Data gaps undermine a widely-used approach for predicting host susceptibility

Background/Question/Methods

The potential for novel biotic interactions is increasing with climate change and human-assisted species introductions. Accurately predicting the susceptibility of hosts to novel pests (for example, trees to insect herbivores) is therefore key to devising effective management and mitigation strategies. To address this challenge, researchers routinely leverage phylogenetic conservatism in host compatibility to help predict host susceptibility. One common approach uses phylogenetic relatedness as a predictor within a logistic regression framework. However, the data used to inform these models tend to consist primarily of observations of compatible host/pest pairs, while being deficient in observations of incompatible pairings. As a result, incompatible pairs must be assumed, and deciding which hosts to include in the analysis as “incompatible” is left up to the researcher. How these decisions impact predictions has not been investigated, despite numerous studies relying on this methodology. Using plant pest data from published analyses and simulations, we compared model predictions under varied scenarios of inclusion of “incompatible” hosts, ranging from a minimal set of compatible hosts’ close relatives to as many potential host taxa as possible.

Results/Conclusions

We found that predictions of host susceptibility are highly sensitive to the phylogenetic configuration of hosts and non-hosts. The apparent phylogenetic conservatism of a pest’s host breadth depends on whether the chosen "incompatible" hosts fall primarily within or outside the minimal clade containing all compatible hosts. As additional “incompatible” hosts are included, the probability that some will fall within this minimal clade increases, which decreases the apparent effect of phylogeny in more inclusive scenarios. In addition, simulations in which all compatible hosts are truly known demonstrate that the model’s accuracy does not necessarily increase as more “incompatible” hosts are added; falsely assuming compatible hosts to be incompatible can yield either over- or underestimates of phylogenetic effect depending on their position in the phylogeny. Correspondingly, both over- and underprediction of the risk a pest poses to novel hosts could result, undermining the practical utility of this approach as a risk assessment tool. We recommend alternative methods to estimating host susceptibility, or, if using the logistic regression approach, conducting sensitivity analyses that quantify the implications of analytical decisions.