Species distribution models rarely predict the biology of real populations

Abstract

Species distribution models (SDMs) are widely used in ecology. In theory, SDMs capture (at least part of ) species' ecological niches and can be used to make inferences about the distribution of suitable habitat for species of interest. Because habitat suitability is expected to influence population demography, SDMs have been used to estimate a variety of population parameters, from occurrence to genetic diversity. However, a critical look at the ability of SDMs to predict independent data across different aspects of population biology is lacking. Here, we systematically reviewed the literature, retrieving 201 studies that tested predictions from SDMs against independent assessments of occurrence, abundance, population performance, and genetic diversity. Although there is some support for the ability of SDMs to predict occurrence (~53% of studies depending on how support was assessed), the predictive performance of these models declines progressively from occurrence to abundance, to population mean fitness, to genetic diversity. At the same time, we observed higher success among studies that evaluated performance for single versus multiple species, pointing to a possible publication bias. Thus, the limited accuracy of SDMs reported here may reflect the best-case scenario. We discuss the limitations of these models and provide specific recommendations for their use for different applications going forward. However, we emphasize that predictions from SDMs, especially when used to inform conservation decisions, should be treated as hypotheses to be tested with independent data rather than as stand-ins for the population parameters we seek to know.

Priority conservation areas and a global population estimate for the Critically Endangered Philippine Eagle derived from modelled range metrics using remote sensing habitat characteristics

“…Here, we also applied our model AOH to calculating a key biological parameter used in IUCN conservation risk assessments, that of a global population estimate (IUCN 2019). However, we stress that our global estimate of 318 pairs (636 mature individuals) is the potential breeding population size based on inferred habitat from SDM outputs which may not always link to population parameters (Lee-Yaw et al 2021). Our global population size of 318 pairs was only slightly less than the current IUCN estimate of 340 pairs (BirdLife International 2018) but greater than an earlier estimate of 88-221 pairs (Krupa 1989).…”

“…Here, we also applied our model AOH to calculating a key biological parameter used in IUCN conservation risk assessments, that of a global population estimate (IUCN 2019). However, we stress that our global estimate of 318 pairs (636 mature individuals) is the potential breeding population size based on inferred habitat from SDM outputs which may not always link to population parameters (Lee-Yaw et al . 2021).…”

Climate biogeography ofArabidopsis thaliana:linking distribution models and individual variation

“…In a synthesis of published experimental studies of 40 species, Lee-Yaw et al (2016) found that on average individual performance measures and distribution-model inferred suitability decline beyond range margins. From 42 studies Lee-Yaw et al (2021) found that 38% identified some predictive ability of distribution models for individual performance. However, many previous studies relied on intensive observations of a small number of populations, while our estimates of performance from thousands of herbarium specimens allowed us to cover most of the species range and observe where specific performance measures did (and did not) correspond to suitability.…”

High resolution species distribution and abundance models cannot predict separate shrub datasets in adjacent Arctic fjords

“…Yet, this 'split-sample' validation approach remains widely used in SDM studies (reviewed in Araújo et al, 2005;Chardon et al, 2020;Santini et al, 2021). Given the plethora of work examining how validation approaches affect model predictive ability (e.g., Araújo et al, 2005;Morrison et al, 1987;Raxworthy et al, 2003;Roberts et al, 2017), it is surprising that models are rarely validated with independent datasets collected within the training spatial region (Lee-Yaw et al, 2021; but see Angert et al, 2018;Elith et al, 2006), such as those from different survey designs. This is especially relevant in the Arctic, where many of the existing data have been collected with varying survey methods (Walker et al, 2016).…”