Models of alien species richness show moderate predictive accuracy and poor transferability

César Capinha,Hanno Seebens,Franz Essl,Ingolf Kühn,Henrique Miguel Pereira

doi:10.3897/neobiota.38.23518

Abstract

Robust predictions of alien species richness are useful to assess global biodiversity change. Nevertheless, the capacity to predict spatial patterns of alien species richness remains largely unassessed. Using 22 data sets of alien species richness from diverse taxonomic groups and covering various parts of the world, we evaluated whether different statistical models were able to provide useful predictions of absolute and relative alien species richness, as a function of explanatory variables representing geographical, environmental and socio-economic factors. Five state-of-the-art count data modelling techniques were used and compared: Poisson and negative binomial generalised linear models (GLMs), multivariate adaptive regression splines (MARS), random forests (RF) and boosted regression trees (BRT). We found that predictions of absolute alien species richness had a low to moderate accuracy in the region where the models were developed and a consistently poor accuracy in new regions. Predictions of relative richness performed in a superior manner in both geographical settings, but still were not good. Flexible tree ensembles-type techniques (RF and BRT) were shown to be significantly better in modelling alien species richness than parametric linear models (such as GLM), despite the latter being more commonly applied for this purpose. Importantly, the poor spatial transferability of models also warrants caution in assuming the generality of the relationships they identify, e.g. by applying projections under future scenario conditions. Ultimately, our results strongly suggest that predictability of spatial variation in richness of alien species richness is limited. The somewhat more robust ability to rank regions according to the number of aliens they have (i.e. relative richness), suggests that models of aliens species richness may be useful for prioritising and comparing regions, but not for predicting exact species numbers.

Highlights

Knowing the distribution patterns of alien species richness is increasingly crucial for assessing and monitoring global biodiversity (Dornelas et al 2014, Tittensor et al 2014, Capinha et al 2015, Latombe et al 2017)
The second approach was a k-fold regional cross-validation (Jiménez-Valverde et al 2011). This approach, which relies on the use of geographically distinct subsets of the data, provides a reliable assessment of the accuracy of the predictions made to new, unrelated, geographical domains ‒ i.e. it assesses the spatial transferability and generality of the relationships identified by the models (Wenger and Olden 2012)
We expect this potential source of error to be of minor importance in our models because random forests (RF), which are known to be susceptible to overfitting (Heikkinen et al 2012, Wenger and Olden 2012, Bahn and McGill 2013), showed consistently better transferability than generalised linear models (GLMs) selected by AICc, a modelling approach that is expected to provide models robust to overfitting (Randin et al 2006; Wenger and Olden 2012)

Summary

Introduction

Knowing the distribution patterns of alien species richness is increasingly crucial for assessing and monitoring global biodiversity (Dornelas et al 2014, Tittensor et al 2014, Capinha et al 2015, Latombe et al 2017). Models of alien species richness show moderate predictive accuracy and poor transferability 79 et al 2006, Nobis et al 2009) The framework for this approach is similar to that of descriptive models relating alien species richness to spatial factors (e.g. Kühn et al 2003, Westphal et al 2008, Pyšek et al 2010, Blackburn et al 2016, Capinha et al 2017, Dawson et al 2017), but it goes one step further by using the identified relationships to make predictions. This would lead to underestimating the potential niche space of the species and would result in biased models, usually leading to incorrect predictions and inflated turnover rates in projections (Pompe et al 2008, Pompe et al 2011) In this context, statistical models directly relating alien species richness with spatial drivers (hereafter referred to as “species richness models”) become relevant as they are less data-demanding. We perform this assessment using a collection of 22 datasets of alien species richness analysed in previous studies

Methods

Results

Discussion

Conclusions