Predicting future spread during an outbreak using species distribution models

Background/Question/Methods

Accurately predicting the eventual spread of a non-native species or novel pathogen facilitates forecasting effects on communities and planning actions to manage or eliminate the outbreak. One set of tools used for these predictions is species distribution models (SDMs), which use information on environmental conditions at presence locations to estimate the suitability of other locations. Although SDMs are increasingly used to predict spread, it is recognized that organisms spreading into new regions violate model assumptions of equilibrium dynamics. One expected effect, supported in recent studies, is that models of the potential range using data from later in the spread process, and therefore closer to equilibrium, will perform better. Given the large set of SDM methods, we designed a study to determine how well and how early different methods could accurately estimate the potential range of an outbreak. We simulated transmission of a model pathogen among locally interacting hosts, with the risk of transmission in each patch determined by variation in environmental variables. Ten SDMs were fit to stochastic realizations of pathogen spread in replicate environments. We compared different pathogen types (e.g. SIS, SIR) and environments with and without deterministic gradients in environmental suitability.

Results/Conclusions

The subset of SDM algorithms that most accurately identified the current distribution of disease at each time step included boosted regression trees, MaxEnt and a newly developed plug-and-play density estimator. The performance on current distribution was markedly higher than accuracy in identifying the potential range (defined as all locations with possible spread, i.e. R₀ >1). The same methods were also best at identifying this potential niche, but were generally inaccurate, with area under the ROC curve usually below 0.6. The presence of a gradient of suitability improved model accuracy. Unlike recent studies, there was not a clear pattern of improved prediction of the potential niche over time. A key factor limiting forecast accuracy was the mobility of infected hosts. Dispersing individuals enabled pathogen spread, but their movement into patches where disease cannot be transmitted confounded the identification of environments where transmission was possible. This issue of source-sink dynamics is an aspect of non-stationarity applicable to many uses of SDMs, but it may be exacerbated in the case of invading organisms. Our results suggest predicting eventual ranges is difficult enough to confound some models, but we identify methods and types of spread where predictions are more likely to be informative.