forests (RF). We demonstrate that AIMs can be structured as spatial models, and capture spatially encoded information to provide context-specific projections of climate impacts. Our results show that AIMs perform comparably to RF models and outperform GLMs in validation tests. Furthermore, AIM projections of future species distributions differ from the projections of RF and GLMs, which we attribute to their spatial nature. We conclude that AIMs are a valuable addition to the climate impact toolkit.