Capturing information from some species abundances to improve predictions of others

Background/Question/Methods

Species distributions are driven by environmental conditions that are often unknown or unmeasured. Because groups of species tend to respond to similar conditions, the abundance of one species can inform the prediction of another, thus helping fill knowledge gaps represented by unmeasured variables. We cannot use some species as predictors of others, because all species are subject to potentially large observation error. We developed conditional prediction in which abundance of “target” species is conditioned on abundances of “incidental” species of known abundance. To account for uncertainty in observed abundances of both target and incidental species, we fit all species jointly in a generalized joint attribute model (gjam), which allows for covariance between responses. Conditional prediction combines information from predictor variables with abundances of incidental species. We use simulations to demonstrate the effectiveness of conditional prediction applied to four types of response data. We then apply conditional prediction to Breeding Bird Survey data to address two issues of conservation interest. First, we show that conditioning on incidental species improves prediction, especially for rare species like Kirtland’s Warbler (Setophaga kirtlandii). Second, we use conditional prediction to understand relationships in abundances of Brown-headed Cowbirds (Molothrus ater), a brood parasite, and its host species.

Results/Conclusions

Simulations show that conditioning on incidental species reduces prediction error of target species by up to 95% for continuous and discrete counts. Brier scores for predicted presence-absence data are improved by up to 50% when conditioning on incidental species. Potential benefits of conditional prediction increase with residual correlation between species; if species abundances are completely explained by covariates, conditioning provides no additional information. With Breeding Bird Survey data, the improvement in RMSE is inversely correlated with species abundance, such that predictions of rare species show the greatest improvement (r = -0.57). RMSE of predictions of Kirtland’s Warblers, which are absent from 99.7% of observations, are improved by 71.2% by conditioning on other species. Finally, we identify species that may be at high risk of nest parasitism due to co-occurrence with Brown-headed Cowbirds. While abundance of Kirtland’s Warblers has a negative relationship with abundance of Cowbirds (-3.8), likely due to management strategies aimed at reducing cowbird abundance in Kirtland’s Warbler breeding grounds, Bell’s Vireo has a positive relationship with cowbird abundance (2.2), suggesting high risk of nest parasitism. Results demonstrate the utility of conditional prediction as an avenue for exploiting information species have to offer on the abundances of others.